ISRS 2026 Technical Program (Draft)

Time Table 1 (Day 1: Wednesday, May 13, 2026)

Time	Hall	Room A	Room B	Room C	Room D	Room E
9:00–10:00	Registration
10:00–10:30	Coffee break
10:30–11:30		Opening ceremony / Keynote1
11:30–13:00	Lunch
13:00–14:00	Poster 1 (Core time)
14:00–15:15		A1: Ocean-1	B1: AI/ML-1	C1: Forestry	D1: Agriculture	E1: Algorithm-1
15:15–15:30	Coffee break
15:30–16:45		A2: Ocean-2	B2: AI/ML-2	C2: Land Cover	D2: Climate	E2: Algorithm-2

Time Table 2 (Day 2: Thursday, May 14, 2026)

Time	Hall	Room A	Room B	Room C	Room D	Room E
9:00–10:00		Keynote2
10:00–10:15	Coffee break
10:15–11:45		A3: Land-S1	B3: GIS-S	C3: Land-S2	D3: Ocean-S1	E3: Algorithm-S1
11:45–13:00	Lunch
13:00–14:00	Poster 2 (core time)
14:00–15:15		A4: Land-S3	B4: Cryosphere/Climate-S	C4: Land-S4	D4: Ocean-S2	E4: Algorithm-S2
15:15–15:30	Coffee break
15:30–17:00		S1: Advanced Ca//Val	B5: Sensor/Platform-S	C5: Land-S5	D5: AI/ML-S	E5: Algorithm-S3

Time Table 3 (Day 3: Friday, May 15, 2026)

Time	Hall	Room A	Room B	Room C	Room D	Room E
9:00–10:15		A6: Cryosphere	B6: Sensor/Algorithm	C6: Disaster	D6: Image Analysis	E6: Atmosphere
10:15–10:30	Coffee break
10:30–11:45		S2: OLaReS	B7: GIS	C7: Infrastructure
11:45–12:00	Short break
12:00–12:45		Closing Ceremony
12:45–13:00	Short break
13:00–14:00		RSSJ council
14:00–14:25	Short break
14:25–15:45		RSSJ General Assembly

Note: After 12:45, the schedule is for the RSSJ members only. ISRS2026 will conclude with the Closing Ceremony.

Session A1 (Chair: TBD)

14:00–14:15	Validation of Multi-mission Altimetry around the Korean Peninsula Using Tide Gauges Do-Hyun HWANG
14:15–14:30	Finding the Optimal data usage of synthetic data on Satellite image ship detection Task Yan-Chen HUANG
14:30–14:45	A Multispectral Fusion Framework Integrating U-Net and SWINIR for Coastal Mangrove Monitoring in Taiwan yiwen LO
14:45–15:00	An Automated Hybrid Framework for Surface Water Extraction: Integrating Spectral-Morphological Techniques with SatMAE Ting-Yu,LAI
15:00–15:15	Estimating tidal flat erosion rates based on landward migration of concentrated tidal channel zones and their controls Jiyoung LEE

Back to time table

Session B1 (Chair: TBD)

14:00–14:15	QueryGIS: A Trustworthy Conversational GeoAI Agent for QGIS Workflow Automation with RAG and Self-Repair Juseong LEE
14:15–14:30	An Automated Scan-to-CAD Framework for Structural Modeling of Bridge Point Clouds Li-Chih CHEN
14:30–14:45	Pothole Detection Under Data Scarcity Using Stable Diffusion and YOLO Cheng-Zong YANG
14:45–15:00	Super Resolution Reconstruction of Kompsat-3A Images Based on a Fine-Tuned Transformer Hongjun YOUN

Back to time table

Session C1 (Chair: Yasumichi Yone?)

14:00–14:15	Spectral Features of Ultra-High-Resolution UAV Hyperspectral Data over Deciduous Broadleaf Forests in Tomakomai, Hokkaido Hideki KOBAYASHI
14:15–14:30	Individual Tree Crown Delineation and Carbon Stock Estimation Based on Airborne LiDAR Data Ting-Hsuan TUNG
14:30–14:45	Evaluation of Satellite Remote Sensing-Based Biomass Estimation for Carbon Credit Calculation under the JCM Yohei TOYOHARA
14:45–15:00	Integrating Environmental Variables into Remote Sensing-Based Forest Type Classification in Taiwan Han-Hui HUANG
15:00–15:15	Integrating Photogrammetry and Deep Learning for Multi-Temporal Change Detection in Rural Landscapes of the Oki Islands, Japan Yasumichi YONE

Back to time table

Session D1 (Chair: Shoko Kobayashi?)

14:00–14:15	Multi-Temporal Remote Sensing Analysis of Land Use Change in High-Altitude Tea Gardens Chin CHENG
14:15–14:30	Prototype Development of Rice Monitoring Methods at Field and Regional Levels in the Republic of the Union of Myanmar Shinya ODAGAWA
14:30–14:45	Integrating satellite remote sensing and meteorological data for early season yield prediction of rice and wheat Akie KOGA, PASCO CORPORATION
14:45–15:00	A Phenology-Driven Approach for Detecting Abandoned Cropland Using HLS Imagery and Interpretable Machine Learning Sinyoung PARK
15:00–15:15	Correlation between field-measured grain moisture content and C-band SAR backscatter: Potential use of Sentinel-1 SAR data to estimate optimal harvest timing in rice crop cultivation Shoko KOBAYASHI

Back to time table

Session E1 (Chair: TBD)

14:00–14:15	Vision-Based Camera Pose Estimation for Bridge Inspection Using a Reusable 3D Feature Map Yi-Chen LEE
14:15–14:30	Comparative Analysis of CNN and Transformer Models for Robust Water Body Detection across Multi-Satellite Sensors Seojin KONG
14:30–14:45	Deep Learning-based Approximation of Building Footprints from High-Resolution Satellite Imagery Yun-Hao LIN
14:45–15:00	LOSS FUNCTION OPTIMIZATION OF DEEP LEARNING MODEL FOR IMAGE ALIGNMENT BETWEEN SATELLITE IMAGERY AND GCP CHIPS Kyeonghwan KIM
15:00–15:15	Analysis of Spatial Resolution Enhancement via Deep Learning-Based Image Super-Resolution for GCP Chip Matching Minkyung CHUNG

Back to time table

Session A2 (Chair: TBD)

15:30–15:45	Investigating a Geospatial Framework for Tidal Flat Sedimentary Facies Analysis Using Satellite Imagery and GIS Data Donguk LEE
15:45–16:00	Micro-topographic Change Analysis of Tidal Flats Using Co-registration and Patch-based Bias Correction of UAV-Derived DEMs Minju KIM
16:00–16:15	Morphological Changes of Tidal Flats in South Korea Using a 40-Year Landsat Time Series Jingyo LEE
16:15–16:30	Spatial Validation of Nighttime Vessel Detections by KIOST VBD using EOG VBD and Ground Truth Data in Korean waters Chan-Su YANG

Back to time table

Session B2 (Chair: TBD)

15:30–15:45	High-Resolution Soil Moisture Estimation over South Korea Using Machine Learning and Integrated Sentinel-1 SAR and Meteorological Data Su-Bin HA
15:45–16:00	A Machine Learning Approach to Spatio-temporal Interpolation of Regional Ionospheric Models for Improved Precise Point Positioning Wei-Chun YU
16:00–16:15	Deciphering Environmental Controls of Total Hydrocarbon Accumulation in Taiwan Using Geospatial Explainable Machine Learning Thia PRAHESTI
16:15–16:30	Integrating Geospatial Analysis and Hybrid Deep Learning Model for Forest Fire Susceptibility Mapping in Sanggau Regency, West Kalimantan, Indonesia Suci RAMAYANTI
16:30–16:45	Accuracy Assessment of SfM/MVS-Based 3D Reconstruction Using RTK-Enabled Smartphone Imagery under Varying GNSS Environments Yonghe LI

Back to time table

Session C2 (Chair: Naoyoshi HIRADE?)

15:30–15:45	Seasonal Land Cover Classification in Japanese Agricultural and Wetland Landscapes via Machine Learning and Multi-Source Sentinel Data Santa PANDIT
15:45–16:00	National-Scale Detection of Golf Courses by Convolutional Neural Network model and Land-Cover Characterization Using Landsat Time Series in Japan Shoki SHIMADA
16:00–16:15	Stability-Controlled Land-Cover Monitoring in Access-Restricted Regions: The DMZ-Adjacent Iron Triangle (Cheorwon–Pyonggang–Kimhwa), DPRK Beom KIM
16:15–16:30	Identifying Suitable Locations for Green Infrastructure Development around the Settlement Zone of Ulaanbaatar City Using Satellite Data Enkhzaya Enkhtaivan
16:30–16:45	DEVELOPMENT OF A JAXA HIGH-RESOLUTION LAND-USE AND LAND-COVER MAP OF SOUTHEAST ASIA USING ALOS-2/PALSAR-2 AND OPTICAL TIME SERIES Naoyoshi HIRADE

Back to time table

Session D2 (Chair: Kazuhito Ichii?)

15:30–15:45	Evaluating the Climate-Adaptive Role of Green Belt Areas Using Satellite-Derived LST and NDMI Yoonji KIM
15:45–16:00	Impacts of Leaf Area Index Data Assimilation on Hydrological Responses to Drought and Flood Events in the Korean Peninsula Hee-Jeong JEONG
16:00–16:15	Multi‑Sensor Assessment of Soil Moisture Seasonality in Semi‑Arid Mongolia Using AMSR2 and Ground Measurements Nozomu HIROSE
16:15–16:30	Toward a Global Hyper-Temporal Land Monitoring Framework Using an International Constellation of Geostationary Meteorological Satellites Kazuhito ICHII
16:30–16:45	ASSESSING THE IMPACT OF CLIMATE CHANGE ON SOLAR RESOURCE IN THE PHILIPPINES USING MODIS AND IN-SITU MEASUREMENT Jeark PRINCIPE

Back to time table

Session E2 (Chair: TBD)

15:30–15:45	An Integrated Framework for Single-Image Automated Reconstruction of Detailed Wireframe Objects Using YOLO, SAM2, and Ray Intersection Geometry Yi Hsuan SU
15:45–16:00	Quantitative Assessment and Algorithm Development for High-Resolution LiDAR-Derived Tidal Creek Extraction in Korea’s West Coast Tidal Flats Ye-Young KIM
16:00–16:15	Overlap-Based Extrapolation for Partial Geometric Correction in GCP-Free Segments of Strip Imagery Jaewon HUR
16:15–16:30	Feasibility Assessment of Depth-based GCPs Weighting Approach for Geometric Correction of BlueBON Satellite Image Hongjin KIM
16:30–16:45	Integrated Analysis of Long-term Field Sediment Data and UAV-Derived Topographic Information for Interpreting Spatial Variability in a Tidal Flat Sehee KIM

Back to time table

Session A3 (Chair: TBD)

10:15–10:30	A Georeferenced Neural Radiance Field Framework for DSM Generation from UAV Imagery Cheng-Hsin LI
10:30–10:45	4-Stream Mid-Level Fusion with nDSM Confidence Gating for Urban Green Space Extraction from RGB and LiDAR-Derived Images Yosuke OMIZU
10:45–11:00	Integrating RTK-GNSS with 3D Gaussian Splatting for Georeferenced Scene Reconstruction Chi-Min CHIU
11:00–11:15	Characterizing Regime Shifts in Mekong Delta Droughts from 1982 to 2025: Multi-Scale Insights from Continuous Wavelet Transform Hong Danh PHAM PHAN
11:15–11:30	Species-Aware Panoptic Segmentation of UAV-SfM Point Clouds Using an Extended SegmentAnyTree Framework Kei YAMAMOTO
11:30–11:45	Rice Plant Coverage Estimation from UAV Data Using a Gaussian Mixture Model Hiroto YOSHIDA

Back to time table

Session B3 (Chair: TBD)

10:15–10:30	An Integrated Spatial Assessment of Urban Fire Risk Using Multi-Source Data: A Case Study of Ulaanbaatar City KHOSTSETSEG.T National Center for Public Health
10:30–10:45	High-Resolution PM2.5 Mapping and Low-Exposure Route Optimization Using Geo-AI and Microsensors Yu-Hsuan LIN
10:45–11:00	Geo-spatial analysis and decision-Making Modeling to identify crime prevention zones in Ulaanbaatar JAVZANDULAM.B
11:00–11:15	Observation-Driven Agent-Based Evacuation Modeling with Time Resolved Wildfire Perimeters JongHwan KIM
11:15–11:30	GeoAI Framework for Urban Flood Resilience Integrating Sentinel 1 SAR Change Detection and H3 Index Location-Based Services AHMAD HABIBULLOH
11:30–11:45	Geo-AI-Based Assessment of 3-D Ultrafine Particle Distribution and Population Exposure: A Digital Twin Approach in Taichung, Taiwan Chia-Wei HSU

Back to time table

Session C3 (Chair: TBD)

10:15–10:30	Rockfall Prediction Revisited: Surprising Strength of Logistic Regression Over Ensemble Models Yun-Ching Chen
10:30–10:45	Automated Landslide Detection via Segment Anything Model Integrating Topographic and Vegetation Features Ling-Chien Kao
10:45–11:00	A Georeferenced Neural Radiance Field Framework for DSM Generation from UAV Imagery Cheng-Hsin LI
11:00–11:15	Investigation of Spatiotemporal Patterns and Associated Factors of Surface Deformation in Taoyuan (2018–2020) Tsung-Chun PENG, National Chengchi University
11:15–11:30	Mapping Taiwan’s Village Foodscapes, 2014–2022: A Bayesian Multilevel Spatial Analysis of Self-rated Health and Happiness Chia-Jung LIN
11:30–11:45	Wide-Area Subsidence Monitoring of Reclaimed Land in Osaka Bay Using Time-Series InSAR Haruka NARUKE

Back to time table

Session D3 (Chair: TBD)

10:15–10:30	Thermal Regulation across Fragmented Mangrove Patch Mosaics: Age- and Environment-Dependent Patterns from 30 Years of Landsat Data in Subtropical Taiwan Pei Wen LIN
10:30–10:45	Preliminary Evaluation of a Bio-Optical Model for Chlorophyll-a Estimation in Shallow Coastal Waters of the Seto Inland Sea Haoran ZHANG
10:45–11:00	SHIP DETECTION AND RE-IDENTIFICATION USING MULTI-TEMPORAL NANOSATELLITE IMAGERY IN BUSAN PORT Seunggi JEONG
11:00–11:15	Deep Learning-Based Vessel Velocity Estimation from Subaperture SAR Images Hyunsun LEE
11:15–11:30	Microsatellite SAR-Based Analysis of Ocean Wave Parameters Using Adaptive Azimuth Cutoff and Phase-Refocusing Methods Minkyeong KIM, Seoul Nat’l Univ.
11:30–11:45	Enhancing Coastline Detection from Sentinel-2 With Support from Sentinel-1 Data in Tongatapu, Tonga Sree Juwel Kumar CHOWDHURY

Back to time table

Session E3 (Chair: TBD)

10:15–10:30	Deep Learning–Based Urban Shadow Detection in True Orthoimages Integrating Multispectral and Height Information Chia-Ying LIN
10:30–10:45	Learning Representative Local Descriptors from Multiple Observations via Contrastive Enhancement Ya-Li LIN
10:45–11:00	An Intelligent Indoor Navigation System Using Visual Place Recognition and Epipolar Geometry for GPS-Denied Environments Ruei-Zeng WANG
11:00–11:15	Fixed Point–Based Image Registration Between Optical and SAR Satellite Images Hajime TAKAHASHI, Akita Univ.
11:15–11:30	STRUCTURED 3D CITY MODELING FROM MULTI-VIEW SATELLITE IMAGERY VIA SPARSE VOXELS RASTERIZATION Hyeonseong CHOI
11:30–11:45	Coarse-to-Fine Iterative Template Matching Using Foundation Model and CNN for Image Registration Gahyun Lee

Back to time table

Session A4 (Chair: TBD)

14:00–14:15	Cross-Sensor Super-Resolution of Himawari AHI Using SNPP VIIRS High-Resolution Imagery Shuma FUKUI
14:15–14:30	L-band SAR Coherence and Intensity Difference Analysis for Burned Area Extraction in the Large-scale Wildfire in Ofunato City Kou KITAGAWA
14:30–14:45	Early Wildfire Detection in South Korea Using Himawari-9 Short-Term Time-Series Mihyeon JEONG
14:45–15:00	Ensemble Stacking Learning for Forest Carbon Stock Estimation from Multi-Source Remote Sensing Data Yi-Chun CHENG
15:00–15:15	An Integrated Tool-Oriented Workflow for Automated Building Detection from Aerial Imagery Using TransUNet and RGB-DSM Fusion Pei-Cheng CHEN

Back to time table

Session B4 (Chair: TBD)

14:00–14:15	Feasibility study of estimating ice thickness distribution for landfast sea ice in Lützow-Holm Bay derived from satellite microwave altimetry data Yurina SATAKE
14:15–14:30	V-Mamba with Cyclic Temporal Encoding for Sea Ice Stage-of-Development Segmentation in SAR Imagery Dabeen Choi
14:30–14:45	Catchment-scale monitoring of subglacial lake activity in the David Glacier region, East Antarctica, using SBAS-DInSAR Taewook KIM
14:45–15:00	Reanalysis-Driven Sub-Seasonal Solar Radiation Forecasting with a 28-Day Lead Time Yejin Lee

Back to time table

Session C4 (Chair: TBD)

14:00–14:15	Landslide Prediction from Remote Sensing Imagery Using Semantic Segmentation: A Case Study in Northern Taiwan TZU-HSIANG HUANG, National Chengchi University
14:15–14:30	A Satellite-Driven Framework for Automated Identification of Suitable Industrial Development Sites via Land Use Classification and Machine Learning Ryota KUBOTA
14:30–14:45	MaxEnt-Based Spatial Risk Assessment for Mitigating Multi-Species Human-Wildlife Conflict in Japan Didier DELGORGE
14:45–15:00	BRIDGE DEFORMATION MONITORING USING THE TIME SERIES INSAR TECHNOLOGY Derrick Muga Mboya
15:00–15:15	Reconstruction of GRACE-Derived Terrestrial Water Storage Anomaly over the United States Using a Bayesian CNN: GLDAS versus NLDAS Ching-Mei, TSENG

Back to time table

Session D4 (Chair: TBD)

14:00–14:15	APPLICATION OF THE BLEACHING SUSCEPTIBILITY INDEX (BSI) AND BOTTOM REFLECTANCE ANALYSIS FOR SATELLITE-BASED CORAL BLEACHING ASSESSMENT IN TONGA Won-Jun CHOI
14:15–14:30	Leveraging satellite remote sensing for hydrodynamic models’ validation and scalable tidal stream resources quantifications ALJBER Morhaf
14:30–14:45	Operational Risk Mapping for Surveillance Prioritization of Suspected IUU Fishing Using AIS-Unmatched SAR Vessel Detections in the Korean EEZ Hyunji LEE
14:45–15:00	Satellite Estimation of Nutrient Variability Across the Seto Inland Sea Using Machine Learning: Assessing Algal Bloom Susceptibility in Osaka Ba Gandhi NAPITUPULU

Back to time table

Session E4 (Chair: TBD)

14:00–14:15	Deep Learning-Based Template Matching with Domain Adaptation for Cross-Domain Satellite–Aerial Image Matching Chanyeop JUNG
14:15–14:30	Integration of Physical Correction Model and LoG Sub-pixel Edge Detection for Image-based Rain Gauge Water Level Measurement Wei-Chi HO
14:30–14:45	A Benchmark Study of Amplitude Filters and Stability Indices for PSC Screening in PSInSAR Yan Akhbar PAMUNGKAS
14:45–15:00	A Graph Laplacian–Based Method for Radiometric Feature Identification in Photogrammetric Point Cloud Guan-Ling LIU
15:00–15:15	Occlusion Region Detection in High-Resolution Stereo Satellite Images Using Multi-Dimensional Relaxation Matching Hyeona KIM

Back to time table

Session B5 (Chair: TBD)

15:30–15:45	Research on Conditions for SAR Jamming Separation: An Approach Based on Phase Characteristic Differences Jaeho BAE
15:45–16:00	Generation of Disparity Maps in Satellite Stereo Matching Using Multi-Asymmetric Patches Chaeyeon LEE
16:00–16:15	Dynamic Collaborative Mapping and Autonomous Frontier Exploration in Heterogeneous Multi-Robot Systems JYUN-YANG HUNG
16:15–16:30	Mamba-Based Semantic Change Detection for Post-Typhoon Building Damage Assessment: A Case Study of Typhoon Danas (2025) in Tainan Pei-Yu WEN
16:30–16:45	Assessing the Vertical Effectiveness of Ground Vegetation on 3D Urban CO2 Distribution using UAV Vertical Sampling and Geo-AI Rui-Zhen YANG
16:45–17:00	Global Navigation Satellite System; Inertial Navigation System; Coarse Alignment; Tightly Coupled Integration; Unmanned Surface Vehicle JiunYo LIN

Back to time table

Session C5 (Chair: TBD)

15:30–15:45	Using LiDAR to Assist in Calculating Carbon Sinks in Plain Afforestation Areas Rui-En HUANG
15:45–16:00	SAR-based Phenology Alignment: A Time-Series Processing Strategy for Rice Growth Monitoring Chih-Chen TSENG
16:00–16:15	Downstream Discharge Prediction Based on Regression Models and LSTM Joon Hyuk CHOI
16:15–16:30	Interpretable Spectral-to-Thermal Framework for Fine-Scale Intra-Urban Thermal Variability from Multispectral Features and Pretrained Geospatial Representations Wonbin KANG
16:30–16:45	IMPROVING LANDSLIDE DETECTION RELIABILITY VIA RANKING LOSS WITH PHYSICALLY-PLAUSIBLE CONSTRAINTS Yeseok LEE
16:45–17:00	Improving Temporally Consistent Building Extraction from Multi-Temporal Satellite Imagery Donghyeon LEE

Back to time table

Session D5 (Chair: TBD)

15:30–15:45	Quantifying Living Environment and Urban Prosperity in Property Tax Assessment Using Landscape Images and Deep Learning Yoshito KOIZUMI
15:45–16:00	MS-Edge MIM: A Masking Strategy for Self-Supervised Learning on Very High-Resolution Satellite Imagery Changhui LEE
16:00–16:15	Enhancing Perceptual Quality and Geometric Features of SkySat Video via SinSR Diffusion Model Pin-han CHEN
16:15–16:30	Autonomous UAV with YOLOv11-Based Gripper System for Real-Time Papaya Ripeness Detection and Harvesting Han-Sheng HSU
16:30–16:45	Study on Building Symbol Extraction and Vectorization Methods from Historical Maps Lewei LIN
16:45–17:00	SimVP-GAN for GK2A Infrared Forecasting: Resolving the Accuracy–Sharpness Trade-off in Weather Satellite Remote Sensing Yeonjun KIM

Back to time table

Session E5 (Chair: TBD)

15:30–15:45	Applicability analysis of monocular depth estimation models for 3D information Inference from remote sensing imagery BAEK JUNSEO
15:45–16:00	Improving Visual SLAM Robustness in Low-Light Environments via Image Enhancement Techniques Yusuke ITO
16:00–16:15	A GIS-Based Multi-Objective Optimization Framework for Age-Sensitive Pedestrian Route Choice FEBRIAN FITRYANIK SUSANTA
16:15–16:30	Pan-Sharpening of KOMPSAT-3A Multispectral Images Using a Transformer-based Super-Resolution Model Jihyeon LIM
16:30–16:45	Building-Level Vacant House Detection Using Exterior Images and Deep Learning: A Cross-Regional Transferability Analysis Taro OKUMURA
16:45–17:00	Near-Real-Time Generation of IMERG-like Precipitation Using GK-2A Satellite Data and Generative Deep Learning Kyung-Hoon HAN

Back to time table

Session S1 (Chair: Hirokazu Yamamoto)

15:30–15:45	Long-term vicarious calibration for the GOSAT series at Railroad Valley Kei SHIOMI
15:45–16:00	Operational Calibration Update of the Advanced Land Observing Satellite-2 and -4 (ALOS-2/-4) Takeo TADONO
16:00–16:15	An Overview of CEOS Analysis Ready Data: Concepts, Implementation, and Future Perspectives Takeo TADONO
16:15–16:30	Extending Mirror Array-Based Co-Registration to High-Resolution Satellite Imagery: Impacts on Geolocation and Band-to-Band Accuracy Muhammad Daniel Iman bin HUSSAIN
16:30–16:45	Past Activities and Future Directions of Accuracy Assessment for Optical Satellite Data: A Case Study of the Phenological Eyes Network (PEN) Taiga SASAGAWA
16:45–17:00	Lunar calibration using a hyperspectral model, SP model Toru KOUYAMA

Back to time table

Session A6 (Chair: TBD)

09:00–09:15	Assessing the Role of Dynamic Opening in Sea Ice Loss Using Passive Microwave Radiometer Observations Jeong-Won PARK
09:15–09:30	Long-Term Regional Variability of Arctic Ocean Color under Changing Sea Ice Conditions Joo-Eun YOON
09:30–09:45	Analysis of long-term grounding line positional variation on Shirase Glacier using SAR interferometry Tsutomu YAMANOKUCHI
09:45–10:00	Ice flow velocity in downstream region of Shirase Glacier related to ice thickness change derived from remotely sensed data Yurina SATAKE
10:00–10:15	Permafrost monitoring in eastern Siberia and Alaska: Research findings using ALOS-1/-2 SAR data and preliminary results from ALOS-4 Takahiro ABE

Back to time table

Session B6 (Chair: Hirokazu Yamamoto)

09:00–09:15	Tree Species Classification of Major Korean Forest Species Using Drone-borne LiDAR Point Cloud Data and Deep Learning Models Jin-Taek Kang
09:15–09:30	Indoor Disaster Investigation using Multi-sensors Built-in Ground Robot Platform and AI Tech. Seongsam KIM
09:30–09:45	Vicarious Calibration of Terra ASTER VNIR using RadCalNet RVUS BOA Reflectance Validated by JAXA GOSAT Campaigns Hirokazu YAMAMOTO
09:45–10:00	Rapid 3D Terrain Reconstruction from a Single Satellite Image Ping-Yuan Yang

Back to time table

Session C6 (Chair: TBD)

09:00–09:15	InSAR-Based Multi-Decadal Investigation of Urban Land Subsidence: Divergent Policy Influences in Jakarta and Bangkok Wahyu Luqmanul HAKIM
09:15–09:30	Surface Deformation Before, During, and After the 2025 Krasheninnikov Eruption, Kamchatka, Russia, from Sentinel‑1 Time‑Series InSAR Wahyu Luqmanul Hakim
09:30–09:45	Nationwide near-surface ground deformation mosaics from multi-year synthetic aperture radar observations across Canada and North America Sergey SAMSONOV
09:45–10:00	Landslide Detection via a U-net-Driven SAR2OPT Approach: Application to the 2024 Hualien Earthquake Wen-Hong CHEN
10:00–10:15	Susceptibility Mapping of Residual Landslide using Multitemporal Imagery and Machine-Learning Methods Tung Cheng, LU

Back to time table

Session D6 (Chair: Masayuki Matsuoka?)

09:00–09:15	Optical Image Reconstruction in Cloud-Contaminated Areas Using Multi-Temporal SAR Data and LSTM Models Yu-Ching WANG
09:15–09:30	An Efficient CNN-Transformer Network for Multitask Learning-Based Semantic Change Detection Aisha JAVED
09:30–09:45	Evaluating GEO–LEO BRDF consistency using Himawari AHI and SNPP VIIRS data Masayuki MATSUOKA
09:45–10:00	A Post-Processing Framework for Occlusion-Robust Multi-Object Tracking in Satellite Video Guided by Kalman Filtering and Single-Object Tracking Po-Wei TSAI
10:00–10:15	Integration of UAV-Derived DSM and Visible Imagery for Terrain and Surface Feature Classification in Open-Pit Mining Areas Wei-Tsun LIN

Back to time table

Session E6 (Chair: TBD)

09:00–09:15	Three Dimensional PM2.5 Characterization in Chiang Mai using Pixel Aligned Himawari 9 AOD Data Katanchalee TAWEEPORNWATTANAKUL
09:15–09:30	Multi-Sensor Validation of the GOSAT-2 CAI-2 Cloud Flag and Developing Targeted Spectral Tests for Missed Clouds Reza SAFABAKHSH
09:30–09:45	Physically-Consistent Cloud Index Sequence Forecasting over the Korean Peninsula Using GK-2A Geostationary Satellite Imagery and a Spatiotemporal Transformer Dongjin KIM
09:45–10:00	Synergy of Sentinel-5P datasets and hybrid CNN-LSTM architecture in assessing Urban Air Quality Hugo Wai Leung MAK

Back to time table

Session B7

10:30–10:45	Spatio-temporal Variation of Human Flow: Investigating the Impact of Space Syntax Metrics via Mobile Phone Data Yun-Zhen CHEN
10:45–11:00	A Geometry-Driven Topology Restoration and Semantic Decoupling Framework for Normalizing OSM-Derived OpenDRIVE Data in Early-Stage Autonomous Vehicle Testing Ogi Dani SAKAROV
11:00–11:15	Synergistic Integration of UAV-LiDAR and 3D Gaussian Splatting for Precision Tree Species Classification and Biomass Estimation in the Tropical University Campus Daisuke SASAKI
11:15–11:30	Development of Spatially Enabled Web Maps and Dashboards Javzandulam BATAA
11:30–11:45	How Conversational AI Reconfigures Retail Geography: Evidence from a GIS-Based Agent Simulation I Gede Brawiswa PUTRA

Back to time table

Session C7 (Chair: TBD)

10:30–10:45	Infrastructure Monitoring in South Korea Using Time Series InSAR: Insights from the K-GMS Platform Khin Myat KYAW
10:45–11:00	Satellite-Echo-Sounder Mapping of Navigation Bottlenecks in the Mekong-Tonle Sap-Bassac Corridor, Cambodia BUNNARATH ouch
11:00–11:15	Detection and Size Estimation of Riverbed Boulders in Mountain Streams from UAV Imagery Using Deep Learning and GIS-Based Automated Labeling Sanae KANG
11:15–11:30	Assessing Dry Port Spatial Efficiency via Geometric Layout Indicators from Remote Sensing Imagery SUN YIRAN
11:30–11:45	Accuracy Evaluation for the Application of Portable LiDAR in Disaster Site Yonghan JUNG

Back to time table

Session S2 (Chair: Yuji Sakuno)

10:30–10:45	Local Adustment of Satellite Chlorophyll-a in Coastal Waters using Machine Learning Techniques: Case of Ise Bay Joji ISHIZAKA
10:45–11:00	Detection of the Distribution Area of Discolored Water Caused by Submarine Volcanoes Mitsuhiro TORATANI
11:00–11:15	Detection of characteristics of time-series in SGLI/GCOM-C observed Chlorophyll-a concentration by information mathematics NAGAI Shin
11:15–11:30	Spatiotemporal Variations of Sea Surface Chlorophyll-a in the Seto Inland Sea, Japan Han Soo LEE
11:30–11:45	Characteristics of satellite water temperature and water quality distribution related to the mass oyster death around Kure City in 2025 Yuji SAKUNO

Back to time table

Poster Session Day 1

P-1	Comparison of field-measured tree heights and canopy height model with machine learning and topographic factors for modelling diameter at breast height Jhe-Syuan LAI
P-3	A Comparative Study of Paired and Unpaired GANs for Optical Image Generation from Polarimetric SAR Hirofumi SHIMAOKA
P-5	A Methodological Study of Geomorphological Survey Approaches for the End Moraine of Imja Glacial Lake Susumu TAKAGISHI
P-7	Improving RGB-Based Discrimination of Japanese Cedar and Cypress through Temporal Optimization Considering Inter- and Intra-Class Variability Yuki SOFUE
P-9	Estimation of Rice Cultivation Water Requirements and Irrigation Water Surplus or Deficit Using Sentinel-2 Satellite Data Chiharu HONGO
P-11	Monitoring Nickel Mining Expansion in Northern New Caledonia using Multi-temporal Sentinel-2 Imagery and Random Forest Classification Yuki SANO
P-13	Damage Assessment Model for Bacterial Leaf Blight in Rice by Integrating Flooding duration Using Sentinel-1/2 Data Takayuki MORI
P-15	Slope Collapse Risk Assessment Based on Topographic Factors in the 2011 Kii Peninsula Heavy Rainfall Disaster Mariko TERAUCHI
P-17	Classification of Flooded and Non-Flooded Paddy Fields Using High-Resolution QPS-SAR Data and Machine Learning Tomoyuki HIROSE
P-19	An Iterative Hypothesis Verification-Based Multi-Agent Framework for Object Activity Inference in High-Resolution Satellite Imagery Hwang junesik
P-21	VALIDATION OF SWOT SATELLITE WATER SURFACE ELEVATION OBSERVATIONS IN THE HAN RIVER BASIN, SOUTH KOREA Keon Ho KIM
P-23	PRELIMINAY STUDY ON GEOSTATIONARY MULTISPECTRAL OBSERVATIONS FROM GOCI-Ⅱ FOR WILDFIRE CHANGE DETECTION Kyung Hwan Shin
P-25	A Cadastral-Aware U-Net Semantic Segmentation Framework for High-Resolution Aerial Crop Mapping in Heterogeneous Agricultural Landscapes Wei-Tsun LIN
P-27	Water Vapour Correction for Improving Landsat-9 Single-Channel LST Retrieval over Northern Taiwan Tri Anggun LESTARI
P-29	Regression-Based Cloud Gap Filling for Sentinel-2 L2A Time Series Masking and Robust Radiometric Harmonization Donggyu KIM
P-31	Localized Grounding-Line Retreat of the Nansen Ice Shelf, East Antarctica, and Its Implications Minji Cho
P-33	Conceptual Design of a UAV-Based Sampling System for Remote Investigation of Tidal Flat Sediments YeongJae JANG
P-35	Secular and Seasonal Vertical Deformation in Taiwan from Long-Term GNSS Time Series Yen-Ru Lai
P-37	Structural-Type and Construction-Year Differences in Building-Level Sentinel-2 Reflectance Koji OGINO
P-39	THIN SEA ICE THICKNESS ESTIMATION USING COMBINED OPTICAL AND SAR DATA Hiroyuki WAKABAYASHI
P-41	Mapping tidal flats using Sentinel-1 and ALOS-2 in the Ariake Sea Tatsuyuki SAGAWA
P-43	Deep Learning-Based Multi-Satellite Sea Fog Detection via Precision Co-Registration and Transformer Segmentation using GK2A AMI and GK2B GOCI-II Jonggu KANG
P-45	A Remote Sensing Framework for Shallow Bathymetry Retrieval from SAR Using Adaptive FFT Windowing Dedy ZULKHARNAIN
P-47	Feasibility of Arctic Iceberg Monitoring Using NEXTSat-2 and the DINOv3 Foundation Framework Seung Hee KIM
P-49	Development of a Satellite-Driven Marine Big Data Framework for Maritime Accident Search and Rescue Support Tae-Sung KIM
P-51	REMOTE SENSING FOR MARITIME SEARCH AND RESCUE IN COASTAL ENVIRONMENTS Jae-Jin PARK
P-53	MONITORING COASTAL TURBIDITY AND UNDERWATER VISIBILITY USING IN-SITU SENSORS AND SATELLITE OPTICAL DATA Yongmyung KIM
P-55	Feasibility study of estimating surface salinity distribution in the Seto Inland Sea using Landsat data Yuji SAKUNO
P-57	Lightweight CNN Ship Detection from Geostationary Ocean Color Imagery (GOCI-II) Won-Kyung BAEK
P-59	VALIDATION OF ICESAT-2 COASTAL BATHYMETHRY PRODUCT IN SHALLOW TROPICAL WATER REGION Anisya Feby EFRIANA
P-61	Assessing Thermal Discharge Impacts from the Hanbit Nuclear Power Plant on Coastal Marine Environment Using GOCI-II Satellite Observations Sungjae PARK
P-63	Assessing Stratus Cloud Variability and SST Coupling Before and After IMO 2020 using Sentinel-2 Hyojin KIM
P-65	Integrating Remote Sensing and Habitat Modeling to Quantify Social–Ecological Impacts of Artificial Mangrove Expansion Su-Fen WANG
P-67	A Deep Learning Approach for Remote Sensing-Based Estimation of PM2.5 Concentration in Urban Areas: A Case Study of Greater Taipei Area Vincent Cletus WISO
P-69	Spatial Assessment of Future Heat Stress Exposure in National Parks of the Republic of Korea under CMIP6 SSP Scenarios Using WBGT Jinhoo HWANG
P-71	CLOUD REMOVAL USING BI-TEMPORAL OPTICAL IMAGERY AND GENERATIVE DEEP LEARNING Soyeon PARK
P-73	Regression-based error refinement for cloud-contaminated regions reconstructed via SAR-to-optical image translation Inseon LEE
P-75	Enhancing the Spatial–Spectral Fusion of Multispectral and High-Resolution Optical Satellite Imagery for Spectral Enrichment of Fine-Scale Observations Yeseul Kim
P-77	Detecting Inshore Objects Occupying Public Waters Using SAR Dual-polarization Seung Jun Lee
P-79	Development of an Algorithm for Estimating Near-Surface Specific Humidity for GPM/GMI -Effectiveness of high-Frequency Channels- Yuta KYONO
P-81	Correction of Elevated Road Distortion in Optical Satellite Orthoimagery using Integrated Elevation Models Chien-Liang LIU
P-83	Comparative Analysis of R-CNN-Based Detection Performance for Vehicle in High-Altitude UAV Imagery Seungchan LIM
P-85	Benchmarking Crack Segmentation Models on Public Datasets with UAV Field Tests and Imaging Conditions Jaehoon JEONG
P-87	Development of a SAR K-ARD Module to Standardize KOMPSAT Satellite Imagery Su-Bin HA, Pukyong National University
P-89	Reversible Compression Framework for Streaming 3D-LiDAR Point Clouds Using Multi-Layer Range Image Processing Masafumi NAKAGAWA
P-91	A Framework for Quality Classification and Enhancement of MMS Image Data Ryunghyeok IM
P-93	Retrieval-Augmented Multimodal Learning for Robust Plant Disease Classification Yuan-Chia CHAN
P-95	Monitoring Spatio-temporal Land Cover Changes in Kangerlussuaq, Greenland, Using DeepLab v3+ and Time-series Satellite Imagery Sungjae LEE
P-97	Methodological Framework for Estimating Livestock Mortality: SWOT and Problem Tree Analysis, and Correlation–Regression Analysis Otgonjamts GOMBOJAV
P-99	Building and Road Change Detection from Heterogeneous KOMPSAT and NEONSAT Imagery Han OH
P-101	Systematic Evaluation of Geospatial Foundation Model Transfer for High-Resolution Land-Cover Segmentation in NEONSAT Imagery Juyoung Kim
P-103	Evaluating the effectiveness of continual pre-training of foundation model for analyzing small SAR satellite data Toru KOUYAMA
P-105	Machine Learning–based Spatial Classification of High-Risk Wildfire Areas in the Forest–Urban Interface Seunghyun Hong
P-107	Tracking Invasive Alien Species in Northeast Asia: An Ensemble Modeling Approach Youngjae YOO
P-109	Watershed-scale Characterization Typology for Wildfire Management Sangwook LEE
P-111	Integrating DEM‑Based Wetness and High‑Resolution Imagery in a GIS Workflow for Agricultural Drainage Prioritization Yuta SHIMIZU
P-113	Changes in the shortest routes and distances for four industries in Miyagi Prefecture Hiroto WAKI
P-115	Accuracy assessment of DEM derived from SpaceEye-T1 tri-stereo imagery Youngran MOON
P-117	Geomorphic Interpretation of Korean Tidal Flats Based on Cross-Validation of Sentinel-1 SAR Backscatter-Derived and TanDEM-X InSAR DEMs Hanah PARK
P-119	JAXA’s L-SAR MISSIONS AND THOSE PPP PROJECTS Shinichi SOBUE
P-121	Spatial Resolution Impact Assessment for Long-term Population Monitoring of Adélie Penguin Guano Jinku PARK
P-123	A Study on the Operational Status of a Nuclear Facility in an Inaccessible Area using Time-Series Analysis of Multi-Optical Satellite Imagery Gayeon HA

Poster Session Day 2

P-2	Floristic Diversity and Spectral Characterization of Vegetation in Tormik Valley, District Skardu, Pakistan: A Field-Based and Sentinel-2 Remote Sensing Approach Naseema KHAN
P-4	Evaluating YOLO-Based Object Detection for Automated Landslide and Debris-Flow Damage Assessment Using Post-Disaster Imagery Kosei UENO
P-6	Development of a Low-Cost Meteorological Satellite Receiving Station for Meteor-M Imagery Tzu-Yu CHEN
P-8	Regional Ionospheric Characterization via Multi-point Signals of Opportunity and Distributed Digital Beacons Randson Huang
P-10	Voxel Entropy Applications for Individual Tree Detection in High-definition Forest Point Clouds Kasparas KARLAUSKAS
P-12	Comparative Evaluation of Supervised Classification Algorithms for Land Use Mapping in Support of Asian Elephant Habitat Modeling Zichen MIN
P-14	Wildfire Burned Area Mapping Using Early-fusion Deep Learning Change Detection Based on U-Net Model with Sentinel-2 Imagery Amalia Gita AYUDYANTI
P-16	Radiometrically Preserving Super-Resolution of Coarse-Resolution Satellite Imagery for Post-Disaster Damage Mapping Angel MATOS
P-18	Assessment of Damage Detection on the Metropolitan Expressway Caused by the 2011 Great East Japan Earthquake Using TerraSAR-X Imagery Masaya KUBOTA
P-20	Backscattering and Texture Characteristics of ALOS-2 PALSAR-2 Images for Buildings Damaged by the 2016 Kumamoto Earthquake Yuriko ASAKA
P-22	Spatial Extension of Tetracorder-Derived Mineral Signatures to Multispectral Sensor Coverage in Cuprite, Nevada Hideki TSUBOMATSU
P-24	Initial Evaluation of a Cross‑Sensor Reflectance Transformation Algorithm Using Landsat 7 and 8 Time‑Series Data at Flux Sites Yuma HORIKAWA
P-26	Classification Criteria for burned areas in Punjab, India, using Sentinel-2/MSI data Nana MIZOGUCHI
P-28	Estimation of Mountain Slope Deformation Using Time-Series InSAR Analysis with ALOS-2/PALSAR-2 Masaya KOIKE
P-30	TIME-SERIES ANALYSIS FOR URBAN CHANGE MONITORING BASED ON MULTI-TEMPORAL HIGH-RESOLUTION NANOSATELLITE IMAGERY Taehong KWAK, Seoul National Univ.
P-32	Comparative Evaluation of Direct and Hierarchical Machine Learning Models for Forest Volume Estimation Hiroyuki AOKI
P-34	Dual-Polarization SAR Analysis for Rice Growth Monitoring: Correlation Between Sentinel-1 Backscatter and UAV-Derived Plant Height Yudai TANAKA
P-36	Survey of Solar Panel Distribution Using Sentinel-2/MSI Shiori FUJII
P-38	Deep Learning–Based Landslide Mapping Using Sentinel-1 SAR Time Series and DEM-Derived Features Byunggun KIM
P-40	Applicability of Idle Farmland Category Discrimination Using Random Forest with Satellite Imagery and GIS Data Konomi OTOMO
P-42	Observation of Typhoon Cold Wake Recovery Time in the Waters Surrounding Taiwan Using Himawari 8/9 Yung-Szu LIANG
P-44	Statistical Analysis of Anomalous Transitions in the Kuroshio Surface Currents Hsiang LEE
P-46	A pixel-Level Comparison of Satellite and Commercial Aircraft Observation Frequencies for Coral Reef Monitoring in the Nansei Islands, Japan Sora TAKAMIYA
P-48	Influence of Southern Landfast Sea Ice on the Flow Variability of Drygalski Ice Tongue in East Antarctica Inferred from Sentinel-1 SAR Imagery Junbeom PARK
P-50	Classification of Landfast Sea Ice in Terra Nova Bay, East Antarctica, Using Sentinel-1 SAR Backscattering and Interferometric Coherence Hyewon SONG
P-52	Measurements of PM10 Spatial Distributions in Seoul using High Spatial Resolution Aerosol Scanning LiDAR Kwanuk Kye
P-54	Automated Traffic Sign Recognition and Photogrammetric Localization for Large-Scale Surveys Hsin-Hung HU
P-56	A method for estimating levee crest heights using AW3D enhanced DEM products Takeru FURUICHI
P-58	Robust Phase Linking for SAR Interferometric Phase Estimation Youngjae CHAE, Sejong Univ.
P-60	Cross-Regional Generalization of a Deep Learning Model for Short-Term Spatiotemporal Population Distribution Forecasting Using Mobile Spatial Statistics Ryuzo TAKAHASHI
P-62	Moving Target Velocity Estimation Using Single-Channel SAR Imagery Juyoung SONG
P-64	Automated Measurement of Intensity Quality in MMS LiDAR Data Seungmin NOH
P-66	Evaluation of the Range–Doppler Model for Capella Spotlight Imagery Across Different Imaging Formations Tee-Ann TEO
P-68	Pre-launch Wavelength and Radiometric Calibration of Hyperspectral Sensors Using an Offner Spectrometer Yuya KOAKUTSU
P-70	Automated method for assessing cyanobacterial levels using image classification based on deep learning model with RGB images Manabu WATANABE
P-72	Benchmarking FAST-LIO2 and LIO-SAM SLAM Systems in Complex Urban Environments Yen-Ru Lai
P-74	Segmentation-Based NN-PSI with Change Point Detection for Nonlinear Deformation Reconstruction Ho-Yeong YOU, Sejong Univ.
P-76	Integrating Multi-temporal SAR Features and Ensemble Feature Selection for High-Precision Forest Type Classification Geng Gui WANG, National Chung Hsing University
P-78	Autonomous relay based navigation and high speed data relay system for satellite constellations Sang-Youn SHIN
P-80	On-board Edge Computing Framework for Real-time Rice Lodging Detection via Multi-task Learning and Multispectral Imagery Hsin-Hung TSENG
P-82	Benchmarking SpectralGPT Against Traditional OBIA-SVM for Solar Site Selection in Fragmented Indigenous Terrains. Hsuan-Yin HUANG
P-84	Bridging Taxonomy Gaps: A Label-Invariant Fusion Strategy for Multi-Dataset High-Resolution Aerial Segmentation Jae Young CHANG
P-86	Case study of the Influence of upstream oil palm expansion on flooding in an Aceh catchment, Indonesia Chihiro NAITO
P-88	Quantifying Multi-Temporal Domain Shift in UAV-Based Sweet Potato Plant Localization Using YOLO Yu-Chun HSU
P-90	Building Damage Mapping Using Dual-Orbit DInSAR and SAR–Optical Data for the 2024 Noto Earthquake Case Study Yifan YANG
P-92	Long-Term Precision Assessment of ICESat-2 Laser Altimetry over the Greenland Ice Sheet (2018–2025) An-Ting Chang
P-94	Wildfire Burn Severity and Scarring in Protected Areas Using Sentinel-2 dNBR for Rapid Post-Fire Mapping: Evidence from the 2024 Belize Wildfires Rosalia DOMINGUEZ
P-96	Assessing Urban Thermal Environments through LCZ Mapping and Satellite-Derived LST Validation of Microclimate Models: A Case Study of Chiayi City Liang-Yun KU
P-98	Integrating MAE-SegFormer and stacking ensemble regression for precision tea tree physiological monitoring based on UAV multispectral imagery Zhong-Han ZHUANG
P-100	Evaluating Arctic Tundra Vegetation Properties Using in-situ Remote Sensing Observation in Svalbard Hibiki NODA
P-102	Leveraging SWOT Observations for Monitoring Inland Water Dynamics in Taiwan Kuo-Hsin TSENG
P-104	Discriminating Snowfall Effects in Post-Earthquake Damage Detection Using Polarimetric Coherence: A Case Study of the 2025 Aomori Prefecture Eastern Offshore Earthquake Fumitaka OGUSHI
P-106	Toward Reliable Bridge Infrastructure Monitoring: Demonstrating the Effectiveness of Complex-Valued SAR Simulation Fumitaka OGUSHI
P-108	Retrieval-Optimized Matched Filtering of Sentinel-2 SWIR Observations for Single-Acquisition Methane Leak Detection and Quantification in Biogas Facilities Harika NULU
P-110	Integrating SAR and Optical Satellite Imagery for Monitoring Land Reclamation and Surface Deformation at a Commercial Port Chia-Che CHANG
P-112	Automatic flood detection using ALOS-2 and ALOS-4: algorithm updates and operation results Masato OHKI
P-114	Landslide detection using Sentinel-1 and ALOS-2, a case study of Noto Earthquake, Japan Shouhao CHIANG
P-116	Establishment of an Intelligent Data Platform for Detecting Nuclear Activities: A Multilingual Ontology and OSINT-GEOINT Fusion Framework Yonhong JEONG
P-118	Spatiotemporal Analysis of Forest Cover Dynamics in Darkhan-Uul Province, Mongolia (2018–2021) Ochirkhuyag LKHAMJAV, National Central University
P-120	Accuracy Validation of ALOS-4 InSAR for Landslide Deformation Monitoring: A Case Study of the Koshio Landslide, Japan Seiya USAMI
P-122	Modeling Land Use and Land Cover Change Dynamics and Prediction Using Machine Learning and QGIS-MOLUSCE Enkhzaya ENKHTAIVAN
P-124	Comparison of multiple strategies for detecting landslide affected areas with optical remote sensing and object-based image analysis Jhe-Syuan LAI

Session A1 Abstracts

A1-1: Validation of Multi-mission Altimetry around the Korean Peninsula Using Tide Gauges

Presenting author: Do-Hyun HWANG

Authors: Do-Hyun HWANG, Yonsei Univ.; Hahn Chul JUNG, Yonsei Univ.; Seo Jin KIM, Yonsei Univ.; Hyongki LEE, Univ. of Houston

Keywords: Satellite Altimetry; Coastal Sea Level; Tide Gauge; Sea Level Anomaly

Coastal sea level monitoring using satellite altimetry is challenged by land contamination and distortion of the return waveform near the coast. These effects reduce both data accuracy and availability. The East Sea, Yellow Sea, and South Sea of Korea surrounding the Korean Peninsula each exhibit distinct oceanographic characteristics, motivating region-specific evaluations of coastal altimetry performance. In this study, we assessed the agreement between sea surface height (SSH) estimates from multiple satellite missions (Jason-3, Sentinel-3A/B, and SWOT) and tide gauge observations around the Korean Peninsula. Prior to the comparison, satellite SSH measurements were converted to sea level anomalies (SLA) by removing a mean sea surface reference, and tide gauge data were treated consistently as anomalies. We focused on the 0-20 km coastal zone to characterize nearshore sea level variability. One-minute tide gauge data were time-matched to satellite overpass times to minimize temporal mismatch. The coastal zones were divided into 0-5 km, 5-10 km, 10-20 km bins, and performance was evaluated using root mean square error (RMSE) and correlation. Based on tide gauge comparisons, we quantified the effective nearshore range over which each altimeter provides reliable SLA estimates and examined retracking approaches for detecting robust coastal sea level signals. These results offer practical guidance for validating coastal SLA products around the Korean Peninsula and provide a baseline for future coastal sea level monitoring.

Back to session list

A1-2: Finding the Optimal data usage of synthetic data on Satellite image ship detection Task

Presenting author: Yan-Chen HUANG

Authors: Yan-Chen HUANG, TASA; Li-Yu CHANG, TASA

Keywords: Synthetic Data, FORMOSAT Imagery, Vessel Detection, Domain Gap, YOLO Object Detection

Synthetic data enables training deep learning model even when real-world data are scarce. With synthetic data, user can potentially generate unlimited amount of data used for training. However, the main limitation is that the appearance of synthetic data does not always match the level of real data. The above problem is called “Domain Gap”. It could reduce the performance of deep learning model. Hence, there is a necessity to find out how much synthetic data should be used in order to achieve better results. Satellite image vessel detection aims to locate vessel in various sizes, types, and formations. Although satellites produce large volumes of imagery every day, certain rare situation cannot be reliably captured in real data. Synthetic data allows users to generate customized satellite images with controlled vessel configurations, which is useful for detecting illegal or uncommon vessels. In this research, we create our own synthetic satellite imagery dataset based on FORMOSAT data and design two experiments using YOLO to investigate optimal synthetic data usage. The first experiment limits the total amount of training data, comparing results with different ratios of real to synthetic data. The second experiment fixes the amount of real data and gradually adds synthetic data to test if large quantities of synthetic data might harm performance. Our first experiment shows that training with mixed real and synthetic data yields an approximate 2–4% improvement in precision and a 1–6% increase in mAP50 compared to training with only real data. The second experiment indicates that when synthetic data is added up to four times the size of the real dataset, precision remains better than using only real data, although the mAP50 score returns to a similar level as the real-only baseline.

Back to session list

A1-3: A Multispectral Fusion Framework Integrating U-Net and SWINIR for Coastal Mangrove Monitoring in Taiwan

Presenting author: yiwen LO

Authors: Yiwen LO, Dept. of Civil Engineering, National Yang Ming Chiao Tung University; Tzuyi CHUANG, Dept. of Civil Engineering, National Yang Ming Chiao Tung University

Keywords: Mangrove detection, Multispectral Fusion, Sentinel-2 L2A, Deep learning, Vegetation indices

With the growing global awareness of environmental protection, blue carbon has gained increasing attention in recent years. Mangrove wetlands, as key blue carbon ecosystems, play a crucial role in climate adaptation and biodiversity conservation, and their conservation value has been widely recognized internationally. In Taiwan, mangroves are mainly distributed along western estuarine regions, including Taipei, Tainan, and Kaohsiung, with Tainan City hosting the most extensive mangrove coverage. Traditional mangrove monitoring relies heavily on field-based surveys, which, although relatively accurate, are constrained by high labor and time costs, limiting their applicability for large-scale and high-frequency monitoring. In contrast, remote sensing techniques enable rapid and large-area observation and have become one of the most widely adopted approaches for mangrove monitoring. Among current optical satellites, SPOT-6 imagery offers high spatial resolution; however, its limited accessibility and high acquisition costs hinder its use for long-term, large-scale mangrove monitoring. Conversely, Sentinel-2 L2A imagery is freely available, offers high temporal resolution, and provides extensive spatial coverage, making it a primary data source for mangrove observation. Nevertheless, its surface reflectance values are sensitive to sensor characteristics and regional atmospheric conditions, limiting its accuracy in Taiwan’s coastal environments. To enhance the regional spectral applicability of Sentinel-2 imagery, this study proposes a deep learning–based multispectral fusion framework. Calibrated SPOT-6 imagery from the Center for Space and Remote Sensing Research (CSRSR) at National Central University is used as a regional spectral reference. By integrating U-Net and SWINIR deep learning models, a multispectral fusion model is developed to transform Sentinel-2 imagery into a version that better reflects local spectral characteristics. The proposed approach effectively improves the accuracy of vegetation indices, including NDVI and MVI, and enhances the reliability of mangrove class detection.

Back to session list

A1-4: An Automated Hybrid Framework for Surface Water Extraction: Integrating Spectral-Morphological Techniques with SatMAE

Presenting author: Ting-Yu,LAI

Authors: Ting-Yu,LAI; Fang-Shii,NING

Keywords: Surface water extraction;spectral water indices;mathematical morphology;SatMAE;pseudo-labeling;change detection

Surface water monitoring is essential for hydrological modeling and disaster management. Although optical remote sensing allows for large-scale observation, extracting water bodies accurately is often difficult due to environmental noise like terrain shadows, urban structures, and cloud cover. Previous studies show that combining spectral water indices (e.g., Normalized Difference Water Index [NDWI], Modified Normalized Difference Water Index [MNDWI]) with mathematical morphology can filter out this noise and preserve geometric features. While these analytical methods provide highly reliable, physically grounded extractions, their reliance on visible optical pixels inherently limits their ability to reconstruct fragmented water targets under heavy cloud obscuration. Furthermore, while deep learning models can address this reconstruction issue, training them requires pixel-level annotations, which are costly and scarce. To address the lack of training data and extend the capabilities of established physical rules, this ongoing research proposes an automated framework that integrates physical remote sensing algorithms with SatMAE, a foundation model based on masked autoencoders. Established spectral and morphological techniques are utilized as a baseline to automatically generate pseudo-labels from multispectral satellite imagery. These pseudo-labels are then used to fine-tune a pretrained SatMAE model for water segmentation tasks. By utilizing SatMAE’s multispectral representations and masked patch reconstruction, the proposed framework aims to fill the observation gaps caused by cloud fragmentation and shadow misclassification. It is expected to improve the accuracy of temporal change detection in complex river systems, offering an automated approach for high-resolution remote sensing applications.

Back to session list

A1-5: Estimating tidal flat erosion rates based on landward migration of concentrated tidal channel zones and their controls

Presenting author: Jiyoung LEE

Authors: Jiyoung LEE, Korea Ocean Satellite Center, Korea Institute of Ocean Science & Technology (KIOST); Keunyong KIM, Korea Ocean Satellite Center, Korea Institute of Ocean Science & Technology (KIOST); Hoi-Soo JUNG, Marine Domain & Security Research Department, Korea Institute of Ocean Science & Technology (KIOST); Sungan KIM, Department of Geosciences, University of Bremen, Bremen, Germany; Joo-hyung RYU, Korean Ocean Satellite Research Center, KIOST, Republic of Korea

Keywords: Cheonsu Bay; Tidal flats; Sediment Starvation; Tidalflat Erosion; Tidal Creek Migration; Airborne LiDAR DEM

Tidal flats on Korea’s west coast exhibit spatially variable erosional and depositional tendencies that depend on local environmental settings, including tidal range, basin geometry, hydrodynamic forcing, and sediment availability. However, for many macrotidal systems in the Yellow Sea, the relative importance of these controls remains insufficiently quantified. Here, we investigate the Cheonsu Bay tidal flat in the Yellow Sea, Korea, a semi-enclosed embayment modified by reclamation, to quantify recent erosion and diagnose geomorphic indicators that constrain dominant controls. We derived a 0.5-m DEM from October 2023 airborne LiDAR and calculated slope gradients using a central-difference algorithm. Cross-sections extracted along representative transects were compared with oyster reefs, shelly sand bars or cheniers, and boundaries between high- and low-density tidal creek zones mapped from high-resolution imagery. Landward migration was measured by delineating (i) concentrated tidal-creek zones and (ii) shelly-bar ridges or cheniers. For each mapped line, we quantified the landward migration distance, D, from multi-temporal low-tide imagery and measured from the most recent mapped position to the low-tide shoreline. We then converted horizontal migration to vertical lowering using a geometric relationship constrained by the local tidal-range height, yielding an estimate of annual erosion rate of ~49 mm yr⁻¹. Consistent with this, slope profiles reveal abrupt breaks that coincide with creek-density transitions and chenier/shelly-bar occurrence, and progressively develop concave-upward shapes consistent with erosion-dominated tidal-flat geometry. This magnitude is too large to be explained by regional relative sea-level rise or typical tidal-flat subsidence alone, implying that human modification of the system—particularly river regulation, dikes, and reclamation—has reduced sediment supply, driving sediment starvation and disrupting the tidal-flat sediment budget. These findings indicate sustained erosion in Cheonsu Bay that has measurably altered tidal-flat morphology, underscoring sediment supply as a significant control that must be accounted for in coastal management and restoration.

Back to session list

Session B1 Abstracts

B1-1: QueryGIS: A Trustworthy Conversational GeoAI Agent for QGIS Workflow Automation with RAG and Self-Repair

Presenting author: Juseong LEE

Authors: Juseong LEE, Image Eng. Research Center, 3DLabs Co. Ltd, Republic of Korea; Byeongjun HYUN, Image Eng. Research Center, 3DLabs Co. Ltd, Republic of Korea; Won-Woo SEO, Image Eng. Research Center, 3DLabs Co. Ltd, Republic of Korea; Sooahm RHEE, Image Eng. Research Center, 3DLabs Co. Ltd, Republic of Korea

Keywords: AI, GeoAI Agent, LLM, QGIS, PyQGIS, Workflow Automation, RAG

With the rise of large language model (LLM) services such as ChatGPT and Google Gemini, GIS and remote sensing users increasingly express workflows in natural language (e.g., raster clipping, index calculation, and zonal statistics). However, LLM-generated scripts are often fragile because they depend on project-specific state—layers, fields, and CRS settings—leading to repeated manual fixes. This matters in QGIS, a widely used desktop GIS, where many remote-sensing users find workflow steps and scripting hard to use. This paper presents QueryGIS, a conversational GeoAI agent implemented as a QGIS plugin–server system that converts user requests into executable PyQGIS/Processing workflows using retrieval-augmented generation (RAG), self-repair, and systematic logging. When a user submits a request in the plugin UI, QueryGIS builds a structured JSON context from the current project state, including layer types, CRS, providers and sources, geometry types, feature counts, extents, and short attribute samples; for raster layers, it includes extent corner coordinates, band counts, and pixel dimensions. The request and context are sent to the server, where LLM-based generator produces runnable code. To improve correctness, the server retrieves set of similar code exemplars from a curated library via vector retrieval. QueryGIS executes the code in QGIS and applies a closed-loop recovery strategy: it starts in a lightweight instruction-only mode and escalates to a RAG-enriched mode upon failures, expanding context and prompt depth. When execution errors occur, the plugin sends the failing code, error message, and original request to a debugging endpoint and retries. Runtime wrapper intercepts calls to core QGIS API objects and processing runs to reduce minor API mismatches, while every runs—success or failure—are stored as JSONL logs and visualized through a monitoring dashboard. Through this workflow, QueryGIS enables conversational remote-sensing automation in QGIS.

Back to session list

B1-2: An Automated Scan-to-CAD Framework for Structural Modeling of Bridge Point Clouds

Presenting author: Li-Chih CHEN

Authors: Li-Chih CHEN, National Cheng Kung University; Jian-Yeou Rau, National Cheng Kung University

Keywords: Point cloud; Scan-to-CAD; Bridge modeling; Semantic segmentation; Machine learning

This study proposes an automated framework for converting bridge point clouds into structured CAD models by integrating machine learning-based semantic segmentation with parametric geometric reconstruction. First, feature engineering is conducted to extract multi-dimensional geometric descriptors from bridge point cloud data, including spatial coordinates, normal vectors, curvature, density, and directional indicators. These features are used to construct discriminative feature vectors for structural component identification. A Random Forest classifier is then employed to perform supervised semantic segmentation, enabling the classification of detailed bridge components such as piers, girders, and decks with improved robustness and separability across varying geometric forms. Following classification, segmented component-wise point clouds are processed using the Point2CAD (P2C) system for geometric fitting and parametric modeling. Through surface fitting, topology inference, and parametric reconstruction, structured CAD geometries are automatically generated. Compared with traditional workflows relying on manual segmentation or purely geometry-based fitting, the proposed approach leverages structure-oriented feature design to mitigate component confusion and over-merging issues, thereby improving reconstruction accuracy and automation level. The results demonstrate that the proposed framework provides a feasible Scan-to-CAD solution for bridge engineering applications and establishes a technical foundation for BIM integration and digital twin development.

Back to session list

B1-3: Pothole Detection Under Data Scarcity Using Stable Diffusion and YOLO

Presenting author: Cheng-Zong YANG

Authors: Cheng-Zong YANG,National Taiwan Univ.; Jen-Yu HAN,National Taiwan Univ.; ChangXin You, National Taiwan Univ.

Keywords: Pothole Detection, Stable Diffusion, DreamBooth, YOLO, Synthetic Data Augmentation

With the rapid expansion of transportation infrastructure, road safety has become a critical concern. Potholes are common pavement defects that threaten driving safety and vehicle integrity. However, developing robust detection models remains challenging due to the scarcity and limited diversity of annotated pothole datasets, which often results in overfitting and poor generalization under real-world conditions. This study proposes a pothole detection framework designed specifically for data-constrained environments. We introduce a diffusion-based synthetic data augmentation pipeline leveraging Stable Diffusion, enhanced with ControlNet for structural guidance and LoRA/DreamBooth fine-tuning to preserve pothole-specific visual characteristics. This approach enables the generation of diverse yet semantically consistent synthetic samples, enriching long-tail distributions across road textures, and viewpoints. The detection backbone is based on YOLOv8, selected for its strong accuracy and real-time inference capability. Starting from 800 real-world images, we generate an additional 400 synthetic samples, forming a dataset of 1,200 images split into training, validation, and testing sets at an 8:1:1 ratio. Evaluated on the held-out test set using COCO-style metrics, the proposed augmentation strategy consistently improves performance compared to the baseline trained on real data only. The Box F1-score increases by 0.02, mAP@50 by 0.05, and mAP@50–95 by 0.02. The final model achieves a Box F1-score of 0.91 and an mAP@50 of 0.92, while maintaining an average inference time of 76 ms per image. Overall, the results demonstrate that diffusion-driven synthetic augmentation effectively mitigates dataset scarcity and reduces the synthetic-to-real domain gap, leading to improved robustness and- generalization without compromising real-time efficiency. Future work will further explore cross-domain validation and adaptive strategies for deployment under diverse environmental conditions.

Back to session list

B1-4: Super Resolution Reconstruction of Kompsat-3A Images Based on a Fine-Tuned Transformer

Presenting author: Hongjun YOUN

Authors: Hongjun YOUN; Taejung KIM

Keywords: Super-Resolution, Modulation Transfer Function, KOMPSAT-3A, Data Consistency, Physics-guided Learning, Point Spread Function

High-resolution optical satellite imagery is fundamental data for remote sensing. However, the inherent trade-off between swath width and spatial resolution limits the acquisition of high-resolution data over a wide area. While super-resolution (SR) techniques have been introduced to address this limitation, conventional models trained on general image domains often fail to account for the specific spatial frequency response of satellite sensors. Such discrepancies result in non-physical distortions and synthetic artifacts, including ringing and localized block patterns. They undermine data reliability in remote sensing applications where pixel-level quantitative analysis is critical. In this study, we propose a modulation transfer function(MTF)-based SR training and inference framework designed to physically simulate the sensor characteristics of the KOMPSAT-3A (K3A) satellite. First, we establish a forward degradation model—integrating an MTF-based point spread function (PSF), area-based downsampling, and Poisson-Gaussian mixed noise—to generate synthetic low-resolution data from original panchromatic images. Regarding the network architecture, we fix the low-frequency components to the upsampled results of the input image and restrict the predicted residuals within the MTF passband. This approach effectively suppresses brightness drift and high-frequency over-compensation typically encountered during patch-based training. The loss function is composed of three components: supervision loss, data consistency (DC) loss, and frequency band-limit loss. We ensure that the reconstructed results remain within a physically explainable range defined by the observation model. Experiment results on K3A patches showed a 22% gain in edge sharpness using the Tenengrad index and an 11% improvement in texture detail using the Laplacian variance index over the bicubic baseline. This physical error reduction confirms that the enhanced details were not artificial artifacts but consistent with the sensor model and that the proposed model significantly suppressed ringing artifacts. Consequently, the framework achieved an optimal balance between visual clarity and physical validity by keeping high-frequency recovery within a reconstructible range.

Back to session list

Session C1 Abstracts

C1-1: Spectral Features of Ultra-High-Resolution UAV Hyperspectral Data over Deciduous Broadleaf Forests in Tomakomai, Hokkaido

Presenting author: Hideki KOBAYASHI

Authors: Hideki KOBAYASHI, JAMSTEC; Tatsuro NAKAJI, Hokkaido Univ.; Roxanne LAI, JAMSTEC; Tomoko AKITSU, Ibaraki Univ. ;Fujio HYODO, Okayama Univ.; Juan Augusto Heins Herrera OLLACHICA, JAMSTEC/Soka Univ.;Victor S. KUWAHARA, Soka Univ.;Hibiki NODA, NIES;Kyaw Kyaw Htoo, Hokkaido Univ.;Tomoki Morozumi, NIES;Masato Hayashi, NIES

Keywords: Hyperspectral sensing; UAV imaging; Deciduous forest; Spectral features; Leaf-to-satellite scaling

In recent years, the development of spaceborne hyperspectral imagers has advanced rapidly, and a variety of sensors have been launched by different countries and institutions. For example, Japan’s hyperspectral sensor HISUI (Hyperspectral Imager Suite), mounted on the International Space Station (ISS), has been continuously observing global terrestrial and coastal regions for several years since its launch in December 2019. In addition, the deployment of hyperspectral satellites such as Germany’s EnMAP and Italy’s PRISMA has progressed, alongside the development and operation of commercial hyperspectral satellite missions. While these advances have accelerated the practical use of hyperspectral satellite data, several challenges remain in their interpretation. In particular, for deciduous broadleaf forests, which are the focus of this study, hyperspectral satellite data are expected to enable the estimation of forest structural properties and photosynthetic functions. However, even the highest-resolution spaceborne hyperspectral data currently available are limited to spatial resolutions of approximately 30 m. As a result, our understanding of how leaf-level spectral information is spatially averaged and represented in satellite observations remains insufficient. In this study, we aim to acquire hyperspectral data at much finer spatial resolutions than those provided by satellites by using a newly developed UAV-mounted hyperspectral imager to conduct low-altitude flights over forested areas. By integrating these UAV-based hyperspectral observations with ground-based measurements, including leaf-level spectral reflectance and physiological and structural vegetation characteristics, we seek to systematically characterize observed spectral features and to provide a foundation for estimating ecosystem functions and vegetation properties from future spaceborne hyperspectral data. In this presentation, we report the results of analyses examining the relationships among ground-based observations, UAV-based hyperspectral data, and satellite data, using UAV observations conducted in the Tomakomai Experimental Forest, Hokkaido, Japan, from 2023 to 2025.

Back to session list

C1-2: Individual Tree Crown Delineation and Carbon Stock Estimation Based on Airborne LiDAR Data

Presenting author: Ting-Hsuan TUNG

Authors: Ting-Hsuan TUNG, National Yang Ming Chiao Tung University; Tee-Ann TEO, National Yang Ming Chiao Tung University

Keywords: Airborne LiDAR; Individual Tree Crown Delineation; Aboveground Biomass; Carbon Stock Estimation

Under the advancement of net-zero emission targets, forest carbon stock estimation has become a critical indicator for climate governance. While airborne LiDAR provides high-resolution 3D spatial information, traditional individual tree crown (ITC) delineation methods—largely developed for temperate forests—often struggle with the structural complexity and mixed species composition of subtropical environments like Taiwan. This study proposes an integrated analytical framework using airborne LiDAR point clouds from Green Island, Taiwan, to bridge the gap between ITC delineation and carbon stock estimation. The methodology utilizes a watershed algorithm combined with local maxima detection to identify treetops, further enhanced by a post-processing workflow that incorporates airborne LiDAR return counts to identify multi-layered canopies and correct over-segmentation. The tree parameters such a tree height and crown area can be extracted from airborne LiDAR data. Results from a 250,000 m^2 study area identified 2,686 trees. Post-processing effectively refined structural parameters, reducing the standard deviation of crown area from 66.57 m^2 to 60.86 m^2, demonstrating improved consistency. Extracted height and crown area were used to develop aboveground biomass (AGB) prediction models. The analysis revealed that species-specific regression models achieved significantly higher explanatory power (R2 = 0.64) compared to generalized models (R2 = 0.52), highlighting the necessity of accounting for species variation in subtropical carbon assessments. Despite some commission errors in dense canopies and right-skewed biomass distributions, the optimized workflow offers a robust, scalable solution for high-resolution forest monitoring. By linking precise geometric extraction with regionally adapted allometric models, this research provides a technical foundation for enhancing the accuracy and applicability of carbon sink assessments in complex ecological landscapes.

Back to session list

C1-3: Evaluation of Satellite Remote Sensing-Based Biomass Estimation for Carbon Credit Calculation under the JCM

Presenting author: Yohei TOYOHARA

Authors: Yohei TOYOHARA, Mitsubishi Electric; Toshiki FUJINO, Mitsubishi Electric; Naoto OMORI, Satellite Data services; Takeo TADONO, JAXA; Hiroto KIKUCHI, JAXA; Sachiyo TAKAHASHI, MUFG Bank; Masato YANO, Mitsubishi UFJ Research and Consulting; Ayano KIDO, Archeda

Keywords: Satellite remote sensing; Above-ground biomass map; LULC; Carbon credit; JCM

In recent years, the global framework for climate change mitigation and the expansion of corporate initiatives toward carbon neutrality have driven rapid growth in the carbon credit market. In particular, demand for nature-based credits derived from forest conservation, reforestation, and land-use change has been increasing. However, ensuring transparency and reliability in the quantification methods remains a critical challenge. Traditional approaches that rely on on-site surveys face significant cost and time constraints, especially for large-scale monitoring and long-term tracking, which have hindered the efficiency of credit issuance. In contrast, satellite remote sensing technology offers the advantage of enabling extensive and frequent observations, allowing for the quantitative assessment of changes in land cover and forest biomass. In this study, a method for estimating biomass volume from satellite data and land cover classification information provided by the Japan Aerospace Exploration Agency (JAXA) was applied to the carbon credit calculation methodology of the Joint Crediting Mechanism (JCM) in Cambodia. Specifically, emission factors and assessment areas were estimated from satellite-derived data, and the resulting carbon credit values were compared and evaluated against credits calculated from ground-based measurements. The results indicated that the satellite-based approach has the potential to improve the reliability of carbon credit estimation and to streamline the issuance process. This method is considered to play an important role in international carbon credit systems and climate change mitigation frameworks.

Back to session list

C1-4: Integrating Environmental Variables into Remote Sensing-Based Forest Type Classification in Taiwan

Presenting author: Han-Hui HUANG

Authors: Han-Hui HUANG, Department of Forestry, National Chung Hsing University (NCHU); Yu-Chun HSU, Department of Forestry, National Chung Hsing University (NCHU)

Keywords: Sentinel-2; Forest type classification; Environmental variables; National-scale mapping; SHAP

Forest type classification based solely on spectral information may encounter limitations in environmentally heterogeneous regions, where vegetation communities exhibit similar spectral responses under different geographic conditions. This study evaluates the effect of integrating environmental variables into remote sensing-based forest type classification at national scale in Taiwan, using a 250 m grid aligned with Taiwan’s 4th National Forest Inventory (NFI). Two ensemble models with identical architectures (LightGBM + ExtraTrees) were developed. The baseline model utilized multi-temporal Sentinel-2 imagery acquired between 2020 and 2024, from which phenological and harmonic features were derived using vegetation indices. The extended model incorporated additional environmental variables, including terrain, soil, and climate factors. Model robustness was assessed using both stratified cross-validation and Leave-One-Region-Out spatial cross-validation to examine cross-regional transferability across Taiwan. Results indicate that integrating environmental variables improved overall classification performance by approximately 23% under stratified validation. Under spatial cross-validation, national-scale generalization performance increased by over 22%, with overall misclassification reduced by roughly 11%. Regional analysis further showed that performance gains were more evident in topographically complex eastern Taiwan, approaching a 17% improvement relative to the spectral-only model. SHAP-based interpretation revealed that the baseline spectral model was primarily driven by vegetation indices, with Enhanced Vegetation Index (EVI), Normalized Difference Red Edge (NDRE), and Normalized Difference Moisture Index (NDMI) ranking as the three most influential predictors. After environmental integration, model influence shifted toward elevation, soil organic carbon, and regional temperature variables. The findings in this stage indicate that environmental variables provide geographically structured context that complements spectral information, contributing to measurable improvements in nationwide forest type classification in Taiwan.

Back to session list

C1-5: Integrating Photogrammetry and Deep Learning for Multi-Temporal Change Detection in Rural Landscapes of the Oki Islands, Japan

Presenting author: Yasumichi YONE

Authors: Yasumichi YONE, Shimane Univ.; Yuta CHUJO, Shimane Univ.; Kazuma YAMASAKI, Shimane Univ.; Erina TAKAHASHI, Shimane Univ.; Hiroyuki OGUMA, Shinshu Univ.

Keywords: Deep Learning; Photogrammetry; Multi-Temporal Aerial Photographs; Land Cover Change Detection; Landscape Dynamics; Forest Management; Farmland Abandonment

In recent decades, mountainous rural regions in Japan have experienced rapid depopulation and aging, resulting not only in farmland abandonment but also in substantial shifts in forest utilization and management practices. These socio-economic transformations have significantly altered rural landscape structures. Understanding the long-term processes underlying such land cover changes is essential for sustainable land and forest resource management. This study investigates 50-year land cover dynamics in the Oki Islands (Okinoshima Town, Shimane Prefecture), Japan, by integrating photogrammetric reconstruction and deep learning–based image analysis within a GIS framework. Multi-temporal aerial photographs from five periods were processed to generate orthophotos and terrain-derived datasets, which were combined into composite images. A U-Net semantic segmentation model was applied to classify land cover into categories representing forested areas, agricultural land, transitional vegetation, and other land uses. This integrated approach enabled consistent classification across different decades and facilitated landscape-scale change detection. The analysis reveals a long-term expansion of forest-dominated areas associated with farmland abandonment and forest maturation, alongside more recent increases in transitional or disturbed forest areas potentially linked to renewed harvesting and management activities. These findings suggest that rural landscape transformation reflects not only ecological succession but also shifts in forest policy and utilization strategies. Comparison with official agricultural and forestry statistics indicates general consistency in long-term trends while highlighting differences between administratively defined land use and physically observed land cover. This study demonstrates that integrating photogrammetry and deep learning with multi-temporal aerial photographs provides an effective framework for analyzing long-term rural landscape dynamics.

Back to session list

Session D1 Abstracts

D1-1: Multi-Temporal Remote Sensing Analysis of Land Use Change in High-Altitude Tea Gardens

Presenting author: Chin CHENG

Authors: Chin CHENG, National Chung Hsing University; Ming-Der YANG, National Chung Hsing University

Keywords: Deep Learning; Precision Agriculture; Satellite Imagery; Tea Plantation Change; Land Use

High-altitude tea plantations are vital economic crop production areas in mountainous regions, where their land-use patterns and development transitions exert a critical influence on hillside environmental management and sustainable operations. In recent years, alongside the adjustment of tea planting areas and the increasing intensity of mountainous land use, mastering spatial distribution and land-use change trends has become a pressing issue. This study is based on multi-temporal remote sensing data, collecting SPOT satellite imagery from 2015 to 2025. Following image fusion and the calculation of vegetation indices, a multi-year analytical database was established to conduct an analysis of land-use changes in high-altitude tea plantations. Through supervised classification methods, a dataset of seven land-use categories was created—including tea plantations, forest land, cropland, buildings, roads, water bodies, and bare land—to compare spatial distribution differences and transition relationships across different years. Regarding model construction, this research adopts a deep learning architecture combining convolutional neural networks and long short-term memory networks to perform land-use classification and temporal feature modeling, with performance evaluated through training, validation, and testing sets. The results show that the model achieved an overall pixel accuracy of 0.8458 and a Macro F1-score of 0.7299 in the validation data; forest land and water bodies yielded the best identification results with F1-scores of 0.89 and 0.97, respectively, while the F1-score for the tea plantation category reached 0.71, demonstrating the model’s effectiveness in distinguishing tea plantations from other vegetation types.

Back to session list

D1-2: Prototype Development of Rice Monitoring Methods at Field and Regional Levels in the Republic of the Union of Myanmar

Presenting author: Shinya ODAGAWA

Authors: Shinya ODAGAWA, RESTEC; Asei KANEKO, Nippon Koei Co., Ltd.; Koki UDA, Nippon Koei Co., Ltd; Cho THANDA NYUNT, Myanmar Koei International LTD.

Keywords: Precision agriculture, Rice monitoring, SuperDove, Sentinel-2, Yield prediction

The importance of rice cultivation in Myanmar is not only as the staple food for its people, but also as an export commodity for earning foreign currency. Furthermore, the appropriate management of rice productivity in Myanmar is also important for the food security of Asian countries, including Japan. The objective of this study is to develop an efficient and cost-effective rice monitoring method utilizing satellite data in Myanmar. This study attempted to develop not only a detailed rice monitoring method at the field level using high-resolution satellites, but also a regional-level rice monitoring method using medium-resolution satellites. This study conducted field surveys during the rainy seasons from 2023 to 2025 in two areas (Taikkyi and Twantay) within the Yangon Region, one of Myanmar’s major rice-producing areas. During the field survey, we measured plant height and stem count, which are important for monitoring rice crops. Furthermore, just before the harvest period, rice samples were collected and rice yields were measured. This study utilized SuperDove which is one of the high-resolution satellites. This satellite can capture images nearly every day through formation flying with multiple satellites, enabling imaging even during Myanmar’s rainy season when clouds are constantly present. Applying linear regression and generalized additive models to the results of field surveys and surface reflectance data obtained from SuperDove enabled estimation of plant height and yield with an accuracy of approximately 0.7 to 0.8 correlation coefficient. Next, based on this estimation model, we developed a prototype method for monitoring rice paddies at the regional level using Sentinel-2. The target regions are Bago Region, Yangon Region, and Ayeyarwady Region, which are Myanmar’s major rice-growing areas. As a result, we were able to predict rice yields at the regional level consistent with existing statistical data.

Back to session list

D1-3: Integrating satellite remote sensing and meteorological data for early season yield prediction of rice and wheat

Presenting author: Akie KOGA, PASCO CORPORATION

Authors: Akie KOGA, PASCO CORPORATION; Nobuyuki KOBAYASHI, Smart Link Hokkaido Co.,Ltd; Kumi IIDA, PASCO CORPORATION; Atsushi KIMURA, PASCO CORPORATION

Keywords: Early-season crop yield estimation; PlanetScope; Vegetation indices; RedEdge

Accurate crop yield prediction is essential for accelerating decision-making and strengthening food security. As climate change intensifies production risks such as heat stress, the need for methods that capture yield variability earlier and with higher reliability is increasing. Remote sensing enables spatially extensive and frequent monitoring; however, yield prediction studies for major crops such as rice and winter wheat often rely on vegetation indices observed during late growth stages (e.g., heading or grain filling), limiting their usefulness for the early-season forecasts required for policy and supply–demand management. This study developed yield prediction models for direct-seeded rice and winter wheat in Hokkaido, Japan, using vegetation indices derived from PlanetScope imagery and meteorological indicators up to the flag leaf stage. Huber regression was employed to enhance robustness. Group Leave-One-Out (Group LOO) validation showed that predictive performance did not reach practical levels for either crop, likely due to limited sample size and data imbalance. However, by identifying features consistently selected across Group LOO training sets and retraining models using the full dataset, we obtained final models exhibiting significant linear relationships between yield and key spectral and meteorological indicators. Selected features included CI (Chlorophyll Index Red Edge), NDVI (Normalized Difference Vegetation Index), NDRE (Normalized Difference RedEdge Index), and accumulated temperature for rice, and CI, OSAVI (Optimized Soil Adjusted Vegetation Index), and mean solar radiation for wheat, highlighting the utility of RedEdge-based indices alongside near-infrared measures. These final models achieved R² values of 0.702 for rice and 0.438 for wheat. Overall, the findings demonstrate the potential of combining early- to mid-season remote sensing information with statistical modeling for earlier-stage yield estimation, while also showing that generalization performance remains constrained by sample size and data variability. Expanding datasets across regions and years will be essential for improving model generalizability and operational utility.

Back to session list

D1-4: A Phenology-Driven Approach for Detecting Abandoned Cropland Using HLS Imagery and Interpretable Machine Learning

Presenting author: Sinyoung PARK

Authors: Sinyoung PARK, Department of Forest Resources, Kookmin University; Sanae KANG, Department of Forest Resources, Kookmin University; Byungmook HWANG, Department of Forest Resources, Kookmin University; Gwan YUN, Department of Forest, Environment, and Systems, Kookmin University; Dongwook W. KO, Department of Forest, Environment, and Systems, Kookmin University, Forest Carbon Graduate School, Kookmin University

Keywords: SHapley Additive exPlanations; XGBoost; balanced bagging classifier; phenology-based detection; remote sensing

The expansion of abandoned cropland, driven by urbanization, demographic changes, and reduced agricultural profitability, poses both ecological and agricultural concerns that require systematic monitoring. Remote sensing provides an effective way to track cropland status through phenological signals captured in time-series spectral data. However, persistent cloud contamination and spatially fragmented field patterns remain major challenges for satellite-based detection, particularly in monsoon-influenced regions. To overcome these limitations, this study compiled monthly spectral and radar time-series composites spanning January to September by integrating Harmonized Landsat and Sentinel-2 (HLS) imagery with Sentinel-1 SAR data in agricultural landscapes of South Korea. In addition to the monthly index values, statistical features capturing temporal variability across seasonal periods were derived to characterize phenological patterns that differentiate rice paddy, upland fields, and abandoned cropland. Classification was performed using an eXtreme Gradient Boosting (XGBoost) model coupled with a Balanced Bagging Classifier (BBC) to account for class imbalance. The model achieved an accuracy of 0.84, a Cohen’s kappa of 0.71, and an F2-score of 0.84. Model interpretability was examined using SHapley Additive exPlanations (SHAP), which revealed that NIR variability during the early growing season (May–June) and SWIR2 reflectance in winter (January) were the most influential predictors, reflecting distinct vegetation dynamics of managed versus unmanaged fields. Notably, SHAP feature importance was interpreted in the context of crop phenological stages, providing ecologically meaningful explanations for abandoned cropland detection. The proposed framework, combining phenology-driven multi-sensor features with interpretable machine learning, offers a transferable approach for monitoring cropland abandonment across monsoon-influenced agricultural regions. This study was conducted with the support of the R&D program for Forest ScienceTechnology (project no. RS-2025-02214405) provided by Korea Forest Service (Korea Forestry Promotion Institute).

Back to session list

D1-5: Correlation between field-measured grain moisture content and C-band SAR backscatter: Potential use of Sentinel-1 SAR data to estimate optimal harvest timing in rice crop cultivation

Presenting author: Shoko KOBAYASHI

Authors: Shoko KOBAYASHI, Tamagawa Univ.

Keywords: C-band SAR; Sentinel-1; Agriculture; Rice plant; Grain; Vegetation water content; Harvest time

Increases in temperature associated with climate change can potentially damage rice crops. Accurate prediction of the optimal time for harvesting therefore represents a key precautionary countermeasure that can mitigate against the damage caused by heat exposure. The remote sensing of grain moisture content, which is an important indicator for accurate harvest timings, has been rarely studied using microwave Synthetic Aperture Radar (SAR). Here we examine the relationship between crop biophysical parameters and C-band SAR backscatter in order to determine the correlation between field-measured grain moisture content and satellite SAR backscatter. This study takes advantage of the specific properties of microwave radar signals, and, in particular, their sensitivity to moisture content. We first pre-processed Sentinel-1 data taken in 2020 and 2022 to extract radar backscatter values within rice crop fields. Partial least squares regression was used to determine the relation between field-measured crop biophysical parameters and backscatter data. This statistical analysis revealed that co-polarized radar backscatter strengths obtained with shallow incidence angles exhibited a significant correlation to grain moisture content and leaf and tiller moisture content (R2 = 0.669), with grain moisture showing the highest variable importance. On the other hand, the correlation with these biophysical parameters was not statistically significant for backscatter at the steeper incidence angle. The ripening phenomenon restricts stem elongation in order to furnish nutrients to the grain so that the physical structure of the plant changes little and grain moisture substantially decreases. This most likely explains why backscatter signals obtained at shallower incidence angles, which are more susceptible to interaction with the crop canopy layer, can be strongly affected by grains. Our study indicates the great potential of synoptic C-band SAR observations for the estimation of grain moisture content in rice and thus points the way toward improved prediction of optimal harvest timings.

Back to session list

Session E1 Abstracts

E1-1: Vision-Based Camera Pose Estimation for Bridge Inspection Using a Reusable 3D Feature Map

Presenting author: Yi-Chen LEE

Authors: Yi-Chen LEE; Ya-Li LIN; Jiann-Yeou RAU

Keywords: Vision-based localization；3D feature map；Camera EOPs estimation；Bridge inspection

Bridge inspection images acquired by unmanned aerial vehicles (UAVs) are often collected under challenging conditions where reliable GNSS information is unavailable or degraded. Moreover, single-view inspection images typically lack explicit three-dimensional (3D) geometric context, limiting their effective use for spatially grounded analysis such as damage localization or semantic interpretation. This study presents a vision-based camera pose estimation framework that leverages a reusable 3D feature map constructed from an existing UAV bridge image dataset. In the proposed pipeline, a 3D feature map is first established by triangulating matched key-points across multiple views, where each 3D point is associated with at least two representative local feature descriptors. Given a newly acquired inspection image, two-dimensional image features are matched against the 3D feature map using a learned feature matching strategy, enabling robust 2D–3D correspondences. The camera exterior orientation parameters (EOPs) are then estimated via a Perspective-n-Point (PnP) formulation. The proposed framework is designed to operate without reliance on GNSS measurements and is particularly suited for bridge inspection scenarios with constrained accessibility or signal obstruction. The study focuses on system design and pipeline integration, and preliminary results demonstrate the feasibility of accurate pose recovery from a single image in real bridge inspection environments. By estimating the EOPs of each inspection image, this study establishes 2D-3D correspondences, allowing crack detection results to be localized on the actual bridge 3D model and enabling semantic crack analysis and image-based condition assessment.

Back to session list

E1-2: Comparative Analysis of CNN and Transformer Models for Robust Water Body Detection across Multi-Satellite Sensors

Presenting author: Seojin KONG

Authors: Seojin KONG, Image Eng. Research Center, 3DLabs Co. Ltd, Republic of Korea; Wonwoo SEO, Image Eng. Research Center, 3DLabs Co. Ltd, Republic of Korea; Sooahm RHEE, Image Eng. Research Center, 3DLabs Co. Ltd, Republic of Korea

Keywords: Water Body Detection; High-resolution Satellite Imagery; Deep Learning; CNN and Transformer; Semantic Segmentation

With the increasing availability of high-resolution satellite imagery, accurate water body extraction has become critical to mitigate errors caused by shadows, reflections, and vessel interference. Water bodies exhibit significant spatio-temporal variability due to rainfall, hydrological dynamics, and urban expansion, making rapid and stable automated detection essential. This study aims to develop a robust deep learning framework for reliable and automated water body detection in high-resolution optical satellite imagery. We conducted a comprehensive performance comparison between CNN-based (UNet, DeepLabV3+) and Transformer-based (SegFormer) models, employing MobileNetV3, MobileViT, and MiT as encoders to balance accuracy and computational efficiency. The models were trained on the GLH-Water dataset, a large-scale benchmark, and evaluated using Accuracy and Intersection over Union (IoU). In particular, the generalization performance and robustness were verified by applying multiple satellite images with varying resolutions to the model. Results demonstrated a model structure capable of stable detection in complex surface environments and strong cross-satellite generalization. This study provides a critical preprocessing framework to reduce water-induced noise, enables the prior detection of water body regions, allowing noise caused by water surfaces to be effectively reduced. This is meaningful in that it can be utilized as a preprocessing step to improve the accuracy of subsequent image analysis. In addition, it is expected to contribute to various remote sensing applications such as disaster monitoring, water resource management, and urban environment analysis in the future.

Back to session list

E1-3: Deep Learning-based Approximation of Building Footprints from High-Resolution Satellite Imagery

Presenting author: Yun-Hao LIN

Authors: Yun-Hao LIN, National Central University; Fuan TSAI, National Central University

Keywords: Building Extraction; Footprint Approximation; Deep Learning; Mask R-CNN

Building is one of the primary structural elements in metropolitan areas, exhibiting clear geometric patterns and spatial distributions that can be captured using high-resolution satellite imagery. Deep learning provides an effective framework for automating remote sensing analysis, and the Mask R-CNN model has been demonstrated to have a great potential in building instance extraction. Recent studies have focused on improving the geometric quality of extracted building polygons to produce smoother and more regular shapes, while enhancing model robustness in complex urban environments. This study uses orthorectified high resolution (Pleiades and Pleiades Neo) satellite imagery to approximate building footprints. Although orthorectification reduces terrain-induced distortions, it does not fully correct relief displacement, resulting in offsets between rooftops and their ground footprints. In addition, orthorectification often removes or obscures the original viewing geometry information, making footprint estimation more challenging. These issues become more critical in dense urban areas, where overlapped high-rise and low-rise buildings introduce significant occlusion. To address these challenges, a two-stage framework is proposed. First, a Mask R-CNN model with a ResNeXt-101 backbone and Feature Pyramid Network (FPN) is pre-trained on an opensource dataset and subsequently fine-tuned with local satellite imagery. Next, variations in satellite viewing geometry are used to approximate building footprints using a multi-view geometric approach. The results suggest that combining deep learning–based rooftop detection with multi-view geometry can help mitigate displacement effects and improve the alignment between rooftops and their ground footprints. This framework provides a viable approach to urban building mapping in complex metropolitan environments.

Back to session list

E1-4: LOSS FUNCTION OPTIMIZATION OF DEEP LEARNING MODEL FOR IMAGE ALIGNMENT BETWEEN SATELLITE IMAGERY AND GCP CHIPS

Presenting author: Kyeonghwan KIM

Authors: Kyeonghwan KIM, Chungbuk National University; Jaewan CHOI, Chungbuk National University; -; -

Keywords: deep learning; loss function; image alignment; GCP chips; orthorectification; template matching

While researches on high-precision satellite ortho-imagery generation have been actively conducted, satellite images inherently contain geometric distortions caused by high altitude and variations in sensor attitude. To address this, orthorectification using GCP (Ground Control Point) chips has been widely utilized. However, conventional area-based matching methods such as Zero-Normalized Cross-Correlation (ZNCC) and Relative Edge Cross-Correlation (RECC) often experience a significant degradation in matching accuracy under complex conditions involving relief displacement, shadows, and spectral variations. To overcome these limitations, this paper proposes a deep learning-based image alignment framework that exploits features extracted from non-linear road areas, which are relatively less affected by relief displacement and possess distinct morphological characteristics. In the proposed method, GCP chips based orthoimages were first transformed using Rational Polynomial Coefficients (RPC) to approximate the geometric conditions of the target satellite imagery. These transformed chips and the satellite images are then fed into a deep-learning model that combines U-Net, for robust local feature extraction, and Mamba, for efficient global context modeling. The matching points are subsequently determined by performing template matching on the feature maps generated by the deep learning model to identify the coordinates with the highest correlation. Especially, this study comparatively analyzes the impact of various loss functions including distance-based and heatmap-based losses to maximize alignment performance and proposes an optimal loss function configuration.

Back to session list

E1-5: Analysis of Spatial Resolution Enhancement via Deep Learning-Based Image Super-Resolution for GCP Chip Matching

Presenting author: Minkyung CHUNG

Authors: Minkyung CHUNG, Seoul National University of Science and Technology; Jaewon HUR, TelePIX; Doochun SEO, Korea Aerospace Research Institute; Youkyung HAN, Seoul National University of Science and Technology

Keywords: Image Super-Resolution (SR); Ground Control Point (GCP) Chip; Template Matching; KOMPSAT-3; KOMPSAT-3A

Ground Control Point (GCP) chips are fixed-size reference image patches with known geographic coordinates that are used for automatic geometric correction of satellite imagery. Although high-resolution GCP chips are critical for achieving high positional accuracy, their acquisition is often constrained by limited access to airborne imagery, particularly in restricted or inaccessible regions, as well as by the limited availability of very-high-resolution satellite data. This study investigates the effectiveness of deep learning-based image super-resolution (SR) for improving the template matching performance of GCP chips under varying spatial resolution conditions. To this end, original GCP chips extracted from KOMPSAT-3 imagery were synthetically degraded to simulate lower spatial resolutions. The degraded chips were then enhanced using a pretrained SR model and matched to target KOMPSAT-3A imagery using template-based matching. The matching accuracy was evaluated at independent check points and quantified using root mean square error (RMSE) and success rates based on pixel error thresholds (≤3 pixels and ≤5 pixels). By comparing degraded chips and their super-resolved counterparts across multiple resolution levels, the study analyzes the spatial resolution range in which SR contributes to improved matching robustness and positional precision. The experimental results indicate that SR can improve matching performance when the spatial resolution of the super-resolved chips is comparable to that of the target imagery. However, its effectiveness decreases when the spatial resolution discrepancy becomes excessive, possibly due to geometric resampling during coordinate transformation. These findings provide preliminary evidence regarding the practical applicability of SR for generating and updating high-resolution GCP chips using existing satellite imagery.

Back to session list

Session A2 Abstracts

A2-1: Investigating a Geospatial Framework for Tidal Flat Sedimentary Facies Analysis Using Satellite Imagery and GIS Data

Presenting author: Donguk LEE

Authors: Donguk LEE; Keunyong KIM; Hyeong-Tae JOU; JOO-Hyung RYU*

Keywords: Tidal flat, Sedimentary facies, Remote sensing, GIS, Regional-scale monitoring

Tidal flat sedimentary facies are critical indicators for understanding coastal hydrodynamic energy and environmental stability. While on-site surveys provide precise, point-source data, their application to macro-scale monitoring is often constrained by the physical hazards and limited accessibility inherent in tidal environments, as well as the prohibitive time and labor costs required for extensive field sampling. Consequently, there is an increasing demand for integrated analytical frameworks that combine remote sensing and GIS data to facilitate a sustainable, regional-scale monitoring system for the coastal areas of the Republic of Korea. This study proposes an approach to analyze the spatial distribution of sedimentary facies by synthesizing multi-source geospatial information. The primary inputs include national-scale Geographic Information System (GIS) datasets, such as coastline configurations and river networks, which reflect the geomorphological context. In addition, multi-spectral satellite indices are employed to capture the physical and multi-temporal properties of the intertidal surface. Utilizing the Tidal Flat Ecosystem Survey Data from the Korea Marine Environment Management Corporation (KOEM) as a reference point, the study examines the feasibility of regional-scale facies characterization, with preliminary assessments indicating that the integrated variables can effectively reflect the macroscopic spatial trends of sedimentary environments. Integrating morphological GIS features with satellite observations provides a foundational methodology that complements traditional field surveys by offering a macro-scale spatial perspective. This framework is expected to facilitate more informed coastal management and contribute to the systematic, long-term environmental assessment of tidal ecosystems.

Back to session list

A2-2: Micro-topographic Change Analysis of Tidal Flats Using Co-registration and Patch-based Bias Correction of UAV-Derived DEMs

Presenting author: Minju KIM

Authors: Minju KIM, KIOST; Joo-Hyung RYU, KIOST; Keunyong KIM, KIOST

Keywords: UAV; Tidal flats; Bias correction; Micro-topographic change

This study aims to quantitatively analyze micro-topographic changes in the Hwangdo tidal flat using UAV-derived digital elevation models (DEMs). Tidal flats are highly dynamic environments sensitive to tidal and sedimentary processes, where subtle morphological changes can occur over short time scales. While UAV-derived DEMs enable high-resolution topographic mapping, systematic errors and spatial distortions may arise from the distribution of ground control points (GCPs) and image processing conditions. These errors can lead to misinterpretation of actual topographic changes in DEMs of difference (DoD). To address these issues, this study proposes a workflow that reconstructs consistent multi-temporal DEMs through co-registration and patch-based bias correction, followed by the extraction of statistically significant topographic changes. UAV-derived DEMs acquired in October 2022, 2024, and 2025 were analyzed. First, elevation errors were assessed using RTK measurements, revealing that localized outliers, particularly in peripheral areas, can adversely affect bias correction. To minimize their influence, a common analysis area was defined to ensure stable bias correction. Subsequently, co-registration was applied to eliminate inter-temporal positional discrepancies, followed by patch-based bias correction to remove residual local biases and generate consistent DEMs. DoD analysis was then conducted, and a threshold of detectable change (TCD) based on error propagation was applied to extract statistically significant changes. The results revealed distinct erosion and deposition patterns across different time periods, with pronounced changes observed near tidal channels and in specific zones. This study presents a reconstruction-based framework integrating DEM co-registration and patch-based bias correction, ensuring inter-temporal consistency and enabling reliable, uncertainty-aware topographic change analysis. The proposed approach enhances the reliability and robustness of high-resolution tidal flat monitoring using UAV data.

Back to session list

A2-3: Morphological Changes of Tidal Flats in South Korea Using a 40-Year Landsat Time Series

Presenting author: Jingyo LEE

Authors: Jingyo LEE, KIOST(Korea Institute of Ocean Science & Technology); Keunyong KIM, KIOST(Korea Institute of Ocean Science & Technology); Joo-Hyung Ryu, KIOST(Korea Institute of Ocean Science & Technology) *Corresponding author

Keywords: Landsat series, tidal flat, waterline method, DEM

Tidal flats are dynamic coastal geomorphic systems that provide essential ecosystem services, including carbon sequestration, biodiversity conservation, and coastal hazard mitigation. However, establishing nationally consistent, long-term topographic datasets has been constrained by the high cost and irregular acquisition of field surveys and airborne LiDAR. This study utilized the time-series archive of the Landsat program to generate 40 years (1984–2024) of tidal-flat distribution maps for the entire South Korean coastline and to quantitatively evaluate long-term morphodynamic changes. To ensure temporal consistency, imagery acquired under diverse tidal conditions was systematically selected, and cloud-contaminated scenes were removed to construct multi-temporal datasets. The Normalized Difference Water Index (NDWI) was applied to delineate the tidal-flat–water boundary, and stable waterlines were repeatedly extracted using the Otsu thresholding algorithm. Tidal elevations derived from in situ observations and tidal model outputs were assigned to each contemporaneous waterline, generating elevation-referenced contour samples. Based on these samples, continuous five-year interval Digital Elevation Models (DEMs) of tidal flats were reconstructed. The resulting DEM time series enabled classification of tidal-flat morphological types and quantification of elevation change (Δh), volumetric change (ΔV), and erosion–accretion patterns. The results reveal pronounced regional heterogeneity in spatiotemporal dynamics. Estuarine and development-intensive coasts exhibit persistent areal loss and morphological simplification, whereas sediment-dominated regions show localized expansion and periodic variability. This study demonstrates the feasibility of long-term, satellite-based national coastal monitoring and presents a scalable analytical framework that can contribute to tidal-flat conservation and sustainable coastal management.

Back to session list

A2-4: Spatial Validation of Nighttime Vessel Detections by KIOST VBD using EOG VBD and Ground Truth Data in Korean waters

Presenting author: Chan-Su YANG

Authors: Chan-Su YANG, Korea Institue of Ocean Science & Technology; Sree Juwel Kumar CHOWDHURY, Korea Institute of Ocean Science & Technology; Jae-Seong AHN, Kyungil University

Keywords: VIIRS, Nighttime vessels; EOG VBD

Remote sensing of nighttime lights provides an effective means of detecting fishing vessels at sea. The Visible Infrared Imaging Radiometer Suite (VIIRS) Day/Night Band (DNB) onboard the Suomi National Polar-orbiting Partnership (SNPP) satellite has enhanced nighttime detection capability. The Earth Observation Group (EOG) at the Colorado School of Mines developed a vessel detection algorithm based on VIIRS data and maintains the VIIRS Boat Detection (VBD) database. However, in Korean waters, limitations of the EOG VBD have been identified, including over-detections caused by noise. Therefore, several modifications and parameter adjustments were introduced to the EOG VBD algorithm to enhance detection reliability, referred to as KIOST (Korea Institute of Ocean Science & Technology) VBD. Modifications include adjustments to the Spike Median Index threshold, refined no-data handling, application of a minimum brightness threshold, and land masking. The study area covers waters surrounding the Korean Peninsula (30°N - 40°N, 120°E - 135°E). Spatial consistency of KIOST VBD was evaluated using EOG VBD as a reference. EOG detections were temporally filtered within ±1 minute of KIOST VBD for alignment. A two-stage spatial matching procedure was applied. First, EOG detections within a 0.01° radius of each KIOST VBD were selected. Geodesic distances were calculated, and pairs < 0.75 km were retained. One-to-one matching in ascending distance assigned Flag 1 to shortest pairs. Unmatched detections underwent second-stage matching within a 0.02° radius. A 750 m buffer accounted for positional uncertainty. Candidates within the buffer were treated as matches; those outside were evaluated by distance from the buffer boundary, retaining pairs within 0.75 km. Shortest pairs were assigned Flag 2. Results indicate KIOST VBD reduces over-detection observed in EOG VBD while maintaining spatial consistency. Vessel positions were validated using ground truth (V-Pass and AIS) data near Busan Port, confirming the robustness of the vessel detections.

Back to session list

Session B2 Abstracts

B2-1: High-Resolution Soil Moisture Estimation over South Korea Using Machine Learning and Integrated Sentinel-1 SAR and Meteorological Data

Presenting author: Su-Bin HA

Authors: Su-Bin HA, Pukyong National University; Ye-Young KIM, Pukyong National University; Gwang-Woo OH, Pukyong National University; Seung-Kuk LEE, Pukyong National University

Keywords: Sentinel-1; Synthetic Aperture Radar; Random Forest; soil moisture; high-resolution mapping

Soil moisture is a key state variable in the land surface water, regulating infiltration and runoff generation, evapotranspiration, and land–atmosphere feedbacks. It is also essential for agricultural monitoring and for assessing water-related hazards such as droughts and floods. In South Korea, producing spatially continuous, high-resolution soil moisture remains challenging because in situ measurements are sparse and provide limited spatial coverage. To address this gap, this study develops a Random Forest (RF) model to estimate surface soil moisture by integrating Sentinel-1 dual-polarization Synthetic Aperture Radar (SAR) observations with meteorological, soil, land-cover, and topographic predictors. Sentinel-1 GRD imagery was processed through a standardized workflow to obtain terrain-corrected backscatter (VV and VH) and the VH/VV ratio. Volumetric soil moisture observations at ~10 cm depth were collected from the Rural Development Administration (RDA) monitoring network and time-matched to satellite acquisitions. Physically unrealistic values were removed, and winter periods were excluded from evaluation to reduce freeze–thaw effects. Meteorological variables from the Korea Meteorological Administration (KMA) included air temperature and multi-timescale antecedent precipitation totals, which were spatially interpolated and resampled to the SAR grid. Additional predictors describing soil properties, land cover (one-hot encoded), and topography were included to represent environmental heterogeneity. The RF model was trained with matched datasets from 2023–2024 and tested using an independent year (2025). The model achieved R = 0.83, RMSE = 4.50 vol.%, and MAE = 3.37 vol.%, indicating robust performance beyond the training period. Land-cover-stratified validation showed the best accuracy in rice paddies, followed by forests, upland croplands, and grasslands. Seasonal validation was also stable (R > 0.8), with the highest accuracy in spring, followed by summer and autumn. Overall, the proposed multi-source RF framework provides a practical basis for generating high-resolution soil moisture maps across South Korea, supporting operational applications in agriculture and drought/flood risk assessment.

Back to session list

B2-2: A Machine Learning Approach to Spatio-temporal Interpolation of Regional Ionospheric Models for Improved Precise Point Positioning

Presenting author: Wei-Chun YU

Authors: Wei-Chun YU, National Cheng Kung Univ.; Ming YANG, National Cheng Kung Univ.; Feng-Yu CHU, National Cheng Kung Univ.

Keywords: Global Navigation Satellite System; Precise Point Positioning; Regional Ionospheric Model; Machine Learning

Precise Point Positioning (PPP) is a high-precision positioning technique based on the Global Navigation Satellite Systems (GNSS). By utilizing a single GNSS receiver in combination with precise satellite orbit and clock products, as well as various correction models for atmospheric and instrumental errors, PPP is capable of mitigating major error sources and achieving centimeter-level positioning accuracy. Owing to its independence from local reference stations, PPP offers significant flexibility and wide-area applicability. However, one of its primary limitations is the relatively long convergence time required to reach high accuracy, which restricts its effectiveness in time-sensitive or real-time applications. To address this limitation, incorporating ionospheric information as an external constraint within PPP processing has proven to be an effective strategy for accelerating convergence and enhancing positioning performance. In Taiwan, the National Land Surveying and Mapping Center (NLSC) routinely generates real-time Regional Ionospheric Models (RIM) using data collected from a dense network of GNSS Continuously Operating Reference Stations (CORS). These high-resolution regional ionospheric products provide valuable information that can be integrated into PPP to reduce ionospheric-related uncertainties and improve overall solution stability and accuracy. In recent years, rapid advances in machine learning and artificial intelligence have opened new possibilities for geospatial modeling and data-driven prediction. Motivated by these developments, this study proposes the application of machine learning techniques to perform spatio-temporal interpolation of the RIM. By capturing complex spatial and temporal patterns in ionospheric variations, the proposed approach aims to generate more accurate ionospheric corrections. The performance of the machine learning-based interpolation method is evaluated against conventional interpolation techniques to assess its effectiveness in improving PPP convergence speed and positioning accuracy.

Back to session list

B2-3: Deciphering Environmental Controls of Total Hydrocarbon Accumulation in Taiwan Using Geospatial Explainable Machine Learning

Presenting author: Thia PRAHESTI

Authors: Thia PRAHESTI, National Cheng Kung University; Aji Kusumaning ASRI, National Cheng Kung University; Yu-Ting ZENG, National Cheng Kung University; Shih-Chun Candice LUNG, Academia Sinica, National Taiwan University; Chih-Da WU, National Cheng Kung University, National Health Research Institutes, National Chung Hsing University, Kaohsiung Medical University, Chang Gung University of Science and Technology

Keywords: Ensemble Modeling; Environmental Controls; Geospatial Explainable Machine Learning; Interpretable Spatial Mapping; Total Hydrocarbon (THC)

Total hydrocarbon (THC) accumulation represents a persistent challenge in urban air quality management due to its complex environmental regulation and spatial heterogeneity. While predictive models can estimate concentration levels, understanding the environmental controls underlying THC accumulation and their spatial expression remains essential for mechanistic insight. This study deciphers the dominant environmental drivers of THC accumulation in Taiwan using a Geospatial Explainable Machine Learning framework that integrates predictive modeling with interpretable spatial mapping. THC observations from nationwide monitoring stations were combined with atmospheric chemistry, land-use patterns, spatial data layers, and topographical and meteorological dynamics to capture multi-scale environmental influences. Five machine learning algorithms, Extreme Gradient Boosting Regression, Light Gradient Boosting Machine Regression, Gradient Boosting Regression, Random Forest Regression, and CatBoost Regression were developed to model nonlinear relationships among predictors. The most robust models were integrated within an ensemble framework to enhance generalization and stability, with validation procedures ensuring model reliability. Beyond predictive performance, SHAP (Shapley Additive Explanations) was employed to quantify variable importance and interaction effects, enabling driver-informed interpretation. The framework generates interpretable spatial maps that reveal environmental control regimes and hotspot structures across Taiwan, illustrating how chemical, meteorological, and land-use factors jointly regulate THC accumulation patterns. By coupling ensemble modeling, explainable analytics, and spatial mapping, this study advances a mechanism-oriented approach that supports evidence-based air quality management and strategic environmental planning.

Back to session list

B2-4: Integrating Geospatial Analysis and Hybrid Deep Learning Model for Forest Fire Susceptibility Mapping in Sanggau Regency, West Kalimantan, Indonesia

Presenting author: Suci RAMAYANTI

Authors: Suci RAMAYANTI, Physics Study Program, Faculty of Mathematics and Natural Sciences Education, Universitas Pendidikan Indonesia, Bandung 40154, West Java, Indonesia; Muhammad Fulki FADHILLAH, Physics Study Program, Faculty of Mathematics and Natural Sciences Education, Universitas Pendidikan Indonesia, Bandung 40154, West Java, Indonesia; Mimin IRYANTI, Physics Study Program, Faculty of Mathematics and Natural Sciences Education, Universitas Pendidikan Indonesia, Bandung 40154, West Java, Indonesia; Lilik HASANAH, Physics Study Program, Faculty of Mathematics and Natural Sciences Education, Universitas Pendidikan Indonesia, Bandung 40154, West Java, Indonesia; Ida HAMIDAH, Automotive Engineering Education Study Program, Faculty of Engineering and Industrial Education, Universitas Pendidikan Indonesia, Bandung 40154, West Java, Indonesia

Keywords: Kalimantan; Forest Fire Susceptibility; Climate Action; FIRMS; CNN

Kalimantan is one of the regions in Indonesia that is highly vulnerable to forest fires, particularly during the dry season. Mapping forest fire susceptibility is essential for effective risk mitigation and management. This study aims to identify and map forest fire–susceptible areas in Sanggau Regency, West Kalimantan, by integrating geospatial analysis with hybrid deep learning models. Historical forest fire data from July to October 2023 were obtained from MODIS, NOAA-20, and Suomi NPP satellite imagery through the Fire Information for Resource Management System (FIRMS) to develop inventory map. Sixteen geospatial factors related to forest fire occurrence, including topographic, optical index, climatic, and anthropogenic variables, were compiled and selected using the Frequency Ratio (FR) method. The Variance Inflation Factor (VIF), Information Gain Ratio (IGR), and Information Gain Average Entropy (IGAE) methods were applied to evaluate feature importance and remove redundant variables. Forest fire susceptibility modeling was performed using a Convolutional Neural Network (CNN) and two hybrid models optimized with Independent Component Analysis (ICA) and the Grey Wolf Optimizer (GWO). The resulting susceptibility map was classified into five categories: very low, low, moderate, high, and very high. Model performance was evaluated using the Area Under the Receiver Operating Characteristic Curve (AUC). A comparative analysis was conducted between the CNN-base and the hybrid models (CNN–ICA and CNN–GWO) to determine the optimal approach. The results indicate that the CNN–GWO model achieved the highest predictive performance, with an AUC value of 0.912, followed by CNN–ICA (AUC = 0.883) and the standard CNN-base (AUC = 0.824). Elevation, land use, NDMI, and wind speed consistently emerged as the most influential factors across all feature selection iterations. These findings provide valuable spatial information to support forest fire mitigation planning and to enhance public awareness of fire risk in the study area.

Back to session list

B2-5: Accuracy Assessment of SfM/MVS-Based 3D Reconstruction Using RTK-Enabled Smartphone Imagery under Varying GNSS Environments

Presenting author: Yonghe LI

Authors: Yonghe LI, Nihon Univ.

Keywords: RTK-GNSS positioning; smartphone images; Structure from Motion (SfM); Multi-View Stereo (MVS); 3D model reconstruction; positioning accuracy

This study evaluates the factors influencing the accuracy of 3D models reconstructed by Structure from Motion and Multi-View Stereo (SfM/MVS) using smartphone images augmented with high-precision RTK-GNSS positioning. The objective is to clarify how GNSS reception environments and image acquisition strategies affect camera positioning accuracy and the resulting 3D model quality. Field experiments were conducted on pedestrian spaces under three different conditions: open-sky, sparsely tree-obstructed, and building-tree mixed environments. Images were acquired using an iPhone 15 Pro equipped with the LRTK Phone device, and 3D models were generated with Agisoft Metashape. The results show that in open-sky and sparsely tree-obstructed environments, stable RTK fix rates close to 100% were obtained, enabling centimeter-level accuracy in both horizontal and vertical directions without the use of ground control points. In contrast, in the building-tree mixed environment, fix rates decreased to below 50%, leading to significant degradation in model accuracy. Nevertheless, even under low fix rates, practical accuracy was achieved in some cases depending on the distribution and quality of fix solutions. Additional experiments under open-sky conditions revealed that image acquisition strategies had limited influence on horizontal accuracy, whereas vertical accuracy was more sensitive to shooting direction diversity and the number of acquisition paths. Bidirectional acquisition and increased course numbers improved geometric stability and reduced height errors. These findings demonstrate that RTK-assisted smartphone photogrammetry can provide practical 3D modeling performance, while highlighting the importance of GNSS environment assessment and optimized image acquisition design for reliable results.

Back to session list

Session C2 Abstracts

C2-1: Seasonal Land Cover Classification in Japanese Agricultural and Wetland Landscapes via Machine Learning and Multi-Source Sentinel Data

Presenting author: Santa PANDIT

Authors: Santa PANDIT, IIS The University of Tokyo; Kazuo OKI, IIS The University of Tokyo, and KUAS, Kyoto; Timothy DUBE, University of the Western Cape, South Africa; Salem Ibrahim SALEM, Plymouth Marine Laboratory, Prospect Place, Plymouth, UK; Tadanobu OKUMURA, Wildlife Management Office Inc., Tokyo; and Masayasu MAKI, Fukushima University, Japan

Keywords: diverse-landscape; cross-seasonal; dual-sentinel; ecosystem classification; machine learning

This study assessed land use and land cover (LULC) in the Oze wetland and Hatase agricultural fields using Random Forest (RF) and Support Vector Machine (SVM).Sentinel-1 Synthetic Aperture Radar (SAR) and Sentinel-2 optical data from2023–2024 were processed into seasonal median composites. Input features included SAR backscatter coefficients (vertical transmit–vertical receive, vertical transmit–horizontal receive, and their ratio), Sentinel-2 bands (10 m resolution), and vegetation indices: Normalized Difference Vegetation Index (NDVI), Normalized Difference Built-up Index (NDBI), Bare Soil Index (BSI), and Modified Normalized Difference Water Index (MNDWI). Training and testing data were derived from high-resolution Planet Scope and drone imagery. Models were implemented in Python(Google Colab). Results showed RF consistently outperformed SVM, achieving kappa scores of 81%–83% in Oze and 79%–81% in Hatase, while SVM failed to exceed 80%.RF’s robustness for seasonal LULC mapping highlights its potential to support monitoring and sustainable land management in cloud-prone wetland–agriculture systems.

Back to session list

C2-2: National-Scale Detection of Golf Courses by Convolutional Neural Network model and Land-Cover Characterization Using Landsat Time Series in Japan

Presenting author: Shoki SHIMADA

Authors: Shoki SHIMADA, Earth Observation Research Center, Japan Aerospace Exploration Agency; Kei OYOSHI, Earth Observation Research Center, Japan Aerospace Exploration Agency

Keywords: Grassland ecosystem, Forest ecosystem, Solar PV, UNet, Machine Learning

Golf remains a popular sport in Japan, with more than nine million players in 2024, although participation has declined by approximately four million over the past 25 years. This decline has reduced the economic sustainability of golf courses, leading to their gradual closure. Despite this trend, a detailed national-scale map of golf courses is not currently available. Moreover, golf courses have recently attracted attention for their potential role in maintaining grassland ecosystems, while abandoned courses may be increasingly converted to other land uses, including solar photovoltaic (PV) installations following the 2012 Feed-in Tariff (FIT) policy. This study aims to (1) produce national-scale golf-course maps using Landsat imagery and a convolutional neural network (CNN), (2) quantify land-cover characteristics within golf courses with a focus on grassland ecosystems, and (3) assess land-cover transitions of declining golf courses. Pan-sharpened 15 m spatial resolution Landsat-7/8 RGB images were prepared for two periods: 2008–2011 (pre-PV boom) and 2022–2024 (post-PV boom). A total of 1,443 image patches (256 × 256 pixels) with manually annotated golf-course masks were prepared. A UNet architecture with an EfficientNet-B7 encoder achieved an average intersection-over-union (IoU) score of 0.80. The trained model was applied nationwide, identifying 2,146 golf-course polygons (2,245 km²) in the pre-PV period and 2,059 polygons (2,161 km²) in the post-PV period. Grassland was the dominant land-cover type within golf-course boundaries and accounted for 15–20% of total prefectural grassland area in regions such as Chiba, and Kanagawa. Land-cover change analysis revealed that approximately 15% of converted golf-course areas transitioned to solar PV, a disproportionately high share given PVs occupy only ~0.15% of Japan’s land area. These results demonstrate that golf courses can be reliably detected using satellite time series and play a significant role in grassland provision, while recent PV expansion is increasingly reshaping abandoned golf-course landscapes.

Back to session list

C2-3: Stability-Controlled Land-Cover Monitoring in Access-Restricted Regions: The DMZ-Adjacent Iron Triangle (Cheorwon–Pyonggang–Kimhwa), DPRK

Presenting author: Beom KIM

Authors: Kihun KIM, Institute for North Korean Studies, Dongguk University; Beom KIM, Florida State University

Keywords: Land-Cover Change; Stability Filtering; Proxy Supervision; Remote Sensing; DPRK

Land-cover monitoring in access-restricted regions such as North Korea lacks ground truth and is prone to mixed-pixel and boundary-driven instability. Apparent land-cover “change” may reflect algorithmic reallocation rather than environmental processes. This study presents an automated workflow that separates defensible change signals from classification noise. Results are reported for Cheorwon County, with extension to the full Iron Triangle underway. Proxy labels from 0.5 m imagery (Google Earth) were cross-checked; 285 Cheorwon tiles (95 for validation) were used to fine-tune SatlasPretrain on a nine-band Sentinel-2/Sentinel-1/DEM stack. Annual maps were derived from season-consistent Aug–Sep composites. Model outputs were stabilized using hierarchical hybrid segmentation integrating Random Forest refinement with index- and terrain-informed post-processing. Spatial transfer on 101 proxy-labeled Kimhwa tiles (~40 km from Cheorwon) achieved 4-class deep learning segmentation mIoU = 0.721 and pixel accuracy = 0.847. To control temporal instability, Flip is defined as the number of land-cover class transitions per pixel between 2020 and 2025. Pixels with Flip ≤ 1 (allowing one spurious reassignment) account for 72.59% of the study area. Transition matrix analysis within this stability-controlled domain indicates strong persistence, with cropland and woody vegetation showing 87.42% and 79.78% persistence, respectively, while exchanges between waterbody and wetland classes concentrate along riparian boundaries. Within stability-controlled pixels, Theil–Sen trend analysis identifies a statistically significant decrease in waterbody area (−0.13 km² yr⁻¹, p = 0.026) and a concurrent increase in wetland area (+0.11 km² yr⁻¹, p < 0.001), indicating gradual hydrological reconfiguration, consistent with a shift from open water to wetland under uncertainty control rather than algorithmic artifacts. The workflow is implemented as an automated pipeline enabling surveillance across the Iron Triangle and other border regions.

Back to session list

C2-4: Identifying Suitable Locations for Green Infrastructure Development around the Settlement Zone of Ulaanbaatar City Using Satellite Data

Presenting author: Enkhzaya Enkhtaivan

Authors: Enkhzaya ENKHTAIVAN, Centre for Policy Research and Analysis; Bayanmunkh NOROVSUREN, Centre for Policy Research and Analysis; Bolormaa BATSUURI, National University of Mongolia; Angarag ZUNDUI, Baljinnyam BATTSOLMON, Tsend-Ayush BAASANJAV

Keywords: Geospatial, green space, NDVI, Satellite image, AHP

In Ulaanbaatar, 49% of Mongolia’s total population resides, and green space per capita is 5.9 m². Expanding green infrastructure in and around the city’s settlement zone is important for mitigating multiple urban challenges, including air pollution, the urban heat island effect, flood risk, and the decline in environmental quality. This study assesses the status of green infrastructure around the settlement zone of Ulaanbaatar and identifies suitable areas for expansion using satellite-based information. Using Sentinel-2 data, key factors including land use, urban density, elevation, slope, and the drainage network were derived and integrated within a Geographic Information System (GIS). Suitability for green infrastructure expansion was evaluated through the Analytic Hierarchy Process (AHP) and classified into five categories: highly unsuitable, unsuitable, moderately suitable, suitable, and highly suitable. Spatial data analysis further suggests a high potential for increasing green infrastructure in the sparsely vegetated mountainous areas along the city’s northern outskirts, as well as in wetlands and along riverbanks. If green infrastructure is expanded by 45,785.39 ha, per-capita green space could increase to approximately 9 m². In the future, we will conduct research to identify tree species suitable for establishing green infrastructure and will present the findings at the decision-making and planning levels.

Back to session list

C2-5: DEVELOPMENT OF A JAXA HIGH-RESOLUTION LAND-USE AND LAND-COVER MAP OF SOUTHEAST ASIA USING ALOS-2/PALSAR-2 AND OPTICAL TIME SERIES

Presenting author: Naoyoshi HIRADE

Authors: Naoyoshi HIRADE, RESTEC/Tsukuba Univ.; Koichi IMAMURA, RESTEC; Sota HIRAYAMA, JAXA EORC; Ryo SATAKE, RESTEC; Fumi OHGUSHI, RESTEC; Kenlo NISHIDA NASAHARA, Tsukuba Univ.; Takeo TADONO, JAXA EORC

Keywords: ALOS-2; Sentinel-2; LULC; classification; SouthEastAsia;

This study presents a JAXA High-Resolution Land-Use and Land-Cover (HRLULC) map of Southeast Asia for 2023. The dataset was publicly released in September 2025 through the Japan Aerospace Exploration Agency (JAXA) Earth Observation Research Center (EORC) to support regional environmental monitoring, ecosystem assessment, and sustainable development efforts. This product is called 2023SEA_v25.09. 2023SEA_v25.09 was produced by integrating multi-sensor satellite observations from ALOS-2/PALSAR-2 and Sentinel-2/MSI, together with the ALOS World 3D digital surface model (DSM) and topographic slope. SAR is especially valuable in the humid tropics because it is not affected by cloud cover and provides stable structural information. For ALOS-2/PALSAR-2, WBD inputs were prepared as a six-period sequence (HH, HV, and local incidence angle) to capture seasonal patterns, while FBD features were included as a single-period input to add complementary SAR information. Sentinel-2/MSI Level-2A surface reflectance data were summarized into six bi-monthly periods using a weighted median composite to obtain stable seasonal features under frequent tropical cloud cover. Supervised classification was performed using SACLASS2.5, designed for multispectral and multi-temporal feature representations. A 15-class scheme was developed through expert consultations to reflect the regional forestry and agricultural characteristics of Southeast Asia. The scheme includes region-specific categories such as Single-crop Paddy Field, Multi-crop Paddy Field, Rubber Tree Plantation, and Oil Palm Tree Plantation, which are often not clearly separated in global products. Accuracy assessment using an independent test dataset (5,688 points) across 619 geographic tiles achieved an overall accuracy of 94.59% with a kappa coefficient of 0.94 (the evaluation is an independent expert-labeled dataset). The 2023SEA_v25.09 offers higher spatial resolution and a more detailed class scheme than existing global LULC products, providing a robust baseline for forest biomass estimation, agricultural monitoring, and land-use change analysis in Southeast Asia.

Back to session list

Session D2 Abstracts

D2-1: Evaluating the Climate-Adaptive Role of Green Belt Areas Using Satellite-Derived LST and NDMI

Presenting author: Yoonji KIM

Authors: Yoonji KIM, Korea Univ.; Youngjae YOO, Korea Univ.; Sangwook LEE, Korea Univ.; Jinhoo HWANG, Korea Univ.

Keywords: LST; NDMI; Green Belt; Urban Heat Island; Climate Resilience; Land-Use Change

Rapid urban expansion under the climate crisis has intensified surface heat stress and ecological degradation, raising critical questions regarding the climate-adaptive role of peri-urban conservation areas. Although Green Belt areas (Restricted Development Zone) are not legally designated as protected areas, they function as spatial buffers that restrict urban sprawl, preserve green infrastructure, and sustain ecosystem services. From a functional and ecological perspective, they may therefore be conceptualized as de facto protected areas contributing to urban climate resilience. This study evaluates the climate-adaptive function of Green Belt areas using satellite-based remote sensing data. Land Surface Temperature (LST) and the Normalized Difference Moisture Index (NDMI) were derived from multi-temporal Earth observation imagery to quantify thermal and ecological conditions, respectively. LST, retrieved from thermal infrared bands, represents surface radiant temperature and serves as a key indicator of urban heat stress. NDMI, calculated from near-infrared and shortwave-infrared reflectance, captures vegetation moisture content and provides insight into ecosystem health and hydrological stability. By integrating these two remote sensing–based indicators, the study assesses both thermal and ecological dimensions of climate resilience in a spatially explicit manner. A comparative analysis was conducted between maintained and released Green Belt zones to examine the impacts of land-use change. The results indicate that maintained areas exhibit moderated LST increases and relatively stable NDMI values, whereas released areas show intensified surface warming and declining vegetation moisture conditions. Spatial pattern analysis further reveals differentiated thermal vulnerability and ecological degradation associated with development intensity. These findings demonstrate the capacity of remote sensing indicators to provide robust, spatially continuous evidence for evaluating the climate-regulating functions of conservation-oriented land-use policies. The study highlights the importance of integrating Earth observation–based indicators into urban planning and supports the recognition of Green Belt areas as critical climate-resilient infrastructure in rapidly urbanizing regions.

Back to session list

D2-2: Impacts of Leaf Area Index Data Assimilation on Hydrological Responses to Drought and Flood Events in the Korean Peninsula

Presenting author: Hee-Jeong JEONG

Authors: Hee-Jeong JEONG, Department of Earth System Sciences, Yonsei University; Hahn Chul JUNG, Department of Earth System Sciences, Yonsei University; Seung Hwan LEE, Department of Earth System Sciences, Yonsei University; Yeosang YOON, Hydrological Sciences Laboratory, NASA Goddard Space Flight Center

Keywords: Leaf Area Index, Data Assimilation, Land Surface Modeling, Drought and Flood Events, Hydrological Responses

Understanding hydrological responses to extreme events, such as droughts and floods, is crucial for improving water resource management in the face of intensifying climate variability. In regions with pronounced seasonal vegetation dynamics, such as the Korean Peninsula, vegetation–hydrology interactions play a key role in regulating land surface hydrological responses. Satellite-based data assimilation (DA) provides an effective framework for incorporating vegetation information into land surface models by systematically integrating observational constraints into model simulations. In particular, the Leaf Area Index (LAI), as a key indicator of vegetation growth and stress, enables a more realistic representation of vegetation dynamics when assimilated into land surface models, thereby improving the simulation of seasonal hydrological and energy processes. In this study, we investigated the impacts of LAI data assimilation on hydrological and vegetation responses to extreme events across the Korean Peninsula. Simulations were performed using the Korea Land Data Assimilation System (KLDAS), built on the NASA Land Information System (LIS) framework with the Noah-MP model. By assimilating 500 m MODIS LAI (MCD15A3H), we improved the representation of vegetation dynamics at a 1 km spatial resolution for the period 2002–2024. Compared to the Open-Loop (OL) simulation without data assimilation, the overestimation of LAI was substantially reduced, leading to improved estimates of key hydrological variables. In particular, significant improvements in evapotranspiration and soil moisture were observed over agricultural and cropland regions, where vegetation dynamics have a strong control over land–atmosphere water exchange. Additional analyses under drought and flood conditions revealed that LAI data assimilation effectively modulates hydrological responses by realistically representing vegetation stress processes. These results highlight the potential of LAI-DA to improve reliable assessments of hydrological impacts under drought and flood conditions and to support sustainable water resource management in the Korean Peninsula.

Back to session list

D2-3: Multi‑Sensor Assessment of Soil Moisture Seasonality in Semi‑Arid Mongolia Using AMSR2 and Ground Measurements

Presenting author: Nozomu HIROSE

Authors: Nozomu HIROSE, National Institute of Technology, Matsue College; Kentarou AIDA, JAXA; Ichirou Kaihotsu, Hiroshima Univ.

Keywords: Soil Moisture; AMSR2; Mongolian Plateau; In-situ Field Observation; Seasional Change

This study analyzes long‑term soil moisture variability on the semi‑arid Mongolian Plateau and assesses the reliability of satellite‑derived estimates. In‑situ measurements at 3 cm were compared with AMSR2 and SMOS data set from 2015–2025, alongside MODIS NDVI and snow‑cover products. AMSR2 reproduced seasonal patterns but consistently overestimated moisture under dry conditions, underestimated values above 15%, and showed no clear response to spring snow cover.

Back to session list

D2-4: Toward a Global Hyper-Temporal Land Monitoring Framework Using an International Constellation of Geostationary Meteorological Satellites

Presenting author: Kazuhito ICHII

Authors: Kazuhito ICHII, Chiba Univ.; Yuhei YAMAMOTO, Chiba Univ.; Wei LI, Chiba Univ.; Beichen ZHANG, Chiba Univ.; Wei YANG, Chiba Univ.; Rui FU, Chiba Univ.; Misaki HASE, Taiga SASAGAWA, Chiba Univ.; Masayuki MATSUOKA, Mie Univ.; Hiroki YOSHIOKA, Aichi Prefectural Univ.

Keywords: Geostationary Satellite, Terrestrial Monitoring, Himawari-8/9, Carbon Cycle, Drought, Phenology

Recent advances in third-generation geostationary meteorological satellites have enabled unprecedented high-frequency observations of the terrestrial environment. Platforms such as Himawari-8 / Himawari-9, GOES-16, GK-2A, and FY-4A provide sub-hourly multispectral observations covering most land areas of the globe. These systems allow us to observe land surface processes continuously throughout the day, rather than relying on one or two observations per day from polar-orbiting satellites. In this presentation, we introduce our international initiative to build a coordinated global framework for land monitoring based on multiple geostationary satellites. The project focuses on harmonizing data from different satellite systems, applying consistent atmospheric and angular corrections, and producing analysis-ready datasets for terrestrial applications. We present recent results on (1) data pre-processing of datasets including atmospheric and BRDF correction, (2) monitoring vegetation phenology and daily photosynthetic activity, (3) detecting vegetation stress during heatwaves and droughts, and (4) integrating satellite observations with flux measurements and data-driven ecosystem models to improve estimates of gross primary production and evapotranspiration. Because geostationary satellites continuously observe the same region throughout the day, they enable direct monitoring of diurnal changes and rapid ecosystem responses that are difficult to capture with conventional satellites. We actively welcome collaborative and application-oriented studies using our geostationary datasets. Potential research topics include carbon–water cycle dynamics, agricultural monitoring, ecosystem resilience, and early warning of climate extremes.

Back to session list

D2-5: ASSESSING THE IMPACT OF CLIMATE CHANGE ON SOLAR RESOURCE IN THE PHILIPPINES USING MODIS AND IN-SITU MEASUREMENT

Presenting author: Jeark PRINCIPE

Authors: Jezleer MONTAJES, University of the Philippines Diliman; Rosalie REYES, University of the Philippines Diliman; Jeark PRINCIPE, University of the Philippines Diliman

Keywords: solar radiation, climate change, Philippines, MODIS, GCM

Solar radiation (SR) drives Earth’s climate; however, it may be influenced by climate change. Global circulation models (GCMs) use mathematical equations to simulate the effects of climate change. In the Philippines, climate change assessments have been conducted for rainfall and temperature, but a comprehensive assessment for SR is still lacking. This study evaluates the impact of climate change on SR through statistical downscaling of small- and large-scale variables. Moderate Resolution Imaging Spectroradiometer (MODIS)-derived SR (DSR) and in situ weather measurements are used as small-scale variables, while baseline and future GCM projections serve as large-scale variables. Climate projections indicate an increase in SR ranging from 120 to 224 W/m²/day during the remaining years of the 21st century. The City of Dagupan shows the lowest increase (120 W/m²/day) under the Shared Socioeconomic Pathway (SSP) 2-4.5 mid-future scenario (2015–2050), whereas Davao City exhibits the highest increase (224 W/m²/day) under the same scenario. Locations closer to the equator generally show greater increases than those farther away. Small-scale variables strongly influence future projections; higher MODIS DSR values correspond to higher projected SR. Other weather factors, such as rainfall, show an inverse relationship with SR projections. A decrease in rainfall is associated with an increase in SR. Davao City, which records the highest projected SR increase, also shows the largest projected rainfall decrease (−3.60 mm/day) under SSP2-4.5 mid-future. Projected increases in SR have both positive and negative sectoral impacts. In agriculture sector, higher SR may enhance crop yields. Renewable energy systems may benefit from increased electricity generation, and both agricultural systems and power production could be optimized through agrivoltaic systems. However, increased SR may also pose health risks to humans, including sunburn and immunodepression, and may alter animal behavior and thermoregulatory responses.

Back to session list

Session E2 Abstracts

E2-1: An Integrated Framework for Single-Image Automated Reconstruction of Detailed Wireframe Objects Using YOLO, SAM2, and Ray Intersection Geometry

Presenting author: Yi Hsuan SU

Authors: Yi Hsuan SU, National Yang Ming Chiao Tung University; Tzu Yi CHUANG, National Yang Ming Chiao Tung University

Keywords: Single-image photogrammetry; Wireframe reconstruction; Uncertainty propagation; Geometry-aware deep learning

Single-image reconstruction of façade details is geometrically indeterminate due to limited observational redundancy. Although existing 3D building wireframe models provide structural priors, detailed façade elements, such as doors, windows, and ventilation units, are often omitted or manually constructed, resulting in incomplete and non-reproducible geometry. This study presents an integrated framework for automated reconstruction of detailed wireframe objects from a single image by fusing deep learning perception with photogrammetric ray intersection geometry under explicit structural constraints. Based on an existing 3D building wireframe model, YOLO performs object localization, and SAM2 refines object boundary segmentation. Extracted image features are projected into object space via ray-intersection geometry, constrained by the known structural configuration. To address single-image uncertainty, an uncertainty modeling scheme quantifies image-space detection variance and propagates it to object-space coordinates via analytical error propagation. The covariance of reconstructed vertices is derived from the first-order linearization of the ray intersection equations, enabling a quantitative assessment of positional reliability and geometric confidence. The proposed AI-geometry fusion framework eliminates manual point measurement while enforcing spatial determinacy via structural constraints, yielding closed, topologically consistent wireframe representations. Experimental results demonstrate stable reconstruction performance and measurable accuracy under single-image conditions, validating the feasibility of geometry-aware AI automation for detailed 3D building refinement.

Back to session list

E2-2: Quantitative Assessment and Algorithm Development for High-Resolution LiDAR-Derived Tidal Creek Extraction in Korea’s West Coast Tidal Flats

Presenting author: Ye-Young KIM

Authors: Ye-Young KIM, Pukyong National University; Su-Bin HA, Pukyong National University; Seung-Kuk Lee, Pukyong National University

Keywords: High-resolution LiDAR; Tidal flats; Tidal creeks;Airborne LiDAR;Drone LiDAR;

Tidal flats evolve rapidly under tides, sediment transport, and fluvial inputs, and tidal creeks are the main corridors that route water and sediment and structure erosion–deposition patterns. Because creek networks can shift quickly with changing hydrodynamics and sediment supply, reliable monitoring requires fine-scale extraction and quantitative description of creek networks. Many existing products, however, are derived from coarse DEMs that under-detect narrow, branching channels and limit robust geometric analysis. This study develops and validates a high-resolution LiDAR-based tidal creek extraction algorithm using 0.5 m DEMs from airborne and drone LiDAR over two representative sites (Hwangdo and Gomso Bay) in Korea’s west coast tidal flats. The algorithm applies a watershed-based sequence—sink correction, flow direction and accumulation, threshold-based channel delineation, connectivity-oriented pruning, and raster-to-vector conversion—to produce creek centerlines. Parameters are tuned via sensitivity analyses to support multi-scale extraction, from main channels to fine tributaries visible in LiDAR topography. We quantify network structure using density and connectivity metrics and estimate creek width and depth from elevation cross-sections sampled perpendicular to centerlines. Field measurements of creek width are used to evaluate LiDAR-derived estimates and constrain parameter choices. Cross-platform consistency is examined by comparing network metrics and width/depth distributions between airborne and drone outputs within each site. Results show that high-resolution LiDAR yields more complete and coherent creek networks than low-resolution DEM-based delineations, particularly by preserving small branches and network continuity. Standardized vector outputs and attributes enable interannual comparison and integration with sediment-facies or habitat data to support long-term monitoring and management of Korea’s west coast tidal flats.

Back to session list

E2-3: Overlap-Based Extrapolation for Partial Geometric Correction in GCP-Free Segments of Strip Imagery

Presenting author: Jaewon HUR

Authors: Jaewon HUR; Han-Gyeol KIM; Hongjin KIM

Keywords: Geometric Correction, Strip Imagery, Polynomial Extrapolation, GCP-deficient regions

Strip imagery acquired by BlueBON may include segments where reliable ground control points (GCPs) cannot be secured due to clouds, water bodies, or low-contrast surfaces. Consequently, geometric correction is possible only for limited portions of the strip. In segments where correction fails, accurate image-to-ground correspondence based on true GCPs is unavailable. Pseudo GCPs are therefore generated to predict corresponding positions and enable geometric correction. However, their generation requires prior estimation of two-dimensional image correspondences. This study addresses partial geometric correction and proposes an overlap-based extrapolation framework to estimate geometric relationships in uncorrected segments. The overlapping region between a corrected and an uncorrected segment is defined as an anchor region. Within this region, the relationship between image and ground coordinates along the strip line direction is approximated using a polynomial model. The model is extrapolated to non-overlapping regions to predict additional correspondences for two-dimensional registration. Image coordinates (c, r) are sampled at regular grid intervals, and corresponding ground coordinates (x, y) are derived using SRTM DEM elevation values (z) to generate pseudo GCPs. These pseudo GCPs are used to re-estimate the rational polynomial model (RPC) and perform geometric correction of the failed segment. Correction performance is evaluated using phase correlation between the corrected result and an orthorectified image. Residual behavior under different DEM roughness conditions is analyzed to assess terrain sensitivity. Results confirm that pseudo GCP-based correction remains stable in GCP-deficient regions. The framework quantitatively evaluates the feasibility and limitations of overlap-based extrapolation for partial geometric correction of strip imagery.

Back to session list

E2-4: Feasibility Assessment of Depth-based GCPs Weighting Approach for Geometric Correction of BlueBON Satellite Image

Presenting author: Hongjin KIM

Authors: Hongjin KIM, TelePIX; Han-Gyeol KIM, TelePIX; Jaewon HUR, TelePIX

Keywords: Geometric correction, GCPs, Depth estimation, BlueBON, CubeSat

With the increasing utilization of CubeSat images, the need for precise geometric correction is growing. However, radiometric and geometric inconsistencies between satellite images often degrade the extraction of reliable ground control points (GCPs) through matching techniques. To address this, this study assesses the feasibility of a method that utilizes deep learning-based depth estimation to verify GCPs and adjusts weights for rational polynomial coefficients (RPCs) estimation for BlueBON satellite images. Initially, matching points were extracted using SuperPoint and LightGlue, and a second-order polynomial transformation was applied to provide a coarse geometric alignment. Subsequently, image patches corresponding to the matching points were extracted, and dense correspondences were obtained using the LoFTR and grid-based pseudo GCPs generation. To enhance the robustness of matching results, DepthAnythingV2 is employed to assess the 3D structural consistency of correspondences, thereby effectively filtering out geometric outliers. We performed RPCs estimation by assigning increased weights to GCPs with a depth map correlation exceeding a threshold between the image patches. The proposed method was validated using five datasets, with 10 check points (CPs) extracted from Google Maps for each dataset. Experimental results demonstrated significant accuracy improvements, achieving a geometric accuracy with a root mean square error (RMSE) of less than 2 pixels. Furthermore, qualitative evaluation confirmed a marked reduction in geometric distortion within the generated orthoimages. Consequently, this study validates the efficacy of integrating 3D structural constraints into the GCPs refinement process. By effectively compensating for the relatively low image quality and limited ancillary data inherent to CubeSat systems, this approach provides a robust solution for highly accurate geolocation.

Back to session list

E2-5: Integrated Analysis of Long-term Field Sediment Data and UAV-Derived Topographic Information for Interpreting Spatial Variability in a Tidal Flat

Presenting author: Sehee KIM

Authors: Sehee KIM, Korea Institute of Ocean Science & Technology (KIOST); Keunyong KIM

Keywords: Tidal flat; Sediment variability; Coastal geomorphology; Long-term monitoring

Gyeonggi Bay, located on the west coast of Korea, is a representative macro-tidal environment with a tidal range of approximately 6–9 m and is one of the major coastal regions in East Asia where extensive intertidal flats are well developed. The Songdo and Siheung tidal flats have undergone continuous land reclamation and urban development, leading to complex interactions between anthropogenic disturbances and tidal-driven sedimentary processes. This study integrates long-term field-based sediment data collected from 2018 to 2025 with high-resolution UAV-derived orthophotos and a Digital Elevation Model (DEM) acquired in 2025 to investigate spatial relationships between sediment characteristics and micro-topographic structures in the Songdo and Siheung tidal flats. The Songdo tidal flat was classified into two sub-areas (Area A and Area B) based on sedimentary characteristics. Grain size composition, total organic carbon (TOC), acid volatile sulfide (AVS), and trace metal concentrations were analyzed to evaluate long-term variability. The results indicate that Area A of Songdo and the Siheung tidal flat have persistently maintained fine-grained, silt-dominated sedimentary environments, with occasional exceedances of environmental guideline levels (TEL) for AVS and selected trace metals. In contrast, Area B of Songdo exhibits relatively stable sand-dominated sediment characteristics. UAV-DEM analysis revealed distinct elevation gradients and microtopographic variability associated with main tidal channels, and these subtle elevation differences closely correspond to the spatial distribution of surface sediment facies. This integrated assessment provides quantitative insights into geomorphology–sediment interactions in a macro-tidal urban coastal environment and supports precise monitoring and management strategies for rapidly developing tidal flats.

Back to session list

Session A3 Abstracts

A3-1: A Georeferenced Neural Radiance Field Framework for DSM Generation from UAV Imagery

Presenting author: Cheng-Hsin LI

Authors: Cheng-Hsin LI, National Cheng Kung University; Farhan Ardianzaf PUTRA, National Cheng Kung University; Chao-Hong LIN, National Cheng Kung University; Bo-Yi LIN, Taiwan Space Agency ; Yu-Lin TSAI, Taiwan Space Agency

Keywords: Digital Surface Model; Neural Radiance Fields; Photogrammetry; 3D Reconstruction; Georeferencing; Orthographic UAV Imagery

In remote sensing, the Digital Surface Model (DSM) serves as a pivotal digital representation of the physical world, providing a foundation for urban planning, environmental monitoring, and spatio-temporal analysis. DSMs are commonly generated from Unmanned Aerial Vehicle (UAV) imagery via Structure-from-Motion (SfM) and Multi-View Stereo (MVS) algorithms, which ensure rigorous consistency in both coordinate systems and metric scale. However, photogrammetry-based methods face significant image matching challenges when scenes lack sufficient reliable feature points. In the domain of deep learning, Neural Radiance Fields (NeRFs) offer a novel paradigm for reconstructing 3D scenes from multi-view image sets. By incorporating ray-based information into neural networks, NeRFs represent complex environments as a continuous implicit field of color and volume density. In this paper, we propose a georeferenced NeRF-based framework designed to achieve high-precision geospatial alignment and ensure the geometric consistency. In the future, we will aim to eliminate spurious sparse geometries and refine the topographic accuracy of the generated DSM.

Back to session list

A3-2: 4-Stream Mid-Level Fusion with nDSM Confidence Gating for Urban Green Space Extraction from RGB and LiDAR-Derived Images

Presenting author: Yosuke OMIZU

Authors: Yosuke OMIZU, Toho Univ.; Yuki SHIBATA, Toho Univ.; Yuu NAKAJIMA, Toho Univ.; Misuzu KAKIZAKI, Toho Univ.

Keywords: Deep Learning; Semantic Segmentation; airborne LiDAR; Urban Green Space; Multimodal Fusion; nDSM Confidence Gating

Urban Green Spaces (UGS) mitigate the urban heat island effect and reduce urban floods. However, despite their recognized importance, efforts to accurately extract UGS remain limited. While deep learning models are actively developed in remote sensing, their application in Japan is hindered by distinct urban structures and a lack of large-scale datasets. Therefore, it is essential to develop a method capable of adapting to Japan’s urban structures and extracting UGS even with limited data. We used 0.2 m resolution images covering a 12-ha area, consisting of an aerial orthoimage and airborne LiDAR-derived images (nDSM, Number of Returns, and Intensity) from 16 points/m^2 data. To integrate these multimodal inputs, we proposed a UNetFormer-based four-stream mid-level fusion architecture. It employs independent ResNet-18 encoders for each input stream, fusing features at each encoder stage via Stage Fusion Blocks. These blocks learn stream-specific gating weights via sigmoid-activated convolutional layers to adaptively modulate the contribution of LiDAR-derived features at different feature levels. Furthermore, we introduce an nDSM confidence gating module that down-weights nDSM features out of the training distribution, stabilizing inference on data from unseen regions. We pre-trained the RGB encoder and decoder on the ISPRS Potsdam dataset (five classes) and the model was fine-tuned on data from a low-rise exclusive residential district in Tama City, Tokyo. Evaluating the proposed method across five land use districts yielded an average IoU of 0.650 (SD=0.060) for green spaces, confirming its effectiveness. This represents an IoU improvement of 0.057 over an RGB-only baseline, and 0.139 over a multimodal baseline without the nDSM confidence gating module. These findings suggest that the proposed method is a promising for practical UGS extraction. Furthermore, our study indicates the utility of increasingly open LiDAR data in Japan, which is expected to encourage data release and broader use across diverse fields.

Back to session list

A3-3: Integrating RTK-GNSS with 3D Gaussian Splatting for Georeferenced Scene Reconstruction

Presenting author: Chi-Min CHIU

Authors: Chi-Min CHIU, Dept. of Civil Engineering, National Yang Ming Chiao Tung University, Taiwan; Tee-Ann TEO, Dept. of Civil Engineering, National Yang Ming Chiao Tung University, Taiwan

Keywords: 3D Gaussian Splatting; Digital Cadastral Mapping; 3D Reconstruction

3D Gaussian Splatting (3DGS) has recently attracted significant attention in 3D computer graphics and photogrammetry, particularly offering efficient rendering and high-quality 3D reconstruction. While 3DGS is widely used for general scenes, its application in high-precision cadastral mapping remains largely unexplored. This study addresses this gap by developing a structured workflow to apply 3DGS to land surveying using a smartphone. To ensure the spatial accuracy required for land administration, the proposed method integrated RTK-GNSS within the orientation determination process to determine absolute camera poses for 3DGS model initialization. We performed a comparative analysis of different 3DGS methods to evaluate their geometric accuracy and reconstruction quality for cadastral mapping. Overall, these results provide the technical validation needed to adopt 3DGS in cadastral mapping workflows. This study establishes a foundational framework that offers a scalable path toward the modernization of 3D cadastral mapping.

Back to session list

A3-4: Characterizing Regime Shifts in Mekong Delta Droughts from 1982 to 2025: Multi-Scale Insights from Continuous Wavelet Transform

Presenting author: Hong Danh PHAM PHAN

Authors: Hong Danh PHAM PHAN; Hone-Jay CHU; Mohammad Adil AMAN

Keywords: Drought, Precipitation, Mekong River Delta, Continuous Wavelet Transform

Drought in the Mekong River Delta (MRD) is a sophisticated climatic phenomenon driven by multi-scale interactions, primarily reflected through significant precipitation variability. As a global agricultural hub, the MRD’s vulnerability to climate extremes necessitates monitoring tools that go beyond traditional station-based metrics, which often overlook the dynamic evolution and frequency-dependent nature of these events. This study introduces a multi-scale spatiotemporal framework utilizing Continuous Wavelet Transform (CWT) to isolate and map drought “hotspots” across diverse temporal scales from 1981 to 2024. Using data of Climate Hazards Group InfraRed Precipitation with Station (CHIRPS) satellite precipitation anomalies, the research employs a complex Morlet wavelet to decompose hydrological signals into three distinct bands: sub-annual (< 1 year), annual (1 year), and inter-annual (2–8 years) cycles. By extracting Wavelet Real Coefficients and spatiotemporal visualization over the study area, this methodology establishes a robust physical distinction between moisture deficits (negative phases/drought) and surpluses (positive phases/flood). The resulting isosurface visualizations reveal that drought hotspots manifest with unique morphological characteristics across scales. Findings indicate that the sub-annual band is dominated by fragmented, localized, and transient weather noise with rapid phase transitions, allied with the common behavior and characteristics of the “flash drought” weather condition in the MRD. Conversely, the 2–8 year inter-annual band captures massive, spatially-cohesive drought clusters synchronized with ENSO oscillations. Analysis of milestone years, such as 1987-1988, 2015-2016, and 2019-2020 reveals distinct spatial propagation patterns, with coastal regions exhibiting significantly higher energy accumulation compared to inland areas. This multi-scale approach transitions from static and traditional snapshots to a dynamic diagnostic tool, providing high-fidelity insights into the rhythmic evolution and regional vulnerabilities of extreme drought in the MRD.

Back to session list

A3-5: Species-Aware Panoptic Segmentation of UAV-SfM Point Clouds Using an Extended SegmentAnyTree Framework

Presenting author: Kei YAMAMOTO

Authors: Kei YAMAMOTO, Shimane Univ.; Yasumichi YONE, Shimane Univ.; Akitsugu IMAOKA, Hiroshima Prefectural Government

Keywords: UAV, LiDAR, Individual Tree Detection, Deep Learning, SfM

Tree-level species mapping from UAV photogrammetry is increasingly demanded for forest inventory application. A common workflow derives a DSM/CHM from Structure-from-Motion (SfM) and applies a local maximum filter (LMF) to detect treetops; however, LMF is sensitive to crown overlap in dense stands, and rasterization can discard fine 3D canopy-surface structure, making results dependent on grid resolution and search-window settings. To overcome these limitations, we developed a species-aware panoptic segmentation model that directly exploits SfM point clouds for simultaneous individual-tree delineation and conifer species classification by extending SegmentAnyTree (SAT), with a view toward future multispectral SfM point clouds that could enable broadleaf species discrimination. We prepared a labeled dataset of sugi (Cryptomeria japonica) and hinoki (Chamaecyparis obtusa) from three sites in Shimane and Hiroshima, Japan (160 trees in total). RGB imagery was acquired using Zenmuse P1 and Mavic 3M. SfM point clouds were processed to extract single-tree regions, assign tree IDs and species labels, and compute normalized RGB features (norm_r, norm_g, norm_b) used together with XYZ as model inputs. To ensure both species were present within the training sampling radius (8 m), single-tree point clouds were spatially shifted and merged to create mixed-species training scenes. SAT was fine-tuned from a LiDAR-pretrained checkpoint with a three-class semantic head (non-tree, sugi, hinoki) and evaluated on held-out test data from the same stand. In intra-stand tests, the model achieved a mean IoU of 0.61, with class-wise IoU of 0.94 (sugi) and 0.87 (hinoki). Instance-level F1 averaged 0.78, with species-wise F1 of 0.85 (sugi) and 0.70 (hinoki). Preliminary cross-stand experiments indicate that species classification can degrade under stand-level domain shift, even when instance delineation remains viable. These findings highlight domain robustness as a key barrier to operational deployment of species-aware panoptic segmentation from SfM point clouds, motivating further work on generalization-oriented training and evaluation.

Back to session list

A3-6: Rice Plant Coverage Estimation from UAV Data Using a Gaussian Mixture Model

Presenting author: Hiroto YOSHIDA

Authors: Hiroto YOSHIDA, Mie Univ.; Masayuki MATSUOKA, Mie Univ.

Keywords: Unmanned Aerial Vehicle; Paddy fields; Point cloud; Multispectral image; Gaussian Mixture Model

The use of Unmanned Aerial Vehicle (UAV) for agricultural field monitoring has attracted increasing attention. Agricultural management is shifting from field-level control to within- field precision management, and ultimately toward plant-level monitoring. However, the spatial resolution of UAV imagery remains insufficient to directly capture conditions at the individual plant level. Therefore, it is necessary to establish methods for estimating sub-pixel information from UAV observations. The goal of this study is to estimate rice canopy coverage at the individual plant level using a Gaussian Mixture Model (GMM). As an initial step, this study aims to establish a method for determining Gaussian components using UAV data acquired over paddy fields. Accurate estimation of Gaussian components representing rice, weeds, soil, and water is essential for high-resolution subpixel unmixing of UAV-derived imagery. Height information derived from point cloud data generated through Structure-from-Motion (SfM) processing of UAV-derived imagery is utilized to enhance the reliability of endmember selection. Point clouds and an orthomosaic images of the target field were generated using commercial SfM software. Height information derived from the point clouds was used to identify pure rice pixels, while other components were identified from the orthomosaic images. Reflectance values were extracted from multispectral images comprising the red, green, red-edge, and near-infrared bands. The density distribution of reflectance values in feature space is examined, and high-density regions are regarded as representative spectral characteristics of each surface component. These Gaussian components were defined as endmembers and applied to the GMM to evaluate their validity. By integrating height information with spectral characteristics, the proposed method enables more reliable endmember determination and contributes to improved sub-pixel analysis in paddy field monitoring.

Back to session list

Session B3 Abstracts

B3-1: An Integrated Spatial Assessment of Urban Fire Risk Using Multi-Source Data: A Case Study of Ulaanbaatar City

Presenting author: KHOSTSETSEG.T National Center for Public Health

Authors: AMARBAL.A Centre for policy research and analysis of capital city, Ulaanbaatar, Mongolia; BATSUREN.B Centre for policy research and analysis of capital city, Ulaanbaatar, Mongolia; KHOSTSETSEG.T National Center for Public Health; NURMAA.D Centre for policy research and analysis of capital city, Ulaanbaatar, Mongolia

Keywords: Fire risk zoning, Analytic Hierarchy Process (AHP), Hexagonal tessellation, Geospatial analysis

The aim of this study is to identify fire risk zoning in Ulaanbaatar City through an evidence-based spatial assessment by integrating multi-source statistical and geospatial data. The novelty of the research lies in the integrated evaluation of diverse datasets within a unified analytical framework, including population concentration—particularly high-density hexagonal grid units and child age groups (0–1, 2–5, and 6–17 years)—fire emergency calls and object-based fire incident records, the locations and storage capacities of fuel stations and warehouses, electrical power substations, soil moisture conditions, and building characteristics. Furthermore, the spatial relationships between the locations and capacities of fuel stations and the distribution of the child population were examined in detail. In terms of materials and methods, multi-source input datasets were aggregated into hexagonal tessellation units and converted into raster format. A Multi-Criteria Decision Analysis (MCDA) approach was applied to evaluate fire risk across the study area. The relative weights of the evaluation criteria were determined using the Analytic Hierarchy Process (AHP), and an integrated fire risk index was calculated. The general form of the index is expressed as: FIRE RISK = 0.4255·F + 0.2269·P + 0.1804·N + 0.0933·D + 0.0476·C + 0.0264·R. As a result, a city-wide fire risk zoning map for Ulaanbaatar was produced, identifying 37 hexagonal units as high-risk zones, 186 units as medium-risk zones, and 604 units as no-risk zones. The analysis shows that 34% of the city’s total child population resides in high-risk areas, while 46.3% lives in medium-risk areas. These findings provide critical evidence for enhancing child safety, supporting preventive urban planning, and optimizing the allocation of fire prevention and response resources. Overall, this study serves as a foundational, evidence-based assessment to support decision-making for measuring, monitoring, and managing urban fire risk

Back to session list

B3-2: High-Resolution PM2.5 Mapping and Low-Exposure Route Optimization Using Geo-AI and Microsensors

Presenting author: Yu-Hsuan LIN

Authors: Yu-Hsuan LIN, National Cheng Kung Univ.; Yu-Ju LIN, National Cheng Kung Univ.; Chia-Wei HSU, National Cheng Kung Univ.; Yu-Chia LIN, National Cheng Kung Univ.; Chih-Da WU, National Cheng Kung Univ.

Keywords: Geo-AI; Microsensors; PM2.5; Route Optimization; Air Pollution Exposure; Machine Learning

Rapid urbanization has intensified spatial variability in urban air quality, posing elevated exposure risks for active commuters such as pedestrians and cyclists. Traditional navigation systems typically prioritize the shortest distance or travel time, overlooking the spatial heterogeneity of air quality and its associated health impacts. This study addresses this limitation by developing an exposure-aware routing framework for Taiwan’s Central Air Quality Zone. A comprehensive dataset spanning September 2021 to June 2024 was integrated, comprising PM2.5 measurements from 11 regulatory monitoring stations and over 2,000 microsensors, alongside meteorological, land use, and transportation network variables. A three-step Geo-AI-based spatial modeling framework was developed. First, Random Forest Regression (RFR) was applied to calibrate microsensor data, achieving a coefficient of determination (R2) of 0.94. Second, Extreme Gradient Boosting Regression (XGBR) was selected for spatial estimation, attaining an R2 of 0.96 and generating high-resolution (50 × 50 m) PM2.5 concentration maps. This framework enables fine-scale representation of pollution heterogeneity for exposure-aware network analysis. Finally, Dijkstra’s algorithm was employed to compute and compare shortest-distance and minimum-exposure routes. Validation using 3,000 randomly generated origin-destination pairs demonstrated that the proposed routing strategy achieved an average exposure reduction of 3.08%, with a maximum decrease of 26.4% in areas exhibiting substantial pollution variability. These findings highlight the potential of integrating microsensor networks with Geo-AI spatial modeling for high-resolution environmental mapping and exposure-aware routing. The proposed framework provides a practical basis for balancing mobility efficiency with health protection in high-density urban environments.

Back to session list

B3-3: Geo-spatial analysis and decision-Making Modeling to identify crime prevention zones in Ulaanbaatar

Presenting author: JAVZANDULAM.B

Authors: BATSUREN.B, Centre for policy research and analysis of capital city, Ulaanbaatar, Mongolia; NURMAA.D, Centre for policy research and analysis of capital city, Ulaanbaatar, Mongolia; JARGALSAIKHAN.T, Centre for policy research and analysis of capital city, Ulaanbaatar, Mongolia; JAVZANDULAM.B, Mongolian University of Science and Technology B.BOLORMAA, National University of Mongolia, School of Arts and Sciences

Keywords: Geo-spatial; Reported crimes; NDVI; Impact factor; AHP; CPTED

This study, covering the six central districts of Ulaanbaatar City, integrates data on the ten most frequently reported types of crimes in public areas in 2024 (GPD, 2024), population distribution (NSO, 2024), vegetation index (NDVI) derived from Sentinel satellite imagery (SCLD, 2024; CPC, 2024), and surveillance camera locations (IARRO, 2024). A geo-spatial analysis (Jeffery, 1971; 1977) was conducted using ArcGIS Pro decision-making modelling tools to identify priority areas for implementing CPTED (Crime Prevention Through Environmental Design) (Saaty, 1977). The data were aggregated into 64-hectare hexagonal grids and converted into raster datasets. Thirteen selected factors were weighted and evaluated using the AHP (Analytic Hierarchy Process) decision-making model. In 2024, a total of 32,123 crimes (GPD, 2024) were recorded in Ulaanbaatar, of which 7,490 cases (23.3%) occurred in public spaces. High-risk areas were mainly concentrated in the central parts of Bayanzurkh, Sukhbaatar, Bayangol, and Chingeltei districts. The analysis revealed a tendency for higher crime density in areas with low NDVI values, indicating a lack of green spaces. According to the AHP results, “surveillance camera coverage” (20.21%) was ranked as the most influential factor, followed by “robbery” (12.31%), “circumstances leading to suicide” (11.53%), “causing moderate bodily harm” (11.52%), and “traffic regulation violations” (10.37%). The consistency ratio (CR) was 0.1, indicating a high level of consistency in the weighting results. The study area was classified into five hazard levels (I–V). The Level I zone, identified as the “high-risk area” where CPTED should be prioritized, accounted for 3% of the total area, followed by Level II – 7%, Level III – 11%, Level IV – 27%, and Level V – 52%, respectively.

Back to session list

B3-4: Observation-Driven Agent-Based Evacuation Modeling with Time Resolved Wildfire Perimeters

Presenting author: JongHwan KIM

Authors: JongHwan KIM, Kookmin univ; Seunghyun HONG, Kookmin univ; Heejung NAM, Kookmin univ; Seyeon HWANG, Kookmin univ; Wanmo KANG, Kookmin univ

Keywords: Wildfire; Agent-based modeling; Dynamic evacuation; Remote sensing; Network accessibility

Wildfires in the forest–urban interface can disrupt transportation networks and constrain evacuation options as hazard conditions evolve. While many evacuation studies rely on static wildfire footprints, relatively few incorporate time-resolved wildfire observations and capacity-constrained shelter operations within a behaviorally explicit modeling framework. Accounting for these dynamic interactions is critical for a realistic assessment of evacuation feasibility. This study applies a time-dependent agent-based modeling (ABM) framework driven by wildfire perimeter delineations and firelines derived from operational wildfire mapping. Discrete, irregularly time-stamped wildfire boundaries were compiled into an irregularly sampled time series and used to update hazard states dynamically without predictive wildfire spread modeling. At each simulation step, road segments intersecting the evolving perimeter and buffer zones were classified as inaccessible or high-risk, thereby imposing time-varying constraints on evacuation routes. Agents exhibited heterogeneous response delays and mode-specific travel speeds (vehiclebased vs. pedestrian evacuation) and adaptively rerouted under changing network accessibility. Shelter locations and capacity limits were represented to enable admission or redirection as occupancy evolved, and congestion was quantified using segment-level traffic accumulation. Scenario experiments indicate that integrating time-resolved wildfire perimeters and shelter capacity constraints substantially alters evacuation performance compared with static hazard assumptions. Delayed departures and slower travel modes resulted in longer travel times and greater exposure durations, while progressive network disruption concentrated congestion and exposed vulnerable links not detected under static representations. Overall, the proposed framework provides a geospatially explicit foundation for coupling observed wildfire evolution with behaviorally differentiated evacuation modeling to support evacuation and shelter planning in wildfire-prone forest–urban interface regions. * This study was conducted with the support of the R&D program for Forest Science Technology (project no. RS-2025-25438293) provided by Korea Forest Service (Korea Forestry Promotion Institute).

Back to session list

B3-5: GeoAI Framework for Urban Flood Resilience Integrating Sentinel 1 SAR Change Detection and H3 Index Location-Based Services

Presenting author: AHMAD HABIBULLOH

Authors: Ahmad HABIBULLOH, National Taiwan University of Science and Technology; Firdausi Zahara GANDES, Institut Teknologi Bandung; Jyun-Ping JHAN, National Taiwan University of Science and Technology

Keywords: GeoAI, H3 Index, Change Detection Methodology, Urban Flood Resilience, Location-Based Service

Urban flooding is a critical issue in populated metropolitan areas, where population growth intensifies hazard exposure and strains emergency response capacities. However, existing evaluations often overlook the intersection of hazard exposure and availability of shelters and healthcare, restricting actionable information. This study develops a comprehensive, GeoAI-enhanced system with automated satellite flood detection using location-based services (LBS) for Jakarta, Indonesia. A change detection methodology is utilized to determine flood extents by processing Sentinel-1 Synthetic Aperture Radar (SAR) imagery through Google Earth Engine. The framework differentiates transient flooding from permanent water bodies by analyzing the variation in radar backscatter between pre- and post-flood images. After incorporating SAR, the 2025 Flood Risk Index (FRI), emergency shelter locations from the Indonesian National Disaster Management Agency (BNPB), and healthcare facility points of interest from OpenStreetMap, the data is normalized using the H3 hexagonal geospatial index across 267 administrative units. A multi-criteria assessment of past floods is completed to categorize risk into five. These profiles are made using standard distance intervals: critical (0.4–334.4 m), high (334.4–757.4 m), moderate (757.4–1038.6 m), low (1038.6–1424.0 m), and minor (1424.0–2500.0 m). Spatial analysis depicted that several grid cells with a high flood risk consist of zero shelter or healthcare facilities. The integration of the insights and the FRI highlights areas where high flood risk coincides with limited access to emergency infrastructure. To connect the gap between data and action, these risk profiles are deployed via a location-aware Telegram LBS chatbot. This tool equips citizens with personalized risk scores, localized mitigation techniques, and supply checklists activated by real-time GPS (Global Positioning System) coordinates. This multipurpose framework provides a scalable model for urban resilience, providing government agencies with a dynamic basis for infrastructure enhancement and providing residents with immediate, actionable GeoAI-driven intelligence.

Back to session list

B3-6: Geo-AI-Based Assessment of 3-D Ultrafine Particle Distribution and Population Exposure: A Digital Twin Approach in Taichung, Taiwan

Presenting author: Chia-Wei HSU

Authors: Chia-Wei HSU, National Cheng Kung Univ.; Jun-Jun SU, National Cheng Kung Univ.; Rui-Zhen YANG, National Cheng Kung Univ.; Candera WIJAYA, Agricultural Engineering Research Center; Yu-Cheng CHEN, National Health Research Institutes; Shih-Chun Candice LUNG, Academia Sinica; Ta-Chih HSIAO, Academia Sinica; Chao-Hung LIN, National Cheng Kung Univ.; Chih-Da WU, National Cheng Kung Univ.

Keywords: UAV; Geo-AI; Digital Twin; PM₀.₁; 3-D Mapping; Spatial Analysis

Despite the prevalence of high-rise urban structures in Asia, air quality monitoring remains primarily two-dimensional (2-D), leaving the vertical dynamics of pollutants absent from public risk awareness. While ultrafine particles (PM₀.₁), characterized by their extremely small size, pose significant health risks, they are not widely monitored internationally. This study integrated 3-D sampling with Geospatial Artificial Intelligence (Geo-AI) to develop a high-resolution 3-D estimation and mapping framework for PM₀.₁. An unmanned aerial vehicle (UAV) platform equipped with a high-precision P-Trak counter was deployed in Taichung, Taiwan, to capture horizontal and vertical pollution profiles across complex land-use types. These data were integrated with multi-source geospatial datasets, including 3-D building models, meteorological variables, and localized emission sources. The SHapley Additive exPlanations (SHAP) method was first employed to identify and rank the significance of key spatial predictors, which were subsequently utilized to fit machine-learning models. The resulting model demonstrated robust predictive performance, with an R² of 0.95 in training and exceeding 0.85 in external validation, 10-fold cross-validation, and stratified validation. The 3-D estimation results of PM₀.₁ revealed distinct non-linear vertical stratification, with significant hotspots identified not only in ground-level traffic zones but also at mid-to-high altitudes due to horizontal transport. Through population-weighted exposure calculations, this study further visualized exposure risks of PM₀.₁ across different floor levels, enabling urban planners and public health authorities to identify high-risk zones by distinguishing between absolute concentration levels and actual human exposure.

Back to session list

Session C3 Abstracts

C3-1: Rockfall Prediction Revisited: Surprising Strength of Logistic Regression Over Ensemble Models

Presenting author: Yun-Ching Chen

Authors: Yun-Ching Chen, Department of Civil Engineering, National Taipei University of Technology; Walter Chen, Department of Civil Engineering, National Taipei University of Technology

Keywords: rockfall prediction; logistic regression; random forest; machine learning

Predicting rockfall occurrence accurately is vital to highway safety and reduces risks in mountainous regions. Machine learning (ML) is widely used for rockfall susceptibility mapping, and ensemble methods such as random forests (RF) are often considered the most effective. However, using an extensive rockfall dataset from highways across Taiwan, the logistic regression (LogReg) outperformed the RF and eight other common classifiers (such as support vector machine, gradient boosting, multilayer perceptron, and k-nearest neighbors) according to the ROC (receiver operating characteristic) curves, with an AUC (area under the curve) of 0.9349. Even though the RF performed well, the LogReg model demonstrated greater robustness despite its simplicity, suggesting that this dataset is better suited to linear decision boundaries. These results urge a re-examination of baseline models in geoscience and show that systematic benchmarking is essential to prevent simple yet highly effective approaches from being overlooked in rockfall hazard assessment.

Back to session list

C3-2: Automated Landslide Detection via Segment Anything Model Integrating Topographic and Vegetation Features

Presenting author: Ling-Chien Kao

Authors: Ling-Chien Kao; Walter Chen

Keywords: Segment Anything Model; Landslide; Image Segmentation; Automatic Point Prompts; Google Earth Engine

This study utilizes Google Earth Engine (GEE) to establish a cloud-based automated workflow, selecting the Laonung River basin in southern Taiwan as the study area. Within areas above 500 meters in elevation, 100 squares of 300×300 meters were randomly sampled as analysis and validation units, and Sentinel-2 and SPOT 6 images were cropped according to these squares. Subsequently, using slope, curvature, and Normalized Difference Vegetation Index (NDVI) as interpretation criteria, point prompts required for the Segment Anything Model (SAM) were programmatically generated to automatically segment landslide areas and produce landslide masks. This approach aims to reduce the time cost of manual point selection and delineation while enhancing workflow consistency. To verify the reliability of the method, this study also employed Computer Vision Annotation Tool (CVAT) combined with SAM to perform semi-automated delineation as a reference, comparing its consistency with the results from the automated point prompt delineation. The results show that for Sentinel-2, the Intersection over Union (IoU) was 0.366 and the Dice coefficient was 0.536, while for SPOT 6, the IoU was 0.420 and the Dice coefficient was 0.592. These findings indicate that this workflow possesses a certain level of interpretative capability and can serve as a foundation for future large-scale applications and disaster monitoring.

Back to session list

C3-3: A Georeferenced Neural Radiance Field Framework for DSM Generation from UAV Imagery

Presenting author: Cheng-Hsin LI

Authors: Cheng-Hsin LI, National Cheng Kung University; Farhan Ardianzaf PUTRA, National Cheng Kung University; Chao-Hong LIN, National Cheng Kung University; Bo-Yi LIN, Taiwan Space Agency ; Yu-Lin TSAI, Taiwan Space Agency

Keywords: Digital Surface Model; Neural Radiance Fields; Photogrammetry; 3D Reconstruction; Georeferencing; Orthographic UAV Imagery

In remote sensing, the Digital Surface Model (DSM) serves as a pivotal digital representation of the physical world, providing a foundation for urban planning, environmental monitoring, and spatio-temporal analysis. DSMs are commonly generated from Unmanned Aerial Vehicle (UAV) imagery via Structure-from-Motion (SfM) and Multi-View Stereo (MVS) algorithms, which ensure rigorous consistency in both coordinate systems and metric scale. However, photogrammetry-based methods face significant image matching challenges when scenes lack sufficient reliable feature points. In the domain of deep learning, Neural Radiance Fields (NeRFs) offer a novel paradigm for reconstructing 3D scenes from multi-view image sets. By incorporating ray-based information into neural networks, NeRFs represent complex environments as a continuous implicit field of color and volume density. In this paper, we propose a georeferenced NeRF-based framework designed to achieve high-precision geospatial alignment and ensure the geometric consistency. In the future, we will aim to eliminate spurious sparse geometries and refine the topographic accuracy of the generated DSM.

Back to session list

C3-4: Investigation of Spatiotemporal Patterns and Associated Factors of Surface Deformation in Taoyuan (2018–2020)

Presenting author: Tsung-Chun PENG, National Chengchi University

Authors: Tsung-Chun PENG, National Chengchi University; Shih-Yuan LIN, National Chengchi University

Keywords: SBAS-InSAR, GTWR, Spatiotemporal analysis, Taoyuan.

This study applies a two-stage workflow to (i) map time-series surface deformation and identify annual deformation hotspots in the Taoyuan area from 2018 to 2020, and (ii) examine spatiotemporally varying associations between deformation and candidate explanatory factors. In the first stage, Sentinel-1 SAR imagery is processed using an SBAS-InSAR approach to derive deformation time series and annual cumulative deformation for each year. The results indicate localized uplift and subsidence hotspots with distinct interannual differences in both magnitude and spatial extent during 2018–2020. Hotspot coverage accounts for 47.61%, 42.76%, and 66.95% of the study area in 2018, 2019, and 2020, respectively. Over 2018–2020, the maximum deformation velocity reaches 12.54 mm/yr for uplift and -6.07 mm/yr for subsidence. In the second stage, a 100-m gridded spatiotemporal dataset is constructed by linking SBAS-derived deformation metrics with natural and anthropogenic variables, including building weight, groundwater level, normalized difference vegetation index (NDVI), distance to MRT lines, distance to the roads, pond density, population density, rainfall, topographic wetness index (TWI), and average shear-wave velocity (Vs30). A geographically and temporally weighted regression (GTWR) model is then fitted to estimate location- and time-specific coefficients. To focus interpretation on well-explained locations, only grids with local model fit exceeding a threshold (local R² > 0.5) are retained for subsequent factor-deformation analysis. The estimated coefficients show that the dominant associated factors vary across subregions and years: rainfall (+) dominates subsidence in 2018, whereas groundwater level emerges as the primary driver in subsequent years, exhibiting a positive association (+) with both uplift and subsidence in 2019, and a negative association (-) in 2020. These findings provide an empirical, spatially explicit basis for interpreting deformation patterns and prioritizing monitoring in a rapidly developing metropolitan region.

Back to session list

C3-5: Mapping Taiwan’s Village Foodscapes, 2014–2022: A Bayesian Multilevel Spatial Analysis of Self-rated Health and Happiness

Presenting author: Chia-Jung LIN

Authors: Chia-Jung LIN; Stephan VAN GASSELT

Keywords: Food environment, foodscape, self-rated health, happiness, subjective well-being, multilevel models, BYM2, nighttime lights, spatial analysis, Taiwan

Unhealthy diets are a major driver of noncommunicable diseases, yet evidence on how local foodscapes relate to subjective well-being in dense East Asian settings remains limited, and spatial clustering is often ignored. This study estimates associations between village-level food environments and self-rated health and happiness in Taiwan from 2014 to 2022. We link restricted-use Taiwan Social Change Survey (TSCS) microdata to a national village–year panel built from business tax registration records, population and socioeconomic indicators, historical village boundaries, and satellite-derived nighttime lights. Using industry codes to identify food-related businesses, we construct village–year counts from approximately 870,000 tax-registration records covering food retail and food service, which are used to derive outlet density, composition, and diversity measures. To capture broader local activity and amenity intensity beyond registered outlets, we incorporate VIIRS annual nighttime-light composites as a village-level proxy for urban vitality and after-dark economic activity. TSCS is fielded twice yearly with about 1,800 respondents per wave, yielding roughly 32,000 observations with village identifiers, self-rated health and happiness, and individual covariates. We fit individual-level Bayesian multilevel models in R with year fixed effects and village-level BYM2 spatial random effects to account for adjacency-based dependence and unobserved heterogeneity under unbalanced village coverage. We expect healthier retail profiles, especially greater availability of fresh food outlets and supermarkets relative to convenience and fast-food options, to be associated with better self-rated health. and we expect higher densities of eating and drinking venues to be positively associated with happiness via amenity and psychosocial pathways. By integrating nationwide administrative business records, satellite-based vitality indicators, and repeated survey data at the village scale within a spatial multilevel framework, this study provides a scalable approach to map community foodscapes and quantify how everyday commercial landscapes relate to both physical and subjective well-being in Taiwan.

Back to session list

C3-6: Wide-Area Subsidence Monitoring of Reclaimed Land in Osaka Bay Using Time-Series InSAR

Presenting author: Haruka NARUKE

Authors: Haruka NARUKE, Department of Civil and Environmental Engineering, Institute of Science Tokyo; Masashi MATSUOKA, Department of Architecture and Building Engineering, Institute of Science Tokyo; Fumitaka OGUSHI, Terra Phase, Inc.

Keywords: ALOS-２/PALSAR-2; Sentinel-1; land subsidence; MT-InSAR; infrastructure monitoring; seabed foundation

Land subsidence continues to threaten the structural integrity of Osaka Bay’s vital artificial islands and airports even 30 years after reclamation. The seabed foundation of this area is soft and complex, attracting many geological experts to study this area. However, the settlement characteristics during construction and the precise mechanisms of the settlement process remain unclarified. This study aims to characterize wide-area ground deformation using multi-temporal InSAR (MT-InSAR) and identify the dominant factors of long-term subsidence by integrating satellite data with geological and construction records. We conducted MT-InSAR analysis using 42 ALOS-2/PALSAR-2 scenes (2014/10/05–2025/09/07) and 264 Sentinel-1 scenes (2015/11/18–2025/09/14), calibrating the resulting subsidence velocities with GNSS data. Results revealed that significant subsidence has been confirmed in the Phase 2 construction area of Kobe Port Island, the southeastern part of Rokko Island, Yumeshima, etc. Major differential subsidence was identified through strain analysis and corroborated by field observations of structural cracks and deformation. Based on our findings, we propose that two factors primarily drive this differential subsidence. First, strong correlations with soft ground conditions (Vs < 350m/s) and upper Pleistocene clay thickness suggest that consolidation of deep clay layers (30 m to several hundred meters) including secondary consolidation of quasi-overconsolidated clay significantly influences long-term deformation. Second, differences in ground improvement: disparities were linked to the implementation, timing, and specific types of sand drain methods employed during reclamation. This study demonstrates that MT-InSAR, integrated with subsurface data, enables the quantitative evaluation of complex mechanical behaviors, such as the secondary consolidation of quasi-overconsolidated Pleistocene clay. Furthermore, it enables the evaluation of the long-term effectiveness of settlement mitigation through ground improvement based on actual observation data. SAR-based analysis serves as an essential framework for elucidating the behavior of reclaimed land on deep-water soft ground, thereby contributing to sustainable coastal development strategies.

Back to session list

Session D3 Abstracts

D3-1: Thermal Regulation across Fragmented Mangrove Patch Mosaics: Age- and Environment-Dependent Patterns from 30 Years of Landsat Data in Subtropical Taiwan

Presenting author: Pei Wen LIN

Authors: Pei Wen LIN, National Chengchi university

Keywords: Mangroves; Carbon sinks; Global Warming; Landsat; LST; GEE

With record-level global temperature rise in 2024 and 2025, as reported by the World Meteorological Organization (WMO) in the State of the Global Climate report, global warming has become one of the most critical global challenges, particularly for Asia, which is warming twice as fast as the global average. As industrial carbon emissions are a major driver of global warming, nature-based solutions to store carbon and mitigate impacts have gained widespread support. Mangrove ecosystems act as significant carbon sinks due to long-term carbon storage in biomass and soils. They may also provide additional ecosystem services through a potential cooling effect; however, direct evidence of their effectiveness remains limited. Using the WMO climatological standard normal (1991–2020), we integrated long-term mangrove distribution data from the Mangrove Dynamics in China (MDC) dataset with Landsat-derived land surface temperature (LST) data processed on the Google Earth Engine (GEE) platform to quantify cooling effects of fragmented mangrove patches across lagoons, riverbanks, and nearshore coastal environments in subtropical Taiwan. Because mangrove distributions have continuously changed over the period, climate-regulating functions may not respond as rapidly as spatial distribution dynamics. We used age-classified mangrove patches to assess cooling capacity. Land surface temperatures over water bodies and built-up areas were calculated as reference baselines representing the coolest and hottest conditions, respectively. We further developed a Temperature Regulation Ratio (TRR) to quantify relative cooling and thermal stabilization effects. Results indicate that mangrove cooling is more pronounced along riverbanks than in lagoons. Mature mangroves were associated with an average temperature reduction of about 2°C, whereas younger patches showed limited or even reverse cooling capacity. These findings suggest that mangrove maturity and environmental context strongly shape thermal regulation performance, underscoring the importance of conserving mature stands and incorporating age structure into coastal climate adaptation and nature-based solution strategies.

Back to session list

D3-2: Preliminary Evaluation of a Bio-Optical Model for Chlorophyll-a Estimation in Shallow Coastal Waters of the Seto Inland Sea

Presenting author: Haoran ZHANG

Authors: Haoran ZHANG, Hiroshima Univ.; Yuji SAKUNO, Hiroshima Univ.

Keywords: chlorophyll-a, bio-optical model, coastal waters

Accurate estimation of chlorophyll-a (Chl-a) in optically complex coastal waters is challenging due to the combined effects of phytoplankton, suspended sediments, and water itself. Previous studies in the Seto Inland Sea have primarily relied on empirical relationships between in-situ spectral reflectance and measured Chl-a; however, such approaches are often limited by site-specific conditions. This study extends previous work by applying and evaluating a bio-optical model proposed by Oyama (2009) in shallow coastal waters of the Seto Inland Sea, Japan. In-situ spectral reflectance data covering 400–900 nm and surface Chl-a values were collected at five stations in Mitsu Bay in 2023. Instead of directly regressing reflectance against Chl-a, the observed spectra were analyzed using the bio-optical modeling framework to separate contributions from phytoplankton, non-phytoplankton suspended sediments, and clear water components. The derived coefficients representing phytoplankton-related optical components were then compared with measured Chl-a concentrations to assess the applicability of the model in this turbid coastal environment. Preliminary analysis indicates that the bio-optical modeling approach can better represent the optical complexity of the study area compared with conventional empirical methods based solely on reflectance band relationships. Although the current results are still under validation, the model shows potential for improving Chl-a estimation in shallow and optically complex coastal waters. This study provides an initial evaluation of a bio-optical model in the Seto Inland Sea and establishes a basis for further verification using additional in-situ datasets and satellite observations, such as Sentinel-2 data, in future work.

Back to session list

D3-3: SHIP DETECTION AND RE-IDENTIFICATION USING MULTI-TEMPORAL NANOSATELLITE IMAGERY IN BUSAN PORT

Presenting author: Seunggi JEONG

Authors: Seunggi JEONG, Seoul nationial university; Taehong KWAK, Seoul national university; Yongil KIM, Seoul national university

Keywords: Ship Detection, Ship Re-Identification, Nanosatellite Imagery

Ship detection and tracking are essential for applications such as maritime surveillance and monitoring. Conventional ship tracking predominantly relies on Automatic Identification System (AIS) data. However, AIS-dependent ship-monitoring approaches are constrained by intentional signal deactivation and deliberate attempts to evade surveillance. In such scenarios, satellite imagery provides a viable alternative for ship tracking without dependence on AIS. To address these issues, this study aims to introduce a deep learning-based ship detection and Re-Identification (Re-ID) model using multi-temporal nanosatellite images. In the detection stage, a YOLO-OBB model was employed, and a domain-specific training dataset was constructed to fine-tune the model for improved detection accuracy. In the re-identification stage, visual embedding feature vectors were extracted from each detected ship image using a ship Re-ID model. The similarity between ship embeddings was then evaluated to determine whether ships observed at different times correspond to the same ship. Experimental results show that the detection model achieves reliable ship detection performance on PlanetScope imagery in Busan port, particularly for medium-sized ships with diverse orientations. Furthermore, ship re-identification results using multi-temporal images demonstrate that ships can be successfully re-identified across time using optical satellite imagery. Nevertheless, some challenges remain in detecting ships close to ports and re-identifying small-sized ships under the spatial resolution constraints of PlanetScope Imagery.

Back to session list

D3-4: Deep Learning-Based Vessel Velocity Estimation from Subaperture SAR Images

Presenting author: Hyunsun LEE

Authors: Hyunsun LEE, Seoul Nat’l Univ.; Junwoo KIM, Seoul Nat’l Univ.; Duk-jin KIM, Seoul Nat’l Univ.

Keywords: Velocity estimation, SAR, Deep learning, Automatic Identification System

Estimating vessel velocity is important for maritime surveillance and moving vessel detection and can support moving target refocusing in synthetic aperture radar (SAR) imagery. In SAR imagery, vessel velocity results in a Doppler shift. Conventional SAR-based methods estimate vessel velocity by measuring Doppler shifts or analyzing ship wakes. However, wake-based methods require clearly visible wake patterns, and Doppler-based approaches are sensitive to observation geometry, limiting reliable velocity estimation under rough sea conditions. Although deep learning methods have recently been applied to vessel velocity estimation, most approaches rely on full-aperture SAR imagery, where motion information is not explicitly separated. In this study, we propose a deep learning-based method that estimates vessel velocity and direction from a single SAR image by exploiting subaperture SAR imagery. From 0.25 m resolution SAR SLC data, the azimuth frequency spectrum was divided into Doppler frequency subbands to generate subaperture images. Moving vessels show positional and structural variations among the subaperture images. To capture these variations, we used a ResNet-based regression model that learned relationships between subaperture images. The network regressed the vessel velocity vector by modeling subtle variations in vessel position and shape across subaperture SAR images. For training, ground-truth labels were derived from AIS (Automatic Identification System) data, and the AIS-provided course over ground (COG) and speed over ground (SOG) were converted into velocity components. The results indicate that subaperture SAR imagery combined with deep learning enables direct estimation of vessel velocity from a single SAR image, suggesting potential applicability to real-time maritime surveillance and vessel tracking systems.

Back to session list

D3-5: Microsatellite SAR-Based Analysis of Ocean Wave Parameters Using Adaptive Azimuth Cutoff and Phase-Refocusing Methods

Presenting author: Minkyeong KIM, Seoul Nat’l Univ.

Authors: Minkyeong KIM, Seoul Nat’l Univ.; Duk-jin KIM, Seoul Nat’l Univ.

Keywords: Significant Wave Height, Wave direction, Azimuth cutoff, Phase-Refocusing

Significant wave height (SWH) and wave direction are key parameters that represent the dynamic characteristics of ocean waves. These parameters provide practical value for various applications, including ship-route optimization and disaster response. The growing deployment of microsatellite-class high-resolution SAR constellations has enabled near–real-time observations, increasing the importance of research that leverages such capabilities. However, SAR-based wave retrieval still faces major challenges, including the sensitivity of azimuth cutoff estimation and inherent directional ambiguity (180° ambiguity), and we propose methods to mitigate these issues. In this study, we estimate high-precision SWH and wave direction using single-look complex (SLC) imagery acquired by the X-band Umbra SAR satellites with a spatial resolution of approximately 0.25–0.5 m. The analysis uses imagery over the seas around Korea and Japan, collocated with ocean meteorological buoys. SWH estimation strongly depends on the azimuth cutoff, which is highly sensitive to changes in the region of interest (ROI) size and pixel spacing. To obtain a more robust azimuth cutoff estimate for X-band SAR imagery, we empirically identified an optimal median-filter window size ranging from 10–50 m. By adaptively applying this window size according to the pixel spacing, we achieved a more stable and accurate SWH estimation. Wave Direction estimation from a single SAR image suffers from an inherent 180° ambiguity. To resolve this, we apply a Phase-Refocusing method. Specifically, we first derive wave direction in the radar coordinate system by accounting for the flight direction and viewing geometry, then convert it to a geographic reference using the satellite heading angle, and finally determine the true wave direction by applying a fore–aft discrimination based on the Phase-Refocusing results. The retrieved wave parameters are quantitatively validated through comparisons with in situ buoy observations and reanalysis datasets. The favorable results demonstrate the feasibility of extracting precise ocean-state information from microsatellite-class, high-resolution SAR observations.

Back to session list

D3-6: Enhancing Coastline Detection from Sentinel-2 With Support from Sentinel-1 Data in Tongatapu, Tonga

Presenting author: Sree Juwel Kumar CHOWDHURY

Authors: Sree Juwel Kumar CHOWDHURY, Korea Institute of Ocean Science and Technology; Chan-Su YANG, Korea Institute of Ocean Science & Technology

Keywords: Coastline; Sentinel-2; Segmentation; Adaptive thresholding

Coastline information is essential for environmental monitoring, coastal planning, and hazard assessment in Pacific Island nations such as Tonga. However, automated coastline extraction from optical satellite imagery remains challenging due to heterogeneous coastal conditions, including sandy beaches, mangroves, shallow waters, artificial structures, and rocky cliffs, which often cause spectral confusion between land and water. Cloud cover further reduces the reliability of single-date optical observations. This study presents a coastline extraction framework based primarily on Sentinel-2 Level-2 reflectance (L2R) data derived from Level-1C imagery using atmospheric correction. The methodology integrates object-based image segmentation, multi-index spectral analysis, adaptive thresholding, and selective synthetic aperture radar (SAR) support. First, Sentinel-2 imagery is segmented into superpixels using a simple linear iterative clustering algorithm to reduce pixel-level noise and enhance spatial coherence in heterogeneous coastal zones. Superpixels are classified into land and water using spectral statistics derived from the normalized difference vegetation index (NDVI) and the modified normalized difference water index (MNDWI), supported by green and shortwave infrared information to improve discrimination in shallow, bright water areas. To mitigate cloud-related errors, a cloud mask is generated using a combination of NDVI and reflectance thresholds from Sentinel-2 Bands 2, 3, 4, and 11. In areas where coastline delineation is hindered by cloud contamination, temporally close Sentinel-1 GRD data are incorporated. Sentinel-1 preprocessing is performed in SNAP, including radiometric calibration, speckle filtering, and terrain correction. SAR-derived land–water separation is applied only within cloud-affected regions to recover missing coastline segments, ensuring continuity without replacing the optical-based extraction elsewhere. The integration of object-based classification, multi-index analysis, adaptive thresholding, and selective optical–SAR complementarity improves coastline stability and reduces cloud-induced gaps. The results demonstrate that the proposed framework enhances the reliability of automated coastline extraction across diverse coastal environments in Tonga.

Back to session list

Session E3 Abstracts

E3-1: Deep Learning–Based Urban Shadow Detection in True Orthoimages Integrating Multispectral and Height Information

Presenting author: Chia-Ying LIN

Authors: Chia-Ying LIN,National Chengchi Univ; Shih-Hong CHIO,National Chengchi Univ

Keywords: True orthoimages; Urban shadow detection; Deep learning; Res-UNet; Multispectral information; Digital Height Model

Compared with conventional orthoimages that are geometrically corrected only for terrain relief, true orthoimages incorporate object height information to eliminate building displacement and leaning effects, making them particularly suitable for high-density urban environments. Nevertheless, extensive shadows cast by tall buildings reduce spectral reflectance and increase land-cover ambiguity, thereby complicating image interpretation and analysis. To improve shadow detection accuracy in urban true orthoimages, this study proposes an automated deep learning framework integrating multispectral and height information. The Xinyi District of Taipei City, characterized by dense high-rise buildings, was selected as the experimental area. A Res-UNet architecture was adopted, in which the encoder was constructed based on a residual network to alleviate gradient degradation during deep training, while the decoder employed a U-Net structure with skip connections to preserve spatial details and shadow boundaries. In addition, near-infrared (NIR) bands and a Digital Height Model (DHM) were incorporated into the input features, enabling the model to simultaneously learn spectral attenuation patterns and height-induced occlusion relationships. Experimental results indicate that the proposed approach achieves an Intersection over Union (IoU) of 0.9447 and an F1-score of 0.9715 on validation data, with an overall accuracy of 0.9831. These findings suggest that integrating multispectral and geometric information within a deep learning framework effectively enhances shadow detection performance in high-density urban environments.

Back to session list

E3-2: Learning Representative Local Descriptors from Multiple Observations via Contrastive Enhancement

Presenting author: Ya-Li LIN

Authors: Ya-Li LIN, National Cheng Kung Univ.; Yi-Chen LEE, National Cheng Kung Univ.; Jiann-Yeou RAU, National Cheng Kung Univ.

Keywords: Local feature descriptors, Descriptor representativeness, Contrastive learning, Feature matching, Vision-based systems

Local feature descriptors play a central role in a wide range of vision-based systems, including visual localization, mapping, and structure-from-motion pipelines. While these descriptors are typically treated as stable representations, the same physical point or landmark is often observed repeatedly under diverse viewpoints, illumination conditions, and imaging configurations. As a result, multiple descriptors associated with a single physical point can exhibit substantial intra-instance variability. This variability becomes particularly problematic when a single representative descriptor is required for efficient matching or map construction, as naive aggregation strategies frequently fail to maintain sufficient discriminability due to high inter-instance similarity. To address this issue, this paper introduces a lightweight descriptor enhancement framework based on contrastive learning, designed to improve descriptor representativeness under multi-observation conditions. Exploiting naturally available associations between descriptors and their corresponding physical points, a small multilayer perceptron (MLP) is trained to project individual descriptors into a more discriminative embedding space. In this space, descriptors originating from the same physical point are encouraged to become more compact, while descriptors from different points are effectively separated. Importantly, the proposed training strategy requires no additional manual annotations and can be readily applied on top of existing feature extractors. Once enhanced, simple aggregation strategies such as mean or median aggregation regain their effectiveness for constructing representative descriptors from multiple observations. The proposed approach is evaluated on a UAV-based bridge inspection dataset. Experimental results demonstrate improved intra-point compactness and inter-point separability, which in turn lead to more reliable feature matching in downstream tasks. Overall, the results suggest that learning-based descriptor enhancement provides a practical and broadly applicable solution to the challenge of multi-observation descriptor representation in vision-based systems.

Back to session list

E3-3: An Intelligent Indoor Navigation System Using Visual Place Recognition and Epipolar Geometry for GPS-Denied Environments

Presenting author: Ruei-Zeng WANG

Authors: Ruei-Zeng WANG, National Chengchi University; Fang-Shii,NING, National Chengchi University

Keywords: Indoor Navigation；Visual Place Recognition；Epipolar Geometry；AKAZE Feature Matching；A* Path Planning

The purpose of this study was to develop an intelligent indoor navigation system capable of achieving decimeter-level localization accuracy in GPS-denied environments without relying on wireless infrastructure such as Wi-Fi or Bluetooth beacons. The research was conducted on the sixth floor of the General Building of Colleges at National Chengchi University, covering 25 pre-surveyed reference points across three corridor zones. A multi-stage visual localization pipeline was implemented, integrating Visual Place Recognition (VPR), epipolar geometry triangulation, and AI-assisted semantic recognition. Coarse localization was performed via AKAZE feature matching against a panoramic image database, followed by eight-direction refinement using lazy loading. The highest-priority method applied epipolar geometry with absolute scale recovery, supported by three fallback methods—weighted triangulation, scale-ratio estimation, and Perspective-n-Point (PnP) solving. Path planning employed the A* algorithm with cubic spline smoothing, while a Pure Pursuit controller managed real-time path tracking. Room numbers were recognized using the Gemini 2.5 Flash vision-language model. The results revealed that the system achieved a localization accuracy of approximately ±1.5 meters under the primary epipolar method. The multi-method fusion architecture with sanity validation demonstrated reliable performance across all corridor zones, requiring no fewer than twenty feature matches and deviating no more than five hundred pixels from the nearest VPR reference point. The principal conclusion was that a camera-only indoor navigation system can achieve practical localization without any wireless infrastructure by combining hierarchical visual feature matching with epipolar geometry. Future studies are recommended to extend the database to additional floors and integrate inertial measurement units (IMUs) to improve localization continuity under poor visual conditions.

Back to session list

E3-4: Fixed Point–Based Image Registration Between Optical and SAR Satellite Images

Presenting author: Hajime TAKAHASHI, Akita Univ.

Authors: Hajime TAKAHASHI, Akita Univ.; Hisatoshi TORIYA, Akita Univ.; Tsuyoshi ADACHI, Akita Univ.; Narihiro OWADA, Akita Univ.

Keywords: Image registration, Multimodal remote sensing, Optical images, Synthetic Aperture Radar (SAR) images

Recent advances in remote sensing technologies have enabled the acquisition of diverse and high-resolution earth surface information. Consequently, increasing attention has been directed toward methods for integrating multi-sensor data. Deep learning–based image registration between optical satellite images and Synthetic Aperture Radar (SAR) images has emerged as a promising approach for such integration. However, conventional methods rely on unsupervised learning, which extract unstable feature points due to seasonal vegetation changes. To address this limitation, this study aims to extract fixed points based on supervised learning and to establish reliable image registration using deep learning. In this study, fixed points were selected from buildings, roads, and rocks, which are less affected by seasonal vegetation changes. Fixed point-based image registration facilitates optical and SAR image integrations across a wide variety of conditions. A two-step framework consisting of fixed point extracting and matching was adopted on THE SEN 1-2 dataset to evaluate performance of the proposed method. For fixed point extracting, a Convolutional Neural Network (CNN) backbone which has strong feature representation capability was employed. For fixed point matching, a Cross-Fusion Matching Network (CFMN) based on a Vision Transformer (ViT) architecture was utilized, which is effective in evaluating inter-image similarity. Experimental results demonstrated the effectiveness of the proposed method in both fixed point extracting and matching. Fixed point extracting network achieved Localization Error (LE) was 1.79 pixels. Fixed-point matching network achieved a Mean Matching Accuracy (MMA) was 0.04 within a threshold of 3 pixels. Furthermore, proposed extracting network exhibited higher extracting accuracy (LE) than conventional methods. In contrast, proposed matching network exhibited lower correctly matched fixed points within the threshold (MMA) than conventional methods due to noise inherent to SAR images. The results presented in this study provide an important foundation for multi-sensor data integration.

Back to session list

E3-5: STRUCTURED 3D CITY MODELING FROM MULTI-VIEW SATELLITE IMAGERY VIA SPARSE VOXELS RASTERIZATION

Presenting author: Hyeonseong CHOI

Authors: Hyeonseong CHOI, Seoul National University; Wonbin KANG, Seoul National University; Gyeongwon OH, Seoul National University; Yongil KIM, Seoul National University

Keywords: Remote Sensing, SVRaster, 3D City Modeling, 3D Reconstruction, Rational Polynomial Camera

High-fidelity 3D city models are increasingly required for accurate representation and analysis of complex urban environments. While recent breakthroughs in advanced 3D reconstruction paradigms, such as Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS), have achieved remarkable success in visual reconstruction, their integration into the satellite remote sensing workflow remains technically constrained. Since the learned scene representations are often implicit or unstructured, making them difficult to integrate with Geographic Information Systems (GIS) that require deterministic, coordinate-indexed, and queryable geometry. Moreover, in many satellite-domain pipelines, the Rational Polynomial Camera (RPC) model is primarily utilized for auxiliary orientation or coarse calibration, rather than serving as a rigorous geometric constraint. These limitations preclude the direct extraction of absolute geometric coordinates, which is a prerequisite for high-precision geospatial applications. To bridge the gap between visual reconstruction and geometric accuracy, this paper proposes a unified optimization framework that rigorously integrates the RPC model into SVRaster— a structured representation that supports efficient spatial partitioning and coordinate-aware data processing capabilities. By embedding RPC consistency directly into the optimization objective, we establish a robust mapping between the reconstructed 3D representation and a global coordinate system, ensuring that the resulting models are both radiometrically consistent and geometrically accurate. Experimental evaluations on urban scenes demonstrate stable georeferencing accuracy and cross-view spatial consistency, indicating that the proposed RPC-integrated SVRaster substantially enhances the GIS-readiness and geospatial interpretability of satellite-derived 3D city models. Consequently, our method provides a reliable foundation for georeferenced applications such as urban monitoring and large-scale environmental analysis.

Back to session list

E3-6: Coarse-to-Fine Iterative Template Matching Using Foundation Model and CNN for Image Registration

Presenting author: Gahyun Lee

Authors: Gahyun Lee; Chanyeop Jung; Youkyung Han

Keywords: template matching; deep learning; foundation model; Iteration; feature concatenation

Large-scale self-supervised foundation models (FMs) pretrained on natural images have recently been widely adopted across various vision tasks. Among them, DINOv2 provides stable object- and scene-level semantic representations, making it suitable for multi-sensor image matching scenarios involving temporal, illumination, and radiometric variations. However, the patch tokens of DINOv2 do not represent individual pixels but instead encode contextual information from neighboring regions, which may lead to spatially unstable correspondences despite semantic similarity. To address this limitation, we propose a coarse-to-fine template matching framework that first performs coarse matching using semantic features from the FM and then refines correspondences by concatenating FM and CNN features for fine matching. In the coarse stage, a frozen DINOv2 backbone extracts patch-level features from reference and target images, and a global cosine similarity heatmap identifies candidate correspondence regions. Rather than selecting only the highest response, the top-k candidates are retained to alleviate ambiguity caused by repetitive structures. In the fine stage, local refinement is performed around each candidate. High-level semantic features from DINOv2 are fused with low-level structural features from a ResNet backbone to form normalized representations. A dense all-pairs correlation volume, capturing similarity scores between all spatial locations of the reference and target feature maps, is computed under a 3×3 multi-anchor scheme to improve robustness in texture-deficient areas. Based on the current displacement estimate, local correlation neighborhoods are iteratively sampled and processed by a Gated Recurrent Unit (GRU)-based refinement module, which predicts residual offsets that are cumulatively applied to update correspondence coordinates. This iterative refinement facilitates stable convergence and achieves subpixel-level localization accuracy. Experimental results demonstrate that the proposed framework consistently outperforms conventional and state-of-the-art approaches in satellite–aerial image registration.

Back to session list

Session A4 Abstracts

A4-1: Cross-Sensor Super-Resolution of Himawari AHI Using SNPP VIIRS High-Resolution Imagery

Presenting author: Shuma FUKUI

Authors: Shuma Fukui, Mie Univ; Masayuki MATSUOKA, Mie Univ.

Keywords: Super-Resolution; Himawari/AHI; SNPP/VIIRS; Near-Infrared band; Convolutional neural network

The Advanced Himawari Imager (AHI) aboard the Himawari-8 and -9 geostationary meteorological satellites provides high-frequency, wide-area observation data. However, its limited spatial resolution necessitates super-resolution processing for detailed environmental analyses. While Convolutional Neural Networks (CNNs) have been employed for this purpose, a persistent challenge remains: when super-resolving 1.0-km resolution images, accuracy in the near-infrared (NIR) band is consistently lower than in the visible band. This degradation is primarily attributed to the conventional reliance on high-resolution visible band images as the sole source of training data. To address this limitation, this study proposes a method to improve super-resolution accuracy by incorporating high-resolution NIR imagery from the Visible Infrared Imaging Radiometer Suite (VIIRS) aboard the polar-orbiting Suomi NPP satellite as training data. Specifically, to enhance AHI’s NIR band (Band 4, 1.0-km resolution), we utilize VIIRS Band I2, which shares a nearly identical central wavelength. The native 375-m resolution VIIRS data are resampled onto a 500-m grid to serve as ground truth for training. To ensure model reliability and minimize parallax errors, we construct a precisely synchronized dataset over Japan by strictly pairing AHI and VIIRS observations with minimal time differences. By implementing this approach, we aim to achieve NIR band super-resolution accuracy comparable to that of the visible band. Furthermore, this method is expected to be highly applicable for improving the super-resolution accuracy of longer-wavelength bands (Band 5 and beyond), which have a coarser native spatial resolution of 2.0 km.

Back to session list

A4-2: L-band SAR Coherence and Intensity Difference Analysis for Burned Area Extraction in the Large-scale Wildfire in Ofunato City

Presenting author: Kou KITAGAWA

Authors: Kou KITAGAWA, Nihon Univ.; Takashi NONAKA, Nihon Univ.; Tomohito ASAKA, Nihon Univ.

Keywords: ALOS-2

In recent years, both the frequency and extent of wildfires in Japan have been increasing. Large wildfires make slopes unstable and raise the risk of landslides. These events also change forest environments and can affect animal behavior. Some reports describe changes in the areas they use of large mammals such as bears. These examples show the wide impacts of wildfires. Ground based observations cannot cover broad areas for long periods, so satellite remote sensing is essential. Synthetic Aperture Radar can observe burned areas even when clouds or smoke are present. The extraction of burned areas for cases with carbonized tree trunks that remain standing has not been fully examined. The purpose of this study is to extract burned areas based on coherence and intensity‑difference analyses using L‑band SAR data from the PALSAR‑2 sensor onboard ALOS‑2, focusing on a large‑scale forest wildfire that occurred in Ōfunato City, Iwate Prefecture, in February 2025.This wildfire showed extensive burning at the base of tree trunks, and many trunks remained standing. Field observations made on April 7, 2025, were conducted to examine the consistency between the changes detected by SAR and the actual damaged area. The results show low coherence in both burned and unburned areas for image pairs taken pre- and post-wildfire. Seasonal changes in unburned areas caused tree growth and changes in canopy structure, and these changes reduced coherence. Coherence derived from during- and post-fire image pairs showed stability in areas with carbonization reaching the upper trunks, as tree growth was stopped by carbonization. Coherence remained relatively stable even with carbonization limited to the base. Coherence in unburned areas decreased in the same way as in the pre- and post-fire pairs because seasonal growth continued. Based on these findings, this study attempts to extract burned areas with higher reliability.

Back to session list

A4-3: Early Wildfire Detection in South Korea Using Himawari-9 Short-Term Time-Series

Presenting author: Mihyeon JEONG

Authors: Mihyeon JEONG, University of Science and Technology; Yeji KIM, Korea Aerospace Research Institute; Daewon CHUNG, Korea Aerospace Research Institute

Keywords: Himawari-9, Early wildfire detection, Spectral difference, Time-series analysis, Topographic heterogeneity

Geostationary satellites, such as Himawari-9, provide 10-minute temporal resolution essential for early wildfire detection. However, single-band time-series models have limitations in topographically heterogeneous areas like coastlines, where sub-pixel registration shifts often cause abrupt fluctuations in background temperatures, leading to false alarms. To address these limitations, this study proposes a pixel-isolated early detection algorithm that stabilizes background temperatures by utilizing multi-band spectral difference (TSWIR - TTIR) information. This approach effectively cancels out background temperature distortions occurring across both bands due to topographic heterogeneity, thereby securing a stable background baseline even in coastal regions. To prevent background temperature estimation errors in heterogeneous terrain, the algorithm avoids spatial averaging and tracks the independent temporal history of each pixel. The background temperature is estimated using a short-term mean that excludes fire-affected observations, providing a reliable baseline for anomaly detection. In the detection process, physical filters utilizing thresholds and spectral difference information are applied based on the GK2A wildfire detection algorithm (ATBD), as GK2A shares similar sensor characteristics with Himawari-9. This study was conducted using the Gyeongbuk and Sancheong wildfires of March 2025 and validated through comparisons with high-resolution VIIRS active fire products and WLF products. Experimental results demonstrated that the spectral difference filter reduces the standard deviation of data caused by topographic heterogeneity. Furthermore, an optimal temporal window size was established through experiments to achieve a low false alarm rate while maximizing early detection capability. These results confirm that the proposed methodology resolves background temperature correction issues and enables reliable early fire detection in topographically complex environments.

Back to session list

A4-4: Ensemble Stacking Learning for Forest Carbon Stock Estimation from Multi-Source Remote Sensing Data

Presenting author: Yi-Chun CHENG

Authors: Yi-Chun CHENG, Department of Geomatics, National Cheng Kung University, Tainan 70101, Taiwan; Chien-Hao SUNG, Department of Geomatics, National Cheng Kung University, Tainan 70101, Taiwan; Jun-Jun SU, Department of Geomatics, National Cheng Kung University, Tainan 70101, Taiwan; Yu-Yao CHANG, Department of Geomatics, National Cheng Kung University, Tainan 70101, Taiwan; Wan-Yu LIU, Department of Forestry, National Chung Hsing University, Taichung City, 402, Taiwan; Chih-Da WU, Department of Geomatics, National Cheng Kung University, Tainan 70101, Taiwan

Keywords: ALS; SAR; Multispectral Imagery; Geo-AI; Ensemble Learning

Precise estimation of forest carbon pools is critical to tracking ecosystem carbon fluxes and guiding international climate action. In recent years, numerous studies have estimated forest aboveground biomass or carbon-related indicators using multi-source remote sensing data and a variety of machine learning algorithms. However, different algorithms exhibit distinct strengths and limitations. Ensemble learning can integrate complementary advantages across multiple models, potentially reducing bias and variance and improving the accuracy and stability of estimates. Therefore, this study focuses on evaluating the feasibility of an ensemble stacking model to enhance the accuracy and stability of forest carbon stock estimation. This study utilized carbon measurements from 109 field survey plots as the response variable and combined multi-source remote sensing data, including Landsat 8 optical imagery, Sentinel-1 C-band SAR, and airborne laser scanning (ALS). Predictor variables were derived from spectral reflectance, vegetation indices, SAR backscatter, and canopy structural metrics related to the canopy height model (CHM). Five machine learning regressors were implemented: CatBoost Regressor (CBR), Gradient Boosting Regressor (GBR), Light Gradient Boosting Machine Regressor (LGBMR), Random Forest Regressor (RFR), and XGBoost Regressor (XGBR). To construct the ensemble framework, CatBoost, Random Forest, and XGBoost were selected as base learners based on their robust preliminary performance, with individual training R² values ranging from 0.611 to 0.955 and testing R² values ranging from 0.555 to 0.576. By integrating these complementary algorithms, the final ensemble model achieved significant performance improvements, yielding R² values of 0.951 and 0.746 on the training and testing sets, respectively. These results demonstrate that the ensemble stacking framework effectively mitigates the limitations of individual algorithms, providing a more robust and accurate approach for forest carbon stock estimation.

Back to session list

A4-5: An Integrated Tool-Oriented Workflow for Automated Building Detection from Aerial Imagery Using TransUNet and RGB-DSM Fusion

Presenting author: Pei-Cheng CHEN

Authors: Pei-Cheng CHEN, National Yang Ming Chiao Tung Univ.; Tee-Ann TEO, National Yang Ming Chiao Tung Univ.

Keywords: Building change detection, TransUNet, RGB-DSM Fusion, Aerial imagery, Cross-region generalization

As smart city initiatives in Taiwan demand increasingly precise and frequent urban updates, high-resolution aerial imagery has become a primary data source. However, traditional automated methods often struggle with the diversity of Taiwanese rooftop styles, ranging from dense concrete structures to various roofing materials. To bridge this gap, we developed a comprehensive, tool-oriented pipeline that leverages the TransUNet architecture, effectively combining local feature extraction from convolutional layers with global contextual modeling via transformer-based self-attention. A key focus of this research is the investigation of RGB and Digital Surface Model (DSM) fusion, demonstrating that the inclusion of elevation data is vital for distinguishing structural changes from spectral noise. The proposed tool manages an end-to-end process, from bi-temporal semantic segmentation to the production of area-regularized vector maps, ensuring practical utility for GIS professionals and land management authorities. Methodologically, the model was trained on a massive dataset of 1,048,575 image patches from Hsinchu, Taiwan, and was subjected to stringent generalization tests, including a cross-region evaluation in Tainan, Taiwan. The experimental results highlight the robust performance of the RGB–DSM fusion approach, which achieved an F1-score of 90.82% in temporal-independent tests and a remarkably high 81.45% in spatial-independent tests—a stark contrast to the 55.03% recorded by the RGB-only model. These findings showcase the tool’s superior ability to generalize across geographically distinct regions, providing a stable, reliable, and automated solution for real-world building change monitoring and digital map maintenance in Taiwan’s diverse geographic contexts.

Back to session list

Session B4 Abstracts

B4-1: Feasibility study of estimating ice thickness distribution for landfast sea ice in Lützow-Holm Bay derived from satellite microwave altimetry data

Presenting author: Yurina SATAKE

Authors: Yurina SATAKE, Nihon Univ.; Kazutaka TATEYAMA, KITAMI Institute of Technology; Tsutomu YAMANOKUCHI, Remote Sensing Technology Center of Japan; Shigeru AOKI, Hokkaido Univ.; Kazuki NAKAMURA, Nihon Univ.

Keywords: Landfast sea ice; thickness; Lützow-Holm Bay; Altimeter; CryoSat-2; EM

Lützow-Holm Bay in East Antarctica is mostly covered with the landfast sea ice, which surrounds the terminus of Shirase Glacier. The landfast sea ice frequently broke up and flowed away from the bay, consequently, the terminus of Shirase Glacier broke away. In addition, the Shirase Glacier flow was accelerated as a large-scale breakup of the landfast sea ice. This shows that the landfast sea ice plays a role in a buttress that suppresses the direct drainage from Shirase Glacier into the bay. Since observations of ice thickness in the bay are still insufficient, it is necessary to investigate the thickness of landfast sea ice to understand the interaction between the movement and the thickness of glacier. Satellite remote sensing is available for extensive and periodic observations, therefore, this study used the microwave altimetry data from CryoSat-2 satellite to estimate the ice thickness distribution in Lützow-Holm Bay under the assumption of hydrostatic equilibrium. The ice thickness distribution was generated using a 5 km grid by stacking CryoSat-2 data from December 2022 to February 2023 which was validated by the measurement results used an Electromagnetic induction (EM) instrument onboard the icebreaker Shirase by the 64th Japanese Antarctic Research Expedition. The measured ice thickness from the EM (mean±SD) was 2.00±0.38 m along the cruise track of the icebreaker Shirase, and the estimated ice thickness from the CryoSat-2 was 1.34±0.44 m. Hence, the estimated ice thickness was underestimated by 0.66 m compared with the measured ice thickness. Since the thickness derived from the EM was measured the sum of snow depth and ice thickness, that underestimated result could be comparable. We therefore discuss the result of comparison between measured and estimated ice thicknesses and the characteristic of landfast sea ice in Lützow-Holm Bay.

Back to session list

B4-2: V-Mamba with Cyclic Temporal Encoding for Sea Ice Stage-of-Development Segmentation in SAR Imagery

Presenting author: Dabeen Choi

Authors: Dabeen Choi, Seoul National Univ.; Duk-jin Kim, Seoul National Univ.

Keywords: Synthetic Aperture Radar; Sentinel-1; Sea Ice; Deep Learning; Segmentation; Arctic Ocean

Under ongoing Arctic warming, the reduction in Arctic sea ice has increased shipping activity along the Northern Sea Route (NSR), making accurate sea ice monitoring essential for navigational safety. Conventional passive microwave satellite observations are constrained by coarse spatial resolution, and recent deep learning approaches for SAR-based sea ice mapping also present limitations: CNN-based models have restricted receptive fields, while Transformer-based models typically incur high computational costs. To address these issues, we propose a sea ice Stage-of-Development (SoD) segmentation method based on V-Mamba, a State Space Model (SSM) architecture. V-Mamba models long-range dependencies with linear computational complexity and, through its 2D scanning mechanism, captures both fine-scale textures and regional ice distributions in SAR imagery. We defined four SoD classes and trained the model on a newly compiled high-resolution dataset. The input features were expanded by adding the polarization difference (HH - HV) to the Sentinel-1 HH and HV channels. To account for the seasonal evolution of sea ice physical properties, we introduced a cyclic temporal encoding scheme. By converting date information into sine and cosine channels, these were integrated with the polarimetric features as a five-channel input. This design enables the model to distinguish temporal contexts even when backscatter values are similar. Additionally, Dense Conditional Random Field (DenseCRF) post-processing was applied to reduce speckle noise and improve spatial coherence along ice boundaries. The experimental results demonstrate that V-Mamba with cyclic encoding improves accuracy and yields more stable performance than configurations using only polarimetric features. By explicitly incorporating seasonal variability, the proposed approach enables a single model to operate across seasons without separate seasonal models. Furthermore, DenseCRF refines boundary delineation and suppresses residual noise. This research provides a practical technical foundation for near-real-time NSR route planning and contributes to long-term analyses of sea ice variability and navigational risk assessment.

Back to session list

B4-3: Catchment-scale monitoring of subglacial lake activity in the David Glacier region, East Antarctica, using SBAS-DInSAR

Presenting author: Taewook KIM

Authors: Taewook KIM, Kangwon National University; Hyangsun HAN, Kangwon National University

Keywords: subglacial lake; synthetic aperture radar; differential interferometric synthetic aperture radar (DInSAR); Small BAseline Subset (SBAS); David Glacier

Differential Interferometric Synthetic Aperture Radar (DInSAR) has been used to detect ice sheet elevation anomalies due to subglacial lake activity by differencing two interferograms under an assumption of steady ice flow. However, conventional DInSAR is typically applied to individual subglacial lakes because it requires a reference interferogram that minimize elevation changes, making catchment-scale analysis of subglacial hydrology and ice dynamics. In this study, we develop a framework integrating the Small Baseline Subset (SBAS) with DInSAR to detect ice sheet elevation anomalies across a catchment. For a region of the David Glacier catchment in East Antarctica, we generated Sentinel-1 interferograms with 12-, 24-, and 36-day temporal baselines from 2017 to 2022. We constructed a reference interferogram by taking the median of all interferograms to represent glacier flow signal. We then applied DInSAR technique by differencing the reference from each interferogram and applied the SBAS technique to the resulting differential interferograms. The proposed SBAS-DInSAR approach revealed localized displacement anomalies over subglacial lakes, as well as broader anomalies along the main stream of the glacier.

Back to session list

B4-4: Reanalysis-Driven Sub-Seasonal Solar Radiation Forecasting with a 28-Day Lead Time

Presenting author: Yejin Lee

Authors: Yejin Lee, Jeonbuk National University; Dongjin Kim, Jeonbuk National University; Jieun Wie, Jeonbuk National University; Corresponding author - Jong-Min Yeom, Jeonbuk National University

Keywords: Surface Solar Radiation; ERA5; ECMWF S2S; Sub-Seasonal Forecasting; Spatiotemporal Modeling; Deep Learning; Renewable Energy

Sub-seasonal (2–4 weeks) solar radiation forecasting is crucial for solar power operations and energy management, yet remains highly uncertain due to the gap between short-term numerical forecasts and seasonal predictions. This study develops a deep learning-based mid-term prediction model that generates 28-day-ahead averages of surface shortwave radiation (SSRD) using atmospheric and oceanic variables from the 7 days prior to the target date, based on ERA5 reanalysis data. Input variables include SSRD, total cloud cover, precipitation, temperature, sea surface temperature, and mid-layer atmospheric fields. Predictions are performed at a global grid scale with a 28-day lead time. The model architecture is based on the U-Net series with extended structures designed to capture spatiotemporal information. Trained on approximately 10 years of reanalysis data, performance was evaluated across different lead times (1–28 days) and weekly intervals, showing stable short-term predictive skill with a gradual decline at longer lead times. A hybrid structure combining the model with ECMWF S2S forecast fields was also designed to supplement numerical forecast information. Variable importance analysis is being conducted to quantitatively assess each factor’s contribution, and model performance is evaluated at both grid and spatial-average levels against ECMWF raw forecasts. Future work includes incorporating ground observation (ASOS) and satellite-based radiation and cloud cover data for regional-scale validation and high-resolution extension experiments. This study demonstrates the potential of data-driven mid-term radiation forecasting from reanalysis data and evaluates the applicability of an integrated framework combining numerical forecasts and deep learning, offering significant implications for sub-seasonal solar energy prediction.

Back to session list

Session C4 Abstracts

C4-1: Landslide Prediction from Remote Sensing Imagery Using Semantic Segmentation: A Case Study in Northern Taiwan

Presenting author: TZU-HSIANG HUANG, National Chengchi University

Authors: TZU-HSIANG HUANG, National Chengchi University

Keywords: Deep Learning, Hazards, Landslide, Prediction, Semantic Segmentation

Landslides are a natural hazard worldwide, resulting in numerous fatalities and economic losses each year. Asia was the most severely affected area, accounting for 40% of the fatalities. For instance, the 2022 Manipur landslide in India and the debris flow in Atami, Shizuoka in 2021, caused 83 deaths in total,were both triggered by a combination of intense rainfall and improper land use. In response to landslide impacts, some international measures have been proposed. There have been three World Conferences on Natural Disaster Reduction (WCNDR), each proposing corresponding strategies based on the circumstances at the time of the conference. Regarding the Sustainable Development Goals (SDGs), SDG 11.5 and 13.1 are the indicators relevant to landslides. These actions demonstrate the growing awareness and attention of it. In Taiwan, earthquakes and typhoons are the main triggering factors, with 1,000 felt earthquakes and four typhoons each year. The complex topography of Taiwan further increases landslide susceptibility. Therefore, a landslide prediction model is required to reduce the losses and to support early preventive measures. This study employs deep learning-based semantic segmentation to train and predict the specified multi-band imagery, which is constructed with DEM, faults, geologically sensitive areas, land use and land cover (LULC), and precipitation. The model was developed using U-Net, a convolutional neural network (CNN) architecture extensively utilized for image segmentation. The prediction results were classified into landslide and non-landslide, and a confusion matrix was employed to evaluate the model performance. Through the landslide prediction model, the spatial distribution of landslides in northern Taiwan can be identified. An 81% hot-spot overlap was calculated between the actual locations and the prediction results. Simultaneously, the LULC model is employed to provide the three bands needed. The model classifies all pixels within the Landsat imagery, and its current classification accuracy has reached up to 88%.

Back to session list

C4-2: A Satellite-Driven Framework for Automated Identification of Suitable Industrial Development Sites via Land Use Classification and Machine Learning

Presenting author: Ryota KUBOTA

Authors: Ryota KUBOTA，Tokyo City Univ．; Kento TERADA，Tokyo City Univ．; Yuki AKIYAMA，Tokyo City Univ．

Keywords: Land Use Classification; Satellite Imagery; Machine Learning; Suitable Development Site; Image Classification

In recent years, efficient and objective site selection for large-scale industrial facilities has become increasingly important due to expanding logistics demand and regional development pressures. One area where data-driven automation remains limited is the selection of suitable development sites, defined as large, contiguous land parcels appropriate for industrial facilities. Currently, site selection relies heavily on manual interpretation of land use information, resulting in substantial time and cost burdens. The Land Use Mesh Data provided by the Japanese government as open data, which divides the entire country into 100 m square grid cells, is widely used in planning practices for development site selection. The dataset classifies each grid cell into land use categories such as urban areas, agricultural land, forest, and water bodies. However, updating and refining this dataset requires extensive manual processing, and updates are typically conducted at multi-year intervals with inconsistent revision timing, thereby limiting temporal responsiveness, scalability, and consistency. This study therefore proposes an integrated framework that automatically updates land use information from satellite imagery and quantitatively evaluates development suitability. We applied prefecture-specific Random Forest models to classify land use across the entire Tohoku region into four categories. Multispectral bands from Sentinel-2 satellite imagery (10–20 m resolution) were used as input features. Datasets were constructed by integrating these spectral features with ground truth labels derived from Land Use Mesh Data. Next, we quantitatively assessed development suitability by combining evaluation indicators derived from government guidelines for industrial site selection. The proposed framework achieved high land use classification performance, demonstrating an overall accuracy of over 98% across all six prefectures in the Tohoku region. The system successfully extracted candidate development sites in a consistent and scalable manner. The findings demonstrate the feasibility of coupling satellite-based land use classification with rule-based suitability assessment to support objective, large-scale industrial site identification.

Back to session list

C4-3: MaxEnt-Based Spatial Risk Assessment for Mitigating Multi-Species Human-Wildlife Conflict in Japan

Presenting author: Didier DELGORGE

Authors: Didier DELGORGE, The Graduate School of Arts and Sciences, the University of Tokyo; Wataru TAKEUCHI, Institute of Industrial Sciences, the University of Tokyo

Keywords: Human-Wildlife Conflict; Machine Learning; MaxEnt Model; Risk Assessment; Conflict Mitigation

Human-wildlife conflict in Japan is a growing concern, with impacts ranging from crop damage and vehicle collisions to human injuries and fatalities. These conflicts are expected to intensify as large mammals expand their ranges driven by socio-ecological factors such as wildlife population recovery, climate change, and land abandonment. These trends highlight the need for smarter, targeted wildlife management strategies. Much of the existing literature on human-wildlife conflict focuses on single species or conflict types, leaving a gap in multi-species risk assessment and integrated, spatial mitigation strategies. This study uses the MaxEnt machine learning model to generate probability distribution maps for Asiatic black bears, wild boars, sika deer, and macaques. These probability outputs were combined with target-specific exposure indicators to quantify risks to agriculture, vehicle collisions, and human safety, forming the foundation for an integrated management recommendation map. Initial MaxEnt results using data from Akita Prefecture reveal species-specific probability patterns, with bears showing higher probabilities in urban areas, and wild boars and deer in forested areas. Effective wildlife management requires strategies that reduce conflict while ensuring the conservation of large mammals. However, as rural communities in Japan shrink and age, human and financial resources to implement such measures is increasingly strained, often leading to actions that are extreme, unsustainable, and offer only temporary relief. These findings provide valuable insights by revealing critical conflict zones and informing spatially explicit decision-making for conflict mitigation and conservation planning.

Back to session list

C4-4: BRIDGE DEFORMATION MONITORING USING THE TIME SERIES INSAR TECHNOLOGY

Presenting author: Derrick Muga Mboya

Authors: Derrick Muga MBOYA, National Central University; Fuan TSAI, National Central University

Keywords: Persistent Scatterer InSAR (PS-InSAR), Bridge deformation monitoring, TerraSAR-X, line of sight, GNSS, structural health.

Monitoring deformation of critical infrastructure in urban environments is essential for ensuring structural integrity and public safety. The Yuan Shan Bridge, located in the Taipei metropolitan area, has experienced significant deformation in recent years, prompting the need for high-precision and continuous monitoring. This study investigates the deformation behavior of the Yuan Shan Bridge using a Persistent Scatterer Interferometric Synthetic Aperture Radar (PS-InSAR) approach applied to TerraSAR-X data acquired between January 2025 and June 2025, with Global Navigation Satellite System (GNSS) observations employed for independent validation. The spatial resolution and short wavelength of X-band TerraSAR-X make it possible to identify persistent scatterers on the bridge deck and supporting structures, which translates to high coherence of millimeter-scale deformation to be determined in the lower-temporal frequency band. The PS-InSAR processing is applied to extract the time-series displacement measurements and average rates of deformation in the direction of the line-of-sight of the satellite. The deformation patterns that result provide a spatial non-uniform behavior throughout the bridge structure, and local segments have strong displacement relative to their neighbors, in accordance with the structural issues reported before. Temporal analysis also shows short-term trends of deformations that may be associated with traffic loading, environmental factors, and underground factors. In order to evaluate the reliability of the PS-InSAR results, deformation estimates are analyzed alongside GNSS displacement tasks of stations in the proximity of the bridge. The analogy shows that both the extent of deformation and the time-dependent characteristics have high consistency, which validates the soundness of the PS-InSAR methodology. The study reveals that combining high-resolution PS-InSAR and GNSS data is effective in monitoring short-term deformations of urban bridge systems and offers a useful understanding of the structural health of the region, early warning systems, and the integration of knowledge into the decision-making of maintenance strategies.

Back to session list

C4-5: Reconstruction of GRACE-Derived Terrestrial Water Storage Anomaly over the United States Using a Bayesian CNN: GLDAS versus NLDAS

Presenting author: Ching-Mei, TSENG

Authors: Ching-Mei TSENG, National Chung Hsing Univ.

Keywords: GRACE/GRACE-FO; Terrestrial Water Storage Anomaly (TWSA); Bayesian Convolutional Neural Network (BCNN); GLDAS/ NLDAS

This study implements a Bayesian Convolutional Neural Network (BCNN) framework to reconstruct and predict the Terrestrial Water Storage Anomaly (TWSA) over the United States by integrating GRACE CSR observations with land-surface model outputs. GRACE-derived TWSA is adopted as the reference, while hydrological and climatic variables from the Global Land Data Assimilation System (GLDAS) and the North American Land Data Assimilation System (NLDAS), both based on the NOAH land surface model, serve as input features for hydrological process representation. The model is trained over April 2002–June 2017 and independently evaluated from August 2018–December 2024 to assess its temporal generalization capability and long-term predictive stability. At the continental scale, the GLDAS-based framework achieves a mean correlation coefficient (R) of 0.88 with an RMSE of 3.64 cm, whereas the NLDAS-based framework demonstrates improved performance (R = 0.89; RMSE = 3.39 cm). Basin-scale analysis across 18 major hydrologic regions further highlights regional variability. In the Upper Mississippi basin, NLDAS (R = 0.92; RMSE = 2.40 cm) outperforms GLDAS (R = 0.91; RMSE = 2.64 cm). Similarly, in the Lower Mississippi basin, NLDAS (R = 0.97; RMSE = 2.71 cm) outperforms GLDAS (R = 0.96; RMSE = 3.06 cm). In the Missouri basin, NLDAS (R = 0.96; RMSE = 1.49 cm) outperforms GLDAS (R = 0.93; RMSE = 2.16 cm). Consistently high correlations are observed in the Ohio and Tennessee basins (R = 0.98), as well as in the Pacific Northwest (R = 0.98). Stable performance in western arid regions, such as the Rio Grande and Lower Colorado basins (R ≈ 0.90 to 0.92), demonstrates robustness across diverse hydroclimatic conditions. Overall, the NLDAS-driven BCNN framework provides higher predictive accuracy across most regions and effectively reproduces GRACE-observed TWS variability, offering a reliable, uncertainty-aware methodology for hydrological reconstruction and regional water resource assessment.

Back to session list

Session D4 Abstracts

D4-1: APPLICATION OF THE BLEACHING SUSCEPTIBILITY INDEX (BSI) AND BOTTOM REFLECTANCE ANALYSIS FOR SATELLITE-BASED CORAL BLEACHING ASSESSMENT IN TONGA

Presenting author: Won-Jun CHOI

Authors: Won-Jun CHOI, Korea Institute of Ocean Science and Technology; Chan-Su YANG, Korea Institute of Ocean Science and Technology

Keywords: Coral Reef Bleaching, Bleaching Susceptibility Index (BSI), Remote Sensing, Long Short-Term Memory (LSTM)

Coral bleaching is a complex ecological response driven not only by short-term thermal anomalies but also by the cumulative effects of heat stress, light conditions, and water quality over time [1]. In recent decades, the increasing frequency and spatial extent of mass coral bleaching events under global climate change have emphasized the need for satellite-based approaches capable of characterizing long-term environmental susceptibility of coral reef systems [2]. Degree Heating Weeks (DHW), provided by NOAA Coral Reef Watch, is widely used to assess bleaching risk; however, its primary focus on short-term thermal stress limits its ability to represent spatial heterogeneity and long-term bleaching susceptibility associated with persistent environmental exposure [3]. During coral bleaching events, the loss of symbiotic zooxanthellae leads to an increase in bottom reflectance in optically shallow reef environments, producing detectable optical signals in multispectral satellite imagery [4]. Recent studies have demonstrated that bottom reflectance can be retrieved from satellite observations using physics-based or semi-analytical shallow-water optical models that minimize water-column effects, enabling large-scale assessment of coral reef optical responses without extensive in situ measurements [5]. Furthermore, changes in bottom reflectance between pre-bleaching and bleaching periods have been proposed as effective indicators of coral bleaching responses [6]. The objective of this study is to apply and evaluate a satellite-based coral bleaching assessment framework in the Pacific Island nation of Tonga by first identifying reef areas characterized by different levels of bleaching susceptibility using a multi-satellite-derived Bleaching Susceptibility Index (BSI), and subsequently analyzing bottom reflectance changes during bleaching events within these susceptibility classes. The BSI integrates multiple environmental variables, including sea surface temperature (SST) and anomalies (SSTA), accumulated thermal stress (DHW), light conditions represented by photosynthetically active radiation (PAR), and water quality indicators such as chlorophyll-a concentration and diffuse attenuation coefficient (Kd), to characterize long-term environmental exposure relevant to coral bleaching susceptibility. Based on the resulting bleaching susceptibility map, reef areas were classified into susceptibility classes, and bottom reflectance changes during bleaching periods were analyzed and compared across these classes. By integrating long-term bleaching susceptibility assessment with observed optical responses during bleaching events, this study complements conventional heat-stress-based approaches and provides a more comprehensive satellite-based framework for coral bleaching assessment. The proposed approach is particularly valuable for data-limited regions such as Tonga, where it can support reef management prioritization, long-term monitoring design, and climate change adaptation strategies based on relative patterns of bleaching susceptibility [7].

Back to session list

D4-2: Leveraging satellite remote sensing for hydrodynamic models’ validation and scalable tidal stream resources quantifications

Presenting author: ALJBER Morhaf

Authors: ALJBER Morhaf, Hiroshima University; LEE Han Soo, Hiroshima University; COLES Danny, University of Oxford

Keywords: Sentinel-1 SAR, Tidal Power, Remote Sensing, Model Validation, Current Velocity

Tidal stream energy (TSE) represents a cornerstone of the global transition to net-zero, offering an absolute reliability in the renewable sector. Unlike the intermittent nature of wind or solar, tidal cycles are predictable over a long time which is crucial for sustainable power production. Despite the TSE potential, commercialisation remains limited due to uncertainty in resource characterisation and flow model validation. The prevailing methodology for assessing tidal sites relies on sparse, point-based in-situ measurements and computationally expensive hydrodynamic models. These models often lack the robust validation required to satisfy investors and policymakers, leading to discrepancies between theoretical energy yield and practical power production. The objective of this study is to investigate the transformative role of satellite remote sensing to bridge the data availability gap. By leveraging high-resolution Earth Observation data such as Sentinel-1 SAR, we can characterise complex tidal flows across vast spatial scales with high precision. This enables the validation of flow models and the accurate translation of theoretical potential into dependable system value. The methodology employs two primary techniques. First, doppler shift retrievals and reflecting intensity is analysed to estimate surface current velocity and tidal constituents. These methods calculate the frequency shift induced by moving ocean water to derive surface currents. Second, backscattering approaches can be used as well to derive tidal elevations by using a reference ground measurement. The results will be validated spatially and temporally using Acoustic Doppler Current Profiler (ADCP) datasets in different geographical locations. The implication of this study contributes a scalable framework for the global standardisation of tidal resource assessment, moving beyond site-specific in-situ constraints to enable a wide-domain validation of hydrodynamic models.

Back to session list

D4-3: Operational Risk Mapping for Surveillance Prioritization of Suspected IUU Fishing Using AIS-Unmatched SAR Vessel Detections in the Korean EEZ

Presenting author: Hyunji LEE

Authors: Hyunji LEE, University of Science and Technology; Sanghyuck HAN, Korea Aerospace Research Institute

Keywords: Sentinel-1; SAR–AIS unmatched detections; IUU fishing; Korean EEZ; Maritime surveillance

Illegal, unreported, and unregulated (IUU) fishing is difficult to monitor when vessel activity is missing from public AIS (Automatic Identification System)-based tracking. This study develops an operational framework that prioritizes suspected high-risk IUU fishing areas (hotspots) in the Korean Exclusive Economic Zone (EEZ) using SAR–AIS unmatched cases, with direct relevance to enforcement and surveillance decision-making. Unlike prior SAR–AIS studies that have mainly advanced vessel detection, association, and anomaly identification, this work focuses on operational prioritization in the Korean EEZ. We estimate an IUU suspicion score by integrating four components: intensity of AIS-unconfirmed vessel activity, likely-fishing vessel among AIS-unconfirmed detections, proximity to EEZ boundaries and marine protected areas (MPAs), and persistence of repeated appearances. We then perform component-wise contribution analysis and sensitivity tests for threshold and weight settings. The workflow standardizes monthly SAR vessel-detection records and computes scores using a common baseline: AIS-unconfirmed detections in each grid cell normalized by the number of SAR overpasses in that cell. Using publicly available Sentinel-1 vessel-detection data for 2025, we produced monthly and annual risk maps and ranked surveillance-priority areas by combining the baseline indicator with the IUU module score. Annual results show relatively higher risk in the West Sea, with additional high-risk zones in parts of the eastern South Sea and southern East Sea. Monthly analysis also indicates within-year variability. For selected surveillance areas, we quantified component-level score contributions to clarify prioritization rationale, and sensitivity analysis showed that the priority ranking remained generally stable, supporting operational reliability. The framework enables grid-level normalization and scoring of SAR–AIS unmatched activity under a unified scheme, allowing comparable risk assessment and actionable surveillance prioritization. It moves beyond descriptive mapping toward decision-support outputs that can be used directly for inspection targeting. Future work will extend the same pipeline to 2023–2025 data and add seasonal and time-of-day statistical analyses.

Back to session list

D4-4: Satellite Estimation of Nutrient Variability Across the Seto Inland Sea Using Machine Learning: Assessing Algal Bloom Susceptibility in Osaka Ba

Presenting author: Gandhi NAPITUPULU

Authors: Gandhi NAPITPULU, Coastal Hazards and Energy System Science Laboratory, Graduate School of Innovation and Practice for Smart Society, Hiroshima University; Han Soo LEE, Transdisciplinary Science and Engineering Program, Graduate School of Advanced Science and Engineering, Hiroshima University; Jae-Soon JEONG, Center for the Planetary Health and Innovation Science (PHIS), The IDEC Institute, Hiroshima University, 1-5-1 Kagamiyama, Higashi-Hiroshima, 739-8529, Hiroshima, Japan; Vinayak Nitin BHANAGE, Center for the Planetary Health and Innovation Science (PHIS), The IDEC Institute, Hiroshima University; Morhaf ALJBER, Coastal Hazards and Energy System Science Laboratory, Graduate School of Innovation and Practice for Smart Society, Hiroshima University; Yuji SAKUNO, Graduate School of Advanced Science and Engineering, Hiroshima University; Nobuhito MORI, Disaster Prevention Research Institute, Kyoto University

Keywords: Algal bloom; DIENOF+; MODIS; Nutrient estimation; Random forest; Sea surface chlorophyll-a

Accurate and spatially continuous nutrient fields are fundamental for resolving coastal biogeochemical dynamics and improving ecosystem modeling in semi-enclosed seas. Yet, satellite-based nutrient estimation remains challenging in optically complex coastal environments influenced by strong terrestrial inputs and anthropogenic activity. This study examines regional differences in satellite-derived nutrient estimation between the inner and outer Seto Inland Sea (SIS) using a Gradient Boosted Regression Trees (GBRT) machine learning framework for the period 2013–2021. MODIS Level-2 observations of sea surface temperature (SST) and chlorophyll-a (SSC) were reconstructed using the Data-Interpolating Empirical Orthogonal Functions Plus (DINEOF+) technique to improve spatial and temporal continuity. These reconstructed variables were used as predictors to estimate key biogeochemical parameters, including dissolved inorganic carbon, dissolved oxygen, nitrate, phosphate, silicate, and pH. Model training and validation were conducted using in situ measurements compiled from the Japan Meteorological Agency (JMA), the World Ocean Database (WOD), the Japan Oceanographic Data Center (JODC), and the Ministry of Land, Infrastructure, Transport and Tourism (MLIT). Performance was evaluated separately for the inner and outer SIS. Results show consistently higher predictive skill in the outer SIS, with greater coefficients of determination and lower error metrics, whereas reduced performance in the inner region is associated with elevated turbidity, strong river discharge, and intensified anthropogenic forcing. The reconstructed nutrient fields were subsequently applied to assess algal bloom susceptibility in Osaka Bay by examining nutrient variability alongside chlorophyll-a dynamics. The analysis reveals that machine learning–derived nutrient products effectively capture spatial gradients and temporal fluctuations linked to bloom development. These findings demonstrate the applicability of GBRT-based satellite nutrient estimation for bloom monitoring and underscore its potential to support coastal ecosystem management and regional biogeochemical modeling in the SIS.

Back to session list

Session E4 Abstracts

E4-1: Deep Learning-Based Template Matching with Domain Adaptation for Cross-Domain Satellite–Aerial Image Matching

Presenting author: Chanyeop JUNG

Authors: Chanyeop JUNG, Seoul National Univ. of Science and Technology; Doochun SEO, National Satellite Operation and Application Center, Korea Aerospace Research Institute; Taeheon KIM, National Satellite Operation and Application Center, Korea Aerospace Research Institute; Changno LEE, Seoul National Univ. of Science and Technology; Jaehong OH, Korea Maritime and Ocean Univ.; Jaewan CHOI, Chungbuk National Univ.; Youkyung HAN, Seoul National Univ. of Science and Technology

Keywords: template matching, deep learning, cross-domain correspondence, domain adaptation

Accurate correspondence estimation between satellite imagery and ground control point (GCP) chips remains challenging due to significant domain discrepancies arising from differences in sensing modalities, spatial resolution, and radiometric characteristics. To address this issue, we propose a domain-adaptive deep learning framework for robust template matching between satellite and aerial images. Training datasets are constructed by applying random translations around keypoint correspondences extracted using SIFT-LightGlue, generating both aerial-aerial (A2A) and satellite-aerial (S2A) image pairs. To improve robustness under practical conditions, additional negative samples are generated by shifting reference patches outside the predefined search range. This allows the model to explicitly learn no-match scenarios when valid correspondences are absent within the search region. The proposed network adopts a residual U-Net architecture that processes reference and target patches to extract dense feature representations. To mitigate domain discrepancies, domain-aware instance normalization is incorporated into the decoder, allowing shared feature learning while adapting domain-specific statistics. A dense cosine similarity map is computed from the extracted feature maps via depthwise convolution, and temperature scaling is applied to produce well-calibrated matching logits. A no-match (dustbin) class is further introduced to explicitly model unmatched conditions and stabilize prediction confidence. The total loss consists of cross-entropy terms with both hard one-hot and Gaussian-based soft supervision, together with a negative mining loss that enlarges the similarity margin between matching and non-matching regions. A domain adaptation strategy is employed by first training the model on A2A data and subsequently adapting it to S2A matching while freezing batch normalization statistics for stable cross-domain generalization. Experimental evaluation on satellite–aerial datasets acquired over diverse land-cover conditions demonstrates reliable cross-domain matching performance, with an average root mean square error (RMSE) of 1.451 pixels.

Back to session list

E4-2: Integration of Physical Correction Model and LoG Sub-pixel Edge Detection for Image-based Rain Gauge Water Level Measurement

Presenting author: Wei-Chi HO

Authors: Wei-Chi HO, National Taiwan University Department of Civil Engineering; Jen-Jer JAW, National Taiwan University Department of Civil Engineering

Keywords: Image-based Water Level Measurement, Rain Gauge, Ground Sample Distance (GSD), Laplacian of Gaussian (LoG), Refraction and Parallax Correction, Surface tension

Accurately measuring water levels is crucial for rainfall analysis when using a rain gauge. In practical rainfall observations, the required accuracy often reaches the millimeter or even sub-millimeter level. In the environment of a pit for collecting ground truth of rain gauge water levels, image-based methods with continuous photography offer significant advantages, including real-time monitoring, automated data collection, and long-term reliability. However, achieving high accuracy remains challenging. The ground sample distance (GSD) typically limits accuracy to the millimeter level due to constraints in sensor resolution and object distance. In addition, image-based water level measurement requires calibration to convert image observations into physical water level estimates, with corrections for viewpoint-related parallax, container refraction, and surface tension effects on the water surface. Relying solely on pixel-level edge detection is insufficient to achieve the expected precision. The Laplacian of Gaussian (LoG) operator, which is based on second-order derivatives, identifies edge points as zero-crossing locations corresponding to sign changes in the LoG response of an image. By treating the discrete LoG response on pixels as a continuous surface, the discrete pixel coordinates constrained by the grid can be transformed into continuous coordinates, enabling more precise zero-crossing detection. It follows that the LoG is employed to detect sub-pixel accurate water surface edges, which are then incorporated into a physical correction model to attain the water level up to sub-millimeter accuracy. Overall, the proposed method offers an accurate ground truth reference of the rain gauge water level in the pit by supporting the physical correction model with sub-pixel edge detection quality.

Back to session list

E4-3: A Benchmark Study of Amplitude Filters and Stability Indices for PSC Screening in PSInSAR

Presenting author: Yan Akhbar PAMUNGKAS

Authors: Yan Akhbar PAMUNGKAS, Center for Space and Remote Sensing Research, National Central University, Taiwan & Department of Urban and Regional Planning, Brawijaya University, Indonesia; Shou-Hao CHIANG, Center for Space and Remote Sensing Research, National Central University, Taiwan

Keywords: InSAR, PSInSAR, StaMPS, persistent-scatterer-candidate, Sentinel-1A, amplitude-filtering

Persistent Scatterer Candidate (PSC) selection is a critical step in multi-temporal InSAR analysis because it directly influences the spatial density and reliability of points used in subsequent deformation measurements. In practice, amplitude-based screening is often represented by the Amplitude Dispersion Index (ADI/DA), yet the effect of different amplitude filtering strategies on PSC selection behavior remains insufficiently evaluated, particularly in StaMPS-based processing. In response, this study proposes a benchmarking framework to evaluate amplitude filtering strategies for PSC screening using Sentinel-1A Descending Orbit, comprising 170 SLC images spanning 2015 to 2022 over Jakarta, Indonesia. Six filtering configurations are proposed: no filtering, Lee, Kuan, Frost, Gamma-MAP, and multi-temporal median filtering. For each configuration, amplitude-based indices (ADI, mean amplitude, robust Amplitude Stability Index/rASI) will be generated using a consistent multi-epoch support criterion to preserve comparability with the conventional ADI selection criterion. PSC masks will then be derived using literature-based thresholds and percentile-based alternatives, and then quantitatively compared using PSC counts, retention fraction (%), statistics (min, max, mean, median, and standard deviation), histogram characteristics, and spatial distribution map. The expected outcome is to identify which filtering and index combinations could preserve physically plausible urban or stable targets while minimizing artifact-prone or spatially incomplete selections. The study is also expected to provide a practical recommendation for complementing the conventional ADI threshold in a reproducible StaMPS preprocessing workflow. Ultimately, this work aims to enhance the reliability of PSC screening and improve the robustness of downstream PSInSAR analysis for urban subsidence monitoring.

Back to session list

E4-4: A Graph Laplacian–Based Method for Radiometric Feature Identification in Photogrammetric Point Cloud

Presenting author: Guan-Ling LIU

Authors: Guan-Ling LIU, National Taiwan Univ.; Jyun-Yi WU, Ministry of the Interior Section Chief; Jen-Jer Jaw, National Taiwan Univ.

Keywords: Photogrammetric Point Cloud；Radiometric Feature；Graph Laplacian；Point Cloud Simplification；Texture Preservation

With the increasing adoption of photogrammetric point clouds in applications such as urban planning and cultural heritage preservation, the rapid growth in data volume has imposed substantial burdens on storage and computational resources. Accordingly, achieving effective data reduction while preserving critical features has become a fundamental challenge in point cloud simplification. Most existing methods primarily emphasize geometric feature preservation, with insufficient consideration of radiometric characteristics. Under high simplification rates, important texture information located in non-geometric feature regions is prone to removal, thereby degrading scene clarity and distinctiveness, and potentially affecting the accuracy of subsequent manual interpretation and automated point cloud analysis. In this study, radiometric features are defined as points located along color transition boundaries, typically corresponding to material variations or semantically meaningful structures, such as road markings, signage, and decorative or textual elements on cultural artifacts. To effectively identify such features, the point cloud is modeled as a graph based on the concept of graph Laplacian weighting scheme. Color information is assigned as node attributes, while the Euclidean distance between a point and its neighboring points is used as edge weight, mitigating the influence of spatially distant points on feature computation. By analyzing color differences between each point and its adjacent neighbors, the radiometric importance of each point within the overall point cloud is quantitatively measured. Compared with conventional local neighborhood–based methods, the proposed graph-based approach alleviates sensitivity to neighborhood scale and enables stable identification of points in regions with significant color transitions, providing a reliable basis for preserving radiometric features in subsequent point cloud simplification strategies.

Back to session list

E4-5: Occlusion Region Detection in High-Resolution Stereo Satellite Images Using Multi-Dimensional Relaxation Matching

Presenting author: Hyeona KIM

Authors: Hyeona KIM, Inha Univ.; Taejung KIM, Inha Univ.

Keywords: Stereo Matching; High-Resolution Satellite Image; 3D Reconstruction; Occlusion Detection; WorldView-3

High-resolution satellite stereo imaging is widely employed for precise large-scale 3D reconstruction. However, it faces extensive occlusions from off-nadir viewing and long baselines. Occlusion regions are not simultaneously observed in both images, which makes stereo matching fundamentally impossible in those areas. As a result, they are a major cause of disparity estimation errors and DSM distortions. Therefore, to improve the reliability of 3D reconstruction process, it is necessary to detect occlusion regions and separate them as low-confidence areas. In this study, we performed stereo matching using the Multi-Dimensional Relaxation approach. We detected occlusion and low-confidence matching regions by analyzing the distribution of disparity continuity along epipolar lines. First, we generated matching candidates by simultaneously applying multiple search windows of different sizes. Next, we computed a hit-score based confidence for each candidate through a relaxation process. A hit-score was assigned to matches satisfying the left–right consistency constraint. High confidence matches were used to construct an epipolar-line-wise hit-score map and estimate a 1D disparity path under ordering and one-to-one correspondence constraints. Then, along each epipolar line, blank segments and abrupt disparity changes were identified as potential occlusion and low-confidence regions. They were accumulated for overall epipolar lines to produce a 2D occlusion mask. The generated occlusion mask was validated using the original stereo satellite images and the overall matching confidence map. The results showed high agreement between the proposed occlusion mask and visually identified occlusion regions, confirming that occlusion and low-confidence regions can be effectively detected through analysis of the HitMap derived from Multi-Dimensional Relaxation-based stereo matching.

Back to session list

Session B5 Abstracts

B5-1: Research on Conditions for SAR Jamming Separation: An Approach Based on Phase Characteristic Differences

Presenting author: Jaeho BAE

Authors: Jaeho BAE, UST-KARI School(University of Science and Technology - Korea Aerospace Research Institute School); Jaemin SHIN, UST-KARI School(University of Science and Technology - Korea Aerospace Research Institute School)

Keywords: Synthetic Aperture Radar(SAR), Jamming, Anti-jamming, Jamming suppression, Phase,

In prior studies on SAR jamming analysis, attempts have been made to identify jamming signals within the two-dimensional domain of transmission (ft) and reception (fr) frequencies. However, when jamming signals are designed to occupy the same frequency bandwidth and time duration as the target echoes, they physically overlap with the signal of interest within this planar representation. In such scenarios, discrimination techniques based on conventional spectral analysis have shown fundamental limitations, as they cannot clearly distinguish the interference from the target signal without a significant loss of information. To overcome these limitations, this study proposes a novel framework focusing on the inherent differences in phase characteristics derived from the physical signal paths. While jamming and target signals may appear spectrally identical, their behaviors in the slow-time domain differ significantly due to their distinct origins. SAR target signals possess a deterministic phase history dictated by the geometric distance variation to the radar platform, whereas externally generated jamming signals lack this coupling and exhibit phase characteristics that are independent of the platform’s motion. Therefore, for effective separation, it is essential to introduce conditions beyond existing frequency variables that can specifically isolate these ‘inter-signal phase differences.’ This presentation defines the logical conditions to distinguish signals based on their phase consistency. We discuss the theoretical derivation of these conditions and the potential to isolate jamming components, thereby ensuring robust SAR imaging performance.

Back to session list

B5-2: Generation of Disparity Maps in Satellite Stereo Matching Using Multi-Asymmetric Patches

Presenting author: Chaeyeon LEE

Authors: Chaeyeon LEE, Inha Univ.; Taejung KIM, Inha Univ.

Keywords: Stereo matching; Dispartiy estimation; Satellite image; 3D reconstruction; WorldView-3

Recent demand for digital twins and three dimensional city models is driving the refinement of three dimensional geospatial data, where stereo derived depth is essential. In practice, depth is estimated from disparity maps computed from stereo pairs. This study generates disparity maps from satellite stereo epipolar images and examines whether combining multi scale windows with asymmetric patches improves matching robustness in challenging regions. Satellite stereo pairs enable efficient wide area mapping, but their large baselines and viewpoint differences often increase matching ambiguity. To handle diverse texture conditions, we adopt multi scale matching windows with five sizes from thirteen by thirteen to five by five. Smaller windows preserve boundaries and discontinuities, while larger windows suppress noise and stabilize matches in weakly textured regions. However, square windows aggregate evidence across the neighborhood, which may mix foreground and background cues near discontinuities and blur object boundaries. To reduce this mixing, we introduce asymmetric patches centered at each pixel, including two vertically split half patches and two diagonal triangular patches. Matching cost is computed using the Census transform, which is robust to radiometric variations. For fair comparison, we keep the feature dimension fixed for both square only and mixed patch settings. We fuse multi patch results by weighted averaging to form feature values without increasing the feature dimension. We evaluate match reliability using hit counts, which accumulates the number of single patch matching with left-right consistency. Experiments were conducted on low texture scenes from the US3D dataset. Experiments showed that under high hit thresholds, asymmetric patch fusion increased the number of valid reliable matches while maintaining similar endpoint error, thereby improving disparity map coverage and usability in satellite stereo matching.

Back to session list

B5-3: Dynamic Collaborative Mapping and Autonomous Frontier Exploration in Heterogeneous Multi-Robot Systems

Presenting author: JYUN-YANG HUNG

Authors: JYUN-YANG HUNG

Keywords: ROS; Heterogeneous Multi-Robot Systems; Collaborative Mapping; Distributed Perception; Sensor Fusion; AMCL

The deployment of heterogeneous multi-robot systems, leveraging the terrain adaptability of quadruped robots alongside the energy efficiency of wheeled robots, holds immense potential for complex environmental exploration. However, integrating closed-source commercial quadrupeds with open-source platforms in dynamic environments presents critical challenges, particularly in coordinate frame synchronization and real-time spatial alignment. This study proposes a robust dynamic framework for collaborative mapping and autonomous exploration using a Puppy Pi quadruped and a TurtleBot3 wheeled robot operating under ROS Noetic. To overcome inherent namespace collisions and independent odometry drift, we developed a decentralized master-slave architecture. The core contributions of this dynamic system are threefold: (1) implementation of Adaptive Monte Carlo Localization (AMCL) enabling the wheeled robot to dynamically track its pose within the continuously updating global map generated by the quadruped, effectively eliminating the reliance on static coordinate bridging; (2) integration of a real-time multi-robot map merging algorithm to seamlessly fuse heterogeneous LiDAR data into a unified spatial representation; and (3) deployment of an autonomous frontier exploration strategy that allocates mapping tasks based on the distinct mobility characteristics of each robot. Experimental results in unstructured environments demonstrate that this dynamic collaborative framework significantly accelerates spatial exploration speed, minimizes mapping blind spots, and exhibits high robustness in heterogeneous sensor data fusion.

Back to session list

B5-4: Mamba-Based Semantic Change Detection for Post-Typhoon Building Damage Assessment: A Case Study of Typhoon Danas (2025) in Tainan

Presenting author: Pei-Yu WEN

Authors: Pei-Yu WEN; Pei-Cheng CHEN; Tee-Ann TEO

Keywords: Semantic Change Detection, Mamba Architecture, Damage Assessment

As extreme weather events intensify due to global climate change, rapid and accurate post-disaster damage assessment is essential for enhancing urban resilience. This study investigates the application of Mamba-based semantic change detection (MambaSCD) for assessing building damage following Typhoon Danas, which struck Tainan in July 2025 with Level 13 gusts, affecting over 16,000 houses. Traditional deep learning architectures, such as CNNs and Transformers, often struggle with either limited receptive fields or quadratic computational complexity when processing high-resolution imagery. To address these limitations, this study adopted a framework utilizing the State Space Model (SSM) and Visual Mamba (VMamba) backbone, which achieves global contextual modeling with linear computational complexity O(N). The experiment utilized 0.15m high-resolution bi-temporal aerial ortho-rectified imagery (captured on April 27 and July 18, 2025). The study area covers a 1/5000 standard scene of the Taiwan eMap, encompassing approximately 716 hectares. The model was initialized with weights pre-trained on the SECOND dataset to enhance feature representation. The experimental results indicate a significant change ratio of 32.79% among 430 building targets. Crucially, MambaSCD successfully categorized semantic transitions into nine classes, effectively differentiating between complete structural collapse (i.e., Building-to-BareLand, 24.82%) and structural modifications without collapse (i.e., Building-to-Building, 28.37%). These findings demonstrate that the Mamba-based architecture provides high-fidelity, actionable intelligence for emergency response and resource allocation, offering a high-performance alternative for smart disaster management in large-scale urban scenarios.

Back to session list

B5-5: Assessing the Vertical Effectiveness of Ground Vegetation on 3D Urban CO2 Distribution using UAV Vertical Sampling and Geo-AI

Presenting author: Rui-Zhen YANG

Authors: Rui-Zhen YANG, Department of Geomatics, National Cheng Kung University; Chia-Wei HSU, Department of Geomatics, National Cheng Kung University; Jo-Pin HUNG, Agricultural Engineering Research Center; Chao-Hung LIN, Department of Geomatics, National Cheng Kung University; Chi-Da WU, Department of Geomatics, National Cheng Kung University

Keywords: UAV; Geo-AI; 3D CO2 Distribution; SHAP; NDVI; CO2 Mitigation

Urban vegetation serves as a critical carbon sink, mitigating global warming. However, traditional 2D monitoring lacks vertical resolution, making it difficult to quantify the impact of ground vegetation on the 3D urban canopy. This study aims to model the 3D distribution of urban CO2 and evaluate the vertical effectiveness of vegetation at different altitudes. We conducted high-resolution vertical sampling from 0 to 120 meters using Unmanned Aerial Vehicles (UAVs) in Taichung City. By integrating ground environmental variables, we applied a Geospatial Artificial Intelligence (Geo-AI) framework to develop a Stacking Ensemble Learning model to capture non-linear interactions. Additionally, 3D Kriging was applied to visualize the spatial structure and CO2 hotspots. To interpret the model, SHAP (SHapley Additive exPlanations) analysis was employed. The results revealed that NDVI was the most dominant predictor (23% contribution) and exhibited a significant negative correlation with CO2, robustly verifying the carbon sink effect. Crucially, this study reveals vertical consistency in vegetation effectiveness; the reduction effect remains stable from the ground up to 120 meters. This suggests that the influence of ground vegetation propagates vertically throughout the urban boundary layer. By overcoming the limitations of traditional 2D observations, this study demonstrates that integrating UAVs with Geo-AI can precisely resolve multidimensional greening benefits, providing robust scientific evidence for future 3D urban climate governance and carbon assessment.

Back to session list

B5-6: Global Navigation Satellite System; Inertial Navigation System; Coarse Alignment; Tightly Coupled Integration; Unmanned Surface Vehicle

Presenting author: JiunYo LIN

Authors: JiunYo LIN, National Cheng Kung Univ.; KaiWei CHIANG, National Cheng Kung Univ.

Keywords: Global Navigation Satellite System; Inertial Navigation System; Coarse Alignment; Tightly Coupled Integration; Unmanned Surface Vehicle

Strapdown inertial navigation systems (SINS) are widely used for marine and unmanned surface vehicle (USV) navigation due to their autonomy and continuity. However, under realistic maritime environments, vessels are subject to persistent six-degree-of-freedom (6-DOF) motions induced by waves, currents, and hull dynamics. These motions violate the stationary assumptions of conventional coarse alignment methods, which rely solely on gravity and Earth rotation, resulting in slow convergence and large heading errors. To address this issue, this study proposes a North–East–Down (NED)-based anti-swaying coarse alignment algorithm that formulates the alignment directly in the NED frame, the proposed approach avoids intermediate coordinate transformations and numerical approximations, improving computational stability and compatibility with downstream navigation and control modules. Real vessel trajectory data are used to simulate inertial measurement unit (IMU) outputs, capturing realistic coupling motions. Frequency-domain analysis shows dominant low-frequency dynamics below 1 Hz, corresponding to moderate sea states. The alignment method determines the body-to-navigation attitude using gravity and Earth rotation vectors, along with their cross product, forming an orthogonal reference triad. Wave-induced translational accelerations are filtered to isolate gravity, whose slow conical motion around Earth’s spin axis enables direct north estimation without iterative optimization or external aiding. Numerical simulations incorporating realistic IMU stochastic errors demonstrate fast and stable convergence. Roll and pitch errors remain within ±0.2°, while yaw converges to within ±1° in under 100 s. Compared with conventional analytic coarse alignment, the proposed method significantly reduces heading errors under dynamic conditions. When applied to initialize a tightly coupled GPS/INS system, the resulting navigation accuracy closely matches that of ideal initialization and substantially outperforms traditional methods. The results confirm that the proposed algorithm provides a practical and efficient coarse alignment solution for USV navigation under real sea motion.

Back to session list

Session C5 Abstracts

C5-1: Using LiDAR to Assist in Calculating Carbon Sinks in Plain Afforestation Areas

Presenting author: Rui-En HUANG

Authors: Rui-En HUANG, National Chengchi Univ.; Shih-Hong CHIO, National Chengchi Univ.

Keywords: plain afforestation；carbon sink；UAV-LiDAR；handheld Lidar；allometric volume equation

The purpose of this study was to address the inaccuracy of carbon sink calculations in plain afforestation areas caused by diverse tree species and management differences. Traditional forest carbon sink estimation relies on manual surveys, which are time-consuming, labor-intensive, and make it difficult to quickly obtain large-scale three-dimensional information. Therefore, this study proposed an innovative mapping workflow integrating unmanned aerial vehicle light detection and ranging (UAV-LiDAR) and handheld Lidar. UAV-LiDAR was utilized to obtain detailed top canopy information, and was combined with handheld Lidar to fill in the vertical structure of trunks and understory trees, constructing a complete, high-density 3D point cloud of the plain afforestation area. Refined tree models were established manually to serve as the true volume values. Meanwhile, three different volume estimation equations were fitted and compared: Schumacher and Hall’s (1933) logarithmic model, Spurr’s (1952) combined variable model, and a modified Spurr’s combined variable model incorporating allometric growth concepts. Subsequently, the biomass and carbon sinks of the plain afforestation areas were accurately estimated based on the Intergovernmental Panel on Climate Change (IPCC) parameters.

Back to session list

C5-2: SAR-based Phenology Alignment: A Time-Series Processing Strategy for Rice Growth Monitoring

Presenting author: Chih-Chen TSENG

Authors: Chih-Chen TSENG, National Taiwan Univ.; Yunung Nina LIN, Academia Sinica; Li-yu Daisy LIU, National Taiwan Univ.; Chih-Yung TENG, Taichung District Agricultural Research and Extension Station; Li-Ting LAI, Taichung District Agricultural Research and Extension Station; Jun-Wei GUO, National Taiwan Univ.

Keywords: SAR; Sentinel-1; Food security; Agriculture; Rice Phenology; Time-Series;

Remote sensing provides an effective framework for large-scale crop monitoring, yield assessment, and precision field management. Synthetic Aperture Radar (SAR) is particularly valuable in cloud-prone regions such as Taiwan due to its all-weather imaging capability. In Taiwan, rice is a dominant crop; however, rice cultivation exhibits substantial phenological heterogeneity caused by variations in varieties, regional temperature gradients and thereby transplanting dates, all together reducing satellite-based yield estimation accuracy. To address this issue, we propose a SAR-based phenology alignment approach using multi-source Sentinel-1 and PAZ data. The approach is developed and validated on 12 experimental rice fields with detailed phenological records, and then applied to large-scale rice cultivation areas in central western Taiwan. We first identify non-crop pixels and remove them using time-domain dispersion of VH amplitude; fields with insufficient sample points are excluded. We then transform the cleaned VH time-series into 12-day temporal difference series to capture phenological dynamics. Based on the difference series, we identify key growth peaks and analyze their relationship with yields. The result shows that approximately 60% of fields exhibited primary VH rate peaks at the tillering stage (~25% of total growth), while subsequent peaks in volumetric scattering power (Pv) corresponded to the heading stage (~70% of total growth). After phenology alignment, critical stages within the time-series of VV, Pv, surface scattering power (Ps), and the total polarized power (SPAN) show stronger association with the final yields (Spearman ρ = 0.6–0.7) as compared to those without alignment. These results suggest that SAR-based phenology alignment may provide reliable insights into rice growth dynamics and offer a scalable approach to synchronize growth stages across diverse fields, supporting large-scale yield estimation with improved accuracy.

Back to session list

C5-3: Downstream Discharge Prediction Based on Regression Models and LSTM

Presenting author: Joon Hyuk CHOI

Authors: Joon Hyuk CHOI, Yonsei Univ.; Hahn Chul JUNG, Yonsei Univ.; Seo Jin KIM, Yonsei Univ.; Hyongki LEE, Houston Univ.

Keywords: Discharge Prediction; River Width; Regression Model; LSTM; ICEYE

Accurate discharge prediction is crucial for mitigating flood damage and water resource management. Acquiring upstream in-situ data in transboundary river basins presents significant challenges. Consequently, satellite-based observation has emerged as a necessary alternative. This study aims to predict downstream discharge by estimating river width using ICEYE satellite imagery. The study focused on the Imjin River basin, a representative transboundary river where data accessibility is limited due to geopolitical factors between North and South Korea. High-resolution ICEYE SAR images collected from 2022 to 2023 under high flow condition were utilized for the analysis. Water bodies were detected by applying the Otsu thresholding method to each image, and the river width was derived by dividing the extracted water area by the river length. To effectively capture the non-linear relationship between river width and discharge, both regression models and Long Short-Term Memory (LSTM) networks were employed. The prediction performance was evaluated under three distinct scenarios with different input variable configurations: (1) using only the main channel width; (2) using the effective river width; and (3) utilizing both the main channel width and floodplain width as separate input features. The models were assessed using R², NSE, KGE, and RMSE. The results indicated that Scenario 3 yielded the most accurate predictions, followed by Scenario 2 and Scenario 1. These findings suggest that incorporating floodplain width is essential for precise discharge estimation. Furthermore, due to the morphological differences between the main channel and the floodplain, distinguishing between these two features provides superior predictive performance compared to using a combined effective width. This approach is expected to contribute to flood mitigation by enabling reliable downstream discharge prediction during sudden rainfall or flooding events.

Back to session list

C5-4: Interpretable Spectral-to-Thermal Framework for Fine-Scale Intra-Urban Thermal Variability from Multispectral Features and Pretrained Geospatial Representations

Presenting author: Wonbin KANG

Authors: Wonbin KANG, Seoul National Univ.; Donghyeon Lee, Seoul National Univ.; Yongil KIM, Seoul National Univ.

Keywords: Urban Heat Island; Land Surface Temperature; Thermal mapping; Multispectral Indices; Interpretability; Geospatial Representations

High-resolution mapping of intra-urban thermal variability is essential for urban climate analysis and heat management. This study develops a spectral-to-thermal framework that derives spectral thermal diagnostics for urban thermal assessment and evaluates spatial agreement with satellite land surface temperature (LST) products across major metropolitan areas in South Korea. Multispectral predictor variables are derived from Sentinel-2 surface reflectance and organized into a physically interpretable feature set comprising multispectral bands, established spectral indices, and local spatial statistics characterizing fine-scale heterogeneity. Thermal reference data are obtained from ECOSTRESS LST and Landsat Level-2 surface temperature products for cross-scene model training and validation. Scene-normalized LST anomalies are also considered to emphasize within-scene thermal contrasts and reduce cross-scene variability. Multispectral patches are encoded using pretrained Prithvi-EO-2.0 representations. A LightGBM regressor is trained to infer thermally coherent spatial patterns on the native multispectral grid. Model performance is assessed using spatially blocked cross-validation and temporally independent validation, with stratified error analysis across land-cover classes and, when available, urban-zone categories. The framework captures coherent intra-urban thermal gradients and hotspot morphology and outperforms single-index proxy baselines. Permutation importance, aggregated into physically grounded feature groups, attributes thermal variability to vegetation condition, moisture sensitivity, built-up intensity, and radiative surface properties. This study infers fine-scale urban thermal patterns by integrating multispectral predictors with pretrained geospatial representations, establishes an interpretable formulation of spectral–thermal relationships, and identifies spatial regimes in which multispectral predictors exhibit closer spatial correspondence with reference thermal patterns.

Back to session list

C5-5: IMPROVING LANDSLIDE DETECTION RELIABILITY VIA RANKING LOSS WITH PHYSICALLY-PLAUSIBLE CONSTRAINTS

Presenting author: Yeseok LEE

Authors: Yeseok LEE, Seoul National Univ.; Gyeongwon OH, Seoul National Univ.; Yongil KIM, Seoul National Univ.

Keywords: Landslide Detection; Change Detection; Morpho-Spectral Confidence Map; Ranking Loss; False Positive Reduction

Recent landslide detection studies have increasingly employed deep learning–based change detection models using multi-temporal remote sensing imagery. However, existing methods frequently misidentify non-semantic changes—such as illumination variations and seasonal vegetation dynamics—as actual landslides, particularly in complex regions where such changes frequently occur. To address this, we propose a ranking loss–based learning framework that incorporates a physically-plausible surface change criterion into the training process. Specifically, we define a Morpho-Spectral Confidence Map (MSCM) that captures surface disruption patterns associated with landslides by integrating brightness shifts, local structural variations, and spatial clustering. By coupling the MSCM with feature differences extracted from the encoder, we introduce a ranking loss that optimizes the relative detection confidence of landslide candidates. Experimental results on the GVLM dataset demonstrate that the proposed method reduces false positives by approximately 50% compared to the baseline, achieving a more conservative yet highly reliable detection performance. These findings suggest that the MSCM-based ranking strategy effectively filters landslide-unrelated change signals and improves detection reliability in diverse environments.

Back to session list

C5-6: Improving Temporally Consistent Building Extraction from Multi-Temporal Satellite Imagery

Presenting author: Donghyeon LEE

Authors: Donghyeon LEE, Seoul National University; Yongil KIM, Seoul National University

Keywords: Polygonization; Building Extraction; Multi-Temporal; Satellite Imagery; Post-Processing

Accurate and temporally consistent building extraction from satellite imagery provides a foundation for analyzing long-term urban dynamics. While recent polygonal building extraction models achieve strong performance on single-date imagery, their predictions often suffer from temporal inconsistency, missing vertices, and unstable polygon shapes when applied independently to time series satellite images. This paper addresses the problem of improving vectorized building representations in multi-temporal imagery by explicitly exploiting temporal coherence. Instead of treating each timestamp separately, we propose a post-processing framework that refines polygonal building extraction results across time. Our method consists of three sequential steps. First, buildings are matched across different timestamps using an Intersection-over-Union (IoU)-based criterion to establish temporal correspondence among the same physical structures, forming a set of polygons for each building. Second, within each building-specific set, corresponding polygon vertices are matched over time based on spatial proximity, enabling the construction of point-level temporal trajectories. Third, the polygon geometry of each building is refined by exploiting these temporal relationships, with a focus on recovering missing vertices, restoring missing polygons, and improving the positional accuracy and temporal consistency of individual points. We validate the effectiveness of the proposed method on multi-temporal satellite imagery with representative building extraction models. Experimental results demonstrate that vector-level temporal refinement produces more reliable and temporally coherent building representations, providing a foundation for downstream urban analysis and monitoring tasks.

Back to session list

Session D5 Abstracts

D5-1: Quantifying Living Environment and Urban Prosperity in Property Tax Assessment Using Landscape Images and Deep Learning

Presenting author: Yoshito KOIZUMI

Authors: Yoshito KOIZUMI, Tokyo City Univ; Kento TERADA, Tokyo City Univ.; Yuki AKIYAMA, Tokyo City Univ.; Katsuhiro AONO, AERO TOYOTA CORPORATION; Manami TAGA, AERO TOYOTA CORPORATION

Keywords: landscape image; deep learning; property tax assessment; qualitative factor

Approximately 180 million land parcels and buildings in Japan are subject to property tax assessment nationwide, which is revised every three years. However, while quantitative factors are systematically evaluated, qualitative factors such as living environment and prosperity -key determinants of roadside land values- do not have clear numerical standards or objective indicators and are currently left to the judgment of appraisers. As a result, appraisers are increasingly dependent on their own experience and subjectivity, leading to increased workloads and inconsistencies in assessment outcomes. Amid workforce shortages caused by population aging and the retirement of experienced appraisers, the real estate appraisal industry faces workforce shortages, increasing the need for data-driven support systems. To address this challenge, this study proposes a deep learning–based framework for quantifying qualitative environmental attributes, including living environment quality and urban prosperity, from street-level landscape images. Landscape image data were classified into four land-use district types, and living environment and urban prosperity were assigned four-level rating scores based on appraisal criteria provided by professional real estate appraisers, enabling the incorporation of expert knowledge into supervised learning datasets. A total of 1,935 street-level images were collected in 2025 using a 360-degree camera in Suita City, Osaka Prefecture. Using these labeled datasets, this study constructed district-specific and cross-district deep learning models via transfer learning with ResNet50 to evaluate environmental quality and prosperity. As a result, the district-specific models achieved accuracy rates of 68% to 71%, while the cross-district model achieved an accuracy rate of 66%, demonstrating stable predictive performance for qualitative appraisal indicators. Prediction accuracy varied across land-use districts, indicating the need for district-specific modeling strategies. These findings highlight the potential of deep learning–based landscape image analysis to support more objective and consistent property tax assessment and to reduce dependency on subjective appraisal practices.

Back to session list

D5-2: MS-Edge MIM: A Masking Strategy for Self-Supervised Learning on Very High-Resolution Satellite Imagery

Presenting author: Changhui LEE

Authors: Changhui LEE, Seoul National University of Science and Technology.; Jinmin LEE, Seoul National University of Science and Technology.; Youkyung HAN, Seoul National University of Science and Technology.

Keywords: Masked Image Modeling, Self-Supervised Learning, Very High-Resolution, Satellite Image

Masked Image Modeling (MIM) has emerged as an effective paradigm for Self-Supervised Learning (SSL), yet most existing approaches rely on uniformly random masking that does not fully exploit structural priors inherent in satellite imagery. We propose an enhanced MIM framework that incorporates two complementary masking strategies, multi-scale masking and edge-aware masking. The multi-scale strategy applies variable block sizes within the random mask grid, encouraging the model to reconstruct both coarse semantic regions and fine grained details. The edge-aware strategy leverages a lightweight Sobel-based prior to bias mask sampling toward high gradient regions, thereby increasing reconstruction difficulty on structurally informative areas such as object boundaries, roads, and man-made patterns. The proposed method is trained using SSL without labels on a combined very high-resolution satellite image dataset composed of MLRSNet and the fMoW RGB subset. To evaluate the learned representations, we conduct image classification experiments under linear probing and full fine-tuning to assess representation quality and adaptability. In addition, semantic segmentation is performed to examine practical applicability in real world scenarios. Experimental results demonstrate that the proposed masking strategies consistently outperform standard fixed block size random masking in both classification and semantic segmentation tasks. The improved representation quality effectively transfers to semantic segmentation, indicating enhanced practical utility in remote sensing applications. These results confirm the effectiveness of the proposed multi-scale and edge-aware masking strategies for SSL in very high-resolution satellite imagery.

Back to session list

D5-3: Enhancing Perceptual Quality and Geometric Features of SkySat Video via SinSR Diffusion Model

Presenting author: Pin-han CHEN

Authors: Pin-han CHEN, National Yang Ming Chiao Tung University; Tsung-Han WEN, National Yang Ming Chiao Tung University; Tee-Ann TEO, National Yang Ming Chiao Tung University

Keywords: SkySat video, Super-resolution, Single-step diffusion model, Perceptual quality

Microsatellite constellations have catalyzed a paradigm shift toward high-frequency Earth observation. Among these, Planet Labs’ SkySat provides sub-meter resolution satellite video with unprecedented revisit capabilities, yet image quality remains constrained by sensor limitations and motion-induced degradation during dynamic sampling. While Multi-Frame Super-Resolution (MFSR) leverages spatiotemporal redundancy, its efficacy is highly sensitive to sub-pixel alignment accuracy; the misregistrations often result in structural distortion. Recent advances in Convolutional Neural Networks, Vision Transformers (ViT), and diffusion-based architectures have improved perceptual quality on general benchmarks, but their application to microsatellite video within a physically grounded remote-sensing framework remains underexplored. This study introduces the Single-step Super-Resolution (SinSR) diffusion model, utilizing deterministic distillation to facilitate high-efficiency inference. Through a teacher-student training paradigm, the generative capacity of a pre-trained latent diffusion model is distilled into a single-step student framework optimized for SkySat’s radiometric characteristics. We employ a multi-objective loss function integrating adversarial, perceptual, and pixel-wise components to enhance structural fidelity. By compressing the multi-step denoising process into a singular direct mapping, SinSR inherits robust generative priors that enable the synthesis of high-frequency textures in a single pass, bypassing traditional sampling bottlenecks. The methodology was validated using a 30-second SkySat video acquired over Kaohsiung City in 2021. The dataset comprises 900 consecutive panchromatic frames at 29.97 fps, featuring an average ground sampling distance (GSD) of 0.885 m. These frames were subjected to 4X super-resolution reconstruction (achieving an effective GSD of 0.221 m). Performance was rigorously evaluated using PSNR, LPIPS, and MUSIQ, alongside Modulation Transfer Function (MTF) analysis. Results demonstrate a 41.6% increase in the MTF50 index—improving from 0.303 cy/px to 0.429 cy/px—indicating a substantial gain in structural clarity. While pixel-wise gains remain moderate, the framework effectively reconstructs fine-scale features beyond physical sensor constraints. This demonstrates the practical feasibility of single-step diffusion models for high-fidelity dynamic monitoring in microsatellite remote sensing.

Back to session list

D5-4: Autonomous UAV with YOLOv11-Based Gripper System for Real-Time Papaya Ripeness Detection and Harvesting

Presenting author: Han-Sheng HSU

Authors: Han-Sheng HSU, National Pingtung University; Po-Yuan YANG, National Pingtung University

Keywords: UAV; YOLO v11; 3D-printed gripper; robotic arm; multi-target recognition; SITL simulation; Raspberry Pi 5

With the growing development of smart agriculture and the increasing shortage of labor, automated fruit harvesting has become a key research focus. While existing UAVs are commonly used for aerial monitoring or pesticide spraying, their direct application in fruit harvesting remains limited, especially for medium-to-large fruits such as papayas. This study proposes an autonomous UAV with a gripper and robotic arm, both fabricated using PLA+ 3D printing, utilizing YOLOv11 for fruit ripeness detection and multi-target recognition, built on a self-assembled F450 drone platform, with papayas as the primary target. The system employs the YOLOv11 model for ripeness detection, t trained on a dataset of public and self-captured papaya images, covering different viewing angles and ripeness stages. Experimental results show that YOLOv11 outperforms YOLOv8 in fruit ripeness detection, achieving a precision of 0.78885, a recall of 0.81883, and an mAP50 of 0.86438, demonstrating its effectiveness in multi-target recognition tasks. This approach helps farmers quickly assess crop growth and ripeness, allowing timely harvesting of a few already-ripe fruits to prevent overgrowth and quality loss. In the future, integrating relevant algorithms could estimate total harvestable fruits across a farm or orchard. SITL simulation experiments show that the UAV can correctly execute the lawnmower-pattern patrol logic, and the gripper module can reliably complete fruit picking. The hardware system integrates the robotic arm and gripper with a Raspberry Pi 5 for image recognition, sending commands to the Arduino to operate the robotic arm and gripper for fruit picking. While fully autonomous picking is still under testing, joystick-assisted operation has been used for verification. This study confirms the feasibility of integrating deep learning-based image recognition with UAV autonomous control for fruit harvesting, and provides a practical reference for future applications, including smaller fruits such as strawberries or blueberries, and multi-UAV cooperative harvesting.

Back to session list

D5-5: Study on Building Symbol Extraction and Vectorization Methods from Historical Maps

Presenting author: Lewei LIN

Authors: Lewei LIN, Tokyo City University; Yuki AKIYAMA, Tokyo City University

Keywords: Historical Map; Map Archive; Historical GIS; Raster Vectorization; Image Tracing

Historical maps are widely recognized as an indispensable resource for historical geography research. However, since these maps are mainly stored and distributed as raster image data, advanced data utilization such as geospatial analysis remains inaccessible to researchers lacking specialized expertise in image processing. This study aims to establish a method for systematically extracting and vectorizing building regions from topographic maps based on modern surveying techniques, with the results expected to contribute to quantitative research using historical maps. We selected eight 1:25,000-scale maps provided by the Geospatial Information Authority of Japan (GSI) as experimental subjects, in which building symbols are present in large quantities and distributed relatively uniformly across the entire map. This study leverages classical image processing approaches to facilitate the efficient generation of training data. K-means clustering was first applied to suppress background (white), water (blue), and terrain-related (brown) pixels. Morphological processing was then used to remove thin linear structures from the remaining regions, successfully isolating regions consisting of building candidates with minimal residual noise. The selection process for a ground-truth dataset of 1,500 256×256 map tiles required approximately 10 hours to complete. The U-Net model was applied as a semantic segmentation architecture well suited for pixel-level classification, using 1,200 pairs of original images with ground-truth masks. Subsequent validation on the remaining 300 images yielded an Intersection over Union (IoU) of 0.8792, compared to an IoU of 0.6458 using only classical methods. Finally, the trained model was applied to the remaining map tiles to extract building regions and polygon representations of buildings were vectorized using the Douglas–Peucker algorithm. This study proposes a method for accurately extracting and vectorizing building regions on historical maps by combining classical image processing techniques with deep learning methods. Future research goals include developing segmentation models targeting other map symbols and enhancing detection accuracy.

Back to session list

D5-6: SimVP-GAN for GK2A Infrared Forecasting: Resolving the Accuracy–Sharpness Trade-off in Weather Satellite Remote Sensing

Presenting author: Yeonjun KIM

Authors: Yeonjun KIM, Sejong Univ.; Sungwook HONG, Sejong Univ. (Corresponding author)

Keywords: SimVP-GAN; Nowcasting; Geostationary weather satellite; Infrared; Brightness temperature

Precise short-term forecasting of geostationary meteorological satellite imagery is essential for effective early warning systems against natural hazards. However, existing deep learning approaches often face a dilemma: convolutional neural network (CNN)-based models, such as Simpler yet Better Video Prediction (SimVP) and Convolutional Long Short-Term Memory, tend to produce blurry predictions, while generative adversarial network (GAN)-based data-to-data (D2D) models often struggle with quantitative accuracy due to training instability. To address these challenges, this study proposes SimVP-GAN, a novel framework predicting level 1 infrared (IR) observations from the Geo-Kompsat-2A (GK2A) satellite up to 6 hours in advance. The proposed architecture integrates a SimVP generator with a 3D convolutional PatchGAN discriminator. This hybrid approach enables the effective capture of spatiotemporal dynamics while enforcing structural coherence. Comprehensive experiments using nine GK2A IR channels demonstrate that SimVP-GAN bridges the gap between accuracy and realism. Compared to D2D models, it achieves significantly higher accuracy, recording the lowest Mean Absolute Error and Root Mean Squared Error in seven channels. Unlike CNN baselines, the model enhances visual sharpness, reducing blur artifacts by approximately 47% according to a novel absolute blur-effect error metric, while successfully reconstructing fine-grained features. Despite minor limitations in water vapor channels, SimVP-GAN offers a superior balance between accuracy and perceptual realism, providing a robust solution for operational weather analysis. This research was supported in part by Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education (RS-2025-16064562), in part by Development Programs “Development of AI Techniques for Weather Forecasting” under Grant KMA2021-00120, and in part by the National Research Foundation of Korea (NRF) through Korean Government (Ministry of Science and ICT (MSIT)) under Grant RS-2025-16064562 (Corresponding author: Sungwook Hong).

Back to session list

Session E5 Abstracts

E5-1: Applicability analysis of monocular depth estimation models for 3D information Inference from remote sensing imagery

Presenting author: BAEK JUNSEO

Authors: BAEK JUNSEO, Seoul National Univ. of Science and Technology; HAN YOUKYUNG, Seoul National Univ. of Science and Technology

Keywords: Monocular Depth Estimation, Remote Sensing, Applicability Analysis

Monocular Depth Estimation (MDE) predicts per-pixel depth from a single RGB image to reconstruct the three-dimensional geometric structure of a scene. In remote sensing, three-dimensional spatial information is typically generated through photogrammetric methods or multi-view image matching processes that exploit parallax between images acquired from different viewpoints. Although these approaches can reconstruct physically consistent spatial information based on geometric triangulation, they require complex processing steps, such as ensuring sufficient image overlap and performing geometric calibration. In contrast, MDE models infer spatial structure from a single image, offering advantages in data acquisition and processing efficiency. However, most existing MDE models have been trained primarily on ground-level images, and their data distributions as well as geometric and radiometric characteristics differ substantially from those of remote sensing imagery acquired from high altitudes and covering large spatial extents. Therefore, it is necessary to verify whether such models can generalize consistently to the remote sensing domain, including Unmanned Aerial Vehicle (UAV), aerial, and satellite imagery. In this study, previously developed MDE models are applied to various remote sensing images. Specifically, Depth Anything V2 and UniDepth V2 models are used as a relative and an absolute depth estimation, respectively. For the relative depth results, structural representation characteristics are examined with a focus on how consistently the shapes of major objects, such as buildings and roads, and their relative height relationships are preserved. For the absolute depth results, the consistency and accuracy of depth values and the absolute scale are visually and qualitatively evaluated based on existing Digital Surface Model (DSM) and GPS-measured elevation data.

Back to session list

E5-2: Improving Visual SLAM Robustness in Low-Light Environments via Image Enhancement Techniques

Presenting author: Yusuke ITO

Authors: Yusuke ITO, Mei Univ.; Masayuki MATSUOKA

Keywords: Deep Learning; Autonomous Driving Car; Illumination Condition; Monocular Camera; Computer Vision

As autonomous driving technology advances, the importance of Simultaneous Localization and Mapping (SLAM) has increased, with Visual SLAM (VSLAM) being particularly notable for its low cost and practicality. However, VSLAM faces a significant challenge in low-light or nighttime environments because the number of extracted feature points decreasing, leading to lower accuracy. This study evaluates the performance improvement of VSLAM in such conditions by applying image enhancement technologies to correct brightness. Specifically, two methods were used: EnlightenGAN, which allows for unsupervised learning without paired images, and Bread, which combines multiple networks. Using these techniques, self-localization was performed with ORB-SLAM3. We used monocular camera images from the Oxford Robotcar Dataset. For nighttime data, we compared the route completion rate by 20 trials and analyzed the positions where the system lost its location. For daytime data, the estimation accuracy was evaluated by comparing the results with the Ground Truth. The results showed that while the original nighttime data had a completion rate of 5%, the EnlightenGAN-enhanced data achieved 70% and tracked longer distances even when it failed. Bread also improved the tracking distance compared to the original data, although it never reached the goal. Most failures occurred at corners where rapid image changes made it difficult to find corresponding points. In contrast, for the daytime data, the Root Mean Squared Error (RMSE) rose from 1.9 m in the original data to 21.0 m with EnlightenGAN and 30.4 m with Bread. This decrease in accuracy likely occurred because enhancing already bright images reduced their contrast, which made it difficult to match feature points. In conclusion, image enhancement is effective for VSLAM in the dark. In the future, we aim to evaluate the performance of the image enhancement module by looking at the correlation between image brightness and the number of feature points.

Back to session list

E5-3: A GIS-Based Multi-Objective Optimization Framework for Age-Sensitive Pedestrian Route Choice

Presenting author: FEBRIAN FITRYANIK SUSANTA

Authors: Febrian Fitryanik SUSANTA, National Cheng Kung University; Universitas Gadjah Mada; Pei-fen KUO, National Cheng Kung University; Tsun-Yuan KU, National Cheng Kung University; I Gede Brawiswa PUTRA, National Cheng Kung University

Keywords: Network Analysis; Route Optimization; Older Adults; Sidewalk; Pedestrian Routing

Most pedestrian routing models optimize for distance or travel time. Although efficient, such approaches do not fully reflect how older adults make walking decisions. For aging populations, safety conditions, sidewalk quality, slope, crossing difficulty, and environmental comfort often shape route choice as much as travel efficiency. However, these dimensions are rarely incorporated into operational network optimization models. This study develops a GIS-based routing framework that compares conventional single-objective shortest-path solutions with multi-objective optimization for age-sensitive pedestrian navigation. Dijkstra’s algorithm is applied to generate shortest-distance and shortest-time routes. In contrast, the Non-dominated Sorting Genetic Algorithm III (NSGA-III) identifies Pareto-optimal routes that simultaneously balance accident risk exposure, walkability conditions, and perceived environmental quality. The model is implemented on the National Cheng Kung University campus in Tainan, Taiwan, using detailed pedestrian network data, slope information, traffic exposure indicators, and survey-based perception measures. Routes generated under identical origin–destination pairs are evaluated in terms of distance, travel time, cumulative risk exposure, and overall walkability performance. Results indicate that while Dijkstra-based routes closely resemble conventional navigation outputs, multi-objective solutions frequently select slightly longer paths that substantially reduce exposure to risk and improve environmental quality. These findings underscore the importance of integrating multidimensional spatial cost structures into pedestrian routing models and offer practical insights for age-sensitive campus planning and urban mobility design.

Back to session list

E5-4: Pan-Sharpening of KOMPSAT-3A Multispectral Images Using a Transformer-based Super-Resolution Model

Presenting author: Jihyeon LIM

Authors: Jihyeon LIM, Inha Univ.; Taejung KIM, Inha Univ.

Keywords: Pan-Sharpening; KOMPSAT-3A; Modulation Transfer Function; Transformer; Super-Resolution

Multispectral satellite images provide rich spectral information, but their spatial resolution is lower than panchromatic images due to constraints in sensor design. Typically, multispectral and panchromatic images are fused using rule-based information injection to generate high-resolution pan-sharpened multispectral images. However, existing pan-sharpening methods may fail to adequately reflect the physical characteristics of satellite sensors, resulting in excessive spatial information injection or spectral distortion. This study proposes a Transformer-based pan-sharpening method that incorporates the sensor characteristics of KOMPSAT-3A into model learning process. To generate training data, pairs of original-resolution and low-resolution patches were prepared. A degradation operator was applied to original multispectral images to generate low-resolution image patches. We constructed the degradation operator based on a Modulation Transfer Function (MTF) derived from the original KOMPSAT-3A images in order to reflect the frequency attenuation characteristics of the sensor. The proposed Transformer’s self-attention operation considered not only the neighborhood of each pixel but also distant patterns. This allowed for the injection of necessary information only without compromising consistency of the overall image. In the spatial branch of the model, edge and texture features were extracted from the panchromatic image. In the spectral branch, inter-band correlations were learned from the multispectral image. The results obtained from the two branches were fused and upsampled to generate a high-resolution multispectral image. The quality of the fused image was evaluated by calculating spatial and spectral distortion indices and comparing them with the results obtained by applying traditional Component Substitution(CS)-based pan-sharpening method. The spatial distortion of the model was approximately 0.0092, which was approximately 64% lower than that of CS. The spectral distortion was approximately 0.0087, which was approximately 2% lower than that of CS. Consequently, the spatial-spectral trade-off that often occurs in traditional pan-sharpening methods was alleviated by model learning that reflected the MTF-based sensor characteristics.

Back to session list

E5-5: Building-Level Vacant House Detection Using Exterior Images and Deep Learning: A Cross-Regional Transferability Analysis

Presenting author: Taro OKUMURA

Authors: Taro OKUMURA, Tokyo City Univ.; Kento TERADA, Tokyo City Univ.; Yuki AKIYAMA, Tokyo City Univ.; Ryo MIYATA, Tokyo City Univ.

Keywords: Deep Learning; Vacant House; Building exterior; 360 Degree Camera; Spatial Integration

In recent years, the number of vacant houses in Japan has increased due to population decline and aging, causing serious urban management issues such as landscape deterioration and disaster risk. Therefore, municipalities across the country have been required to conduct surveys to identify the locations of vacant houses. However, many existing vacant-house survey methods rely on manual visual inspection, making large-scale surveys time-consuming and costly. This study proposes a building-level vacant house detection framework using exterior images and deep learning, and examines its cross-regional transferability by constructing a deep learning model using building exterior images captured by vehicle-mounted cameras and the results of vacant-house surveys, and by automatically extracting visual features related to vacancy conditions. The target areas of this study are Maebashi City in Gunma Prefecture and Tanabe City in Wakayama Prefecture. First, building exterior images were integrated with the results of on-site vacant-house surveys to create labeled datasets of vacant and non-vacant houses. Next, these images were used as training data to construct a convolutional neural network-based classification model using VGG16, and its usefulness was evaluated by examining classification accuracy. In addition, cross-regional validation was conducted to evaluate the generalizability of the model. As a result, the proposed model achieved high classification performance in intra-regional validation (F-scores of 70.3%, 69.6%), whereas performance significantly decreased in cross-regional validation between the two cities (F-scores of 54.6%, 55.1%). This performance degradation suggests that vacancy-related visual features are strongly region-dependent, particularly due to differences in building age distributions shaped by each city’s historical development, including periods of urban expansion and post-war or post-disaster reconstruction, which limit the generalizability of exterior-image-based models across regions. The establishment of this method is expected to reduce the cost and labor of municipal field surveys, thereby contributing to more efficient and lower-cost vacant-house management by municipalities.

Back to session list

E5-6: Near-Real-Time Generation of IMERG-like Precipitation Using GK-2A Satellite Data and Generative Deep Learning

Presenting author: Kyung-Hoon HAN

Authors: Kyung-Hoon HAN, Sejong University, Republic of Korea; Jaehoon JEONG, National Institute of Environmental Research, Republic of Korea; Sungwook HONG, Sejong University, Republic of Korea (Corresponding author)

Keywords: Precipitation Estimation; CGAN; GK-2A; IMERG; Generative Deep Learning

The National Aeronautics and Space Administration (NASA) provides high-quality global precipitation estimates through the Integrated Multi-satellite Retrievals for Global Precipitation Measurement (IMERG), which combines and interpolates data from various satellite constellations. However, the IMERG Final Run, despite its high accuracy, is released with a latency of approximately three months due to extensive processing and calibration, rendering it unsuitable for real-time applications. To address this latency and enhance temporal resolution, this study proposes a Conditional Generative Adversarial Network (CGAN)-based generative deep learning model utilizing data from the GEO-KOMPSAT-2A (GK-2A) satellite. The proposed framework includes pre- and post-processing steps to effectively translate the physical relationships between satellite brightness temperatures and precipitation rates. The model was trained and evaluated using GK-2A brightness temperatures and IMERG Final precipitation data covering the Asia and Oceania regions from September 2019 to December 2022. The generated precipitation data were validated against IMERG Final products and compared with operational GK-2A RR and PDIR-Now products. Furthermore, the model’s capabilities were assessed through a comparative analysis of typhoon cases. Performance metrics for the entire study domain indicated a Probability of Detection of 0.607, a Critical Success Index of 0.482, a Root Mean Square Error of 0.759 mm/hr, and a Correlation Coefficient of 0.671. Although the model showed a slight tendency to underestimate, it successfully simulated precipitation patterns at high temporal resolution. Although the model exhibited a slight tendency toward underestimation, it successfully simulated precipitation patterns with high temporal resolution. By enabling real-time monitoring at 10-minute intervals, this approach is expected to serve as a valuable supplementary tool for weather forecasters and general users in the Asia-Oceania region. This research was supported in part by the Korea Meteorological Administration Research and Development Programs “Development of AI Techniques for Weather Forecasting” under Grant KMA2021-00120, in part by the “Developing Service Platform Technology for AI and Data Convergence” under Grant KMA2021-00122, and in part by the Hyundai Motor Chung Mong-Koo Foundation.

Back to session list

Session S1 Abstracts

S1-1: Long-term vicarious calibration for the GOSAT series at Railroad Valley

Presenting author: Kei SHIOMI

Authors: Kei SHIOMI, JAXA; Rio KAJIURA, JAXA; Atsushi YASUDA, RESTEC; Hirokazu YAMAMOTO, Hiroshima Univ.

Keywords: GOSAT; FTS; Greenhouse gas; Calibration

The GOSAT has been the first dedicated mission for greenhouse gas measurement from space since 2009. Radiometric accuracy is maintained by radiometric change evaluation through on-orbit functions such as solar diffuser, onboard laser and lunar calibration. The other method is a vicarious calibration based on field experiments. The GOSAT team has conducted vicarious calibration experiments for more than a decade in collaboration with the NASA OCO team. Ground-based spectroscopic observations and atmospheric profile measurements define the atmospheric state, and radiative transfer calculations generate theoretical radiances. By comparing these simulations with satellite coincidence observations, on-orbit radiometric gain errors and wavelength-dependent biases are quantitatively evaluated. The primary validation site is Railroad Valley, Nevada, USA, a dry desert region with spatially homogeneous and highly reflective surfaces. Our activities are aimed at establishing and providing radiometric standards for space-borne greenhouse gas sensors: GOSAT, OCO-2, TROPOMI, GOSAT-2, OCO-3, TEMPO, and MircoCarb. This study reports the outcomes of these long-term vicarious calibration activities.

Back to session list

S1-2: Operational Calibration Update of the Advanced Land Observing Satellite-2 and -4 (ALOS-2/-4)

Presenting author: Takeo TADONO

Authors: Takeo TADONO, Japan Aerospace Exploration Agency; Takeshi MOTOHKA, Japan Aerospace Exploration Agency; Masato OHKI, Japan Aerospace Exploration Agency; Osamu ISOGUCHI, Remote Sensing Technology Center of Japan; and Haruya HIRANO, Remote Sensing Technology Center of Japan

Keywords: Calibration; Validation; SAR; ALOS-2; ALOS-4

The Advanced Land Observing Satellite-2 (ALOS-2) is the follow-on mission of the L-band Synthetic Aperture Radar (SAR) of ALOS and the Japanese Earth Resources Satellite-1 (JERS-1). ALOS-2 was launched in May 24, 2014, and currently continues to operate smoothly by the Phased Array type L-band SAR-2 (PALSAR-2) beyond its design life of seven years. It is suitable for monitoring disasters rapidly as well as it can reach the ground partially penetrating through vegetation to obtain information on vegetation and ground surface. The observation capability of PALSAR-2 has been dramatically improved over ALOS/PALSAR e.g., a finer spatial resolution of 1 x 3 m by the Spotlight observing mode, a wider swath width of 490 km by the ScanSAR mode, and a right-and-left looking observable function. ALOS-4, the successor to ALOS-2, was launched on July 1, 2024, aboard the H3 Launch Vehicle Flight No. 3. It is currently continuing observations by its onboard PALSAR-3 L-band SAR. PALSAR-3 not only retains the observation modes of PALSAR-2 but also significantly enhances its observation swath width through the adoption of Digital Beam Forming (DBF) technology. For example, in Stripmap 3 m resolution Mode, PALSAR-3 achieves a 200 km swath width compared to PALSAR-2’s 50 km, while in ScanSAR Mode, it ranges from 350 km to 700 km. Furthermore, ALOS-4 employs the same orbit as ALOS-2, significantly increasing opportunities for SAR interferometric analysis using PALSAR-2/PALSAR-3. Geometric and radiometric calibration is essential for the continuous and quantitative utilization of these satellite observation data. Calibration is essential for the continuous and quantitative utilization of satellite observation data. JAXA continues to perform initial calibration after satellite launch and operational calibration after transition to routine operations. This paper reports the latest results of the ongoing calibration of ALOS-2/-4.

Back to session list

S1-3: An Overview of CEOS Analysis Ready Data: Concepts, Implementation, and Future Perspectives

Presenting author: Takeo TADONO

Authors: Takeo TADONO, Japan Aerospace Exploration Agency; Christopher BARNES, KBR contractor to the U.S. Geological Survey; Jonathon ROSS, Geoscience Australia; Matthew ADAMS, Geoscience Australia; Medhavy THANKAPPAN, Geoscience Australia; Peter STROBL, European Commission Joint Research Centre; Ake ROSENQVIST, soloEO; Matthew STEVENTON, Symbios

Keywords: ARD; Optical; Thermal; SAR; Standardization; CEOS

Committee on Earth Observation Satellite Analysis Ready Data (CEOS-ARD) is a framework for making satellite data easier to understand and use. It has been defined within CEOS enables direct use of satellite data in analysis with minimal additional user preprocessing. Its core concept is to define a common baseline of preprocessing and documentation requirements through, so called, Product Family Specifications (PFS) for optical, thermal, and Synthetic Aperture Radar (SAR) data, including radiometric and geometric corrections, standardized metadata, and pixel-level quality information. Peer-review evaluation confirms whether products demonstrably meet the requirements defined in the PFSs, ensuring consistency and transparency. Set out to accommodate temporal consistency and interoperability across sensors, CEOS-ARD aims to enable robust time-series analysis and multi-mission data integration. The primary benefit is the significant reduction of preprocessing burden for users, allowing researchers and practitioners to focus on scientific analysis rather than data preparation. This also improves reproducibility and supports large-scale processing environments i.e., cloud platforms, Data Cubes, and AI/ML workflows. CEOS provides strategic governance, manages PFS documents, and facilitates discussions on cross-cutting technical issues e.g., interoperability and metadata standards. The CEOS-ARD group also promotes collaboration with international standardization bodies and supports outreach efforts to broaden CEOS-ARD adoption including by the commercial Earth observation sector. The future direction of CEOS-ARD can be described through three complementary scenarios. The first is the establishment of ARD as an international standard for Earth observation data. The second envisions a commercially driven ecosystem in which private-sector providers adopt ARD specifications to enable scalable services and interoperability. The third foresees ARD evolving from a data specification into a cloud-native analysis infrastructure optimized for automated and AI-driven processing. Most likely, these pathways will converge, positioning CEOS-ARD as a foundational framework that supports interoperable, scalable, and analysis-oriented Earth observation systems.

Back to session list

S1-4: Extending Mirror Array-Based Co-Registration to High-Resolution Satellite Imagery: Impacts on Geolocation and Band-to-Band Accuracy

Presenting author: Muhammad Daniel Iman bin HUSSAIN

Authors: Muhammad Daniel Iman bin HUSSAIN, Yamaguchi University Center for Research and Application of Satellite Remote Sensing (YUCARS); Masahiko NAGAI, Yamaguchi University Center for Research and Application of Satellite Remote Sensing (YUCARS)

Keywords: Mirror array; geolocation accuracy; band-to-band registration; high-resolution; PlanetScope

Accurate geolocation and band-to-band alignment are critical for high-resolution Earth observation satellites, yet remain challenging due to sensor-dependent geometric uncertainties and limited availability of reliable high-resolution reference data. Artificial mirror array targets have previously been demonstrated to improve geolocation accuracy for medium-resolution satellite imagery by providing stable, point-like control signals; however, their effectiveness for high-resolution sensors has not yet been systematically evaluated. This study investigates the applicability of mirror-based co-registration to high-resolution satellite imagery and examines its impact on both image-to-image and band-to-band co-registration accuracy. Mirror responses captured in PlanetScope (3–4 m) and high-resolution imagery such as Pleiades and NewSat are used to perform relative co-registration, with Pleiades serving as a high-accuracy reference based on its intrinsic geolocation performance. Independent aerial orthophotography, which does not contain mirror responses but offers superior positional accuracy, is employed to assess absolute geolocation accuracy before and after co-registration. The experimental dataset includes multiple high-resolution satellite acquisitions over a common test site, enabling evaluation of improvements across sensors with differing native geolocation accuracy. Co-registration performance is compared with a mirror-free baseline to quantify the contribution of artificial point sources in the high-resolution domain. In addition, the effect of mirror-based alignment on band-to-band registration accuracy is examined, extending prior work beyond medium-resolution imagery. Preliminary results indicate that mirror arrays significantly improve both geolocation and band-to-band accuracy for high-resolution satellite images. These findings demonstrate the scalability of mirror array-based co-registration and support its use as a practical tool for high-resolution multisensor image harmonization.

Back to session list

S1-5: Past Activities and Future Directions of Accuracy Assessment for Optical Satellite Data: A Case Study of the Phenological Eyes Network (PEN)

Presenting author: Taiga SASAGAWA

Authors: Taiga SASAGAWA, Chiba Univ.; Kenlo NASAHARA, Univ. of Tsukuba

Keywords: In-situ measurement; Time-lapse camera; Sky-radiometer; Vegetation; Phenology

In satellite remote sensing, post-launch calibration is commonly performed using stable reference targets such as desert surfaces. In addition, optical remote sensing of terrestrial vegetation also requires field-based spectral measurements over actual vegetation canopies, and satellite observations are evaluated through comparisons with these ground-based measurements. In this presentation, we introduce the activities of the Phenological Eyes Network (PEN), a long-term in-situ observation network that has been operating since 2003 using a combination of hyperspectral sensors, time-lapse cameras, and sky radiometers. We present an overview of the network development as well as selected research examples that have utilized PEN observations for the validation of satellite-derived vegetation products. Recent advances in satellite observations, including higher temporal frequency, improved spatial resolution, and the increasing availability of multiple satellite platforms, are rapidly changing the framework of satellite data validation. Based on these developments, we discuss future directions for field observations targeting vegetation monitoring, including improved observation strategies and data-sharing frameworks for satellite validation studies.

Back to session list

S1-6: Lunar calibration using a hyperspectral model, SP model

Presenting author: Toru KOUYAMA

Authors: Toru KOUYAMA, AIST; Hirokazu YAMAMOTO, Hiroshima Univ.

Keywords: Lunar calibration, Radiometric Calibration, SELENE, SP

Because of recent rapid increasing of the number of satellites, including hyperspectral sensors, radiometric calibration is more important for ensuring consistency in observed radiance across different sensors. Such consistency is essential for users to perform reliable data analysis. From this perspective, accurate and efficient calibration methods are in high demand. Lunar calibration is one such radiometric calibration technique for optical sensors in space. A key advantage of lunar calibration is that, if the satellite’s configuration and operational constraints permit observation of the Moon, lunar calibration can be performed simply by observing the Moon through comparing the observed lunar brightness with the expected one predicted by a lunar surface reflectance model. A hyperspectral lunar surface reflectance model, known as the SP model, was developed based on hyperspectral observations acquired by JAXA’s lunar orbiter SELENE. This model has been applied to various sensors, not only those for Earth observation but also those supporting space exploration missions, such as asteroid exploration. Because the SP model can generate spatially resolved lunar disk images, lunar calibration using this model can be easily applied even to high resolution sensors that capture only a portion of the Moon within a single frame. In this presentation, we introduce the basic characteristics of the SP model, describe lunar calibration techniques based on the model, and review missions in which the model has been applied. We then discuss how the SP model can contribute to future satellite missions.

Back to session list

Session A6 Abstracts

A6-1: Assessing the Role of Dynamic Opening in Sea Ice Loss Using Passive Microwave Radiometer Observations

Presenting author: Jeong-Won PARK

Authors: Jeong-Won PARK

Keywords: Sea Ice; Polar; AMSR2; Ice Deformation; Ice Fragmentation

The decline of sea ice is driven not only by thermodynamic melting but also significantly by dynamic processes. These two mechanisms operate in a positive feedback loop: thinner and smaller ice becomes increasingly susceptible to drift and fracturing, which in turn accelerates further melting. While thermodynamic melting can be estimated using low-resolution remote sensing data—by leveraging atmospheric and ice surface temperatures—quantifying dynamic processes remains challenging with 25 km-scale passive microwave data. This limitation arises from the insufficient resolution to capture detailed sea ice deformation, such as the specific geometry of leads or ice floes. This study aims to derive dynamic insights by indirectly calculating ice deformation features based on the divergence and convergence of ice motion fields, as well as the degree of fragmentation from AMSR2 channel combinations. Using this approach, we retrieve the degree of the sea ice opening, validate the results against high-resolution data, and analyze the correlation between these dynamic features and sea ice decline rates.

Back to session list

A6-2: Long-Term Regional Variability of Arctic Ocean Color under Changing Sea Ice Conditions

Presenting author: Joo-Eun YOON

Authors: Joo-Eun YOON, KOPRI; Jinku PARK, KOPRI; Jeong-Won PARK KOPRI

Keywords: Arctic Ocean; Ocean Color; Sea Ice; Climate change

Ocean color observations provide valuable information on marine ecosystems by reflecting variability in optically active constituents such as phytoplankton and dissolved organic matter. In the Arctic Ocean, rapid warming and ongoing sea ice decline are altering upper-ocean light conditions and seasonal exposure, potentially affecting ocean color and biological processes. However, long-term Arctic-wide assessments of these changes remain limited. Here, we investigate spatial and temporal variations in Arctic ocean color and their relationship with sea ice changes using satellite-based observations. We analyze ESA Ocean Colour Climate Change Initiative (OC-CCI) data from 1998 to 2024, focusing on chlorophyll-a concentration, hue angle, and remote sensing reflectance (Rrs), together with satellite-derived sea ice concentration. Trend analyses reveal regionally distinct patterns of increasing and decreasing ocean color signals across the Arctic, broadly corresponding to variations in sea ice conditions. These results indicate that Arctic ocean color changes are spatially heterogeneous and influenced by sea ice variability. Our study highlights the importance of integrating sea ice information with multiple ocean color parameters to improve satellite-based monitoring of Arctic marine ecosystem responses to climate change.

Back to session list

A6-3: Analysis of long-term grounding line positional variation on Shirase Glacier using SAR interferometry

Presenting author: Tsutomu YAMANOKUCHI

Authors: Tsutomu YAMANOKUCHI; Chiyuki NARAMA; Kazuki NAKAMURA; Shin SUGIYAMA, Hokkaido Univ.; Shigeru AOKI, Hokkaido Univ.; Whyjay ZHENG, National Central Univ.

Keywords: InSAR; grounding line; glacier; ALOS-4; ALOS-2; CSK; ERS-1

The grounding line is the transition zone where an ice sheet flowing over bedrock enters the ocean and becomes a floating ice shelf or ice tongue supported by seawater buoyancy. The position and geometry of the grounding line are crucial for improving our understanding of the Antarctic ice mass balance. However, owing to the vast spatial extent of Antarctica, grounding-line mapping based solely on field observations is difficult, and satellite remote sensing is therefore widely used. Across the grounding line, ice shelves and ice tongues undergo vertical motion driven by tidal variations. This vertical displacement can be measured as changes in slant-range distance in the radar line-of-sight using Interferometric Synthetic Aperture Radar (InSAR). In this study, we conduct a long-term time-series analysis of the grounding line using InSAR to assess the presence or absence of significant changes. Previous studies have demonstrated substantial inland retreat of grounding lines in West Antarctica. Motivated by these findings, we focus on Shirase Glacier, located in Lützow-Holm Bay, East Antarctica, near Syowa Station, where extensive field observation data are available. Shirase Glacier is one of the fastest-flowing glaciers in Antarctica and plays an important role in understanding the regional mass balance. However, due to its high flow velocity, temporal coherence at the grounding line is easily lost, making InSAR-based analysis challenging. To date, successful InSAR analyses have been limited to ERS-1/2 tandem pairs from 1995–1996 and COSMO-SkyMed (CSK) tandem pairs around 2020, and the grounding-line position has not been accurately measured in recent years. In this study, we perform InSAR analysis of Shirase Glacier using ALOS-2/4 data pairs with a 6-day temporal separation. By extending the analysis period back approximately 30 years, we aim to capture longer-term changes in grounding-line position and geometry.

Back to session list

A6-4: Ice flow velocity in downstream region of Shirase Glacier related to ice thickness change derived from remotely sensed data

Presenting author: Yurina SATAKE

Authors: Kazuki NAKAMURA, Nihon Univ.; Yurina SATAKE, Nihon Univ.; Tsutomu YAMANOKUCHI, RESTEC; Shigeru AOKI, Hokkaido Univ.

Keywords: Glacier; Antarctica, ALOS2; PALSAR-2; CryoSat-2; SIRAL

Shirase Glacier is known for its high flow velocity compared to other Antarctic glaciers, which makes monitoring its significant mass loss crucial. Since 2015, the large-scale breakup of landfast ice has been observed alongside changes in flow velocity, suggesting that landfast ice plays a role in suppressing the direct flow of the glacier. This study therefore investigates the relationship between the flow velocity and the ice thickness of Shirase Glacier, which are directly related to its dynamics. The focus was on the downstream area of Shirase Glacier (seaward from the grounding line: GL). The flow velocity was determined using the offset tracking method applied to ALOS-2/PALSAR-2 HH polarization data. The ice thickness was derived by converting surface elevation from the synthetic aperture radar interferometric (SIN) mode data of CryoSat-2/SIRAL, under the assumption of hydrostatic equilibrium. The analysis period was from 2016 to 2024, which allowed for a combined analysis of the ALOS-2/PALSAR-2 and CryoSat-2/SIRAL data. We examined the annual variations in flow velocity and ice thickness at the floating ice tongue of Shirase Glacier from 15 km downstream of the GL (hereinafter referred to as 15 km DGL) to 60 km DGL from 2016 to 2024. The results showed an increase in flow velocity and a decrease in ice thickness towards the terminus. Additionally, examining the relationship between ice thickness and flow velocity in the downstream region revealed a negative correlation: ice thinned as flow accelerated. Significant differences in ice thickness changes were also observed upstream and downstream of the 35 km DGL boundary. We discuss the potential causes of these differences.

Back to session list

A6-5: Permafrost monitoring in eastern Siberia and Alaska: Research findings using ALOS-1/-2 SAR data and preliminary results from ALOS-4

Presenting author: Takahiro ABE

Authors: Takahiro ABE, Kyoto Univ.; Go IWAHANA, Univ. of Alaska Fairbanks; Yoshihiro IIJIMA, Tokyo Metropolitan Univ.

Keywords: Permafrost, SAR, InSAR, ALOS, ALOS-2, ALOS-4

Permafrost is defined as ground that remains at or below 0 ℃ for at least two consecutive years, and is a key element of the cryosphere. Recent climate change has caused permafrost thawing, significantly impacting their surroundings. Satellite remote sensing techniques allow us to observe this phenomenon on a large scale and monitor surface conditions in harsh environments like the Arctic, where permafrost covers approximately one-quarter of the land area. Synthetic Aperture Radar (SAR) is a microwave remote-sensing technique with the advantage of being able to obtain data regardless of day or night and without any additional equipment. Interferometric SAR is a method to calculate surface displacement using phase information in SAR data. Interferometric coherence is greatly important to detect subtle surface displacement with high accuracy, and L-band SAR is most suitable for monitoring ground surface in the cryosphere in the view of temporal decorrelation. While permafrost thawing occurs in the ground and we cannot observe it directly, monitoring surface displacement in permafrost regions can evaluate its dynamics under current climate change. Since the 2000s, the Advanced Land Observing Satellite (ALOS) and the successor ALOS-2 operated by Japan Aerospace Exploration Agency (JAXA) have carried L-band SAR sensors, which have significantly contributed to clarifying permafrost dynamics. This presentation will introduce recent research findings elucidated through the ALOS series SAR data, and the preliminary results from ALOS-4.

Back to session list

Session B6 Abstracts

B6-1: Tree Species Classification of Major Korean Forest Species Using Drone-borne LiDAR Point Cloud Data and Deep Learning Models

Presenting author: Jin-Taek Kang

Authors: Jin-Taek Kang, National Institute of Forest Science; Jeong-Mook Park, National Institute of Forest Science; Yeon-Ok Seo, National Institute of Forest Science; Chae Jun, National Institute of Forest Science; Lim Chanyoung Park, National Institute of Forest Science; Seung yong Lee, National Institute of Forest Science

Keywords: Drone-borne LiDAR; Tree Species Classification; Deep Learning; PointNet++; Point Cloud Data

This study was conducted to investigate the feasibility of classifying major forest species in Korea and estimating carbon storage at the individual tree level using drone-borne LiDAR. The research focused on four coniferous species (Pinus koraiensis, Larix kaempferi, Cryptomeria japonica, Chamaecyparis obtusa) and three deciduous species (Quercus variabilis, Quercus mongolica, Quercus acutissima). A total of 1,025 individual tree samples were collected from five representative artificial plantation areas across South Korea using high-resolution drone-borne LiDAR systems. To develop an optimized classification framework, a structured training dataset was established with a split of 90% for training/validation and 10% for evaluation. The classification accuracy of four different deep learning architectures was evaluated to identify the most effective AI model for processing 3D point cloud data. Based on the evaluation, the PointNet++ model was selected as the superior architecture, demonstrating an optimal balance between accuracy and computational efficiency with a test accuracy of 92.5% and a processing time of 2.3 seconds. The detailed classification results using the optimized PointNet++ model yielded an average accuracy of 94.8% for coniferous species. Specifically, Chamaecyparis obtusa and Cryptomeria japonica achieved 100%, followed by Pinus koraiensis at 92% and Larix kaempferi at 87%. For deciduous species, the average accuracy was 86.7%, with Quercus variabilis reaching 100%, Quercus acutissima at 83%, and Quercus mongolica at 77%. The results suggest that while the proposed method is highly effective, future improvements in accuracy require the expansion and construction of additional training datasets to better account for the atypical and complex structures of deciduous species. Acknowledgments; This study was supported by the National Institute of Forest Science (NIFoS) as part of the research project “Research on the Development of Technology for Utilizing Digital Forest Resource Information Based on LiDAR” (Project No. FM0101-2024-01-2025).

Back to session list

B6-2: Indoor Disaster Investigation using Multi-sensors Built-in Ground Robot Platform and AI Tech.

Presenting author: Seongsam KIM

Authors: Seongsam KIM; Yonghan JUNG; Seul KOO; Jungwook PARk

Keywords: Disaster Investigation, SLAM, Indoor Mapping, Artificial Intelligence, Disaster Information

Structural collapse accidents in densely populated or enclosed indoor environments have recently become a significant global concern. High-rise buildings, underground facilities, and large construction sites present highly complex conditions that limit the ability of human personnel to safely access disaster scenes. Investigators and first responders are often exposed to secondary collapse hazards, unstable debris, and restricted mobility, making conventional manual response approaches increasingly unsuitable. As a result, there is a growing need for robotic platforms capable of entering hazardous spaces, acquiring situational data, and assisting in the early stages of disaster management. Advances in robotics and artificial intelligence—accelerated by the broader context of the Fourth Industrial Revolution—are transforming numerous industrial sectors, including construction, geospatial engineering, logistics, and public safety. In the disaster management domain, AI-assisted search-and-rescue (SAR) robotics and autonomous mapping technologies represent core developments that support rapid situational awareness in unstable environments. The convergence of LiDAR, optical cameras, inertial sensors, and thermal imaging devices has led to the commercial availability of integrated sensor modules, enabling robots to perform Simultaneous Localization and Mapping (SLAM) in both indoor and outdoor settings. In parallel, deep-learning models now achieve high accuracy in detecting structural defects such as cracks, spalling, debris accumulation, and exposed rebar. Particularly, LiDAR-camera fusion, visual-inertial navigation, and dense SLAM frameworks have emerged as alternatives to georeferencing in GNSS-denied environments. These technologies enable the construction of geospatially coherent 3D representations that support rapid situational assessment, structural damage extraction, and spatial decision-making. This study develops an indoor disaster investigation methodology centered on a multi-sensor ground robot integrating LiDAR, optical imaging, and inertial navigation. The proposed Vision–LiDAR SLAM framework enables precise localization and dense 3D reconstruction, while AI-driven damage detection provides automated semantic enrichment of spatial datasets. The aim is to establish an end-to-end workflow for data acquisition, 3D mapping, damage extraction, and GIS-based spatial reasoning in indoor disaster environments.

Back to session list

B6-3: Vicarious Calibration of Terra ASTER VNIR using RadCalNet RVUS BOA Reflectance Validated by JAXA GOSAT Campaigns

Presenting author: Hirokazu YAMAMOTO

Authors: Hirokazu YAMAMOTO, Hiroshima Univ.; Atsushi YASUDA, RESTEC; Rio KAJIURA, JAXA; Kei SHIOMI, JAXA; Satoshi TSUCHIDA, AIST

Keywords: Vicarious Calibration; Terra ASTER; RadCalNet RVUS; Radiometric Degradation; Railroad Valley; GOSAT vicarious calibration

The Terra ASTER instrument has been operating for over two decades, and radiometric calibration is very important for the consistency of long-term Earth observation records. While the Radiometric Calibration Network (RadCalNet) RVUS site provides TOA reflectance, ensuring the continuity of ASTER’s degradation curve requires a historical vicarious calibration methodology. Therefore, we used bottom-of-atmosphere (BOA) reflectance at RadCalNet RVUS. The legacy AIST vicarious calibration site for ASTER VNIR bands is located 340 m north of RadCalNet RVUS area. This research proposes a hybrid vicarious calibration approach applied to the 2023–2025 period. First, we validate the BOA reflectance provided by RadCalNet RVUS using in-situ ground measurement data collected during JAXA’s GOSAT vicarious calibration campaigns. Second, instead of using the TOA reflectance provided by RadCalNet, we independently calculate TOA radiance using the validated RadCalNet RVUS BOA reflectance combined with the conventional radiative transfer calculation methods. This approach ensures methodological consistency with the historical ASTER calibration data accumulated over the past two decades. The calculated ASTER TOA radiance by radiative transfer code is then compared with actual ASTER observed radiance derived from the ortho-rectified products downloaded via the AIST MADAS system. Finally, we evaluate the affinity of the RadCalNet-based results with the historical trend to determine if this hybrid method can serve as a reliable successor to traditional field campaigns for monitoring ASTER’s radiometric performance in its late operational phase.

Back to session list

B6-4: Rapid 3D Terrain Reconstruction from a Single Satellite Image

Presenting author: Ping-Yuan Yang

Authors: Ping-Yuan Yang, National Chung Hsing Univ.; Ming-Der Yang, National Chung Hsing Univ.

Keywords: WorldView-3, Depth anything, 3D Gaussian Splatting, DEM, Generative AI

The rapid acquisition of 3D Digital Elevation Models (DEMs) in disaster-stricken areas is crucial for informing emergency response decisions. However, traditional remote sensing techniques, such as Interferometric Synthetic Aperture Radar (InSAR) and Multi-View Stereopsis (MVS), are sometimes constrained by satellite revisit cycles, cloud cover, or the need for multi-temporal imagery, posing challenges for immediate mapping. To mitigate the limitations of observational data, this study proposes and evaluates a 3D terrain reconstruction framework based on a single very high-resolution (VHR) optical satellite image (using WorldView-3, with a panchromatic resolution of 0.3m, as an example). In contrast to conventional methods requiring multi-view imagery, this research explores the feasibility of monocular spatial feature extraction and elevation transformation. The study first employs the Depth Anything v2 model for relative depth estimation and, supported by photogrammetric geometry, attempts to utilize solar elevation angles and object shadow lengths for geometric calibration, converting pixel features into a 3D Gaussian Splatting model with physical spatial scales. To address the inherent object occlusion issues in single-view scenarios, this study also preliminarily incorporates generative AI for view synthesis, aiming to improve the geometric and textural details of occluded areas. To investigate the applicability of this framework, it was tested across three high-resolution satellite imagery scenes with varying topographic features. Quantitative evaluation results indicate that the geometrically calibrated 3D models achieved Mean Absolute Errors (MAE) of 4.15m, 1.61m, and 1.26m, and Root Mean Square Errors (RMSE) of 5.79m, 1.92m, and 1.44m, respectively, when compared to the ground truth. Preliminary experimental results suggest that combining a single satellite image with auxiliary geometric information has the potential to construct 3D topographic mapping models in specific scenarios. The findings of this study are expected to provide a flexible technical reference and auxiliary approach for future remote sensing applications in disaster prevention and emergency response.

Back to session list

Session C6 Abstracts

C6-1: InSAR-Based Multi-Decadal Investigation of Urban Land Subsidence: Divergent Policy Influences in Jakarta and Bangkok

Presenting author: Wahyu Luqmanul HAKIM

Authors: Wahyu Luqmanul HAKIM, Kangwon National University; Muhammad Fulki FADHILLAH, Kangwon National University; Sungjae PARK, Kangwon National University; Chang-Wook LEE, Kangwon National University

Keywords: Land Subsidence, InSAR, Jakarta, Bangkok, JERS-1, ALOS-1, Sentinel-1.

Urban land subsidence is a major geohazard in low‑lying coastal megacities, where unsustainable groundwater extraction accelerates ground deformation. This study provides a multi‑decadal comparison of subsidence patterns in two highly vulnerable Southeast Asian cities, Jakarta in Indonesia and Bangkok in Thailand, using a time‑series Interferometric Synthetic Aperture Radar (InSAR) approach. Data from three satellite missions, namely JERS‑1 (1992–1998), ALOS‑1 (2006–2011), and Sentinel‑1 (2014–2023), were integrated to develop a long‑term record of surface deformation and to examine how local policies influence subsidence trends. The results show a clear contrast in the evolution of subsidence in both cities. In Bangkok, the enforcement of the Groundwater Act B.E. 2520 (1977) in 1994 strengthened the regulation of groundwater licensing, usage, and recharge activities. As a result, industrial water extraction declined, and the reliance on surface water increased. Subsidence rates, which reached 2 to 3 cm per year during the JERS‑1 period, were followed by aquifer recovery with uplift reaching up to 3 cm per year between 2007 and 2010 in the ALOS‑1 period. Sentinel‑1 observations later indicated more stable conditions, with subsidence generally remaining within 1 to 2 cm per year in coastal areas. Jakarta presents a different situation because subsidence has continued without clear improvement throughout the observation period. Weak policy enforcement and widespread unregulated groundwater pumping, especially in northern districts, have resulted in deformation exceeding 10 cm per year in several locations. This ongoing trend increases flood risk and threatens critical infrastructure. The contrasting outcomes in Jakarta and Bangkok highlight the essential role of groundwater governance in reducing long‑term subsidence. The findings also emphasize the value of InSAR monitoring for supporting urban resilience and future policy development.

Back to session list

C6-2: Surface Deformation Before, During, and After the 2025 Krasheninnikov Eruption, Kamchatka, Russia, from Sentinel‑1 Time‑Series InSAR

Presenting author: Wahyu Luqmanul Hakim

Authors: Wahyu Luqmanul HAKIM, Kangwon National University; Chang-Wook LEE, Kangwon National University

Keywords: Volcanic Eruption, InSAR, ICOPS, Krasheninnikov Volcano, Kamchatka Peninsula, Russia

The Krasheninnikov volcano, located on the Kamchatka Peninsula in Russia’s Far East, reawakened in 2025 after centuries of quiescence, marking its first confirmed eruption in approximately 475–600 years. The eruption began on 3 August 2025, producing sustained ash plumes rising 5–6 km above the crater and later reaching heights up to 8.5 km, prompting elevated aviation alerts as ash dispersed eastward over the Pacific Ocean. NASA and KVERT reported that the activity followed a powerful 8.8‑magnitude earthquake on 29 July 2025, with interferometric analyses showing measurable ground deformation preceding the eruption, likely indicating magma ascent through a newly formed dike system. This study investigates the before, during, and after eruptive deformation of the Krasheninnikov Volcano system from 1 January 2024 to 31 December 2025 using Sentinel‑1 time‑series InSAR data from both ascending and descending orbits. The use of dual viewing geometries enables improved discrimination between vertical and east–west ground motion components, overcoming the one‑dimensional limitations of single-LOS measurements. The methodology applies multitemporal interferometric processing based on the improved combined scatterers interferometry with optimized point scatterers (ICOPS) to detect deformation trends, interpret pre‑eruptive inflation, quantify co‑eruptive displacement related to magma intrusion and ash venting, and characterize the post‑eruptive relaxation stage. By merging independent line‑of‑sight observations, this analysis aims to reconstruct deformation patterns associated with the earthquake‑triggered reactivation of the volcano and assess the broader volcano–tectonic interactions in the region. The results provide new insights into the dynamics of this rare eruption and demonstrate the capability of InSAR for monitoring remote volcanic systems with limited ground instrumentation.

Back to session list

C6-3: Nationwide near-surface ground deformation mosaics from multi-year synthetic aperture radar observations across Canada and North America

Presenting author: Sergey SAMSONOV

Authors: Sergey SAMSONOV, Canada Centre for Mapping and Earth Observation, Natural Resources Canada

Keywords: Sentinel-1; InSAR; MSBAS; ground deformation; nationwide mosaic

Over nearly a decade, Sentinel-1 has continuously acquired Synthetic Aperture Radar (SAR) data over North America, enabling dense multi-temporal image stacks exceeding 100 acquisitions per frame across most land areas. We processed more than 100,000 images to generate deformation time series and mean velocity maps using a fully automated InSAR workflow deployed on a high-performance computing platform. The system builds on operational processing originally developed for RADARSAT-2 and the RADARSAT Constellation Mission, and was expanded to exploit Sentinel-1’s broader spatial and temporal coverage. This presentation focuses on national-scale deformation mosaics across North America, with particular emphasis on Canada. In Canada, processing was restricted to snow-free seasons to reduce decorrelation and seasonal artifacts, especially in northern regions affected by strong seasonal signals and post-glacial rebound, whereas in southern regions year-round data were processed, reflecting milder climatic conditions and enabling continuous monitoring of deformation related to groundwater extraction, energy development, mining, and landslides. To address challenges associated with incidence-angle variability and long-wavelength deformation, high-pass spatial filtering was applied prior to time-series analysis using an in-house Multidimensional Small Baseline Subset (MSBAS) implementation optimized for very large datasets. A dedicated mosaicking algorithm was developed to seamlessly merge frame-based products into consistent national-scale deformation maps, and combined viewing geometries were used to derive approximate vertical and east–west horizontal motion components. The resulting pilot nationwide mosaics reveal coherent, localized deformation signals superimposed on broader tectonic and glacial isostatic trends, demonstrating the feasibility of operational, continent-scale InSAR monitoring and providing a scalable framework for geohazard detection, infrastructure risk assessment, and continuous ground-motion surveillance across Canada and other parts of North America.

Back to session list

C6-4: Landslide Detection via a U-net-Driven SAR2OPT Approach: Application to the 2024 Hualien Earthquake

Presenting author: Wen-Hong CHEN

Authors: Wen-Hong CHEN; Shou-Hao CHIANG

Keywords: SAR2OPT , Landslide Detection , Synthetic Aperture Radar , U-net , Hualien Earthquake

The devastating April 2024 Hualien earthquake in eastern Taiwan triggered extensive landslides, emphasizing the critical demand for swift and reliable disaster monitoring. While Synthetic Aperture Radar (SAR) guarantees observation regardless of weather conditions, the absence of multispectral features complicates direct landslide identification. To address this challenge, we developed a modified U-net framework designed to translate dual-orbit Sentinel-1 radar signals (VV/VH polarizations) into high-fidelity NDVI maps. Integrating ascending and descending orbits effectively reduces geometric distortions such as layover and shadowing in mountainous terrains. Through an end-to-end training process utilizing overlapping sliding windows, the model accurately reconstructed vegetation patterns directly from SAR inputs. The proposed framework delivered robust structural fidelity and spatial detail, validating that SAR-derived optical proxies can serve as an effective and timely alternative for landslide mapping when cloud-free optical imagery is unavailable.

Back to session list

C6-5: Susceptibility Mapping of Residual Landslide using Multitemporal Imagery and Machine-Learning Methods

Presenting author: Tung Cheng, LU

Authors: Shou Hao, CHIANG, Professor, Center for Space and Remote Sensing Research, National Central University, Taiwan; Tung Cheng, LU, Research Assistant, Center for Space and Remote Sensing Research, National Central University, Taiwan

Keywords: Residual Landslides, Machine-Learning, Convolutional Neural Network, Transformer, Sediment disaster

In Taiwan, the term “residual landslides” describes slopes where massive amounts of landslide-generated failure materials accumulate to form unstable colluvial deposits. These unstable formations pose a major obstacle to effective watershed management and the mitigation of sediment-related disasters. During periods of heavy rainfall, the accumulated materials frequently act as the main sources of debris flows, thereby threatening downstream regions with severe hazards. Moreover, the risks of these sediment disasters are significantly influenced by the inaccessible mountainous landscapes and the escalating extreme rainfall caused by climate change. To address these challenges, this study proposes a machine learning framework for assessing residual landslide hazards. The research focused on the Kaoping River basin in southern Taiwan, using sub-watersheds as the analytical units. The approach involves the development of a three-level predictive model evaluating: (1) landslide versus non-landslide, (2) residual landslide versus non-residual landslide, and (3) residual landslide activity levels. A Convolutional Neural Network (CNN) was utilized to extract features from yearly SENTINEL-2 data spanning from 2019 to 2024. Subsequently, a Transformer model processed these features alongside dynamic data, including rainfall and historical landslides. Model performance was systematically quantified using accuracy, precision, and recall metrics. The model achieved accuracies of 0.96, 0.93, and 0.93 across the three analysis levels. In classifying low, medium, and high activity levels, precision scores were 0.94, 0.74, and 1.0, with recall consistently reaching 1.0 across activity levels. Ultimately, this methodology enables systematic inventory mapping and activity assessment across mountainous regions, thereby enhancing early warning capabilities and informing more effective sediment disaster prevention and management strategies.

Back to session list

Session D6 Abstracts

D6-1: Optical Image Reconstruction in Cloud-Contaminated Areas Using Multi-Temporal SAR Data and LSTM Models

Presenting author: Yu-Ching WANG

Authors: Yu-Ching WANG,National Chung Hsing Univ.; Ming-Der YANG,National Chung Hsing Univ.

Keywords: Radar-Optical Fusion; LSTM; Image Reconstruction; Spectral Fidelity

Agricultural remote sensing in Taiwan is persistently hindered by high cloud cover during the plum rain season, leading to significant gaps in optical satellite time-series. To overcome this limitation, this study utilizes the cloud-penetrating capability of SAR combined with deep learning models to establish a reconstruction framework tailored to the Taiwanese environment. Using a self-collected multi-temporal dataset of local rice paddies (Sentinel-1/2), we implemented a Long Short-Term Memory (LSTM) network for spectral value prediction. The primary contributions of this work include: (1) Validating the effectiveness of various radar feature combinations (the 5-feature set including VV, VH, and RVI) and spatial dimensions (1D point-based vs. 2D spatial-temporal architectures) within local agricultural landscapes ; and (2) Integrating a Spectral Angle Mapper (SAM) loss function to refine the proportional relationships between bands and enhance spectral fidelity. Experimental results demonstrate that the “2D architecture + 5 derived features + 4-band simultaneous output” configuration achieved the best performance, with an RMSE of 0.022 and an SSIM of 0.821. The inclusion of SAM Loss reduced spectral angle error by 11.37%, significantly improving the physical consistency of the reconstructed images. This study establishes a highly reliable reconstruction framework, effectively bridging data gaps caused by cloud cover. Future work will utilize these continuous time-series images to identify rice growth stages and extract phenological characteristics, enabling precise smart agriculture monitoring throughout the entire growing cycle.

Back to session list

D6-2: An Efficient CNN-Transformer Network for Multitask Learning-Based Semantic Change Detection

Presenting author: Aisha JAVED

Authors: Aisha JAVED, Research Center for Civil Engineering Future Talent Seoul National University of Science and Technology Seoul, South Korea; Youkyung HAN, Department of Civil Engineering Seoul National University of Science and Technology Seoul, South Korea

Keywords: Semantic change detection; semantic segmentation; multitask learning; lightweight network; remote sensing images

Semantic change detection (SCD) is a critical task in remote sensing, enabling detailed analysis of land-cover changes by jointly identifying change locations and their semantic categories. Existing convolutional neural network (CNN)-based multitask approaches effectively capture local spatial details but often suffer from fragmented predictions due to limited long-range context modeling. Transformer-based methods alleviate this issue by exploiting global dependencies, yet frequently compromise spatial precision and incur high computational cost. To address these challenges, we propose a lightweight hybrid CNN-Transformer architecture for multitask learning-based SCD. The proposed network employs a shared-weight Siamese CNN encoder to extract multiscale bi-temporal features efficiently, followed by a multiscale change-aware feature modeling module that explicitly enhances temporally consistent representations prior to binary change detection decoding. Transformer-based decoders are adopted for bi-temporal semantic segmentation and change detection, enabling effective integration of global context and local details. Experiments conducted on a representative high-resolution SCD benchmark demonstrate that the proposed method achieves superior semantic consistency in changed regions compared to exiting approaches, while maintaining a compact model size and low computational complexity. These results indicate that the proposed method provides an effective and efficient solution for SCD in resource-constrained remote sensing applications.

Back to session list

D6-3: Evaluating GEO–LEO BRDF consistency using Himawari AHI and SNPP VIIRS data

Presenting author: Masayuki MATSUOKA

Authors: Masayuki MATSUOKA, Mie Univ.; Hiroki YOSHIOKA, Aichi Prefectural Univ.; Wei YANG, Chiba Univ.; Kazuhito ICHII, Chiba Univ.

Keywords: land surface reflectance; bidirectional reflectance distribution function; Himawari AHI; SNPP VIIRS; GEO-LEO data fusion

The integration of geostationary Earth orbit (GEO) and low Earth orbit (LEO) satellite observations offers complementary temporal and spatial information for land surface monitoring. Bidirectional reflectance distribution functions (BRDFs) describe surface reflectance under varying viewing and illumination geometries and are therefore fundamental to accurate surface characterisation. An ideal BRDF model should be applicable to both GEO and LEO data using a single, consistent parameter set. However, our previous studies demonstrated that applying BRDF parameters derived from LEO observations to GEO data resulted in reflectance values that deviated substantially from actual measurements. This study evaluates the consistency of BRDF models and their parameters using data from the Himawari-8/9 Advanced Himawari Imager (AHI) and the Suomi NPP Visible Infrared Imaging Radiometer Suite (VIIRS). The AHI Japan Area dataset was atmospherically corrected and orthorectified to produce daytime reflectance time series. For VIIRS, 16 consecutive daily surface reflectance products provided by NASA were used. Kernel-driven BRDF parameters were estimated from AHI data alone, VIIRS data alone, and a combined dataset. Using these parameter sets, pseudo BRDF-adjusted reflectances were generated for both sensors. The consistency of the adjusted reflectances was assessed with respect to seasonal variation, topography, and land cover. The kernel-driven BRDF model employed in the VIIRS product showed relatively larger discrepancies when applied to AHI observations, particularly under hot-spot conditions in spring and autumn. In mountainous regions, the estimated BRDF parameters exhibited clear terrain dependence, with greater reflectance contrasts between east- and west-facing slopes. These results indicate the necessity of further refinement of BRDF models to achieve more robust GEO–LEO integration in Earth observation.

Back to session list

D6-4: A Post-Processing Framework for Occlusion-Robust Multi-Object Tracking in Satellite Video Guided by Kalman Filtering and Single-Object Tracking

Presenting author: Po-Wei TSAI

Authors: Po-Wei TSAI, National Yang Ming Chiao Tung University; Tzu-Yi CHUANG, National Yang Ming Chiao Tung University

Keywords: Satellite video; Multi-object tracking; Kalman filter; Single object tracking; Long-term occlusion

Video satellites can continuously capture large areas, providing dynamic monitoring information across extensive regions and multiple targets. As a result, the intelligent processing and analysis of satellite video imagery have become an important research topic in remote sensing. However, most existing multi-object tracking (MOT) methods for satellite video adopt the “tracking-by-detection” paradigm, which relies heavily on the temporal continuity of detection results. When targets are obscured by clouds, shadows, or buildings for extended periods, tracking interruptions are common, leading to fragmented trajectories and frequent identity switches. To mitigate this limitation, we introduce a post-processing framework integrating Kalman filtering and a single-object tracking (SOT) algorithm. The Kalman filter is employed to predict target motion states during occlusion periods, enabling the association and merging of fragmented trajectories corresponding to the same target. In addition, the SOT algorithm is utilized to complete trajectories across gaps caused by missed detections, thereby effectively improving trajectory continuity. To validate our approach, we also constructed a multi-object tracking dataset using SkySat satellite video imagery to supplement existing public resources. Experimental evaluations across five distinct regions in three videos indicate that, compared with the original tracking results, the proposed framework achieves average performance improvements of 4.77% in HOTA, 4.53% in MOTA, and 9.24% in IDF1. These results demonstrate the effectiveness of the proposed post-processing approach in enhancing tracking robustness under long-term occlusion conditions.

Back to session list

D6-5: Integration of UAV-Derived DSM and Visible Imagery for Terrain and Surface Feature Classification in Open-Pit Mining Areas

Presenting author: Wei-Tsun LIN

Authors: Wei-Tsun LIN, GIS Research Center, Feng Chia University; Chih-Ling WU, GIS Research Center, Feng Chia University; Chih-Chan CHIANG, GIS Research Center, Feng Chia University; Mei-Ling YEH, GIS Research Center, Feng Chia University; Chen-Yu HAO, GIS Research Center, Feng Chia University; Nai-Yun CHENG, Geological Survey and Mining Management Agency, Ministry of Economic Affairs; Pao-Chi LIU, Geological Survey and Mining Management Agency, Ministry of Economic Affairs; Hsin-Hung SHIH, Geological Survey and Mining Management Agency, Ministry of Economic Affairs

Keywords: Topographic indices; Vegetation indices; Object-based image analysis; Drone; Surface morphology

Mineral resources are essential for economic development, yet mining activities must carefully balance production efficiency, environmental sustainability, and operational safety. Effective monitoring of mining conditions and development progress requires accurate interpretation of terrain morphology and surface characteristics. However, open-pit mining areas are often located in remote and topographically complex regions, making comprehensive field investigations both time-consuming and challenging. This study explores the integration of UAV-derived digital surface models (DSM) and high-resolution visible orthomosaic imagery for terrain and surface feature classification in open-pit mining areas in eastern Taiwan. The study area focuses on non-metallic mineral mines. UAV imagery was processed to generate DSMs and orthomosaic images, from which topographic indices, spectral features, and textural information were extracted. Additionally, vegetation indices derived from visible bands were incorporated to evaluate their contribution to classification accuracy. The classification framework was designed to delineate spatially coherent surface objects and extract object-level features by integrating topographic and spectral information. Major classification categories included slopes, working platforms, bare ground, vegetation, and human-made features associated with mining activities. Several mining sites were used for method development, while independent sites were reserved for validation to assess the framework’s robustness under varying mining conditions. Classification results were validated against manually digitized reference data. Preliminary findings indicate that the integration of topographic attributes with visible spectral information significantly enhances classification performance compared to single-source approaches. This study demonstrates the potential of UAV-based multi-source data integration for scalable and reliable monitoring of open-pit mining environments.

Back to session list

Session E6 Abstracts

E6-1: Three Dimensional PM2.5 Characterization in Chiang Mai using Pixel Aligned Himawari 9 AOD Data

Presenting author: Katanchalee TAWEEPORNWATTANAKUL

Authors: Katanchalee TAWEEPORNWATTANAKUL; Wataru TAKEUCHI; Khin MYAT KYAW; Wilawan ROBROO

Keywords: PM2.5 Biomass burning; Vertical gradients; Sub-pixel variability; Himawari-9 AOD; Complex terrain

Severe PM₂.₅ pollution occurs annually in Chiang Mai, Thailand, during the peak dry-season biomass burning period, when forest fires in surrounding mountainous areas interact with urban traffic emissions. This study reports an intensive field measurement campaign conducted in March 2026 to evaluate how horizontal heterogeneity and terrain-induced vertical gradients within a satellite pixel influence the representativeness of Himawari-9 Level-2 aerosol optical depth (AOD) retrievals under high-pollution conditions. Ground measurements were collected using calibrated portable PM₂.₅ sensors at predefined Himawari-9 satellite pixel locations corresponding to the 5-km spatial resolution of the Himawari-9 aerosol product across urban traffic corridors and mountainous regions influenced by forest fire activity. Each ground sampling point was designed to represent a distinct satellite pixel to enable direct pixel-to-ground comparison. Measurements were repeated in the morning, afternoon, and evening over three consecutive days and temporally matched to the native 10-minute Himawari-9 Level-2 AOD observations to capture diurnal variability. A coordinated city-to-mountain transect design was implemented to characterize vertical gradients between the urban basin and elevated forested areas that are not directly observable from two-dimensional satellite imagery. Urban sites were selected using traffic intensity and emission inventory data to assess local vehicular contributions, while mountain sites were chosen to represent biomass-burning influence from forested highlands. Himawari-9 AOD retrievals were integrated with pixel-aligned ground measurements to examine how heterogeneous surface emissions and vertical atmospheric structure affect satellite-derived aerosol signals. By distinguishing traffic-dominated urban signals from transported biomass-burning plumes, the study provides field-constrained evidence for assessing the limitations and interpretability of geostationary satellite aerosol products in complex terrain.

Back to session list

E6-2: Multi-Sensor Validation of the GOSAT-2 CAI-2 Cloud Flag and Developing Targeted Spectral Tests for Missed Clouds

Presenting author: Reza SAFABAKHSH

Authors: Reza SAFABAKHSH, National Institute for Environmental Studies; Tsuneo MATSUNAGA, National Institute for Environmental Studies

Keywords: GOSAT-2 CAI-2 L2, Cloud flag validation, Triple collocation, Dawid–Skene, Spectral test analysis.

GOSAT-2 XCO/XCH retrievals depend on the CAI-2 cloud mask. We evaluated the CAI-2 L2 cloud flag over Japan and nearby ocean using collocated AHI-9 and VIIRS, without assuming a perfect reference. Categorical Triple Collocation ranked sensor skill, and a Dawid–Skene model produced probabilistic consensus labels from AHI-9 and VIIRS. Visual checks of 500 footprints gave κ=0.88. Against the consensus, CAI-2 reached F1=0.89 and accuracy=0.82; errors were mainly false negatives in thin clouds. True positive (TP)/False negative (FN) radiance features and logistic regression highlighted log (NIR/VIS) and tri-band log-standard deviation for targeted FN recovery. cumulative distribution function and Kolmogorov-Smirnov tests showed separability, reaching 97% and 93% for two tests.

Back to session list

E6-3: Physically-Consistent Cloud Index Sequence Forecasting over the Korean Peninsula Using GK-2A Geostationary Satellite Imagery and a Spatiotemporal Transformer

Presenting author: Dongjin KIM

Authors: Dongjin KIM, Department of Environmental and Energy, Jeonbuk National University,; Jongmin YEOM, Department of Earth and Environmental Sciences, Department of Environmental and Energy, Research Institute for Materials and Energy Sciences, Department of Physics, Jeonbuk National University,

Keywords: GK-2A AMI; Cloude Index; Solar irradiance forecasting; Deeplearning; Nowcasting

Accurate prediction of cloud-field evolution is essential for intraday solar irradiance estimation and renewable energy grid management. Over the Korean Peninsula, monsoon-driven dynamics and orography induce strong spatiotemporal variability; models trained on raw geostationary reflectance can entangle cloud evolution with diurnal illumination changes, reducing skill at multi-hour lead times. We propose a physics-grounded spatiotemporal Transformer to forecast Cloud Index (CI) sequences at 2 km resolution from GK-2A AMI imagery. CI is a bounded [0,1] cloud-opacity proxy computed from VIS reflectance normalized against a monthly clear-sky reference and cloud reflectance ceiling, with all statistics estimated on the training split only. For pre-sunrise periods, a surrogate CI derived from IR105 brightness temperature is calibrated to the VIS-CI scale via monthly quantile mapping. A temporally continuous input is formed by blending VIS- and IR-derived CI using solar-zenith-angle weighting. The model ingests 04:00–09:00 KST at 30-min intervals and predicts 09:30–16:30 KST at 30-min intervals (15 steps; max lead 7.5 h). Auxiliary inputs include a water-vapour channel and solar-zenith-angle maps. The objective combines reconstruction loss with physics-informed regularization enforcing CI boundedness, spatial smoothness, and advection consistency via optical-flow-based warping between consecutive predictions. We train on Jan 2021–Sep 2022, validate on Oct–Dec 2022, and test on full-year 2023. Performance will be assessed using MSE, SSIM, and Fractions Skill Score, compared against persistence, optical-flow extrapolation, ConvLSTM/CNN video baselines, and ablations. This study investigates whether physically consistent CI representations and global spatiotemporal attention can improve cloud-structure prediction, particularly at lead times beyond three hours. This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT)(RS-2025-00515357). This research was supported by Global - Learning & Academic research institution for Master’s·PhD students, and Postdocs(LAMP) Program of the National Research Foundation of Korea(NRF) grant funded by the Ministry of Education (No. RS-2024-00443714).

Back to session list

E6-4: Synergy of Sentinel-5P datasets and hybrid CNN-LSTM architecture in assessing Urban Air Quality

Presenting author: Hugo Wai Leung MAK

Authors: Hugo Wai Leung MAK, The Chinese University of Hong Kong (CUHK) & The Hong Kong University of Science and Technology (HKUST)

Keywords: Sentinel-Air; Sentinel-5P (TROPOMI); CNN-LSTM; Deep Learning; Environmental Monitoring; Air Quality Analytics; Urban Sustainability

Severe air pollution has imposed critical challenges to surrounding environment and public health, especially in densely populated urban clusters. Adverse atmospheric conditions are commonly driven by traffic emissions, industrial operations, and residential activities. Traditional air quality monitoring network rely heavily on sparse ground‑level sensors, thus are often limited by inadequate spatial coverage and missing attributes. Advancement in satellite remote sensing and artificial intelligence techniques have enabled large‑scale, data‑driven air pollution monitoring across the globe. Nevertheless, other limitations could take place, for example, insufficient integration of spatial and temporal characteristics, failure to fully capture and utilize multi-pollutant datasets, and the lack of interpretability in model outputs. These restrict practical possibilities in conducting comprehensive pollution and air quality assessments. This study addresses these technical shortcomings by proposing an innovative deep learning framework, which we call “Sentinel-Air”, to analyze TROPOMI-based Sentinel-5P satellite images, detect and locate urban pollution sources, and infer corresponding estimated pollution concentrations. Sentinel-Air employs a hybrid Convolutional Neural Network (CNN)–Long Short-Term Memory (LSTM) architecture to capture spatial patterns and temporal variations of multiple key pollutants, namely NO₂, CO, SO₂, and CH₄ within prescribed spatial and temporal settings. The model is further enhanced through a pre-processing pipeline, multi-channel data fusion, and model interpretability via Gradient-weighted Class Activation Mapping (Grad-CAM). Corresponding experimental results were inter-compared with baseline and traditional models and approaches, via the use of common statistical metrics, like accuracy, F1-score, RMSE and R2, in terms of categorization of pollutants and estimation of concentration values. It was illustrated that Sentinel-Air surpasses existing frameworks, and can generate interpretable heatmaps and source attribution maps to support actionable environmental insights. This opens new windows in monitoring real-time urban air quality conditions and performing data-driven environmental decision-making, which relates to urban sustainability and smart city development.

Back to session list

Session B7 Abstracts

B7-1: Spatio-temporal Variation of Human Flow: Investigating the Impact of Space Syntax Metrics via Mobile Phone Data

Presenting author: Yun-Zhen CHEN

Authors: Bo-Cheng LIN, National Taipei University; Yun-Zhen CHEN, National Taipei University

Keywords: Space Syntax, Spatio-temporal Dynamics, Mobile Phone Data, Spatial Heterogeneity, Urban Mobility

While urban space constitutes the structural framework of a city, human flow represents the dynamic movement circulating through this geometric network. The heterogeneity of the topological structure dictates flow intensity and spatio-temporal distribution patterns across different time intervals. Within the framework of Space Syntax theory, metrics such as Integration, Choice, and Connectivity significantly influence pedestrian route-choice decisions, thereby modulating human flow and its temporal fluctuations. Leveraging the extensive sample size and spatio-temporal continuity of mobile phone data, this study captures the temporal dynamic characteristics of urban mobility to reflect human flow patterns. The research aims to extract high-resolution human flow features to investigate the spatial heterogeneity between Space Syntax metrics and flow dynamics, clarifying the extent to which urban space influences the expansion and contraction of spatio-temporal human flow. This study employs Multiscale Geographically Weighted Regression (MGWR), utilizing the daily human flow variation as the dependent variable and Space Syntax metrics as the independent variables. The results demonstrate that the impact of spatial configuration on human flow variation is not uniform but exhibits significant spatial non-stationarity. Notably, the study reveals that higher values of Integration—representing superior road accessibility—lead to a drastic reduction in the numerical value of human flow variation.

Back to session list

B7-2: A Geometry-Driven Topology Restoration and Semantic Decoupling Framework for Normalizing OSM-Derived OpenDRIVE Data in Early-Stage Autonomous Vehicle Testing

Presenting author: Ogi Dani SAKAROV

Authors: Ogi Dani SAKAROV, National Cheng Kung University; Kai-wei CHIANG, National Cheng Kung University

Keywords: Autonomous Vehicle Simulation; Digital Twin; OpenStreetMap (OSM); OpenDRIVE; Topology Restoration; Semantic Decoupling, Data Interoperability

Achieving seamless interoperability between heterogeneous geospatial data sources and autonomous vehicle (AV) simulators remains a critical bottleneck in rapid digital twin generation. However, systematic initialization failures are frequently encountered in SUMO–CARLA co-simulation when utilizing OpenStreetMap (OSM)-derived OpenDRIVE (XODR) data due to topology inconsistencies, incompatible traffic semantics, and spatial misalignment, which jointly prevent reliable map parsing and vehicle initialization. To address these challenges, we propose a multi-stage normalization framework grounded in four key contributions. First, geometry-driven topology restoration reconstructs fragmented lane topology using heuristic adjacency to restore drivability continuity. Second, boundary validation filters the network to distinguish between critical structural errors and legitimate map boundaries, ensuring topological validity. Third, semantic decoupling stabilizes initialization by stripping ambiguous traffic control definitions that cause simulator conflicts. Finally, spatial alignment unifies the coordinate systems of both simulators by enforcing a common local reference frame through offset nullification. Empirical validation demonstrates that the proposed multi-stage normalization framework reliably resolves both map-parsing errors and vehicle spawn anomalies. This work establishes a framework for transforming open geospatial data into simulation-ready environments, ensuring stable SUMO-CARLA co-simulation execution. Consequently, this study enables early-stage AV testing and digital twin prototyping without reliance on high-cost, high-precision mapping datasets.

Back to session list

B7-3: Synergistic Integration of UAV-LiDAR and 3D Gaussian Splatting for Precision Tree Species Classification and Biomass Estimation in the Tropical University Campus

Presenting author: Daisuke SASAKI

Authors: Daisuke SASAKI, University of Tokyo; Wataru TAKEUCHI, Institute of Industrial Science, The University of Tokyo; Khin Myat Kyaw, Institute of Industrial Science, The University of Tokyo , Asian Institute of Technology

Keywords: UAV-LiDAR; 3D Gaussian Splatting; Tree Species Classification; Biomass Estimation; Ecological Digital Twin; Web-GIS

Global strategies for climate adaptation and biodiversity conservation are redefining the critical role of Green Infrastructure (GI). University campuses, serving as “Living Labs,” provide essential ecosystem services from carbon sequestration to microclimate regulation. However, accurately quantifying these green assets remains challenging. Estimating biomass requires precise 3D geometry, while automated tree species classification relies on high-resolution 2D-images. Ground-level imagery (e.g., Google Street View) aids species identification but is unavailable for pedestrian-only areas and off-road green spaces. Meanwhile, conventional UAV-LiDAR excels in extracting top-down structural metrics but misses critical ground-level features (e.g., bark texture) strictly required for reliable classification. To resolve this limitation, this study proposes an advanced “Ecological Digital Twin” framework, utilizing the Asian Institute of Technology (AIT) campus as a tropical testbed. We propose a synergistic approach fusing aerial structural data with ground-level visual intelligence. Specifically, we deploy high-density UAV-LiDAR to obtain tree structures (e.g., tree height, crown diameter). Simultaneously, ground-based omnidirectional images are processed through two parallel workflows: (1) automated tree species classification using 2D deep learning models to capture fine textural features, and (2) the generation of 3D Gaussian Splatting (3DGS) models to serve as the Digital Twin’s photorealistic visualization layer. Precisely georeferencing 3DGS models with LiDAR point clouds in a Web-GIS platform (CesiumJS) achieves robust spatial co-registration. By fusing top-down structural geometry with bottom-up high-resolution textures, this hybrid approach overcomes conventional survey blind spots. Literature-backed expectations indicate this data fusion will improve AI-driven tree classification accuracy by 10–15% compared to LiDAR-only baselines. Furthermore, pairing this accurate species classification with LiDAR-derived structural metrics provides the essential data required for robust ecosystem service modeling. Ultimately, successfully integrating these models into CesiumJS demonstrates a scalable solution, delivering both precise quantitative metrics for expert environmental reporting and realistic visualization necessary for evidence-based green infrastructure management.

Back to session list

B7-4: Development of Spatially Enabled Web Maps and Dashboards

Presenting author: Javzandulam BATAA

Authors: Javzandulam BATAA, Mongolian University of Science and Technology; Bataa CHULUUN, Mongolian University of Science and Technology; Davaasambuu NARANTSETSEG, Bat synergy international LLC

Keywords: Keywords: GIS; Spatial Statistical Data; Dashboard; Visualization

The objective of this research is to reduce the time required to extract necessary numerical information from statistical data prepared in Excel by utilizing advanced GIS technologies to process spatially based statistical information and develop web maps and dashboards. The study uses, as its research objects, the statistical dataset titled “Higher Education Academic Year Report 2024–2025” published on the official website of the Ministry of Education in Excel format, as well as the boundary shapefile of Mongolia within a geographic coordinate system. The research was conducted based on GIS methodologies following a technological workflow for developing spatially based web maps and dashboards. The data processing and visualization were carried out using ArcGIS Pro 3.5.4, ArcGIS Online Web Map, ArcGIS Online Dashboard, and Microsoft Excel software. As a result of the study, a spatially based statistical dashboard titled “STUDENTS IN HIGHER EDUCATION INSTITUTIONS: ACADEMIC YEAR 2024/2025 DATA (By Capital City, Province, and District)” was developed. The resulting dashboard includes an interactive map of Mongolia, spatial information on higher education institutions and their student numbers, and enables users to quickly retrieve necessary information by province, capital city, district, and individual institutions. The dashboard incorporates 451 indicators, four serial charts, and four pie charts, providing an integrated visualization environment for spatial statistical data. To objectively evaluate the outcomes of the research within the framework of the proposed objective, the time required for data processing was measured. The development of the dashboard required 24 hours in total. In comparison, extracting the same statistical information directly from Excel and creating serial and pie charts required approximately 48–50 hours, with additional time needed to integrate these outputs into reporting documents such as Word and PowerPoint. From this comparison, it can be concluded that using advanced GIS technologies to process spatial statistical data and develop web maps and dashboards reduces the time required to extract useful numerical information from Excel-based statistical datasets by approximately half. This represents the practical significance of the research.

Back to session list

B7-5: How Conversational AI Reconfigures Retail Geography: Evidence from a GIS-Based Agent Simulation

Presenting author: I Gede Brawiswa PUTRA

Authors: Pei-Fen KUO, National Cheng Kung University; Tsun-Yuan KU, National Cheng Kung University; Chung-Wei SHEN, National Cheng Kung University; Febrian Fitryanik SUSANTA, National Cheng Kung University, Universitas Gadjah Mada ;I Gede Brawiswa PUTRA, National Cheng Kung University

Keywords: Conversational AI; Retail Geography; Agent-Based Modeling; GIS Simulation; Spatial Consumer Behavior

Generative AI chatbots are rapidly becoming a new gateway to retail discovery. As conversational systems integrate shopping and local search functions, they may fundamentally reshape how consumers choose where to shop. Traditionally, urban retail performance has relied on two mechanisms: street-level exposure that generates incidental visits, and platform-based search or advertising that influences consumers during active information seeking. Conversational AI introduces a third pathway by recommending specific merchants before a trip begins, potentially redirecting consumer flows away from physical visibility and traditional search platforms. This study develops a GIS-enabled agent-based model to compare these three customer acquisition mechanisms within a unified simulation environment. Mechanism A represents walk-in demand driven by route traversal and storefront visibility. Mechanism B captures map-based search and promoted listings that influence decisions route. Mechanism C models AI-generated recommendations initiated at trip origin, where consumers may directly follow suggested destinations with reduced exploration. The model is implemented on a stylized street network calibrated to the morphology of Tainan City. Consumer agents are assigned heterogeneous origin–destination pairs and travel schedules. Store choice follows a discrete-choice structure combining travel cost, attractiveness, and information salience. We conduct sensitivity analyses across customer volume, network connectivity, and AI trust levels to evaluate impacts on store visits, spatial concentration, and performance differences between arterial and alley locations. Results highlight how AI recommendations may substitute for or amplify traditional exposure and advertising channels under different urban conditions. The framework provides strategic insights for retailers and policymakers navigating an increasingly AI-mediated retail landscape.

Back to session list

Session C7 Abstracts

C7-1: Infrastructure Monitoring in South Korea Using Time Series InSAR: Insights from the K-GMS Platform

Presenting author: Khin Myat KYAW

Authors: Khin Myat KYAW, The University of Tokyo; Wataru TAKEUCHI, The University of Tokyo

Keywords: Subsidence; Maintenance; Sentinel-1; Ground; Geohazard

National records indicate a growing number of sinkhole events across South Korea over the past decade, particularly in urban areas. These recurring subsurface failures, including sinkholes and land subsidence, are closely associated with underground construction and deteriorating utility systems, posing significant risks to the safety and performance of nearby roads, railways, and expanding transportation networks. This trend highlights the shortcomings of conventional monitoring approaches and reinforces the need for largescale, advanced systems that can prioritize infrastructure maintenance and detect early ground movement before structural damage occurs. The Korea Ground Motion Service (K-GMS) is designed to meet this need by delivering Sentinel-1 based line-of-sight (LOS) ground deformation measurements. Through time series InSAR processing, the system will generate nationwide deformation maps with millimeter-level accuracy at ~30 × 30 m resolution. GNSS observations have already been used to validate the InSAR results, with future enhancements planned to incorporate GNSS-calibrated products for improved precision. In this study, representative case analyses will be presented to demonstrate the capability of K-GMS to detect subtle, early-stage ground deformation along critical transportation infrastructure, including tunnel systems and railway corridors. Such observations will support the advancement of proactive maintenance strategies and the reduction of geohazard risk. Additionally, it will also identify long term subsidence hotspots, where persistent ground lowering can increase flood risk by altering local topography. With frequent satellite revisit cycles and weather independent SAR imaging, K-GMS is expected to serve as a vital resource for continuous assessment of ground stability and the long-term integrity of national infrastructure.

Back to session list

C7-2: Satellite-Echo-Sounder Mapping of Navigation Bottlenecks in the Mekong-Tonle Sap-Bassac Corridor, Cambodia

Presenting author: BUNNARATH ouch

Authors: BUNNARATH ouch; LEE Han Soo

Keywords: inland navigation; bottleneck mapping; Sentinel; bathymetry; hydrotrac II; water level

Seasonal water-level fluctuations and rapid sediment-driven channel changes frequently restrict inland navigation in Cambodia’s Mekong-Tonle Sap-Bassac corridor. These restrictions create recurring “Bottlenecks”, where available depth becomes insufficient for passage, causing delays, cargo limitations, and higher operating costs for ports and vessels. A practical bottleneck map is therefore needed to support survey planning and maintenance decisions. However, field bathymetry alone is expensive and spatially discontinuous, while satellite observations alone cannot directly provide navigation depth. This study develops a satellite-echo-sounder workflow to map navigation bottlenecks using depth-relevant indicators. We combine (i) Sentinel-1 SAR time series to monitor water extent under cloud cover and capture seasonal low-water conditions, (ii) Sentitnel-2 optical imagery to detect exposed bars and turbidity-related features linked to sedimentation, and (iii) single-beam echo-sounder bathymetry collect with a Hydrotrac II system. Water-level (stage) records are used vertically reference bathymetric profiles and to interpret satellite-derived low-water signatures consistently. River reaches are segmented along a common centerline, and reach segment is characterized by satellite-derived persistence metrics (recurrent bar exposure, width construction, and planform adjustment) and by stage-referenced depth statistics from the echo-sounder surveys. We choose this integrated approach because it resolves the key limitation of using either data source alone: satellites provide continuous spatial monitoring, while echo-sounder surveys provide direct depth evidence needed for navigation. The expected outputs are (1) a corridor-scale bottleneck inventory with segment-level minimum-depth constraints, (2) a shoaling susceptibility map identifying reaches where bottlenecks are likely to recur, and (3) a prioritized list of hotspots for targeted re-survey and dredging. These products directly address the navigation reliability problem by enabling evidence-based maintenance planning and reducing uncertainty in where and when depth limitations occur.

Back to session list

C7-3: Detection and Size Estimation of Riverbed Boulders in Mountain Streams from UAV Imagery Using Deep Learning and GIS-Based Automated Labeling

Presenting author: Sanae KANG

Authors: Sanae KANG, Department of Forest Resources, Kookmin University.; Sinyoung PARK, Department of Forest Resources, Kookmin University.; Nuri CHUNG, Department of Forest Resources, Kookmin University.; Byungmook HWANG, Department of Forest Resources, Kookmin University.; Dongwook W. KO, Department of Forest, Environment, and Systems, Kookmin University, Forest Carbon Graduate School, Kookmin University.

Keywords: TerraLab; YOLO; Image segmentation; Drone; Riverbed structure

Riverbed structure, particularly the distribution of boulders, is a key factor in determining the hydraulic characteristics and the ecological health of aquatic microhabitats in streams. However, field surveys in mountain streams are often dangerous and inefficient due to rugged terrain and limited accessibility. To address these challenges, this study proposes a framework combining GIS-based automated labeling and deep learning to detect riverbed boulders and estimate their size from high-resolution UAV imagery, focusing on mountain streams in South Korea. UAV orthoimagery (3.95 cm/pixel) was collected over an approximately 110 m-long target stream. To reduce observer bias and enhance data preparation efficiency, a GIS-based automated segmentation tool developed by TerraLab was applied to label riverbed boulders. Based on the generated dataset, the YOLOv26n algorithm was utilized for instance segmentation, enabling precise identification of boulders and the extraction of their size. The segmentation results demonstrated high overall performance (precison > 0.90, recall > 0.80, mAP50 ≈ 0.90). For diameter estimation, the Mean Absolute Error (MAE) was 12.21 cm, representing 18.40% of the mean diameter of the boulders (66.34 cm). The predicted median and mean values showed strong agreement with the reference measurements. Notably, the mean diameter difference was only 2.22 cm, accounting for 3.34% of the reference mean value. These results demonstrate the reliability of the proposed geometric extraction method. Considering that mountain streams are geomorphologically complex and often hazardous to access, the strong performance of the proposed framework suggests its potential applicability to broader river systems. Furthermore, it is expected that the framework can be extended beyond the riverbed to include riparian zones, thereby supporting more integrated river management and planning. Funding: This study was conducted with the support of the R&D program for Forest ScienceTechnology (project no. RS-2025-02214405) provided by Korea Forest Service (Korea Forestry Promotion Institute). This study was carried out with the support of the “Program for Forest Science Technology (Project No. FE0100-2025-04-2025)” provided by the National Institute of Forest Science.

Back to session list

C7-4: Assessing Dry Port Spatial Efficiency via Geometric Layout Indicators from Remote Sensing Imagery

Presenting author: SUN YIRAN

Authors: SUN YIRAN, the University of Tokyo; Shibasaki Ryuichi, the University of Tokyo

Keywords: Dry port; Satellite image; Faster R-CNN; Remote sensing; Container detection

Dry ports play an important role in extending seaport functions into hinterland logistics systems. However, quantitative analysis of their spatial efficiency remains limited due to the lack of comparable operational data across regions. To address this gap, this research introduces a remote sensing–based perspective that evaluates dry port performance through spatial layout characteristics derived from satellite imagery. High-resolution satellite images are first used to detect and count containers using a Faster R-CNN (FRCNN) framework, allowing the extraction of container-related operational information from imagery. The detected results are then processed within a GIS environment to derive spatial features of container yard areas and port layouts. Using a dataset of 17 representative dry ports in China, geometric layout indicators are calculated and further classified through K-means clustering to identify distinct spatial layout types. These geometric indicators are subsequently integrated with the 2018 CCS dataset, including trade value and trade frequency, to examine the relationship between spatial configuration and trade intensity. The results show that dry port spatial efficiency is heterogeneous and not solely determined by scale. Ports with more compact layouts, balanced geometric proportions, and higher functional allocation to container yards tend to exhibit higher spatial trade intensity. The proposed framework demonstrates that remote sensing–based container detection combined with geometric layout indicators provides a scalable and comparable approach to evaluating dry port spatial efficiency, offering useful insights for logistics infrastructure planning and dry port design.

Back to session list

C7-5: Accuracy Evaluation for the Application of Portable LiDAR in Disaster Site

Presenting author: Yonghan JUNG

Authors: Yonghan JUNG, National Disaster Management Research Institute; Seul KOO, National Disaster Management Research Institute; Jungwook PARK, National Disaster Management Research Institute; Seongsam KIM, National Disaster Management Research Institute

Keywords: Portable LiDAR, Point cloud, 3D modeling, Accuracy assessment, Disaster investigation

Recently, the frequency of complex disasters caused by the interaction between natural hazards and socio-technical factors has been increasing. Complex disasters are characterized by the sequential propagation of terrain collapse, structural damage, and indoor space damage. Accordingly, rapid and safe acquisition of spatial information encompassing both indoor and outdoor environments is required. Ground-based LiDAR and drone-based survey methods have limitations in terms of rapid deployment during the initial response stage. In this context, portable LiDAR devices utilizing GNSS and smartphone LiDAR are easy to operate and have high potential for application at disaster sites. However, quantitative verification of performance accuracy in disaster environments remains insufficient. In this study, the modeling accuracy of facilities using portable LiDAR equipment was evaluated at the Korea Institute of Robotics & Technology (KIRO) disaster simulation test facility to verify its applicability to disaster sites. To evaluate performance according to acquisition conditions (distance and time), the modeling quality generated under nine different acquisition conditions was analyzed, and it was confirmed that the highest-quality model was produced when data were acquired at a distance of 1 m for more than 30 seconds. In addition, point clouds and 3D models of a cross-shaped structure with good GNSS signal reception and a tunnel-type structure shielded on all sides were constructed, and accuracy was evaluated using checkpoints. The results showed that the average RMSE of the 3D models according to GNSS signal reception performance ranged from approximately 0.06 to 0.19 m. Through this study, the strength of rapid data acquisition and model generation in time-critical disaster sites was confirmed, and the potential for application to various disaster site investigations was identified through future research aimed at improving accuracy.

Back to session list

Session S2 Abstracts

S2-1: Local Adustment of Satellite Chlorophyll-a in Coastal Waters using Machine Learning Techniques: Case of Ise Bay

Presenting author: Joji ISHIZAKA

Authors: Joji ISHIZAKA, Nagoya Univ.; Hiroki TSUKAMOTO, Science and Technology; Masataka HAYASHI, Science and Technology; Ryota KIKUCHI, Nagoya Univ.

Keywords: chlorophyll-a, coastal waters, SGLI/GCOM-C, machine learning

In coastal semi-enclosed waters, the accuracy of satellite-derived chlorophyll-a is often poor due to the errors in both atmospheric correction and in-water algorithm. In Japan, the use of satellite chlorophyll-a data is still limited, partly becuase some stakeholders have observed discrepancies between in-site and satellite-derived chlorophyll-a. While the development of more complex algorithms is ongoing, it will take time before they become practical. In the meantime, it is useful to improve existing SGLI/GCOM-C chlorophyll-a products using available in-situ observations of chlorophyll-a. In this study, we aim to improve satellite chlorophyll-a in the Ise Bay, Japan, using simple statistical and machine learning techniques.

Back to session list

S2-2: Detection of the Distribution Area of Discolored Water Caused by Submarine Volcanoes

Presenting author: Mitsuhiro TORATANI

Authors: Mitsuhiro TORATANI, Tokai University; Fumitatsu NAKAMICHI, Tokai University

Keywords: GCOM-C/SGLI, Submarine Volcano, discolored water

Discolored water areas caused by submarine volcanic activity are important indicators for understanding such activity. This study attempts to extract discolored water areas caused by submarine volcanoes from remote sensing reflectance spectra estimated from ocean color satellite sensors, focusing on Fukutoku-Okanoba, Funka-Asane, and Nishinoshima. The satellite sensor used was GCOM-C/SGLI. The extraction method for discolored water areas involves using cluster analysis (Ward’s method) on the remote sensing reflectance spectra to identify spectra characteristic of discoloration. Furthermore, the area classified into each discolored water cluster was calculated, and temporal changes in these areas were investigated. The results confirmed that around Fukutoku-Okanoba, discolored water areas of similar extent persisted both before and after the large-scale eruption in August 2021, indicating ongoing activity. At Fuhka-Asane and Nishinoshima, although the scale was somewhat smaller compared to Fukutoku-Okanoba, continued activity was also confirmed.

Back to session list

S2-3: Detection of characteristics of time-series in SGLI/GCOM-C observed Chlorophyll-a concentration by information mathematics

Presenting author: NAGAI Shin

Authors: NAGAI Shin; Kazunori OGATA; Hiroshi MORIMOTO

Keywords: Chlorophyll-a concentration; fractal; GCOM-C/SGLI; phytoplankton; wavelet transform

Accurate monitoring of the spatiotemporal variability of phytoplankton is important to understand the interaction between the marine ecosystem and human activities under climate change. To this end, we examined the wavelet power spectrum calculated from the GCOM-C/SGLI observed Chlorophyll-a concentration data from January 2019 to December 2025 in seven typical ocean areas. We detected significant power spectra at 0.25, 0.5, and 1.0 year periods but their timings varied depending on the target area. In this presentation, we will discuss the characteristics of these values, focusing on their fractality.

Back to session list

S2-4: Spatiotemporal Variations of Sea Surface Chlorophyll-a in the Seto Inland Sea, Japan

Presenting author: Han Soo LEE

Authors: Han Soo LEE; Gandhi NAPITUPULU; Nobuhito MORI

Keywords: sea surface chlorophyll-a: MODIS: SeaWIFS: Sentinel-2: Marine heat waves: Marine cold spell

Sea surface chlorophyll-a (SSC) is a key indicator of phytoplankton biomass and productivity and is shaped by physical and biogeochemical processes in marine ecosystems. The Seto Inland Sea (SIS), Japan’s largest semienclosed estuary, exhibits complex SSC variability driven by seasonal nutrient dynamics, oceanic intrusions, and atmospheric forcing. This study investigates the spatiotemporal variability in SSC in the SIS from 1998 to 2024 via satellite-derived datasets, with an emphasis on seasonal cycles, long-term changes, and the impacts of extreme thermal events, including marine heatwaves (MHWs) and marine cold spells (MCSs). The results demonstrate pronounced spatial contrasts among subregions: central bays and channels (e.g., Hiroshima Bay, Hiuchi-nada, Harima-nada, and Osaka Bay) exhibit stratification-paced blooms with SSC peaks in spring and autumn, whereas outer gateways and shelf-influenced areas (southern Kii Channel and Tosa Bay) follow a mixing/intrusion-paced regime strongly influenced by upwelling and oceanic intrusions. Change-point analysis revealed that both El Niño–Southern Oscillation (ENSO) phases and typhoon passages triggered abrupt SSC shifts, whereas Hiroshima Bay showed no significant long-term seasonal changes (probability < 0.9). Over the study period, 462 MHWs and 395 MCSs were detected, with MHWs predominantly suppressing SSC (67.53%) and MCSs enhancing it (70.89%). These effects were most pronounced in the inner SIS, where strong stratification and limited mixing amplify phytoplankton responses.

Back to session list

S2-5: Characteristics of satellite water temperature and water quality distribution related to the mass oyster death around Kure City in 2025

Presenting author: Yuji SAKUNO

Authors: Yuji SAKUNO, Hiroshima Univ.

Keywords: GCOM-C; SST; oyster; chlorophyll-a; DO; Seto Inland Sea

The mass death of oysters that occurred around Kure City in 2025 caused serious damage to the Seto Inland Sea fishing industry. The causes of the mass deaths are thought to be high water temperatures in the summer, high salinity, poor nutrition, and low oxygen levels. However, the exact nature of these factors remains unclear. Interview surveys also revealed that the extent of the damage caused by oyster deaths varies greatly depending on the location. Therefore, the authors used satellite data such as the GCOM-C SGLI sensor to examine the deviations of water temperature and chlorophyll-a concentration (Chl-a) from average values in 2025. The results showed a 2-3 degrees deviation in water temperature and a temporary drop in Chl-a in the Kure region. Furthermore, an attempt was made to estimate bottom DO using satellite SST data

Back to session list

Day 1 Poster Abstracts

P-1: Comparison of field-measured tree heights and canopy height model with machine learning and topographic factors for modelling diameter at breast height

Presenting author: Jhe-Syuan LAI

Authors: Cheng-Yu YANG, Feng Chia Univ.; Jhe-Syuan LAI, Feng Chia Univ.; Yung-Chung CHUANG, Feng Chia Univ.; Uen-Hao WANG, Taiwan Forestry Research Institute

Keywords: Diameter at Breast Height; Machine Learning; Random Forest; Regression; Topographic Factors

Diameter at breast height (DBH) is a critical parameter for investigating forest resources and to estimate the related information such as timber volume, biomass, and carbon storage, etc. These results are useful to achieve the Goal 15 (Life on Land) in the United Nations Sustainable Development Goals (SDGs). However, traditional field investigations are too time-consuming and uneconomic to obtain DBH parameters in the large scales and complex environments. With the advancement of remote sensing technologies, canopy height model (CHM) derived from airborne LiDAR provides high-resolution heights of trees to have a potential to instead of field measurements. The relationship between CHM and DBHs can be modelled by machine learning algorithms and geospatial factors in order to estimate timber volumes economically. Taiwania in the Lienhuachih Experimental Forest of Taiwan was selected as the study target. This study integrated the CHM and topographic factors (i.e., aspect, curvature, elevation, and slope) to perform the random forest (RF) based regression, comparing with the field-measured tree heights and typical regression methods (i.e., an independent variable with linear and exponential equations), respectively. The predictions derived from the regression models were validated by root mean square error (RMSE). In addition, these accuracy assessments included single plot validation and multiple plots based cross-validation. Results indicated that RF-based models outperformed typical regressions in all cases. For the single plot validations, RF obtained that the best results modelled by the field-measured tree heights and the CHM were 1.58 and 2.89 cm for RMSE, respectively. In terms of multiple plots based cross-validations, the best results of RF showed 4.83 and 8.36 cm for RMSE obtained from the field-measured tree heights and the CHM, respectively. In conclusion, integrating CHM and topographic factors with machine learning is competitiveness for economically predicting DBH.

Back to poster list

P-3: A Comparative Study of Paired and Unpaired GANs for Optical Image Generation from Polarimetric SAR

Presenting author: Hirofumi SHIMAOKA

Authors: Hirofumi SHIMAOKA,Shizuoka University; Rei SONOBE,Shizuoka University; Kan-ichiro Mochizuki,Pasco; Yudai ICHIKAWA, Shizuoka University

Keywords: SAR,GAN,SuperDove,ALOS-2/PALSAR-2,Polarimetric parameters

Optical satellite imagery plays a vital role in agricultural monitoring but is often limited by cloud cover and illumination conditions. Synthetic aperture radar (SAR) offers an all-weather alternative, and recent advances in deep generative models provide opportunities to reconstruct optical-like imagery directly from SAR data. In this study, we investigated the potential of generating realistic RGB images of croplands using adversarial generative networks (GANs) trained on ALOS-2/PALSAR-2 quad-polarimetric data. A distinctive aspect of our work is the evaluation of not only backscattering coefficients (Gamma nought) but also polarimetric parameters derived from quad-pol decompositions, including the Yamaguchi four-component methods. We selected paired image-to-image translation methods, including FGGAN and Pix2PixHD, as well as unpaired methods, including AttentionGAN, CycleGAN, and PUT. The Structural Similarity Index (SSIM) was employed to evaluate the quality of the generated images. Our results showed that no statistically significant differences were found in the SSIM values of the generated images between the two input images for AttenGAN, CycleGAN, and Pix2PixHD. In the case of PUT, using Yamaguchi-decomposed images produced significantly higher SSIM values across all bands (p < 0.001), confirming the highest image-generation accuracy. The use of polarimetric decomposition is an important factor in generating optical images from SAR data, and extending this approach to other regions and crops as well as to multi‑temporal learning and generation is expected to advance agricultural monitoring technologies.

Back to poster list

P-5: A Methodological Study of Geomorphological Survey Approaches for the End Moraine of Imja Glacial Lake

Presenting author: Susumu TAKAGISHI

Authors: Susumu TAKAGISHI,PASCO Corp./Tohoku Univ.; Chinatsu YONEZAWA,Tohoku Univ.

Keywords: Imja Glacier Lake;GNSS;MADOCA-PPP;DEM;Bathymetry

Imja Glacial Lake in the Nepal Himalaya has expanded in response to glacier retreat, while several small glacial lakes have developed on the downstream end moraine. These lakes reflect internal and surface hydrological–geomorphological conditions of the moraine dam and are therefore relevant to future assessments of moraine stability and potential glacial lake outburst flood (GLOF) hazards. This study presents a methodological investigation of geomorphological survey approaches for the end moraine of Imja Glacial Lake. Field surveys are planned to acquire lake-bottom bathymetry using a compact sonar system and to measure ground control points (GCPs) using GNSS Precise Point Positioning (PPP) along lake margins and on the moraine surface. In addition, a high-resolution satellite stereo digital elevation model constrained by the GCPs will be generated, and its consistency and applicability will be evaluated through comparison with in-situ measurements. The proposed framework provides a practical approach for geomorphological surveying of end moraines under high-altitude, access-limited conditions, supporting future glacial-lake monitoring and hazard-related studies.

Back to poster list

P-7: Improving RGB-Based Discrimination of Japanese Cedar and Cypress through Temporal Optimization Considering Inter- and Intra-Class Variability

Presenting author: Yuki SOFUE

Authors: Yuki SOFUE, Nara Women’s Univ.; Kiyoshi TAKEJIMA, Chubu Univ.; Chinatsu YONEZAWA, Tohoku Univ.; Hiroki MURATA, Kitasato Univ.; Kanako MURAMATSU, Nara Women’s Univ.

Keywords: Tree species classification; Sentinel-2 imagery; Visible-band vegetation index; Transformed Divergence; Support vector machine

Accurate discrimination of tree species is fundamental for forest management, resource assessment, and sustainable harvesting planning. In Japan, Japanese cedar (Cryptomeria japonica) and Japanese cypress (Chamaecyparis obtusa) dominate plantation forests, accounting for nearly 30% of the national forest area. However, as both species are evergreen conifers with similar canopy structure and foliage color, their separation using optical satellite imagery remains challenging. This study evaluates the potential of low-cost RGB imagery from Sentinel-2 for species classification while considering seasonal phenological differences. We calculated the visible-band vegetation index 2G–RBi and assessed class separability for each acquisition date using Transformed Divergence (TD). The date with maximum TD is selected as the optimal timing for classification in our previous work. However, TD reflects only inter-class distance and does not consider within-class variability caused by illumination differences, canopy heterogeneity, or environmental fluctuations, which can reduce reliability. To address this limitation, we introduced a normalized separability metric incorporating the standard deviations of both classes, thereby prioritizing dates with large inter-class differences and stable spectral responses within each species. Multi-temporal Sentinel-2 images acquired from 2018 to 2025 in Mie Prefecture, Japan, were analyzed using this approach. Early summer (May–June) generally showed the highest separability. Compared with conventional TD-based selection, the normalized metric yielded more robust image selection and improved classification performance in several years. For example, Kappa coefficients increased from 0.33 to 0.42 in 2018, from 0.29 to 0.31 in 2019, and from 0.37 to 0.39 in 2021. Support vector machine classification using only RGB information achieved Kappa values up to approximately 0.42. These results demonstrate that incorporating within-class variability into temporal separability analysis enhances the stability and reliability of RGB-based tree species discrimination. The proposed method offers a simple, cost-effective, operationally practical framework applicable to satellite, UAV, and aerial imagery for continuous forest monitoring.

Back to poster list

P-9: Estimation of Rice Cultivation Water Requirements and Irrigation Water Surplus or Deficit Using Sentinel-2 Satellite Data

Presenting author: Chiharu HONGO

Authors: Chiharu HONGO, Chiba Univ.; Takayuki Mori, Chiba Univ.; Eisaku Tamura, Chiba Univ.; Zaina Khoerunnisa Nurul Fath, Desy Nurfitriani, Syamsul Anwar, Reza Ahmad Nurcholis, Water Resources Department of West Java Province, Indonesia

Keywords: Irrigation management, Rice crop, Irrigation water surplus/deficit, Remote sensing

Global climate change and natural disasters are predicted to significantly impact food production worldwide. In regions with distinct dry and rainy seasons, such as Indonesia, recent precipitation changes associated with climate change have led to crop damage from natural disasters such as droughts and floods. Improving cultivation management systems to increase and stabilize rice yields requires effective use of irrigation water and careful review and design of water distribution management plans. This study targeted paddy fields in Bojongsoang District, Bandung Regency, West Java, Indonesia. Using satellite data acquired during the dry season, we estimated each growth stage of rice cultivation, calculated the irrigation water amount required for rice cultivation, and then examined the conditions of excess or shortage of irrigation water. The analysis utilized Sentinel-2 satellite data, daily irrigation water intake data measured at irrigation diversion points, precipitation data acquired from May to November 2024, and irrigation GIS data. To identify the growth stage of each rice plot, the time series vegetation index NDVI was calculated from Sentinel-2 satellite data, and then each growth stage was estimated. To calculate the water volume required for rice cultivation, the number of days required for each growth stage and the rice cultivation water volume per hectare were substituted into the pixels corresponding to each cultivation period estimated from the satellite data. Finally, the difference between the supply water volume and the rice cultivation water requirement was calculated to determine the irrigation water surplus or deficit at the watershed and field levels. The results showed that during the dry season of the target period, with almost no effective precipitation, using only precipitation as the supply water without considering irrigation water resulted in a cultivation water deficit across the entire area. On the other hand, while the incorporation of irrigation water into the supply mitigated the shortage, many paddy fields still showed a supply deficit of approximately 0.5 to 1.0 L/sec/ha during the transplanting, young panicle differentiation, and heading stages, which require higher cultivation water volumes per hectare. The results obtained are expected to support decision-making for improving large-scale water management systems through modeling of the water distribution process that aligns with farmers’ irrigation water demands and farming practices.

Back to poster list

P-11: Monitoring Nickel Mining Expansion in Northern New Caledonia using Multi-temporal Sentinel-2 Imagery and Random Forest Classification

Presenting author: Yuki SANO

Authors: Yuki SANO, Shizuoka Univ.; Rei SONOBE, Shizuoka Univ.

Keywords: Sentinel-2; New Caledonia; Land Cover; Environmental Monitoring; Nickel Mining; Google Earth Engine; Random Forest

New Caledonia is a critical biodiversity hotspot and home to some of the world’s most extensive nickel open-pit mining operations. Monitoring these developments through land-cover (LC) change is essential for environmental oversight. This study quantifies LC transitions in northern New Caledonia from 2021 to 2025. Utilizing Sentinel-2 imagery and Google Earth Engine (GEE), annual median composites were generated to establish a stable, cloud-free observation data. To enhance classification accuracy among spectrally similar classes, topographic features (elevation and slope) from NASADEM and the Modified Soil Adjusted Vegetation Index (MSAVI) were integrated into a Random Forest (RF) classifier. Tuned framework achieved an overall accuracy of 87.03%, with an F1-score of 0.89 for bare land, confirming the efficacy of fusing topographic and spectral data. Change detection analysis identified that approximately 610 hectares of vegetation were converted to bare land, comprising 100.04 ha of forest and 509.58 ha of shrub/grassland. While significant fluctuations occurred within vegetation classes—likely attributable to seasonal moisture variations and spectral confusion—the expansion of bare land remained spatially consistent with existing mining sites. The integration of elevation data proved essential for distinguishing high-altitude open-pit mines from low-altitude coastal mangroves. These findings demonstrate the potential for automated machine-learning workflows to monitor the environmental footprint of mining activities. Future research will focus on improving the distinction between urban and bare land and investigating forest loss driven by non-mining factors, such as wildfires, to further refine environmental impact assessments.

Back to poster list

P-13: Damage Assessment Model for Bacterial Leaf Blight in Rice by Integrating Flooding duration Using Sentinel-1/2 Data

Presenting author: Takayuki MORI

Authors: Takayuki MORI, Center for Environmental Remote Sensing, Chiba University, Japan; Chiharu HONGO, Center for Environmental Remote Sensing, Chiba University, Japan; Kevin Derrian FIRDAUS, Provincial Office of Food Crops and Horticulture of West Java Province, Indonesia; Gunardi SIGIT, Provincial Office of Food Crops and Horticulture of West Java Province, Indonesia; Dadan HYDAYAT, Provincial Office of Food Crops and Horticulture of West Java Province, Indonesia

Keywords: Pest and disease; Flooding duration; Rice paddies; Remote sensing; Data augmentation; Machine learning

Bacterial leaf blight (BLB) is one of the most severe diseases caused by the BLB pathogen, and early control and disease management are essential to suppress its spread. In conventional disease assessment using remote sensing data, vegetation indices have been applied to the assessment of BLB because leaf color changes after disease onset. While epidemiological studies have demonstrated a relationship between flooding duration and BLB occurrence in paddy fields, a lack of methodology has been reported that integrates this environmental factor into remote sensing-based disease assessment models. Therefore, a machine learning model was developed by integrating flooding duration and vegetation indices, enabling the estimation of BLB damage ratio at the harvesting stage. In this study, to overcome the limitation of field-level sample size in field surveys, training data for the satellite-based model was augmented using BLB damage ratio estimated from UAV data. As features, the flooding duration extracted from Sentinel-1 and Sentinel-2 data and vegetation indices at the actual harvesting stage were integrated and utilized. To apply this model across the entire target study area, the harvesting stage for each field was estimated from Sentinel-2 data, and the spatial distribution of the BLB ratio was characterized by using variables from that period as input values. Analysis results indicated that the UAV-based model achieved R² = 0.76 and RMSE = 6.41%, demonstrating sufficient estimation accuracy for the training data used in the satellite data-based model. The satellite-based BLB assessment model with data augmentation achieved higher estimation accuracy (R^2 = 0.38, RMSE = 10.07%) than the model without data augmentation. Introducing the flooding duration into the BLB assessment model improved the accuracy in some years, indicating that the consistency of estimation accuracy was maintained regardless of annual variations in transplanting timing or growing periods. These results suggest that integrating causal environmental drivers into the BLB assessment model can provide a more robust framework for disease monitoring.

Back to poster list

P-15: Slope Collapse Risk Assessment Based on Topographic Factors in the 2011 Kii Peninsula Heavy Rainfall Disaster

Presenting author: Mariko TERAUCHI

Authors: Mariko TERAUCHI, Nara Women’s University; Atsuko NONOMURA, Kagawa University; Kanako MURAMATSU, Department of Environmental Science, Nara Women’s University

Keywords: SPOT

In recent years, the increasing frequency of extreme rainfall has intensified the risk of slope collapses in mountainous regions. Current hazard maps mainly focus on residential areas and are insufficient for mountainous regions where isolated houses exist. In such areas, safe evacuation routes during heavy rainfall are not always clearly identified. Therefore, it is necessary to detect potentially hazardous locations before disasters occur. The 2011 Kii Peninsula Heavy Rainfall Disaster serves as a representative case. Record-breaking rainfall caused by Typhoon Talas triggered numerous slope collapses in southern Nara Prefecture, leading to severe damage. This study aims to investigate the pre-disaster topographic characteristics of collapse-prone slopes in this region. Pre-disaster data included a 10m mesh Digital Elevation Model (DEM) and 2010 imagery from Google Earth Pro to examine topographic characteristics. Post-disaster data consisted of SPOT satellite imagery and 2011 Google Earth Pro images, from which collapse areas were visually interpreted. Analysis was conducted using ArcGIS and ILWIS. Hydrological topographic analysis was used to calculate watershed areas, while topographic quantity analysis was used to calculate topographic factors such as mean curvature. Additionally, as a secondary objective, slope instability under potential seismic conditions was evaluated. Geometric correction was performed on the SPOT imagery, but slight positional discrepancies remained. The results showed that collapse areas roughly correspond to delineated watershed areas. Mean curvature analysis revealed small concave and convex microtopographic features along the collapsed slopes. These findings suggest that pre-existing topographic conditions influenced slope collapses. Future work will apply this analytical framework to other cases to enhance hazard assessment in mountainous regions.

Back to poster list

P-17: Classification of Flooded and Non-Flooded Paddy Fields Using High-Resolution QPS-SAR Data and Machine Learning

Presenting author: Tomoyuki HIROSE

Authors: Tomoyuki HIROSE,PASCO CORPORATION; Atsushi KIMURA,PASCO CORPORATION; Yuta TSUCHIYA,PASCO CORPORATION; Shimpei KUDO,iQPS Inc

Keywords: QPS-SAR;AWD;GHG;Time series analysis

Methane emissions from agricultural activities account for approximately 40% of global anthropogenic methane emissions, with rice cultivation contributing about 8%, particularly in Southeast Asia. In Cambodia, total greenhouse gas (GHG) emissions increased from approximately 27.8 Mt CO2e in 2010 to 42.4 MtCO2e in 2020. In response, the Cambodian government submitted its Long-Term Strategy for Carbon Neutrality (LTS4CN) to the United Nations in December 2021, aiming to achieve carbon neutrality by 2050. According to the LTS4CN, agricultural activities accounted for about 22% of national GHG emissions in 2016 (approximately 18MtCO2e), and emissions from this sector are projected to increase to 34.9 MtCO2e by 2050 under a business-as-usual scenario. Among mitigation measures, the implementation of Alternate Wetting and Drying (AWD) in rice cultivation is emphasized due to its potential to reduce methane emissions by avoiding continuous flooding. To monitor flooded and non-flooded conditions across extensive paddy field areas, satellite-based observation is expected to play an important role. However, observing paddy field water conditions during the rainy season remains challenging because frequent precipitation and persistent cloud cover limit the availability of optical satellite data. This study investigates the potential of a small X-band Synthetic Aperture Radar (SAR) satellite constellation, QPS-SAR, to monitor paddy field water conditions during the rainy season in Cambodia. QPS-SAR offers all-weather observation capability, high spatial resolution of approximately 0.5 m in Spotlight mode, and high-frequency observations enabled by its constellation design. As a result of this study, flooded and non-flooded field conditions were classified using QPS-SAR data and a Random Forest classifier, achieving an overall accuracy of 81%. These results demonstrate that high-resolution and high-frequency SAR data, combined with machine-learning-based classification, can effectively capture temporal variations in paddy field water conditions during the rainy season. The findings indicate that QPS-SAR provides a practical tool for wide-area monitoring of AWD implementation and supports national efforts to reduce agricultural methane emissions in line with Cambodia’s carbon neutrality targets.

Back to poster list

P-19: An Iterative Hypothesis Verification-Based Multi-Agent Framework for Object Activity Inference in High-Resolution Satellite Imagery

Presenting author: Hwang junesik

Authors: Hwang junesik, UST (University of Science and Technology)

Keywords: KOMSAT-3; KOMSAT-3A; Multi-Agent System; Activity Inference; Experience Memory; Model Calibration

This study proposes an experience-based adaptive multi-agent collaborative framework to interpret complex interactions and identify scene-level activities in high-resolution optical satellite imagery. Conventional deep learning and Vision-Language Models often suffer from overconfidence and confirmation bias, leading to unreliable decision-making due to the lack of feedback loops. To address these challenges, we designed an autonomous loop consisting of perception, hypothesis generation, and critic agents that interact iteratively. A key innovation is the Experience Memory System, which accumulates past analytical successes and failures to refine reasoning and suppress overconfident errors. By integrating this system, the framework redefines activity inference as an iterative process of “hypothesis-criticism-revision” rather than a traditional single-shot classification. Experimental results using KOMSAT-3/3A high-resolution (0.5m) satellite imagery demonstrate that the proposed multi-agent framework achieves a Top-3 accuracy of 75.8%, significantly outperforming standard single-agent models at 60.7%. Furthermore, the system effectively enhances the reliability of geographic intelligence by calibrating predictive confidence, achieving an Expected Calibration Error (ECE) of 0.2931 and a Brier Score of 0.3327. These improvements validate the efficacy of the self-evolving mechanism in providing dependable insights for strategic decision-making in noisy and unpredictable defense environments.

Back to poster list

P-21: VALIDATION OF SWOT SATELLITE WATER SURFACE ELEVATION OBSERVATIONS IN THE HAN RIVER BASIN, SOUTH KOREA

Presenting author: Keon Ho KIM

Authors: Keon Ho KIM, Department of Earth System Sciences, Yonsei Univ.; Hahn Chul JUNG, Department of Earth System Sciences, Yonsei Univ.; Joon Hyuk CHOI, Department of Earth System Sciences, Yonsei Univ.; Seo Jin KIM, Department of Earth System Sciences, Yonsei Univ.

Keywords: SWOT satellite, Ka-band Radar Interferometer, Water Surface Elevation (WSE), River water level, Validation

As climate change intensifies, the value of surface water as a resource is increasing, and advanced monitoring technologies are required for the efficient management and securing of surface water resources. Traditionally, surface water information has primarily been obtained from gauging stations. However, despite their high accuracy, limited spatial coverage constrains large-scale monitoring and analysis. Previous satellite remote sensing techniques partially supplemented these limitations by providing spatially continuous observations of surface water extent. However, they either lacked direct information on Water Surface Elevation (WSE) or were limited to one-dimensional, along-track measurements. Launched in 2022, the SWOT satellite is equipped with the Ka-band Radar Interferometer (KaRIn), enabling WSE measurements across a 120 km swath. This enables SWOT to provide sustained two-dimensional observations of WSE, addressing key limitations of previous observation approaches. This study evaluates the WSE observation accuracy and applicability of SWOT Level-2 High-Rate products in the Han River basin, South Korea. The assessment utilized PIXC and RiverSP products from five SWOT passes starting in August 2023, incorporating quality flags to ensure the reliability of the WSE measurements. The calculated WSE values were compared with in-situ water level gauge observations. For rivers satisfying the SWOT-recommended width (> 100 m), the PIXC and RiverSP products achieved WSE accuracies with Root Mean Square Errors (RMSE) of approximately 0.29 m and 0.27 m, respectively. Notably, direct processing of the PIXC product minimized data loss, resulting in more frequent observations across most gauging stations. Furthermore, the PIXC product enabled valid WSE observations in some rivers narrower than 100 m, with RMSE values ranging from 0.1 to 0.3 m. The significance of this study lies in validating SWOT’s WSE observation capabilities in South Korea’s river systems and demonstrating that reach-scale analysis can complement traditional point-based monitoring, providing a broader spatial perspective for water resource assessment.

Back to poster list

P-23: PRELIMINAY STUDY ON GEOSTATIONARY MULTISPECTRAL OBSERVATIONS FROM GOCI-Ⅱ FOR WILDFIRE CHANGE DETECTION

Presenting author: Kyung Hwan Shin

Authors: Kyung Hwan Shin, Department of Earth System Sciences, Yonsei University, Seoul, South Korea; Hahn Chul Jung, Department of Earth System Sciences, Yonsei University, Seoul, South Korea; Seo Jin Kim, Department of Earth System Sciences, Yonsei University, Seoul, South Korea; Do-Hyun Hwang, Greener-Futures Research Institute of Sciences, Yonsei University, Seoul, South Korea; Kyeong-Sang Lee, Korea Institute of Ocean Science and Technology, Korea Ocean Satellite Center, Busan, South Korea; Jong-Kuk Choi, Korea Institute of Ocean Science and Technology, Korea Ocean Satellite Center, Busan, South Korea

Keywords: Wildfire, Multispectral observation, Burn-sensitive vegetation indices, Change detection, GOCI-Ⅱ

Wildfires are one of the most dangerous natural risks, causing significant environmental changes for the land surface and vegetation. Understanding the spatial and temporal characteristics of these changes is essential for quantifying wildfire impacts on terrestrial ecosystems. In general, optical satellite-based wildfire change detection has been performed using polar-orbiting sensors such as the Moderate Resolution Imaging Spectroradiometer (MODIS). The accuracy of change detection requires short revisit intervals for repeated observations. Polar-orbiting sensors are subject to fixed overpass times and data gaps caused by clouds and aerosols, which often limit continuous temporal coverage. Geostationary Ocean Color Imager-Ⅱ (GOCI-Ⅱ) provides notably short revisit time, acquiring up to ten observations per day over the East Asia. GOCI-Ⅱ has multi-spectral bands designed for ocean color applications and land surface reflectance products are also available for calculating burn-sensitive vegetation indices. In this study, the change detection with enhanced temporal information was conducted using GOCI-Ⅱ observations in South Korea. This approach was applied to identify the timing of maximum disturbance based on vegetation indices that were evaluated and validated prior to selection. The uncertainty of timing was validated by comparison with active fire detections from the Advanced Meteorological Imager Fire Flag (AMI-FF), which is derived from satellite thermal anomaly measurements designed to identify burning pixels. The results demonstrate the potential of geostationary sensors for wildfire change detection. This enhanced change detection method supports post-fire impact assessment while providing a scientific basis for wildfire management and policy decision-making.

Back to poster list

P-25: A Cadastral-Aware U-Net Semantic Segmentation Framework for High-Resolution Aerial Crop Mapping in Heterogeneous Agricultural Landscapes

Presenting author: Wei-Tsun LIN

Authors: I-Liang SHIH, National Center for High-performance Computing, National Institutes of Applied Research, Taiwan; Chung-Yen WU, GIS Research Center, Feng Chia University; Ming-Sian CHEN, GIS Research Center, Feng Chia University; Chih-Wei KUAN, GIS Research Center, Feng Chia University; Wei-Tsun LIN, GIS Research Center, Feng Chia University; Mei-Ling YEH, GIS Research Center, Feng Chia University

Keywords: Semantic segmentation; U-Net; Cadastral data integration; Parcel-Level crop mapping; Deep learning; Precision agriculture

Accurate delineation of crop parcels from high-resolution aerial imagery remains a significant challenge in heterogeneous agricultural landscapes characterized by irregular field boundaries, subtle inter-crop texture variations, and mixed land-use patterns. This study introduces a cadastral-aware semantic segmentation framework leveraging the U-Net architecture for automated crop mapping at the parcel scale. The proposed methodology integrates three key components: (1) high-resolution aerial imagery, (2) official cadastral shapefiles, and (3) registered crop-planting records. By spatially aligning imagery with parcel-level cadastral boundaries, a standardized pixel-level training dataset is constructed to enhance label precision and reduce geometric inconsistencies. The U-Net model is trained to perform semantic segmentation for target crops, enabling continuous parcel-shaped predictions rather than coarse object-level representations. Preliminary experiments conducted in agricultural regions of Yunlin, Taiwan, focused on garlic and peanut cultivation. Quantitative evaluation using confusion-matrix-based metrics demonstrates promising classification performance, with an overall accuracy of approximately 76%. Qualitative visualization further highlights the model’s ability to discriminate between target crops, non-target vegetation, and built-up areas. Additionally, pixel-wise confidence mapping is incorporated to enhance interpretability and post-classification quality control. While model refinement and multi-temporal expansion are ongoing, the proposed framework underscores the value of integrating cadastral geospatial constraints with deep learning–based semantic segmentation. This approach provides a scalable foundation for parcel-level crop monitoring, agricultural statistics estimation, and spatial decision support in precision agriculture and remote sensing applications.

Back to poster list

P-27: Water Vapour Correction for Improving Landsat-9 Single-Channel LST Retrieval over Northern Taiwan

Presenting author: Tri Anggun LESTARI

Authors: Tri Anggun LESTARI, National Central Univ.; Tang-Huang LIN, National Central Univ.

Keywords: Landsat-9; Land Surface Temperature; Emissivity; Mono-window; Total Column Water Vapour; ERA5; MODIS

Land surface temperature (LST) retrieval from Landsat often shows systematic bias over heterogeneous urban surfaces because emissivity and atmospheric absorption are imperfectly constrained. This study evaluates how emissivity parameterisation and total column water vapour affect Landsat-9 TIRS Band 10 LST over northern Taiwan using one scene acquired on 22 July 2022 (about 10:00 local time) and 20 ground stations, with 2 m air temperature at overpass time used as the reference. We implement a single-channel mono-window LST framework and compare a baseline Sobrino NDVI-threshold emissivity (SOB) with an enhanced Sobrino combination approach that couples NDVI-based emissivity with a mono-window atmospheric correction explicitly parameterised by total column water vapour. We supply water vapour from ERA5 and from MODIS to test sensitivity to source. To test whether emissivity changes alone can explain the errors, we also evaluated several emissivity-only options, which show only limited improvement. SOB without explicit water vapour correction shows a warm bias (MBE 0.6311 °C) and RMSE 1.5440 °C with R² 0.6570. Adding water vapour largely removes the mean bias and reduces random error. SOB plus ERA5 water vapour achieves MBE 0.0029 °C and RMSE 1.3785 °C (R² 0.6603). SOB plus MODIS water vapour performs nearly identically (MBE 0.0017 °C, RMSE 1.3783 °C, R² 0.6603). Rural stations show much stronger agreement than city stations (R² about 0.93 vs about 0.22), which highlights urban mixed-pixel representativeness limits. Adding water vapour removes most of the systematic bias. The remaining scatter likely comes from mixed land cover within each pixel and from the physical mismatch between LST and 2 m air temperature, since LST represents the radiative temperature of the surface.

Back to poster list

P-29: Regression-Based Cloud Gap Filling for Sentinel-2 L2A Time Series Masking and Robust Radiometric Harmonization

Presenting author: Donggyu KIM

Authors: Donggyu KIM, Pukyong National Univ.; Seungchan LIM, Pukyong National Univ.; Hyojin KIM, Pukyong National Univ.; Jaehoon JEONG, Pukyong National Univ.; Chuluong CHOI, Pukyong National Univ.

Keywords: Sentinel-2; Level-2A; surface reflectance; cloud masking; gap filling; robust regression; kernel density estimation

Sentinel-2 Level-2A (L2A) surface reflectance imagery provides multispectral observations that are widely used for visual interpretation and spatiotemporal analysis. In this study, Bands B2, B3, B4, and B8 were selected as target bands due to their high utility for visualization and land–water feature characterization. A major limitation of optical satellite imagery is occlusion and radiometric distortion caused by clouds. Sentinel-2 L2A products include the Scene Classification Layer (SCL), a pixel-based map generated by Sen2Cor, enabling rule-based identification of cloud-affected pixels. This paper proposes an automated workflow that removes cloud-contaminated observations and reconstructs the resulting large missing areas via regression-based gap filling. All acquisitions are harmonized to a common spatial reference for pixel-wise comparability. Each image is geometrically registered to a master grid, and nearest-neighbor resampling is used to match the target grid; when grids or projections differ, reprojection/warping is applied. Cloud-affected regions are defined not only as visually opaque clouds but also as an extended influence zone reflecting cloud-induced perturbations in DN/reflectance values. A binary cloud mask is produced by combining SCL classes, cloud probability (MSK_CLDPRB), and land/sea layers, followed by a 5×5-pixel dilation to reduce boundary contamination. Isolated non-cloud clusters smaller than 3,000 pixels are removed to improve mask reliability. For radiometric mapping between master and candidate slave images, the method applies low-signal filtering, a wedge-based filter to suppress extreme vegetation/moisture effects, an overlap constraint (>50% valid intersection), Gaussian KDE peak estimation, robust (Huber) regression, orthogonal-residual outlier rejection using MAD, and a covariance-ellipse filter based on Mahalanobis distance. Ranked slave images are then sequentially applied using a linear model to fill missing pixels, iteratively maximizing the overall fill rate while maintaining radiometric consistency. The proposed workflow improves the usability of Sentinel-2 imagery for nationwide, continuous time-series applications under frequent cloud contamination.

Back to poster list

P-31: Localized Grounding-Line Retreat of the Nansen Ice Shelf, East Antarctica, and Its Implications

Presenting author: Minji Cho

Authors: Minji Cho, Kangwon National Univ.; Hyangsun Han, Kangwon National Univ.

Keywords: Nansen Ice Shelf; Grounding line retreat; Ice velocity change; Retrograde basal slope; Ice thickness change

Grounding line retreat is a key dynamic process that reduces basal resistance at the grounding zone and alters ice flow characteristics, thereby influencing ice-sheet stability. In this study, we identified the grounding line position of the Nansen Ice Shelf in Terra Nova Bay, East Antarctica, using differential interferometry of COSMO-SkyMed 1-day interferometric pairs acquired in 2021. We compared the 2021 grounding line with those in 1996 and 2018. The grounding line of the Nansen Ice Shelf remained stable 1996 to 2021. However, a localized sector retreated by approximately 2.2 km between 2018 and 2021. In this sector, surface velocity from ITS_LIVE increased by about 10% 2018, spatially coincident with a retrograde basal slope. Ice shelf thinning was observed from satellite altimetry during 2003-2019. Offshore subsurface ocean temperatures from GLORYS12V1 show a transient warming prior to 2018. Although ice shelf thinning and ocean warming may broadly affect grounding line dynamics, the retreat after 2018 was confined to the localized sector characterized by a retrograde slope. This suggests that local bed geometry could exert a first-order control on grounding line stability of the Nansen Ice Shelf and governs where retreat is initiated.

Back to poster list

P-33: Conceptual Design of a UAV-Based Sampling System for Remote Investigation of Tidal Flat Sediments

Presenting author: YeongJae JANG

Authors: YeongJae JANG, Korea Institute of Ocean Science and Technology; Joo-Hyung RYU, Korea Institute of Ocean Science and Technology; Keunyong KIM, Korea Institute of Ocean Science and Technology; Donguk LEE, Korea Institute of Ocean Science and Technology

Keywords: UAV-based sampling system; Tidal flats; Sediment sampling; UAV Modular system design; Remote investigation

Tidal flats are characterized by soft sediment conditions and tidal constraints, making field access difficult and limiting the efficiency of conventional surveys. Field investigations in such environments have largely relied on manual operations, which impose safety risks and significant physical demands on surveyors. To address these limitations, this study presents a conceptual design of a UAV-based remote sediment sampling system. The proposed system is based on the DJI Matrice 300 RTK platform and is designed with consideration of payload capacity and flight stability, emphasizing a lightweight and modular architecture. The system is divided into an upper module and a lower module centered on a winch mechanism. The upper module interfaces with the UAV via DJI SkyPort, receiving power and communication signals while controlling winch operation and line deployment. The lower module serves as the sampling payload, incorporating an independent power supply and an RF-based control system to operate multiple servo motors for sediment sampling. To improve sampling and flight stability, a multi-line suspension system with two or more lines is employed to suppress payload rotation. The sampling head adopts a gravity-driven penetration approach, eliminating the need for drilling mechanisms and thereby simplifying the structure while enhancing operational reliability. A detachable cartridge-based design is introduced to ensure independent sample storage and minimize cross-contamination, while enabling repeated sampling operations. A fail-safe mechanism is incorporated to allow immediate load release via remote control under abnormal conditions, such as excessive pulling force or obstruction. An initial prototype will be fabricated using 3D printing, and ground-based tests will be conducted to validate the mechanical configuration and operational sequence. The proposed system is expected to improve field accessibility, reduce manpower and operational time, and enable automated, waypoint-based repetitive sampling for tidal flat sediment investigations.

Back to poster list

P-35: Secular and Seasonal Vertical Deformation in Taiwan from Long-Term GNSS Time Series

Presenting author: Yen-Ru Lai

Authors: Yen-Ru Lai,National Chung Hsing University; Ming-En Syu,National Chung Hsing University; An-Ting Chang;National Chung Hsing University

Keywords: Continuous GNSS time series;Vertical crustal deformation;Seasonal deformation

Vertical crustal deformation in Taiwan reflects the combined influences of active tectonics, intense monsoonal rainfall, and extensive groundwater extraction. To quantitatively separate these processes, we analyze long-term continuous GNSS time series using a comprehensive model that simultaneously estimates secular velocity, annual signals, coseismic offsets, postseismic transients, and equipment-related discontinuities (e.g., antenna changes). This approach enables robust isolation of long-term and seasonal deformation components. The estimated secular vertical velocities show strong regional contrasts. Southwestern Taiwan exhibits widespread subsidence reaching −15 to −20 mm/yr, particularly along the Chiayi–Tainan–Kaohsiung coastal plain, consistent with long-term aquifer compaction induced by groundwater over-extraction. Central western Taiwan shows moderate subsidence of −5 to −10 mm/yr. In contrast, central mountainous and eastern Taiwan are near stable or slightly uplifting, with rates between 0 and +5 mm/yr, reflecting tectonic shortening within the active orogenic belt. Seasonal signals further distinguish deformation mechanisms. Annual amplitudes range from 2 to 10 mm, with the largest amplitudes (8–10 mm) concentrated in southwestern Taiwan, coinciding spatially with the main subsidence belt. Across most of Taiwan, the annual phase indicates peak subsidence in June–August, corresponding to the rainy season and consistent with elastic surface loading due to increased water mass. However, stations located within the major subsidence zone show systematic phase shifts of approximately 3–5 months relative to the regional elastic response. This temporal offset suggests a delayed poroelastic response associated with variations in groundwater storage and pressure diffusion within aquifer systems. By jointly interpreting secular velocity (−20 to +5 mm/yr), seasonal amplitude (2–10 mm), and phase (3–8 months), we demonstrate that southwestern Taiwan represents a coupled hydro-mechanical system, where long-term anthropogenic subsidence is superimposed on strong seasonal hydrological forcing with both elastic and poroelastic contributions.

Back to poster list

P-37: Structural-Type and Construction-Year Differences in Building-Level Sentinel-2 Reflectance

Presenting author: Koji OGINO

Authors: Koji OGINO, Institute of Science Tokyo; Toshihiro OSARAGI, Institute of Science Tokyo

Keywords: Sentinel-2; Building footprints; Building-level reflectance; Band reflectance; Structural type; Construction year; Nonparametric tests; Hiroshima, Japan

Datasets containing building attributes, such as structural type and construction year, are fundamental to urban planning and disaster risk management; however, their regional availability varies widely. Although deep learning has been employed to infer these attributes from satellite imagery, the underlying physical basis remains insufficiently understood. This study examines how building-level Sentinel-2 reflectance differs by structural type and construction year in Hiroshima City, Japan, where detailed building data are publicly available. We analyzed a Sentinel-2 Level-2A scene acquired on 21 April 2021, utilizing 10 m resolution bands (2–4, 8) and 20 m resolution bands (11, 12). Buildings were classified into three structural types (wooden, reinforced concrete, and steel) and further stratified into construction-year groups (five for wooden; three for reinforced concrete and steel). Reflectance for each building was extracted by integrating building footprints with imagery through: (i) geometric alignment and masking using the Scene Classification Layer (SCL), (ii) selection of effective pixels based on subpixel occupancy estimated via 10×10 supersampling (occupancy ≥ 0.6), and (iii) calculation of the median of these pixels for robust representative reflectance. Group differences were evaluated using the Kruskal–Wallis test and Holm-corrected Mann–Whitney U tests, with effect sizes quantified by epsilon-squared (ε²). Pronounced differences among structural types were observed in visible bands (bands 2–4; ε² ≈ 0.103–0.126), whereas these differences were negligible in the near-infrared band (band 8; ε² ≈ 0.004). Moderate differences were found in shortwave infrared bands (band 11: ε² ≈ 0.058; band 12: ε² ≈ 0.082), though with greater uncertainty due to coarser resolution and mixed-pixel effects. While differences among construction-year groups within structural types were statistically detectable in some instances, effect sizes were generally small. These findings suggest that structural type contributes more significantly than construction year to building-level Sentinel-2 reflectance patterns.

Back to poster list

P-39: THIN SEA ICE THICKNESS ESTIMATION USING COMBINED OPTICAL AND SAR DATA

Presenting author: Hiroyuki WAKABAYASHI

Authors: Hiroyuki WAKABAYASHI, Tohoku-gakuin Univ.; Kohei CHO, Tokai Univ.

Keywords: thin sea ice; backscattering; Capella; PlanetScope; Sentinel-1

This study investigates an approach for estimating sea ice thickness in thin-ice regions using high-resolution optical and microwave remote sensing data. The study area is Notsuke Bay, located in eastern Hokkaido, Japan, where seasonal sea ice forms under relatively calm coastal conditions. Remote sensing data were acquired from the PlanetScope optical sensor, Sentinel-1 C-band synthetic aperture radar (SAR), and Capella X-band SAR. These satellite datasets were analyzed alongside in situ ice thickness measurements to evaluate their effectiveness for estimating thin ice thickness. Field observations were conducted from Feb.25 to Feb.28 in 2025 at 71 measurement locations across the study area. The observed ice thickness ranged from 0.5 cm to 38.0 cm, with a mean thickness of 14.5 cm and a standard deviation of 9.3 cm, indicating substantial spatial variability within the thin-ice regime. Analysis of PlanetScope data revealed that reflectance in the visible and near-infrared bands exhibited relatively strong correlations with sea ice thicknesses below 20 cm. Except for the near-infrared band, differences in correlation strength among visible bands were small, with coefficients of determination of approximately 0.6. This result suggests that sea ice thickness in thin-ice regions can be reasonably estimated from a single visible band, simplifying data requirements and processing. Furthermore, relationships between ice thickness and VV-polarized backscattering coefficients derived from Sentinel-1 and Capella SAR observations were identified, confirming the sensitivity of SAR backscatter to thin ice thickness. In particular, Capella X-band SAR data acquired at four different incidence angles were examined to assess angular dependence. The results demonstrated that VV-polarized observations acquired at incidence angles near 30° provide the highest sensitivity to ice thickness and are most suitable for thickness estimation. By integrating PlanetScope optical data with Capella X-band SAR data, ice thickness in thin-ice regions was estimated with improved accuracy. The combined optical & microwave approach enabled the estimation of ice thicknesses up to approximately 40 cm at observation sites, achieving a coefficient of determination of approximately 0.76. These results highlight the potential of synergistic use of high-resolution optical and SAR data for monitoring thin sea ice in coastal environments.

Back to poster list

P-41: Mapping tidal flats using Sentinel-1 and ALOS-2 in the Ariake Sea

Presenting author: Tatsuyuki SAGAWA

Authors: Tatsuyuki SAGAWA, Tottori University of Environmental Studies; Toranosuke NAGAHAMA, Tottori University of Environmental Studies; Shinyo KITANO, Sanyo Techno Marine Inc.; Teruhisa KOMATSU, Japan Fisheries Resource Conservation Association

Keywords: tidal flats, SAR, tide, Sentinel-1, ALOS-2

Tidal flats are sandy–muddy environments that are submerged at high tide and exposed at low tide, forming a critical transition zone linking terrestrial and marine material cycles. They provide essential ecosystem services, including habitat provision, water purification, and blue carbon sequestration. However, rapid coastal development has caused widespread degradation and loss of tidal flats worldwide. Accurate mapping and monitoring of their spatial extent and topographic variation are therefore crucial for sustainable coastal management. In Japan, Ministry of Environment has assessed tidal flat distributions, and recent mapping efforts have primarily relied on high resolution optical satellite imagery. A simple approach is to extract tidal flats by subtracting the exposed areas from images taken at high tide and low tide. However, optical satellites are limited to daytime observations and are strongly affected by cloud cover, making it difficult to obtain suitable images at appropriate tidal conditions. In highly transparent waters, tidal flats can be extracted through bathymetric estimation, but tidal flat environments often have high turbidity. In such turbid waters, alternative methods estimate intertidal topography from shoreline displacement under different tide levels, typically assuming a uniform slope, which may not adequately represent complex geomorphology. Therefore, in this study, we investigated a method for extracting tidal flats using synthetic aperture radar (SAR) data from Sentinel-1 and ALOS-2, which can be observed even at night and under cloudy weather, and carried out tidal flat mapping in the Ariake Sea. Analysis of data from six years (2020 - 2025) confirmed that Sentinel-1 and ALOS-2 have more opportunities to observe near low tide than the optical satellite Sentinel-2. The estimated distribution of tidal flats was in good agreement with elevation data obtained from field surveys. Furthermore, we demonstrated that intertidal topography can be also estimated by analyzing multi-temporal SAR images acquired at different tidal levels.

Back to poster list

P-43: Deep Learning-Based Multi-Satellite Sea Fog Detection via Precision Co-Registration and Transformer Segmentation using GK2A AMI and GK2B GOCI-II

Presenting author: Jonggu KANG

Authors: Jonggu KANG, Pukyong National University; Yangwon Lee, Pukyong National University

Keywords: Deep Learning; Multi-Satellite; Sea Fog Detection; GK2A; GK2B

Daytime sea fog poses substantial risks to maritime transportation and coastal operations, particularly over Northeast Asia where seasonal fog events are frequent. Although geostationary satellites provide continuous monitoring capability, single-sensor approaches often suffer from spectral limitations and missed detections under complex cloud conditions. This study proposes a multi-sensor deep learning framework that integrates thermal infrared observations from GK2A AMI and visible–near-infrared measurements from GK2B GOCI-II to improve daytime sea fog detection. To ensure geometric consistency between the two platforms, a deep learning-based dense feature matching algorithm (RoMa) was applied for precise co-registration, enabling construction of a six-channel fused input dataset. State-of-the-art semantic segmentation architectures, including both CNN- and Transformer-based models, were systematically evaluated with Bayesian hyperparameter optimization. Among them, the Swin Transformer achieved the best performance, yielding an Intersection over Union (IoU) of 77.2% and an F1-score of 87.2% on independent test scenes. Compared to single-satellite configurations, multi-sensor fusion significantly improved recall, reducing omission errors in operationally critical fog regions. Additional comparisons with currently operational fog products revealed that the proposed approach mitigates both underestimation and overestimation by leveraging complementary spectral characteristics and spatial context learning. Case analyses further demonstrated robust detection of widespread advection fog and sensitivity to faint localized fog events that are difficult to identify in true-color imagery. These findings indicate that precision co-registration combined with Transformer-based multi-sensor segmentation provides a promising pathway toward next-generation operational sea fog monitoring systems.

Back to poster list

P-45: A Remote Sensing Framework for Shallow Bathymetry Retrieval from SAR Using Adaptive FFT Windowing

Presenting author: Dedy ZULKHARNAIN

Authors: Dedy ZULKHARNAIN, Tohoku University; Bruno ADRIANO, Tohoku University; Erick MAS, Tohoku University; Shunichi KOSHIMURA, Tohoku University

Keywords: PALSAR-2, Sentinel-1, Fast Fourier Transform (FFT), Bathymetry

Synthetic Aperture Radar (SAR) has been widely applied in various fields, including disaster assessment and ocean monitoring. One important application is bathymetry mapping, which can provide valuable depth information in coastal regions where in situ survey data are limited or insufficient in spatial resolution. By analyzing swell wave patterns captured in SAR imagery, the dominant wavelength can be estimated and used as input to the linear dispersion relation to derive water depth. However, two main challenges remain: (1) identifying SAR scenes that contain clearly detectable swell patterns, and (2) improving the accuracy of bathymetry estimation. This study first focuses on swell detection to ensure the availability of suitable swell conditions within SAR images. An adaptive Fast Fourier Transform (FFT) windowing approach is then implemented instead of a fixed window size to better capture spatial variations in wavelength from shallow to deeper waters. The window size is dynamically adjusted based on the distance from the shoreline using linear and hyperbolic tangent scaling functions. In addition, a comparative analysis is conducted using ALOS-2 PALSAR-2 (L-band) and Sentinel-1 (C-band) data to evaluate sensor-dependent performance for SAR-based bathymetry. Preliminary results indicate that the hyperbolic tangent adaptive approach achieves the best performance, with a Root Mean Square Error (RMSE) of 5.32 m and a Mean Absolute Error (MAE) of 4.84 m.

Back to poster list

P-47: Feasibility of Arctic Iceberg Monitoring Using NEXTSat-2 and the DINOv3 Foundation Framework

Presenting author: Seung Hee KIM

Authors: Seung Hee KIM, Korea Polar Research Institute

Keywords: NEXTSat-2; Iceberg detection; DINOv3; SAR; Arctic

This study utilizes Korea’s NEXTSat-2 (N2), a small X-Band Synthetic Aperture Radar (SAR) system developed by the Satellite Technology Research Center. The objective is to detect and monitor icebergs in the Arctic region, specifically in western Greenland. We targeted icebergs in Disko Bay, the outlet of the world’s largest iceberg producer, Jakobshavn Glacier. Detecting small and varying-sized icebergs in single-band X-band SAR imagery is traditionally complicated by speckle noise and complex sea-ice clutter. Furthermore, developing automated deep learning systems is hindered by label scarcity, leaving researchers with a very limited number of training samples. To overcome the limitations of small datasets and traditional intensity-based thresholding, our methodology leverages DINOv3, a state-of-the-art self-supervised Vision Foundation Model (VFM). Instead of training a model from scratch, we utilized a frozen DINOv3 backbone to extract high-quality, semantic-aware dense features. To adapt this framework to our limited dataset, we integrated the backbone with modular object detection headers. By employing parameter-efficient fine-tuning techniques, we were able to successfully bridge the domain gap between natural optical images and SAR data, updating only a fraction of the network to prevent overfitting on our small number of samples. This semantic-first architecture efficiently decoupled iceberg signatures from the chaotic ocean background. Through this advanced detection pipeline, we successfully confirmed the feasibility of utilizing the N2 satellite for robust and automated environmental monitoring in challenging Arctic environments.

Back to poster list

P-49: Development of a Satellite-Driven Marine Big Data Framework for Maritime Accident Search and Rescue Support

Presenting author: Tae-Sung KIM

Authors: Tae-Sung KIM, Korea Research Institute of Ships & Ocean Engineering (KRISO); Yongmyung KIM, Korea Research Institute of Ships & Ocean engineering (KRISO); Jae-Jin PARK, Korea Research Institute of Ships & Ocean engineering (KRISO)

Keywords: big data; sentinel-2; numerical model; maritime accident; search and rescue

Efficient maritime search and rescue (SAR) operations require precise environmental intelligence to maximize the probability of success within the golden time. This study develops an integrated big data framework that leverages high-resolution satellite remote sensing, specifically from sensors like Sentinel-2, as the primary data source. By fusing satellite-derived products—including high-resolution reflectance, turbidity, and sea surface temperature (SST)—with numerical modeling and in-situ measurements, we established a multi-dimensional database to support SAR decision-making. While numerical models provide temporal continuity and in-situ data offer ground-truth validation, the satellite-driven approach ensures the wide-area synoptic coverage and high spatial resolution necessary for pinpointing search areas. The framework processes these heterogeneous datasets to generate actionable intelligence, such as estimating survival windows for missing persons and predicting drift trajectories. This research emphasizes the architecture of the big data platform and its capacity to provide enhanced situational awareness for SAR coordinators. The synergy between high-resolution remote sensing and integrated marine big data offers a robust technological basis for rapid and effective emergency response in complex maritime environments.

Back to poster list

P-51: REMOTE SENSING FOR MARITIME SEARCH AND RESCUE IN COASTAL ENVIRONMENTS

Presenting author: Jae-Jin PARK

Authors: Jae-Jin PARK, KRISO; Kyung-Ae PARK, Seoul National University; Tae-Sung KIM, KRISO; Yongmyung KIM, KRISO

Keywords: Remote sensing; Hyperspectral; Search and rescue (SAR); Vessel detection; Ellipse fitting

Over the past decades, maritime accidents have increased due to the continuous growth of maritime transportation and ship traffic, leading to heightened demands for effective maritime safety and emergency response systems. While the detection of accident-prone vessels is essential for preventing secondary incidents, maritime search and rescue (SAR) operations also require the timely and reliable identification of small floating objects associated with emergency situations on the sea surface. Hyperspectral remote sensing offers significant potential for maritime SAR applications by providing wide-area coverage combined with detailed spectral information, which enables improved discrimination of objects under complex and dynamic marine background conditions. In this study, we propose a hyperspectral remote sensing–based framework for maritime search and rescue, focusing on the detection and characterization of vessels and small floating objects in coastal waters. The proposed framework was demonstrated through controlled experiments conducted under realistic marine environments, from which high-resolution hyperspectral image data were acquired. Vessel detection was performed by extracting edge features from hyperspectral imagery and applying an ellipse-fitting approach, resulting in a vessel length estimation error of approximately 0.44 m. In addition, small floating objects were identified using spectral matching techniques supported by a predefined spectral database. To further enhance detection performance in mixed-pixel conditions, spectral unmixing methods were applied, yielding length estimation errors ranging from 0.08 to 0.17 m for representative small maritime targets. With the rapid advancement of hyperspectral remote sensing technologies and imaging sensors, the proposed framework is expected to support future maritime search and rescue operations by improving situational awareness and enabling more effective detection and monitoring of vessels and small floating objects in both coastal and offshore environments.

Back to poster list

P-53: MONITORING COASTAL TURBIDITY AND UNDERWATER VISIBILITY USING IN-SITU SENSORS AND SATELLITE OPTICAL DATA

Presenting author: Yongmyung KIM

Authors: Yongmyung KIM, KRISO; Jae-Jin PARK, KRISO; Tae-Sung KIM, KRISO; Kyung-Ae PARK, Seoul National University

Keywords: Turbidity; Underwater visibility; Sentinel-2; Remote sensing; Monitoring

Recently, the increasing occurrence of maritime accidents around the Korean Peninsula has emphasized the importance of reliable environmental information for coastal monitoring and maritime safety applications. In particular, turbidity and underwater visibility are key parameters for understanding optical conditions in coastal waters, yet their spatial and temporal variability remains difficult to monitor using in-situ observations alone. This study focuses on monitoring coastal turbidity and underwater visibility through the combined use of in-situ measurements and satellite optical data. We analyze the relationship between turbidity and underwater visibility and propose a turbidity-based underwater visibility retrieval algorithm applicable to the western coast of the Korean Peninsula. Field measurements of turbidity and underwater visibility were conducted at two monitoring stations near Daecheon Port using a turbidity sensor and a Secchi disk, while high-resolution satellite images were simultaneously acquired to obtain surface remote reflectance data. Turbidity in the study area was additionally retrieved from Sentinel-2 satellite imagery using reflectance at 560 nm, and its correlation with measured underwater visibility was examined to assess the spatial distribution of underwater visibility in western coastal waters. The results indicate that turbidity values ranged from 2.0 to 3.0 NTU, whereas underwater visibility varied between 1.5 and 3.5 m. Lower underwater visibility was primarily observed in semi-enclosed bays and around islands, while relatively higher visibility (approximately 3.0 m) was detected along the Daecheon Beach shoreline. These results demonstrate the feasibility of monitoring underwater visibility using integrated in-situ sensors and satellite optical observations. The proposed approach provides an effective framework for coastal water quality monitoring and is expected to support maritime safety and search and rescue operations by improving the understanding of optical conditions in dynamic coastal environments.

Back to poster list

P-55: Feasibility study of estimating surface salinity distribution in the Seto Inland Sea using Landsat data

Presenting author: Yuji SAKUNO

Authors: Yuji SAKUNO, Hiroshima Univ.

Keywords: Landsat; salinity; CDOM, bio-optical algorithm; Seto Inland Sea

In recent years, changes in salinity have been said to be one of the causes of sudden red tides and the mass deaths of farmed oysters in the Seto Inland Sea. However, there are no satellite sensors that specifically measure surface salinity in coastal areas. Meanwhile, high-resolution salinity estimation via CDOM has become popular in recent years. In this study, salinity estimation was attempted using Landsat-series data in the Seto Inland Sea immediately after heavy rain, when salinity distribution becomes prominent. As a result, a high correlation was found between the calculated values of Landsat data and surface salinity. Furthermore, the mechanism behind this was examined using a bio-optical algorithm.

Back to poster list

P-57: Lightweight CNN Ship Detection from Geostationary Ocean Color Imagery (GOCI-II)

Presenting author: Won-Kyung BAEK

Authors: Won-Kyung BAEK

Keywords: Ship detection, GOCI-II, Lightweight CNN

Satellite-based ship detection enables wide-area maritime monitoring, especially where AIS is missing, spoofed, or delayed. Although feasibility of threshold-based ship detection has been reported for the Geostationary Ocean Color Imager-II (GOCI-II), AI-driven ship detection from geostationary ocean color imagery remains largely unexplored. Here we assess the practicality of CNN-based ship detection using GOCI-II (250 m) observations over the Yellow Sea by (1) building an AIS-matched training dataset from clear-sky scenes and (2) training a lightweight CNN that explicitly includes a ship-lookalike class to reduce false alarms. On an independent test set, the proposed model achieved a ship recall of ~0.865, precision of ~0.831, and F1-score of ~0.848, indicating robust detectability under the spatial constraints of ocean color imagery. To demonstrate operational-scale applicability, we applied the trained model to cloud-free slot 7 and slot 10 acquisitions on 8 April 2023, enabling large-area ship mapping and revealing coherent linear detection patterns consistent with major shipping lanes. These results suggest that lightweight CNNs combined with AIS-supervised label construction can extend geostationary ocean color missions toward regional maritime surveillance, offering frequent, wide-swath ship monitoring complementary to SAR/optical polar-orbiting sensors.

Back to poster list

P-59: VALIDATION OF ICESAT-2 COASTAL BATHYMETHRY PRODUCT IN SHALLOW TROPICAL WATER REGION

Presenting author: Anisya Feby EFRIANA

Authors: Anisya Feby EFRIANA, Yonsei Univeristy; Hanh Chul JUNG, Yonsei University; Seo Jin KIM, Yonsei University; Kyung Hwan SHIN, Yonsei University; Masita Dwi Mandini MANESSA, University of Indonesia

Keywords: ICESat-2; ATL24; shallow water; bathymetry; accuracy

Shallow water depth information plays a critical role in the safe navigation of waters, as well as the management of the coasts and marine ecosystem conservation. The coastal areas are among the most complex systems, especially in tropical regions like Indonesia, where coral reefs, seagrass beds, and other benthic habitats coexist alongside intensive human activities. However, bathymetric data collection using ship surveys is often limited by high operational costs and cannot safely approach shallow coastal water. The recent development of the Advanced Topographic Laser Altimeter System Level-3A Coastal and Nearshore Bathymetry product (ATL24) from the Ice, Cloud, and Land Elevation Satellite-2 (ICESat-2) mission opens new avenues for the retrieval of shallow water depths with spaceborne photon-counting lidar. ATL24 provides automated global coastal and nearshore bathymetry in optically clear waters up to approximately 40 m. It is derived from the ICESat-2 green laser (532 nm), which can penetrate a water surface and return photons reflected by the seabed. It also provides useful depth information in areas where data are limited. ATL24 validation in intricate tropical reef environments is still uncommon despite its increasing use. This study used field bathymetric data obtained using Single Beam Echo Sounder (SBES) measurements with 5167 points as reference data to assess the accuracy and usability of ATL24 produced from ICESat-2 in the waters surrounding Derawan and Panjang Island, East Kalimantan, Indonesia. The depths data were then adjusted for tidal variations and harmonized to a consistent vertical datum. The ATL24 bathymetric points were then matched with the SBES measurements and statistical performance metrics. The results of this assessment are expected to provide insights into the applicability of ATL24 in tropical reef environments.

Back to poster list

P-61: Assessing Thermal Discharge Impacts from the Hanbit Nuclear Power Plant on Coastal Marine Environment Using GOCI-II Satellite Observations

Presenting author: Sungjae PARK

Authors: Sungjae PARK, Kangwon National University; Eunseok PARK, Kangwon National University; Bongchan KIM, Kangwon National University; Chang-Wook LEE, Kangwon National University

Keywords: GOCI-II, Thermal Discharge, Sea Surface Temperature (SST), Chlorophyll-a, Coastal Marine Environment

This study investigates the potential environmental impacts of thermal discharge from the Hanbit Nuclear Power Plant on the surrounding coastal waters using Geostationary Ocean Color Imager-II (GOCI-II) satellite products. Sea Surface Temperature (SST), Chlorophyll-a (Chl-a), Total Suspended Sediments (TSS), and Red tide Index (RI) were analyzed within concentric zones of 5 km, 10 km, and 20 km radii centered on the power plant. Daily mean satellite-derived values were compared with in-situ intake and discharge water temperature records provided by the plant operator. Time-series analyses were conducted to examine spatial and temporal variations across zones, and lag-correlation analysis was applied to evaluate potential delayed biological or optical responses to thermal anomalies. The inclusion of TSS allowed discrimination between biologically driven Chl-a variability and turbidity-induced optical effects, while RI was used to identify potential harmful algal bloom (HAB)-related events. Results indicate clear seasonal variability in SST and Chl-a, with localized thermal signals observed near the discharge zone. Statistical comparisons suggest that short-term thermal fluctuations may influence phytoplankton dynamics under specific environmental conditions, although turbidity and external forcings also contribute significantly to observed variability. This integrated satellite-based approach demonstrates the applicability of geostationary ocean color observations for continuous monitoring of anthropogenic thermal impacts on coastal marine ecosystems and provides a framework for assessing ecological responses near coastal power facilities.

Back to poster list

P-63: Assessing Stratus Cloud Variability and SST Coupling Before and After IMO 2020 using Sentinel-2

Presenting author: Hyojin KIM

Authors: Hyojin KIM, Pukyong National Univ.; Donggyu KIM, Pukyong National Univ.; Jaehoon JEONG, Pukyong National Univ.; Seungchan LIM, Pukyong National Univ. ; Chuluong CHOI, Pukyong National Univ.

Keywords: IMO 2020; Stratocumulus; Sea Surface Temperature (SST); Sulfur Oxides (SOx); Sentinel-2

The IMO 2020 regulation may have weakened cloud condensation nucleus (CCN) supply by reducing sulfur oxides (SOx) from ships and secondary sulfate aerosols. This study evaluates changes in SOx emissions from ships in the waters around the Korean Peninsula, using the regulation’s implementation as a baseline, based on satellite observations of stratus cloud characteristics and sea surface temperature (SST) variability. It compares the pre-regulation period (2018-2019) and post-regulation period (2020-2025) for Sentinel-2 MGRS tile 52SED, which includes Busan Port, a major port in South Korea. A stratocumulus influence index is calculated for the marine area based on the SCL from Sentinel-2 Level-2A data. This is aggregated at the patch level to mitigate pixel noise while preventing excessive smoothing of the area. Outliers are identified through periodic component (harmonic) modeling and checked using robust statistics-based diagnostics. Ship activity is represented by an AIS-based traffic intensity index. ERA5 hourly data on single levels and GHRSST MUR v4.1 SST data are used for radiative and SST coupling analysis. In marine areas, SCL-based stratus cloud impact indices are calculated. By using the same zones as AIS vessel traffic frequency, the association between ship-sourced sulfur oxides and stratus clouds is established. Furthermore, robust statistics-based diagnostics are applied to control for variability and outlier effects, enabling comparisons between periods before and after IMO 2020. This study presents an integrated analytical framework to quantitatively verify, using optical observation satellites, the chain reaction that ship SOx reduction under the IMO 2020 regulation can lead to cloud–radiation–SST effects in port areas.

Back to poster list

P-65: Integrating Remote Sensing and Habitat Modeling to Quantify Social–Ecological Impacts of Artificial Mangrove Expansion

Presenting author: Su-Fen WANG

Authors: Su-Fen WANG, National Changhua University of Education; Meng-Fen LEE, National Changhua University of Education

Keywords: InVEST Model, Mangrove Expansion, Habitat Degradation, Coastal Area, Mudflats

Adopting a social–ecological systems (SES) perspective, this study uses remote sensing and habitat modeling to investigate how policy-driven artificial mangrove expansion has reshaped intertidal landscapes and ecosystem functions along the Fangyuan coast of Taiwan. Since 1983, state-led afforestation programs have transformed natural mudflats into mangrove-dominated landscapes, reflecting cross-scale interactions between economic policy and coastal ecology. By integrating multi-temporal satellite imagery, the InVEST Habitat Quality model, and intertidal crab survey data, we analyze landscape transitions and habitat dynamics from 1987 to 2019. Results indicate that mudflats at the Erlin Creek estuary declined by 34.46 ha (19.4%), largely replaced by planted mangroves. Between 2000 and 2009, habitat quality decreased by 12.84%, coinciding with a threefold increase in mangrove area. Spatial analysis shows that mangroves preferentially colonized high-stress mudflats (habitat quality < 0.5), while generating spillover degradation effects on adjacent uninvaded areas (–6.8%). Alteration of sediment structure by dense root systems reduced macrobenthic diversity and waterbird foraging habitats, negatively affecting endemic species such as Xeruca formosensis. These findings reveal a social–ecological trade-off: carbon-oriented afforestation policies enhance selected ecosystem services but undermine the multifunctionality and biodiversity of open mudflat systems. We argue that sustainable coastal governance requires differentiated mangrove management and targeted mudflat restoration to balance ecosystem services within dynamic SES feedback processes.

Back to poster list

P-67: A Deep Learning Approach for Remote Sensing-Based Estimation of PM2.5 Concentration in Urban Areas: A Case Study of Greater Taipei Area

Presenting author: Vincent Cletus WISO

Authors: Vincent Cletus WISO, National Central University; Hsuan REN, National Central University; Tang-Huang LIN, National Central University; N/A

Keywords: PM2.5, Spatiotemporal, Temporal Convolutional Network, Long Short-Term Memory

Air pollution remains a threat, with particulate matter being one of the several key indicators of pollution. Numerous studies show that particulate matter of diameter 2.5 µm or smaller (PM2.5) impacts public health, hence the need for more accurate models. Traditional numerical methods fail to capture pollutants’ spatiotemporal patterns; however, neural networks can learn from environmental data inputs to obtain correlations. In this study, a Temporal Convolution Network – Bi-Long Short-Term Memory (TCN – Bi-LSTM) hybrid model is proposed. TCN- Bi-LSTM is for PM2.5 monthly estimation in 2023 using Taiwan’s Ministry of Environment air quality historical data from July 2015 to December 2022 – including weather parameters and land use data – to produce a 1km PM2.5 map via k nearest neighbors – inverse distance weighting (KNN – IDW). The former model captures causal patterns while the latter handles long-range temporal dependencies. The key findings of the research study are twofold. One is a comparison test of our proposed model against classical models: Random Forest, Long Short-Term Memory (LSTM), and Convolution Neural Network (CNN). Here, the proposed model led with a metric evaluation of R² = 0.796 and root mean square error (RMSE) = 2.161. Finally, our PM2.5 concentration maps for 2023 were verified using Copernicus Atmosphere Monitoring Service (CAMS) global reanalysis data and Civil IoT Taiwan data. Here, the proposed model maps had an anomaly correlation of 0.868 with the former and a metric evaluation of R² = 0.584 and RMSE = 1.975 with the latter. This study’s findings contribute to the bulk of air quality monitoring research. The PM2.5 concentration map helps mitigate the missing data issue, mainly due to cloud cover on satellite observations during aerosol optical depth retrieval. Future work will focus on model generalization for use in an urban area with similar characteristics to the Greater Taipei Area.

Back to poster list

P-69: Spatial Assessment of Future Heat Stress Exposure in National Parks of the Republic of Korea under CMIP6 SSP Scenarios Using WBGT

Presenting author: Jinhoo HWANG

Authors: Jinhoo HWANG, Korea University; Youngjae YOO, Korea University; Yoonji KIM, Korea University; Sangwook LEE, Korea University, Seongwoo JEON, Korea University

Keywords: Extreme heat; Spatial analysis; Climate risk; Human thermal comfort; Visitor safety management

This study quantitatively evaluates the impacts of climate change–induced increases in the intensity and frequency of heat waves on the exploration environment of national parks in the Republic of Korea. Using CMIP6 Shared Socioeconomic Pathway (SSP) climate change scenarios (SSP126–SSP585), the Wet Bulb Globe Temperature (WBGT) index was calculated to assess heat stress exposure and vulnerability across national parks. A physics-based WBGT calculation method proposed by Liljegren (2008) was applied to daily gridded climate data, including air temperature, relative humidity, wind speed, and solar radiation. WBGT time series were constructed for each national park for the period 2021–2100 to compare changes in thermal conditions across scenarios and future periods. Thermal stress characteristics were analyzed by distinguishing between changes in mean thermal conditions and extreme heat exposure, using the number of days with WBGT exceeding 30 °C as a key indicator of heat stress risk. The results indicate that while average WBGT values generally increase across all national parks, differences among SSP scenarios remain within approximately 1 °C. In contrast, the frequency of extreme heat exposure, measured as the number of days exceeding the WBGT threshold of 30 °C, shows substantial variation across both emission scenarios and geographic regions. Under the high-emission SSP585 scenario in the late 21st century (2071–2100), coastal, island-type, and southern lowland national parks exhibit a pronounced increase in heatwave exposure days, whereas alpine and inland national parks display relatively lower vulnerability. These findings suggest that heat-stress vulnerability in national parks is driven primarily by increases in the frequency of extreme heat events rather than by changes in average thermal conditions. This study demonstrates that WBGT-based, scenario-driven spatial assessments can provide a robust scientific basis for heatwave-responsive visitor safety management and climate change adaptation strategies in national parks.

Back to poster list

P-71: CLOUD REMOVAL USING BI-TEMPORAL OPTICAL IMAGERY AND GENERATIVE DEEP LEARNING

Presenting author: Soyeon PARK

Authors: Soyeon PARK, Inha Univ.; Inseon LEE, Inha Univ.; Eui Ho HWANG, K-water Research Institute; No-Wook PARK, Inha Univ.

Keywords: Cloud removal; Deep learning; Image reconstruction; Optical imagery; Sentinel-2

Cloud contamination poses a major challenge to the construction of continuous time-series optical satellite imagery, frequently resulting in information loss and reducing the reliability of environmental monitoring applications. Accordingly, cloud removal techniques that reconstruct cloud-contaminated regions serve as an effective strategy for improving the availability of optical satellite imagery. Cloud removal aims to restore a cloudy image to a cloud-free image by estimating surface reflectance within cloud-contaminated regions. Previous studies have demonstrated that multi-temporal cloud removal approaches outperform their single-image spatial counterparts. From a methodological perspective, deep learning models, such as generative adversarial networks and diffusion models, have been widely adopted to model complex nonlinear temporal relationships. However, learning-based methods commonly rely on multi-temporal training image pairs, which introduces temporal inconsistencies that hinder predictive performance. To address these issues, this work proposes a deep learning-based cloud removal framework that spans from training image construction to post-processing for discontinuity elimination. Synthetic cloudy images are first generated by compositing real cloud patterns onto cloud-free images to ensure temporally consistent training pairs. Furthermore, an additional cloud-free image acquired at a different date is utilized as auxiliary information to compensate for missing surface information. The deep learning model incorporates three inputs: a synthetic cloudy image at the prediction date, a cloud free image at a reference date, and a corresponding cloud mask. After reconstructing the cloud-contaminated regions, Poisson blending is applied to eliminate discontinuities around cloud boundaries. The potential of the proposed framework is demonstrated via cloud removal experiments using multi-temporal Sentinel-2 images acquired in 2022 and 2023 over an agricultural area in Gimje, Korea. The experimental results indicate that the proposed framework provides a practical and flexible solution for optical image reconstruction under cloud contamination, offering strong potential for Earth observation tasks in regions where the availability of cloud-free optical imagery is limited.

Back to poster list

P-73: Regression-based error refinement for cloud-contaminated regions reconstructed via SAR-to-optical image translation

Presenting author: Inseon LEE

Authors: Inseon LEE, Inha Univ.; Soyeon PARK, Inha Univ.; Jaehwa JANG, Inha Univ.; Eui Ho HWANG, K-water Research Institute; No-Wook PARK, Inha Univ.

Keywords: SAR-to-optical image translation, Cloud removal, Residual regression, Boosting, Error correction

Time-series optical satellite imagery is widely used for environmental monitoring and change detection. However, since optical imagery is severely affected by atmospheric conditions at the time of acquisition, constructing continuous optical image time series remains challenging due to cloud contamination. To address this issue, SAR-to-optical image translation has been developed to reconstruct optical imagery in cloud-contaminated regions by combining synthetic aperture radar (SAR) imagery, which is insensitive to weather conditions, with generative deep learning models. Nevertheless, reconstructed cloud-contaminated regions often suffer from substantial errors, highlighting the need for additional error refinement procedures. Previous studies on error correction have primarily employed deterministic regression models, which often fail to account for the complex error structures inherent in reconstructed imagery. In spectrally homogeneous landscapes, such as croplands, reconstruction errors tend to exhibit brightness-dependent biases and heteroscedastic variability across the reflectance range rather than strong local contrasts. Consequently, it is imperative to evaluate and compare regression frameworks that can explicitly model reflectance-dependent error variability for identifying the most robust refinement strategy. In this work, we systematically compare multiple regression-based refinement strategies, including Random Forest (RF), Extreme Gradient Boosting (XGBoost), Natural Gradient Boosting (NGBoost), and Quantile Regression (QR). Both reflectance-level regression and residual regression schemes are evaluated within a unified experimental framework. Furthermore, SAR-derived local features, originally used in the SAR-to-optical translation stage, are incorporated as explanatory variables to investigate their contribution to error refinement. These local SAR features encode structural and spatial context information, enabling the models to better capture systematic reconstruction biases associated with land-cover characteristics. The performance of each regression framework is evaluated using Cosmo-SkyMed SAR and PlanetScope optical imagery acquired over an agricultural region in Gimje, Korea. Experimental results indicate that reflectance-level regression and residual-based regression exhibit similar refinement performance. Boosting-based methods generally demonstrate superior error reduction compared to RF, particularly under reflectance-dependent error conditions. NGBoost and QR additionally provide distributional information that characterizes heteroscedastic uncertainty across brightness ranges. Incorporating SAR-derived local features consistently improves accuracy and spatial consistency across models. These findings suggest that regression-based refinement, particularly when combined with boosting frameworks and SAR local features, serves as an effective and robust post-processing strategy for enhancing cloud-contaminated optical image reconstruction via SAR-to-optical translation.

Back to poster list

P-75: Enhancing the Spatial–Spectral Fusion of Multispectral and High-Resolution Optical Satellite Imagery for Spectral Enrichment of Fine-Scale Observations

Presenting author: Yeseul Kim

Authors: Yeseul KIM, Korea Aerospace Research Institute; Gwan-Young OH, Korea Aerospace Research Institute

Keywords: Spatial-Spectral Fusion; Spectral Enhancement; High-Resolution Imagery; Multisensor Data; Qualitative Validation

Spatial–spectral data fusion aims to integrate complementary information from multi-sensor satellite observations to overcome the inherent trade-off between spatial and spectral resolution. While previous studies have often relied on synthetic datasets that allow direct quantitative comparison with reference data, validation using real satellite imagery poses additional challenges due to the absence of spectrally matched ground truth. This study investigates the practical applicability of spatial–spectral fusion using actual satellite observations, with a particular focus on spectrally enriching high-resolution optical imagery. In this work, multispectral satellite data providing rich spectral information at moderate spatial resolution are fused with high-resolution optical imagery characterized by fine spatial detail but a limited number of spectral bands. The proposed fusion framework exploits cross-sensor spectral relationships derived from the multispectral observations to reconstruct spectrally enhanced reflectance information at the spatial scale of the high-resolution imagery. This approach is designed to infer spectral characteristics that are not directly observed by the high-resolution sensor, such as narrowband spectral responses. Due to the absence of corresponding spectral bands in the high-resolution imagery, direct pixel-wise quantitative comparison with reference data is not feasible. Instead, the fusion results will be evaluated through qualitative visual analysis and indirect validation strategies, including spectral trend consistency, spatial pattern coherence, and comparative analysis against spatially resampled multispectral observations. These analyses are intended to demonstrate that spatial–spectral fusion can be meaningfully applied to real multi-sensor satellite data even in the absence of explicit spectral ground truth. The proposed framework aims to provide a practical pathway for extending the analytical utility of high-resolution optical satellite imagery and to support advanced Earth observation applications where both spatial detail and enhanced spectral information are required.

Back to poster list

P-77: Detecting Inshore Objects Occupying Public Waters Using SAR Dual-polarization

Presenting author: Seung Jun Lee

Authors: Seung Jun Lee, Yonsei Univ.; Hahn Chul Jung, Yonsei Univ.; Do-Hyun Hwang, Yonsei Univ.; Seo Jin Kim, Yonsei Univ.; Gun Hee Song, Yonsei Univ.

Keywords: SyntheticApertureRadar; Dual-polarization, TargetDetection, PublicWaters

South Korea has complex coastal environments, with numerous islands forming an extensive coastal environment in which vast areas of public waters are developed. As more than 20% of the national population resides in coastal cities, the occupation and use of public waters for various purposes are actively carried out nationwide. To ensure public safety and protect the coastal environment, the Korean Ministry of Oceans and Fisheries conduct coastal surveys to continuously monitor coastal facilities including breakwaters, bridges, water architectures and other marine structures. However, field-based monitoring requires substantial time and labor due to the limited accessibility of island regions and the extensive and complex nature of South Korea’s coastal environment. To address these limitations, satellite-based SAR imagery, which can acquire images regardless of the weather conditions, has been explored for efficient monitoring of coastal facilities over wide areas. Nevertheless, practical application remains challenging due to constraints such as land clutter, limited swath width, and insufficient temporal resolution. In this study, we propose a method for high-temporal-resolution monitoring of coastal facilities using dual-polarization information from time-series Sentinel-1 SAR imageries. Target characteristic parameters were employed for facility detection, and a total of 183 Sentinel-1 SAR images acquired over the Busan region from February, 2017 to December, 2023 were used for analysis. The proposed method enabled coastal facility monitoring at 12-day intervals and demonstrated overall high detection performance. However, missed detections occurred for some facilities due to the spatial resolution limitations of Sentinel-1 SAR imagery and the morphological characteristics of coastal structures. Future work is expected to further improve detection performance by incorporating high-resolution SAR data and multi-angle observations.

Back to poster list

P-79: Development of an Algorithm for Estimating Near-Surface Specific Humidity for GPM/GMI -Effectiveness of high-Frequency Channels-

Presenting author: Yuta KYONO

Authors: Yuta KYONO, Hokkaido Univ.; Hiroyuki TOMITA, Hokkaido Univ.

Keywords: Near-Surface Specific Humidity; Microwave Radiometer; 183 GHz Channel; GPM/GMI;

Near-surface specific humidity (Qa) over the ocean is an essential parameter for calculating latent heat flux and understanding air-sea energy exchange. Space-borne microwave radiometers widely used to estimate Qa using a statistical relationship between brightness temperature (Tb) and Qa (Tb-Qa relationship), which varies with the vertical water vapor structure. However, conventional microwave radiometers are limited by their available frequency bands. Consequently, current algorithms do not sufficiently account for variations in the Tb-Qa relationship arising from differences in the vertical water vapor structure. To address this issue, current methods incorporate daily grid data from atmospheric reanalysis as auxiliary information such as water vapor scale height (Hv). This reliance, however, can introduce spatio-temporal discrepancies and uncertainties. The Global Precipitation Measurement (GPM) Imager (GMI) receives high-frequency microwaves, enabling the acquisition of information on upper-level water vapor and capturing the vertical water vapor structure. This study aims to develop and evaluate the first Qa retrieval algorithms designed for the GMI. We examined two classification indices that account for the vertical water vapor structure: the conventional Hv and a newly proposed proxy, dual-bands vertical humidity-index (DVHI), defined as the ratio of Tb at 23v GHz to 183±8v GHz (Tb_23v / Tb_183±8v). We developed multiple regression equations using Tb (19v/h, 23v, 37v/h, 89v/h GHz) and precipitable water as the explanatory variables. Independent in-situ validation confirmed the Hv-based method showed comparable performance to the existing satellite sensor (AMSR2) with superior bias. Our proposed DVHI-based method was nearly equivalent to the Hv-based method, despite slightly lower accuracy. The latter results indicate that using the high-frequency channel of GMI enables comparable estimation with the microwave radiometer alone. Furthermore, this method is applicable to the new sensor (AMSR3), featuring high-frequency channels similar to those of the GMI. This extensibility ensures our new findings provide more comprehensive understanding of the global air-sea energy exchange.

Back to poster list

P-81: Correction of Elevated Road Distortion in Optical Satellite Orthoimagery using Integrated Elevation Models

Presenting author: Chien-Liang LIU

Authors: Chien-Liang LIU, National Central University; Wen-chi CHANG, National Central University

Keywords: Orthorectification; Relief Displacement; Digital Surface Model (DSM); Ultra-High-Resolution (UHR) Imagery; Elevated Roads.

Standard orthorectification of optical satellite imagery typically relies on Digital Terrain Models (DTMs) for elevation data. However, DTMs represent the bare earth surface and exclude the elevation information of man-made structures, such as elevated roads and viaducts. Consequently, relief displacement occurs, causing these structures to appear shifted in the resulting orthophotos—an effect that intensifies with larger off-nadir viewing angles. In flat terrain, this leads to significant positional errors; in rugged terrain, it causes originally smooth road alignments to appear distorted. This phenomenon is particularly pronounced in ultra-high-resolution (UHR) imagery, where the increased spatial detail makes these geometric discrepancies significantly more apparent. This study proposes a method to mitigate these artifacts by integrating elevation data. We utilize road vector maps to extract elevation information for elevated roads from Digital Surface Models (DSMs), which is then merged with the underlying DTM. By using this integrated elevation model for orthorectification, the relief displacement of elevated road structures is effectively corrected, ensuring better alignment with road vector maps. The proposed methodology has been successfully implemented in the Multi Sensor Geocoded Processing System (MSGPS) developed by our center and is currently employed to generate high-quality orthoimagery across Taiwan.

Back to poster list

P-83: Comparative Analysis of R-CNN-Based Detection Performance for Vehicle in High-Altitude UAV Imagery

Presenting author: Seungchan LIM

Authors: Seungchan LIM, Pukyong National Univ.; Donggyu KIM, Pukyong National Univ.; Hyojin KIM, Pukyong National Univ.; Jaehoon JEONG, Pukyong National Univ.; Chuluong CHOI, Pukyong National Univ.

Keywords: UAV; Vehicle Detection; R-CNN; AP; AR; IoU

This study aims to detect vehicles in unmanned aerial vehicle (UAV) imagery using Region-based Convolutional Neural Network (R-CNN)–based object detection models and to quantitatively investigate performance differences according to model architecture and backbone networks. The dataset was collected using a DJI Mavic 3 Enterprise at flight altitudes ranging from 120 to 150 meters. Due to the high acquisition altitude, vehicles occupy only a limited number of pixels. These characteristics pose significant challenges for small-object detection tasks. The original dataset contained 16,474 annotated vehicle instances. However, direct training on full-resolution images led to excessive memory consumption. To address this issue and enhance small-object representation, the images were divided into tiles with approximately 20% overlap. After tiling, the number of vehicle instances increased by approximately 80%, resulting in a total of 29,852 annotated objects. The dataset was split into training and validation sets at a ratio of 75:25 for model development and evaluation. For comparative analysis, six experimental configurations were constructed by combining three detection architectures—Faster R-CNN, Mask R-CNN, and Cascade R-CNN—with two backbone networks, ResNet-50 and ResNet-101. Model performance was evaluated using Average Precision (AP) and Average Recall (AR). AP represents the mean precision computed across multiple Intersection over Union (IoU) thresholds. AR reflects the model’s ability to retrieve object instances under a fixed maximum number of detections per image. All experiments were conducted under identical training and evaluation settings to ensure a fair comparison. This study provides practical insights into selecting appropriate R-CNN variants and backbone depths for small-vehicle detection in high-altitude UAV imagery. The findings contribute to the advancement of drone-based traffic monitoring and aerial surveillance applications.

Back to poster list

P-85: Benchmarking Crack Segmentation Models on Public Datasets with UAV Field Tests and Imaging Conditions

Presenting author: Jaehoon JEONG

Authors: Jaehoon JEONG, Pukyong National Univ.; Seungchan LIM, Pukyong National Univ.; Hyojin KIM, Pukyong National Univ.; Donggyu KIM, Pukyong National Univ,; Chuluong CHOI, , Pukyong National Univ.

Keywords: UAV; pavement crack; segmentation; benchmarking; GSD; domain gap

Automated detection of pavement cracks supports timely maintenance decisions, yet conventional inspection vehicles and specialized survey equipment are costly and difficult to deploy frequently. This study aims to establish a practical, low-cost UAV–AI framework for pavement crack segmentation by (1) benchmarking representative deep segmentation models under a unified protocol and (2) deriving an effective UAV acquisition and processing workflow for real-road deployment. We will train four segmentation models—U-Net++ (ResNet backbone), DeepLabV3+ (ResNet backbone), CrackMamba, and Mask2Former—using six public crack segmentation datasets (denoted A–F). To enable fair comparison, we will standardize input resolution, data splitting (train/validation only), optimization settings, threshold selection, and post-processing. Data augmentation will be applied during training to emulate field variability such as illumination changes, shadows, stains, lane markings, and textured asphalt. Model performance will be assessed on the validation sets using common segmentation metrics (precision, recall, F1-score, crack-class IoU, mIoU, and Dice) together with efficiency indicators (parameter count, computational cost, and inference speed). For field verification, we will acquire UAV imagery over real road sections and manually create pixel-wise ground-truth masks to build an external test set. By applying models trained on public datasets to UAV imagery, we will quantify the domain gap between in-dataset validation and real-world performance and analyze typical failure modes. We will further examine sensitivity to acquisition conditions, including flight altitude, ground sampling distance (GSD), and image overlap, to identify UAV settings that balance accuracy and operational efficiency. The expected outcome is an end-to-end UAV crack inspection framework with evidence-based guidance on model selection and UAV imaging conditions for practical road maintenance applications.

Back to poster list

P-87: Development of a SAR K-ARD Module to Standardize KOMPSAT Satellite Imagery

Presenting author: Su-Bin HA, Pukyong National University

Authors: Su-Bin HA, Pukyong National University; Ye-Young KIM, Pukyong National University; Gwang-Woo OH, Pukyong National University; Dochul YANG, Korea Aerospace Research Institute; Seung-Kuk LEE, Pukyong National University

Keywords: Analysis Ready Data; K-ARD; KOMPSAT; Synthetic Aperture Radar; CEOS

The Republic of Korea plans to develop and operate 18 national satellites by 2026 and up to 70 by 2030. In this context, standardizing imagery products from each satellite and providing Analysis Ready Data (ARD) that can be used without additional preprocessing is essential. ARD enables experienced users to minimize redundant preprocessing and allows non-experts to easily extract physical properties such as surface reflectance and radar backscatter intensity from satellite imagery. To achieve this, the establishment of a Korean ARD (K-ARD) system is crucial. This system will standardize spatial coverage, coordinate systems, grid frameworks, data formats, and product quality levels, ensuring a structured approach to design, development, validation, and operation. K-ARD targets both existing satellites, such as the KOMPSAT series, and upcoming low Earth orbit satellites, aiming to meet the ARD specifications set by the Committee on Earth Observation Satellites (CEOS). This will help reduce user costs and processing time, improve the accuracy of results, and significantly enhance the usability of Korean satellite data. For SAR data products, K-ARD will support the development of various standardized datasets, including Normalized Radar Backscatter (NRB), Ocean Radar Backscatter (ORB), Polarimetric Radar (POL), Interferometric Products (InSAR), and Geocoded Single-Look Complex (GSLC). All data products will meet minimum performance requirements and, where possible, adhere to target specifications. The system will be designed with scalability in mind, initially focusing on the Korean Peninsula and later expanding to global coverage.

Back to poster list

P-89: Reversible Compression Framework for Streaming 3D-LiDAR Point Clouds Using Multi-Layer Range Image Processing

Presenting author: Masafumi NAKAGAWA

Authors: Masafumi NAKAGAWA, Shibaura Institute of Technology

Keywords: 3D-LiDAR, Streaming Point Clouds, Point Cloud Compression, Real-Time Processing, Mobile Mapping System

LiDAR-SLAM and 3D object recognition widely utilize 3D LiDAR. However, processing streaming point clouds acquired over long periods generates enormous data volumes, creating challenges for storage and transmission. Existing size-reduction approaches include lossy compression such as ZIP, background subtraction, and point cloud reduction. Compression based on background subtraction is unsuitable for mobile mapping, while point cloud reduction decreases geometric fidelity and visualization quality in exchange for smaller data size. To reduce data size without degrading positional accuracy, geometric fidelity, or visualization quality, international standards such as Video-based Point Cloud Compression (V-PCC) and Geometry-based Point Cloud Compression (G-PCC), developed by the Moving Picture Experts Group (MPEG), have been proposed. V-PCC prioritizes visualization quality by projecting point clouds onto 2D images and encoding them using video codecs such as H.265/HEVC or H.266/VVC. In contrast, G-PCC directly compresses 3D geometry using techniques such as octree structures and predictive coding, thereby preserving geometric accuracy. Although real-time compression and transmission over commercial networks have been demonstrated, V-PCC may degrade positional accuracy in 3D mapping of streaming LiDAR data, while G-PCC suffers from limited processing efficiency. Therefore, this study proposes a high-speed compression framework that combines V-PCC and G-PCC through multi-layer range image processing to preserve positional accuracy, geometric fidelity, and visualization quality. A waterborne mobile mapping point cloud dataset was used for evaluation. The horizontal scan LiDAR data (VLP32C, Velodyne, 3000 frames, 131 million points) and the oblique scan LiDAR data (VLP16, Velodyne, 3000 frames, 48 million points) were encoded using AVI (uncompressed) and MJ2 in both lossless and lossy modes. The encoding time was approximately 20–25% of the measurement time. The results demonstrate that the proposed framework achieves real-time compression of massive 3D-LiDAR streaming point clouds.

Back to poster list

P-91: A Framework for Quality Classification and Enhancement of MMS Image Data

Presenting author: Ryunghyeok IM

Authors: Ryunghyeok IM, Korea Research Institute for Human Settlements; Kyongah CHOI, Ryunghyeok IM, Korea Research Institute for Human Settlements

Keywords: MMS; NR-IQA; Crowdsourcing; ImageEnhancement; SVM

Ground-based Mobile Mapping System (MMS) imagery is a critical data source for road asset management and geospatial production. However, real-world driving conditions—such as abrupt illumination changes in tunnels, motion blur from vibrations, and sensor noise—frequently degrade image quality. These degradations hinder downstream analyses, including road-region estimation and facility segmentation, thereby limiting the practical usability of MMS data. This study proposes a robust quality enhancement framework that integrates No-Reference Image Quality Assessment (NR-IQA)-based diagnosis with task-oriented usability evaluation. Experiments were conducted on 3,551 MMS images. To quantitatively characterize quality variations, multiple NR-IQA indicators—including Laplacian Variance, Tenengrad, Entropy, Histogram Spread, and BRISQUE—were computed to capture sharpness, information content, and perceptual quality. To link low-level indicators with operational utility, the road-region ratio was utilized as a Region of Interest (ROI) indicator. Using NR-IQA feature vectors, a Support Vector Machine (SVM) binary classifier was trained to predict image suitability for subsequent processing. The SVM accurately identified unsuitable frames, achieving a precision of 0.9715, a recall of 0.9852, and an overall accuracy of 0.9634. Following this diagnosis, a deep learning–based enhancement method was applied to images classified as low-quality. This refinement stage significantly improved quality in challenging scenes, such as tunnel transitions and overexposed conditions. Through this process, 247 out of 362 initially unsuitable frames were restored to suitable quality. Consequently, the valid data ratio for the entire dataset increased from 89.97% to 97.04%, representing a 7.07 percentage point improvement. These findings demonstrate that the proposed framework offers an actionable strategy for quality management and usability improvement in real-world MMS processing pipelines.

Back to poster list

P-93: Retrieval-Augmented Multimodal Learning for Robust Plant Disease Classification

Presenting author: Yuan-Chia CHAN

Authors: Yuan-Chia CHAN, Academy of Circular Economy, National Chung Hsing University; Ming-Der YANG, Department of Civil Engineering, National Chung Hsing University

Keywords: Plant disease diagnosis; multimodal learning; retrieval-augmented inference; precision agriculture; deep learning

Effective plant disease diagnosis requires robustness against long-tailed class distributions and noisy or unreliable user inputs. This study presents SMART (Structured Multimodal Agricultural Retrieval-augmented Transformer), a novel framework designed for real-world deployment by integrating semantically precise linguistic guidance with retrieval-augmented inference. Through rigorous evaluation on 37,586 field images covering 48 disease categories, this study challenges the common assumption that more complex textual descriptions necessarily improve multimodal learning. Results demonstrate that elaborate LLM-generated narratives essentially introduce semantic noise, while structured linguistic representations consistently outperform them in large-scale diagnostics. Benchmarking against 22 contemporary general foundation models reveals a clear specialization gap in fine-grained agricultural diagnostics, which SMART effectively bridges, achieving near-perfect classification performance (top-1 accuracy > 0.99) and surpassing state-of-the-art CNN baselines by a substantial margin (F1 > 0.97 versus approximately 0.90). Crucially, SMART incorporates a retrieval-augmented inference mechanism based on K-nearest neighbor soft voting, providing fault-tolerant robustness when textual inputs are incorrect, incomplete, or entirely missing, improving mean average precision by 16.8% to 29.1% under erroneous input conditions. By resolving both visual ambiguity and long-tailed distribution challenges, SMART supports a practical human-in-the-loop diagnostic workflow and establishes a foundation for integration into field-scale precision agriculture systems.

Back to poster list

P-95: Monitoring Spatio-temporal Land Cover Changes in Kangerlussuaq, Greenland, Using DeepLab v3+ and Time-series Satellite Imagery

Presenting author: Sungjae LEE

Authors: Sungjae LEE, Kore Polar Research Institute; Seung Hee KIM, Kore Polar Research Institute

Keywords: Deeplab v3+, Greenland, Arctic, Landsat-8/9

Environmental shifts in the Arctic driven by climate change have emerged as a significant global concern; in particular, the retreat of the Greenland Ice Sheet and the subsequent expansion of vegetation are exerting profound impacts on the Earth’s environmental systems. Over the past 30 years, Greenland has experienced a substantial glacial retreat of approximately 28,707 km2, accompanied by a 111% increase in vegetation cover and rising meltwater discharge. These transitions can accelerate permafrost degradation and increase greenhouse gas emissions by altering surface albedo and permafrost infiltration rates. While traditional remote sensing techniques face limitations in accurately detecting and classifying complex, fine-scale vegetation changes, more precise and automated analytical methods are increasingly necessitated. In this study, we employed DeepLab v3+, a state-of-the-art deep learning model, to classify land cover in the Kangerlussuaq region and monitor change patterns using long-term time-series data. The dataset was constructed using Landsat 8/9 satellite imagery with less than 10% cloud cover, preprocessed through pansharpening to achieve a spatial resolution of 15 meters. The model was trained to classify seven distinct land cover classes, including glaciers, vegetation, reservoirs, and sand. To enhance feature extraction, multiple spectral indices such as NDVI and NDWI, along with band ratios, were incorporated as auxiliary input variables. The experimental results yielded a mean Intersection over Union (mIoU) of approximately 0.8. Although this figure is slightly lower than the 0.9 threshold often reported in conventional remote sensing studies, these initial results demonstrate significant potential for refinement through the future integration of additional training data. This model is expected to provide a scientific foundation for monitoring Arctic environmental changes and assessing the long-term impacts of climate change.

Back to poster list

P-97: Methodological Framework for Estimating Livestock Mortality: SWOT and Problem Tree Analysis, and Correlation–Regression Analysis

Presenting author: Otgonjamts GOMBOJAV

Authors: Otgonjamts GOMBOJAV, Centre for Policy Research and Analysis; Amarbal AVIRMED, Centre for Policy Research and Analysis; Enkhjargal NATSAGDORJ, Centre for Policy Research and Analysis; Oyunbileg ERDENEBADRAKH, Centre for Policy Research and Analysis

Keywords: Livestock mortality, SWOT шинжилгээ, Problem tree analysis.

This study examines the systemic causes of abnormal livestock mortality in five provinces of Mongolia by integrating SWOT and Problem Tree analyses with quantitative correlation–regression modeling and GIS-based spatial visualization. Using province-level seasonal data from 2011 to 2024—including livestock mortality, temperature, precipitation, wind speed, herder population, and hay production—an Ordinary Least Squares (OLS) regression model was developed and validated using 5–6 consecutive years of data. By substituting each province’s ten-year average temperature into the model, the 2025 projection estimates approximately 100,924 livestock losses nationwide. A consistent negative relationship between temperature and mortality was observed, indicating that colder climatic conditions increase animal losses. The temperature factor was statistically significant in Sukhbaatar, Tuv, Khovd, and Umnugovi provinces, while relatively weak in Khuvsgul. RMSE and residual diagnostics reveal that although the model captures the overall pattern, it does not fully explain extreme-year fluctuations, implying residual influences from factors such as dzud recurrence, pasture and water scarcity, animal health conditions, and management practices. The integration of SWOT and Problem Tree results led to the formulation of policy strategies along the O→W (opportunity–weakness) and O→T (opportunity–threat) linkages, identifying short-, medium-, and long-term intervention pathways for livestock risk mitigation and climate resilience in Mongolia’s pastoral sector.

Back to poster list

P-99: Building and Road Change Detection from Heterogeneous KOMPSAT and NEONSAT Imagery

Presenting author: Han OH

Authors: Han OH, Korea Aerospace Research Institute; Juyoung KIM, University of Science and Technology; Seojin PARK, Ajou University; Yeonjo JANG, Chungnam National Univ.; Jimmer HUR, Chungnam National Univ.; Jin KIM, Pusan National Univ.; Ahhyun JEON, Hanseo Univ.; Chaewon LEE, Ewha Womans Univ.

Keywords: Heterogeneous Change Detection; Multi-sensor Image Registration; Semantic Segmentation; KOMPSAT; NEONSAT

This study proposes a heterogeneous satellite image–based change detection framework for buildings and roads using KOMPSAT and NEONSAT imagery. Change detection across multi-sensor platforms is significantly more challenging than single-sensor approaches due to geometric misalignment, spectral inconsistencies, and variations in acquisition conditions. Accurate identification of infrastructure changes, however, is essential for disaster response, urban expansion monitoring, and land management, requiring robust and geometry-aware methodologies. To address these challenges, we first train a semantic segmentation model for buildings and roads based on a convolutional encoder–decoder architecture to obtain stable object-level representations. A deep learning–based image registration method is then applied to achieve precise geometric alignment between bi-temporal heterogeneous images. The aligned pairs are processed through a Siamese network to estimate potential changes. Instead of relying solely on direct feature differencing, initial change candidates are generated by integrating segmentation outputs with registration-derived overlap masks. A confidence-aware refinement module further corrects uncertain regions, enabling a hybrid rule-learning strategy that improves robustness against residual misregistration and segmentation ambiguity. The segmentation model was trained using 1,270 labeled image patches (1024 × 1024 pixels) from both sensors. For change detection, a ground-truth dataset was constructed from 150 co-registered KOMPSAT–NEONSAT pairs. To compensate for view-angle–induced geometric displacement, building changes were determined through mask dilation and area-overlap analysis. For roads, multi-stage spatial tolerance filtering was introduced to mitigate registration errors and labeling inconsistencies, ensuring reliable discrimination of true structural changes. Experimental results indicate that the proposed framework achieves stable and consistent performance in heterogeneous environments, demonstrating its practical applicability to infrastructure change monitoring. Future work will focus on dataset expansion and further optimization of the model architecture and training strategies.

Back to poster list

P-101: Systematic Evaluation of Geospatial Foundation Model Transfer for High-Resolution Land-Cover Segmentation in NEONSAT Imagery

Presenting author: Juyoung Kim

Authors: Juyoung KIM, University of Science and technology, Korea; Han OH, Korea Aerospace Research Institute

Keywords: Geospatial Foundation Models; Land-cover Segmentation; Transfer Learning; NEONSAT; High-resolution Satellite Imagery

As part of Korea’s national space program, an 11-microsatellite constellation equipped with high-resolution optical payloads (NEONSAT) is currently under development and is scheduled for completion by 2027. NEONSAT aims to enable rapid response to disasters and emergencies and to support continuous monitoring of national terrestrial resources and environmental change. Recent advances in artificial intelligence have highlighted the potential for automated change detection at wide-area scales in Earth observation. However, the effectiveness of change detection critically depends on the quality of the preceding land-cover segmentation stage. Reliable segmentation in sub-meter imagery typically requires large-scale pixel-wise annotations, which are costly and time-consuming to acquire. Recently, self-supervised geospatial foundation models (GFMs) pretrained on large collections of satellite imagery have demonstrated strong transferability under limited labeled-data settings. Nevertheless, systematic analyses of how GFM transfer characteristics affect high-resolution land-cover segmentation performance remain limited. In this study, we applied the pretrained encoder of an open-source GFM to a small-scale NEONSAT land-cover dataset and conducted a systematic investigation of factors affecting transfer performance. We evaluated multiple configurations including segmentation decoder architectures, the extent of encoder adaptation, data augmentation strategies, loss functions, learning-rate schedules, and regularization methods. Through controlled experiments across diverse design combinations, we analyzed both the individual contributions and interaction effects of these components. The results show that pretrained GFM representations achieve competitive segmentation performance even with limited annotations. In addition, we find that the trade off between encoder adaptation and decoder design substantially influences final accuracy. These findings support the feasibility of GFM-based land-cover segmentation in label-scarce scenarios and provide practical design guidelines for integrating foundation models into automated change-monitoring pipelines.

Back to poster list

P-103: Evaluating the effectiveness of continual pre-training of foundation model for analyzing small SAR satellite data

Presenting author: Toru KOUYAMA

Authors: Ryu SUGIMOTO, AIST; Toru KOUYAMA, AIST; Chiaki TSUTSUMI, AIST; Ali CAGLAYAN, AIST

Keywords: Foundation model, Continual pre-training, SAR, L-band, X-band

In recent years, the number of SAR satellites has rapidly increased due to advances in small satellite technologies, enabling more regular and frequent SAR-based monitoring of the Earth’s surface. As the impact and volume of SAR data continue to grow, AI-based analysis has become increasingly important, while both efficiency and accuracy remain essential considerations in AI model development. In this context, the concept of a foundation model is particularly promising. Once such a model is developed, it can be adapted to various downstream tasks using smaller datasets and reduced computational cost through additional training, i.e., finetuning. Furthermore, continual pre training may allow a foundation model originally trained on one SAR frequency band (e.g., L band) to be effectively extended to data from a different band (e.g., X band), although further experimentation is required to fully assess the feasibility of cross band continual pre training. To investigate the benefits of continual pre training for small SAR satellite data, we conducted a series of experiments that varied the training cost, i.e. the number of epochs in the continual pre training phase and the dataset size used during finetuning. We first built an original foundation model using ALOS 2/PALSAR 2 (L band) data, and then continued pre training it with small satellite SAR (X band) data. In the finetuning phase, we selected building segmentation as a downstream task for evaluating model performance. The results show that continual pre training clearly improves building segmentation performance compared with training a model from scratch, even when the pre training band (L band) and continual pre training band (X band) are different. We also found that models with relatively short continual pre training (e.g., 200 epochs) achieved accuracy comparable to models trained for much longer durations (400, 600, and 800 epochs), indicating the potential for reducing computational cost for utilizing AI model through continual pre training.

Back to poster list

P-105: Machine Learning–based Spatial Classification of High-Risk Wildfire Areas in the Forest–Urban Interface

Presenting author: Seunghyun Hong

Authors: Seunghyun Hong, Kookmin Univ.; Heejung Nam, Kookmin Univ.; Seyeon Hwang, Kookmin Univ.; Jonghwan Kim, Kookmin Univ.; Wanmo Kang, Kookmin Univ.

Keywords: Wildfire, Forest disaster, Risk analysis, Wildland-urban interface, Wildfire evacuation

Wildfires increasingly occur in forest–urban landscapes, where accumulated fuels and human activities elevate ignition risk. As wildfire frequency rises under climate change, delineating high-risk areas is critical for impact reduction and effective evacuation and suppression planning. This study identified high-risk wildfire areas in South Korea using an Extreme Gradient Boosting (XGBoost) model trained on multi-source geospatial predictors. Wildfire occurrence records from 2011 to 2025 were used as the response variable. Predictor variables included above-ground biomass density (AGBD) estimated by integrating GEDI LiDAR observations with Sentinel-1 SAR and Sentinel-2 optical data, along with forest type maps and landscape diversity and fragmentation metrics derived from FRAGSTATS. All predictor variables were resampled to a 30 m spatial resolution for consistency. To mitigate spatial overfitting, spatial block cross-validation was implemented during model evaluation. Grid cells within the upper quartile of predicted wildfire probability were classified as high-risk areas. SHAP (Shapley Additive Explanations) was used to assess the contribution of each predictor. Using SHAP-derived importance scores, influential predictors were identified and subsequently used in K-means clustering to classify high-risk areas into distinct spatial patterns. The XGBoost model achieved stable performance under spatial block cross-validation. SHAP results indicated that AGBD and proximity to urban areas were dominant drivers of wildfire risk. K-means clustering identified three major high-risk spatial typologies, with urban-exposed interface zones emerging as the most widespread pattern. These results inform risk-type–specific evacuation planning strategies and provide an objective spatial basis for establishing road control and management priorities during wildfire events. They also provide quantitative support for targeted forest fuel management strategies in high-risk areas. * This study was conducted with the support of the R&D program for Forest Science Technology (project no. RS-2025-25438293) provided by Korea Forest Service (Korea Forestry Promotion Institute).

Back to poster list

P-107: Tracking Invasive Alien Species in Northeast Asia: An Ensemble Modeling Approach

Presenting author: Youngjae YOO

Authors: Youngjae YOO, Korea University; Sangwook LEE, Korea University; Yoonji KIM, Korea University; Jinhoo HWANG, Korea University

Keywords: Invasive species; Ensemble modeling; Species distribution; GIS; Management

This study aims to support the management of invasive alien species in Northeast Asia by analyzing major national monitoring lists and applying species distribution models (SDMs). We reviewed Korea’s ecosystem‐disturbing and alert species, China’s key invasive species, and Japan’s regulated and invasive plant databases, and identified six species requiring common attention across the three countries. Among them, four terrestrial species—Ambrosia artemisiifolia, Ambrosia trifida, Mikania micrantha, and Sicyos angulatus—were selected for modeling. Occurrence records were obtained from the Global Biodiversity Information Facility (GBIF), and only data from 1980 to 2010 were used to match the climate baseline period. SDMs were developed using the BIOMOD2 package in R with an ensemble approach to reduce algorithm‐specific bias. Eight algorithms (GLM, GBM, CTA, ANN, FDA, MARS, RF, and XGBOOST) were applied. The dataset was divided into 80% for training and 20% for validation, and five-fold cross‐validation was conducted. Model performance was evaluated using TSS and AUC, and ensemble models were constructed by averaging models with TSS over 0.7. Environmental variables were selected based on a review of 19 previous SDM studies, considering usage frequency, data availability, and accessibility. The final variables included bioclimatic factors, land use and land cover, elevation, slope, aspect, and distance to roads. Topographic and distance variables were derived using ArcGIS. The results highlight country‐specific management implications. In Korea, Mikania micrantha has not yet widely spread, indicating high priority for preventive management. In China, all target species require pathway‐based control strategies considering large spatial scales. In Japan, the target species exhibited region‐specific distribution patterns, particularly for M. micrantha in Okinawa and Sicyos angulatus in northern coastal areas. This study provides integrated insights into current and potential distributions of invasive species in Northeast Asia and supports proactive and region‐specific management strategies.

Back to poster list

P-109: Watershed-scale Characterization Typology for Wildfire Management

Presenting author: Sangwook LEE

Authors: Sangwook LEE, Korea Univ.; Yoonji KIM, Korea Univ.; Jinhoo HWANG, Korea Univ.; Youngjae YOO, Korea Univ.; Rae-Ik JANG, Korea Univ.; Seongwoo JEON, Korea Univ.

Keywords: wildfire, hierarchical clustering, consensus clustering, forest management, watershed-based classification

As climate change intensifies both the scale and severity of wildfires, the need for region-specific management strategies has become increasingly evident. However, conventional administrative boundary–based response systems have inherent limitations in capturing the spatial and ecological characteristics of wildfire dynamics. To address these limitations, this study proposes a wildfire typology framework based on standard watershed units that reflect topographic and hydrological continuity rather than administrative or grid-based units. Using nationwide data from South Korea, environmental factors including topography, climate, and fuel characteristics, as well as wildfire history variables such as occurrence frequency, burned area, and spread characteristics, were clustered independently and subsequently integrated using a consensus clustering approach. The results indicate that a six-cluster solution (k = 6) yields the most stable and robust cluster structure, exhibiting distinct spatial patterns corresponding to major geographic regions such as the Baekdudaegan highlands, the east coast, the west coast, and Jeju Island. In particular, the east coast is characterized by a wildfire typology dominated by large-scale fire risk, whereas the metropolitan region exhibits a typology driven primarily by high fire occurrence frequency, highlighting clear regional differences in underlying wildfire risk mechanisms. The proposed watershed-based wildfire typology extends beyond conventional risk classification schemes and provides a scientific basis for regionally tailored wildfire management strategies, including surveillance deployment, suppression resource allocation, and fuel management. By integratively considering both environmental vulnerability and historical fire impacts, this framework has the potential to enhance the effectiveness and efficiency of national wildfire management systems.

Back to poster list

P-111: Integrating DEM‑Based Wetness and High‑Resolution Imagery in a GIS Workflow for Agricultural Drainage Prioritization

Presenting author: Yuta SHIMIZU

Authors: Yuta SHIMIZU, National Agriculture and Food Research Organization; Chinatsu YONEZAWA, Tohoku Univ.; Tetsuro SHISHIDO, Tohoku Univ.; Shigetoshi AKASAKA, Tohoku Univ.; Shin-ichiro OGURA, Tohoku Univ.; Mizuhiko NISHIDA, Tohoku Univ.

Keywords: Topological Wetness Index; Groundwater; World View; Grassland; DEM; GIS

Efficient identification of drainage‑poor zones and groundwater discharge hotspots in humid agricultural regions is essential for prioritizing drainage interventions and thereby improving land productivity. Focusing on the grasslands of the Kawatabi Field Science Center, Tohoku University (northeastern Japan), we applied a Geographic Information System (GIS)‑based workflow that integrates (i) a 1‑m‑resolution, Digital Elevation Model (DEM)‑derived Topographic Wetness Index (TWI), (ii) winter drone imagery capturing accelerated snowmelt, and (iii) high‑resolution WorldView‑2/3 near‑infrared (NIR) imagery acquired shortly after rainfall or immediately after sowing, when surface reflectance is governed by soil moisture rather than vegetation. The study area is located on a fan that consists of volcanic ash and gravelly deposits over discontinuous clay and peat layers, producing spatially heterogeneous hydrologic behavior. TWI delineated geomorphologically convergent areas susceptible to shallow groundwater rise or surface saturation; these terrain‑based predictions were evaluated against drone snowmelt signals and satellite NIR responses. Field verification at ten wet or discharge sites found eight co‑located with high‑TWI zones, indicating that DEM‑based wetness is a reliable first‑order indicator for drainage‑risk mapping. Drone‑detected melt patterns further refined the spatial delineation of active groundwater discharge, while satellite NIR consistently decreased in wet areas; imagery from the 2017–2018 sowing period provided especially clear moisture contrasts, enabling effective ranking of drainage problem areas. By integrating these multi‑sensor indicators within a GIS environment, the workflow achieves reproducible, wide‑area, and low‑cost evaluation of moisture‑related risks, and offers a practical basis for prioritizing drainage improvements, designing field‑level countermeasures, and supporting long‑term water‑management strategies in agricultural landscapes. This study demonstrates the feasibility of operational deployment across humid agricultural regions, where reproducible, GIS‑based screening can accelerate evidence‑based drainage planning and improve land productivity as well as disaster prevention attributable to excess water.

Back to poster list

P-113: Changes in the shortest routes and distances for four industries in Miyagi Prefecture

Presenting author: Hiroto WAKI

Authors: Sayaka YOSHIKAWA, Nagasaki Univ.; Hiroto WAKI, Nagasaki Univ.

Keywords: Road Network Analysis, Road Centerline, Interchanges, GIS, Python

The Great East Japan Earthquake that occurred on March 11, 2011, had great impact on industrial activities in the affected areas. For example, loss of seafood facilities due to the tsunami and disruptions in logistics caused by the damaging transportation networks by the earthquake. To achieve recovery from this disaster, a new highway, called the Reconstruction Highway, was constructed along the coastal areas of the Tohoku region. Improved accessibility between new interchanges and facilities has promoted efficient goods transportation and contributed to industrial recovery efforts. This study clarified the changes in the shortest routes and distances between facilities and interchanges from 2011 to 2023 by industry type and region, using GIS. The target areas were Sendai, Ishinomaki, and Kesennuma Cities and their surrounding areas in Miyagi Prefecture. First, facility location data for four industries (construction, seafood, steel, and paper pulp) were obtained using OpenStreetMap. Second, we proposed an efficient analysis method to handle the massive calculation load involved in route searches. Third, road centerline data were extracted considering road widths suitable for truck traffic. Finally, we calculated the shortest routes and distances between facilities and interchanges and compared the results by industry type and region. We further revealed the distinction of the area with improved accessibility, considering road widths. The results show significant reductions in the shortest distances to facilities in Kesennuma City. This reduction was caused by the development of reconstruction roads. Different types of industries and locations influenced the timing of changes in the shortest routes and the load of the shortest distance changes. Furthermore, the analysis of the shortest routes and distances considering truck traffic revealed detours or increases in the shortest routes for some facilities. These methods are applicable to other regions, and the results can be used as indicators for evaluating the practicality of road reconstruction.

Back to poster list

P-115: Accuracy assessment of DEM derived from SpaceEye-T1 tri-stereo imagery

Presenting author: Youngran MOON

Authors: Youngran MOON, SI Imaging Services

Keywords: SpaceEye-T; VHR; Tri-stereo; DEM; Mapping

The SpaceEye-T1 satellite, launched in March 2025, is a very high resolution Earth observation satellite with a ground sampling distance (GSD) of 25 cm. It supports both in-track stereo and tri-stereo imaging modes, enabling robust three-dimensional terrain reconstruction by minimizing occlusions in complex landscapes. In this study, a digital elevation model (DEM) was generated from SpaceEye-T1 tri-stereo imagery, and its geometric accuracy was quantitatively evaluated. For accuracy assessment, Bern, Switzerland where reliable reference elevation and image data were available was selected as the validation area. Sensor modeling was performed using Rational Polynomial Coefficients (RPCs), and stereo image matching was applied to generate a digital surface model (DSM). The DSM was subsequently processed using morphological filtering to produce a DEM suitable for terrain analysis. The generated DEM was compared with reference data to perform a quantitative analysis of both horizontal and vertical positional accuracy. The results demonstrate that SpaceEye-T1 tri-stereo imagery can generate high-quality DEM products with a few-decimeter-level geometric performance. The superior spatial resolution enables precise terrain reconstruction, confirming its suitability for high-accuracy topographic applications.

Back to poster list

P-117: Geomorphic Interpretation of Korean Tidal Flats Based on Cross-Validation of Sentinel-1 SAR Backscatter-Derived and TanDEM-X InSAR DEMs

Presenting author: Hanah PARK

Authors: Hanah PARK, Department of Earth System Sciences, Yonsei Univ.; Hahn Chul JUNG, Department of Earth System Sciences, Yonsei Univ.; Hee-Jeong JEONG, Department of Earth System Sciences, Yonsei Univ.; Yun Seong LEE, Department of Earth System Sciences, Yonsei Univ.; Dongjae KWON, Department of Earth System Sciences, Yonsei Univ.

Keywords: Tidal flats; Intertidal topography; Synthetic Aperture Radar (SAR); Sentinel-1; Multitemporal backscatter analysis; TanDEM-X; InSAR DEM; Coastal geomorphology

Tidal flats along the western and southern coasts of the Korean Peninsula show complex surface forms shaped by strong tides, active sediment movement, and spatially variable surface conditions. However, describing detailed elevation patterns across these tidal flats is difficult because they are frequently inundated by seawater and direct field-based elevation measurements are limited. In this study, we examine the surface topography of Korean tidal flats by comparing two independently generated intertidal Digital Elevation Models (DEMs). The first DEM is derived from Sentinel-1 Synthetic Aperture Radar (SAR) backscatter using a pixel-based, multitemporal analysis, while the second DEM is produced using SAR Interferometry (InSAR) processing of TanDEM-X bistatic SAR image pairs. The Sentinel-1 dataset includes acquisitions from 2016 to 2025 and covers major tidal flats in southwestern South Korea. Temporal changes in SAR backscatter due to intertidal flooding and exposure were analyzed at the pixel level, with tidal flat areas identified using the Jenks Natural Breaks (JNB) method. A logistic model was used to describe the nonlinear relationship between SAR backscatter intensity and tide height, enabling estimation of relative elevation patterns within the intertidal zone. The Sentinel-1-based DEM was compared with a TanDEM-X InSAR-derived DEM to assess how similar the two approaches represent intertidal topography. The comparison focuses on broad geomorphic features, such as intertidal channels and changes in elevation from the shoreline to offshore areas, while also examining differences in estimated elevation values. Differences between the two DEMs are interpreted in terms of SAR sensitivity to surface roughness and inundation conditions, as well as the effects of temporal averaging in multitemporal observations. This study indicates that multitemporal SAR backscatter analysis can capture meaningful patterns in intertidal surface topography, and that comparisons with InSAR-derived DEMs provide a useful way to interpret tidal flat geomorphology in coastal areas where field measurements are limited.

Back to poster list

P-119: JAXA’s L-SAR MISSIONS AND THOSE PPP PROJECTS

Presenting author: Shinichi SOBUE

Authors: Shinichi SOBUE, JAXA

Keywords: ALOS-2, ALOS-4, ALOS PPP

The Advanced Land Observing Satellite-2 (ALOS-2) is a follow-on mission from ALOS launched on May 24, 2014 to be in operation to contribute various application area such as rapid response for disaster, etc. Furthermore, to improve data accessibility and to stimulate archived data usage, JAXA had been proof of concept (POC) activities using ALOS-2 archive data in cooperation with local governments, universities and private sectors as ALOS-2 phase A public private partnership (PPP) during 2022 and 2024. From 2025, after successful launch of ALOS-4 in 2024, JAXA implements ALOS PPP phase B during 2025 and 2027. JAXA already selects 16 demonstration projects in operation. This paper describes status of ALOS PPP project phase A summary with accomplishment as well as phase B status.

Back to poster list

P-121: Spatial Resolution Impact Assessment for Long-term Population Monitoring of Adélie Penguin Guano

Presenting author: Jinku PARK

Authors: Jinku PARK, Korea Polar Research Institute; Yongsik JEONG, Korea Polar Research Institute; Joo-Eun YOON, Korea Polar Research Institute; Jeong-Hoon KIM, Korea Polar Research Institute

Keywords: Adélie Penguin; Remote Sensing; Cape Hallett; Guano Detection; Apparent Density; Resolution-dependent Accuracy

Adélie penguins (Pygoscelis adeliae) are key indicators of the Antarctic ecosystem, and remotely sensed guano stains offer an efficient means of estimating breeding abundance. Landsat medium-resolution (MR) imagery enables long-term monitoring through its archive. However, the effect of spatial resolution on detection accuracy and abundance estimates remains unquantified. This study evaluated resolution-dependent errors using unmanned aerial vehicle (UAV; 0.1 m), very high-resolution (VHR; 1.24 m, WorldView-4), and MR (30 m, Landsat 8) imagery acquired concurrently over the Cape Hallett colony in November 2018. UAV-based delineations were used as reference data to assess guano detection accuracy and to perform Monte Carlo and Poisson-based simulations. VHR imagery achieved high accuracy (F1 = 0.73; omission = 21%), whereas MR imagery showed large commission (59%) and omission (21%) errors. Detection errors rose nonlinearly with coarser resolution, reaching 97% at 30 m. Abundance estimates were sensitive to apparent density; using the Antarctic mean (0.34 nests m⁻²) sometimes gave lower errors than the UAV-based value (0.44 nests m⁻²). These results provide a quantitative framework linking spatial resolution, detection error, and estimation uncertainty, emphasizing the need for UAV-based validation and colony-specific calibration for reliable long-term monitoring of Antarctic penguin populations.

Back to poster list

P-123: A Study on the Operational Status of a Nuclear Facility in an Inaccessible Area using Time-Series Analysis of Multi-Optical Satellite Imagery

Presenting author: Gayeon HA

Authors: Gayeon HA; Dongjin KIM; Byungmarn KHO; Minsoo KIM

Keywords: North Korea; Optical Satellite Imagery; Time-Series Analysis; Nuclear Nonproliferation

This study aims to assess the operational status of the Experimental Light Water Reactor (ELWR) at the Yongbyon Nuclear Complex in North Korea, where physical access is restricted, by employing time-series analysis of high-resolution optical satellite imagery. A total of over 200 multi-source images acquired between October 2023 and February 2026 were analyzed. The discharge status of the reactor’s cooling system was selected as the primary observable indicator to indirectly infer the operational condition of the reactor, and its temporal variations were systematically tracked. The analysis incorporated characteristics typically associated with different operational phases of a nuclear reactor. Structural expansions of the facility and changes in surrounding infrastructure activities identified through satellite imagery were examined in conjunction with discharge patterns. The observed patterns were categorized into distinct operational phases, and the facility’s operational status was inferred based on the characteristics of each category. Based on the derived operational patterns, the study further explores, through scenario-based assessment, the possibility that the facility may serve strategic purposes beyond electricity generation, including potential nuclear material production. Continuous time-series analysis of high-resolution optical satellite assets provides foundational data for estimating the operational status of key facilities in inaccessible areas and for examining potential operational intentions from a nuclear nonproliferation perspective.

Back to poster list

Day 2 Poster Abstracts

P-2: Floristic Diversity and Spectral Characterization of Vegetation in Tormik Valley, District Skardu, Pakistan: A Field-Based and Sentinel-2 Remote Sensing Approach

Presenting author: Naseema KHAN

Authors: Naseema; Mehboob ALI; Kanako MURAMATSU; Abdul MANAN

Keywords: Karakoram Range, Vegetation Mapping, Remote Sensing, Floristic Composition, Sentinel-2, Gilgit-Baltistan, Vegetation Indices

This study integrates Sentinel-2 satellite remote sensing and ground-truthing botanical surveys to illustrate the vegetation of Tormik Valley of Gilgit-Baltistan (elevation 1,800–4,500 m). During the field study, 78 species were identified across 46 families, with the dominant life form being chamaephytes (30 species, 38.46%). This indicates extreme ecological conditions typical of high altitude. Phytosociological studies of the three stands revealed that Trichodesma indicum (IVI = 20.65) was Stand II dominant, while Artemisia vulgaris (IVI = 12.85) and Amaranthus retroflexus (IVI = 11.58) were dominant in Stand I and Stand III, respectively. Nine cloudless Sentinel-2 MSI (L1C and L2A) images of the complete 2024 seasonal cycle were processed using ESA SNAP. Spectral reflectance was determined at 18 field-registered GPS locations at the corresponding phytosociological sampling points. For each date, vegetation indices (NDVI, CIgreen, NDRE, and EVI) were computed. Stand 2 was the only one to validate field measured IVI with the highest NDVI (0.36) and CIgreen (1.01) in summer. Across all stands, the red-edge spectral feature (705–783 nm) was most dominant in August–September. A clear seasonal phenological pattern emerged: NDVI near zero in winter (January–March), progressive spring increase, August peak, and autumn decline through December. The strong compatibility between satellite-derived vegetation indices and field-based IVI rankings demonstrates the effectiveness of integrating traditional phytosociology with Sentinel-2 remote sensing. This study provides the first combined spectral-floristic baseline for Tormik Valley and establishes a reproducible observation framework for biodiversity assessment in the data rare Karakoram region.

Back to poster list

P-4: Evaluating YOLO-Based Object Detection for Automated Landslide and Debris-Flow Damage Assessment Using Post-Disaster Imagery

Presenting author: Kosei UENO

Authors: Kosei UENO Shizuoka Univ; Rei SONOBE Shizuoka Univ; Naoto FUKUTANI

Keywords: Disaster damage assessment;Remote sensing;Deep learning;Object detection;YOLO;Landslide detection;Post-disaster imagery

Accurate and timely assessment of disaster damage is essential for effective emergency response and recovery planning. In many cases, conventional damage assessment methods rely on field surveys and visual interpretation of aerial photographs, which require considerable time and labor. These approaches are often difficult to apply when access to affected areas is limited due to infrastructure damage. From the perspective of remote sensing, satellite and aerial imagery provide an effective means for rapid and wide-area observation immediately after a disaster, including regions that are inaccessible from the ground. Recent advances in deep learning have significantly improved the performance of image recognition and object detection in remote sensing applications. Among various models, YOLO (You Only Look Once) is a representative object detection approach that can analyze an entire image in a single inference, enabling both high processing speed and reasonable detection accuracy. These characteristics are particularly suitable for time-critical disaster monitoring using remote sensing data. However, disaster-related images are characterized by diverse damage patterns, complex backgrounds, and varying imaging conditions, which can reduce detection performance. Therefore, it is necessary to evaluate the applicability and limitations of YOLO using actual disaster imagery. This study aims to investigate the effectiveness of a YOLO-based object detection model for automated disaster damage detection from remote sensing images. The model was trained and evaluated using post-disaster imagery, and detection performance was quantitatively assessed using Intersection over Union (IoU) and mean Average Precision (mAP). The Landslide Segmentation Dataset, an open dataset with annotated landslide areas, was used for training and validation. YOLO11 was implemented on Google Colaboratory with an NVIDIA Tesla T4 GPU. The results indicate clear performance differences among damage classes. The landslide class achieved relatively high detection accuracy with an AP of 0.504, whereas the debris-flow class showed very low performance with an AP of 0.068. The overall recall across all classes was approximately 61%. These findings highlight both the potential and current limitations of YOLO-based disaster damage detection in remote sensing and provide insights for future improvements toward practical disaster monitoring applications.

Back to poster list

P-6: Development of a Low-Cost Meteorological Satellite Receiving Station for Meteor-M Imagery

Presenting author: Tzu-Yu CHEN

Authors: Tzu-Yu CHEN, National Taiwan Ocean University; Chung-Ru HO, National Taiwan Ocean University

Keywords: QFH antenna; RTL-SDR; LRPT; Meteor-M; NDVI

Cloud imagery from meteorological satellites is essential for weather monitoring, typhoon tracking, and disaster warning. However, commercial receiving systems are costly and complex. This study demonstrates the design and implementation of a low-cost meteorological satellite ground station using a self-built Quadrifilar Helix (QFH) antenna and an RTL-SDR receiver. The QFH antenna, constructed from PVC and coaxial cables, was optimized for 137.9 MHz circularly polarized signals to receive LRPT (Low-Rate Picture Transmission) signals from Meteor-M N2-3/N2-4 satellites. Open-source software, including SDR#, WXtoImg, and SatDump, was used for real-time reception, decoding, and image generation. During the experiment, visible, near-infrared, and shortwave infrared imagery were received from Meteor-M satellites. Geometric correction was performed using polynomial regression models to reduce distortions caused by Earth curvature and satellite scanning geometry. A second-order model effectively corrected the imagery, ensuring high-precision georeferencing. Corrected images were further processed with land-sea and cloud masks, followed by the computation of the Normalized Difference Vegetation Index (NDVI) to analyze vegetation distribution and health across East Asia. Results show clear NDVI contrasts between arid and vegetated regions, demonstrating the potential of LRPT imagery for reliable ecological and environmental monitoring. This study confirms that a fully functional Meteor-M satellite receiving system can be constructed for under USD $130 using accessible materials and open-source tools. The low-cost approach significantly lowers the entry barrier for satellite remote sensing research and education, providing practical applications in climate studies, disaster observation, and ecological monitoring.

Back to poster list

P-8: Regional Ionospheric Characterization via Multi-point Signals of Opportunity and Distributed Digital Beacons

Presenting author: Randson Huang

Authors: Randson Huang, National Taipei University of Technology (NTUT); Hower Huang, National Taipei University of Technology (NTUT); Yang-Lang Chang, National Taipei University of Technology (NTUT)

Keywords: Ionospheric; Beacon

Traditional ionospheric sounding systems, while highly accurate, often suffer from low spatial sampling density and high operational costs. This paper presents a novel, cost-effective remote sensing framework that utilizes a distributed network of time-synchronized, narrowband digital beacons as Signals of Opportunity (SoO) to characterize regional ionospheric dynamics. By deploying a high-sensitivity software-defined radio (SDR) receiver at the National Taipei University of Technology (NTUT) campus, we systematically monitor multi-path propagation from a dense grid of synchronized transmitters. While the system is capable of detecting global long-distance propagation, this study specifically focuses on regional geographical baselines to reconstruct localized ionospheric parameters. The methodology leverages the high coding gain and precise temporal synchronization of these beacons to extract Signal-to-Noise Ratio (SNR) fluctuations and Doppler characteristics. By processing these “Big Data” signal logs, we demonstrate the ability to estimate Maximum Usable Frequency (MUF) variability and detect transient anomalies, such as Sporadic E (Es) events, across a high-resolution observation grid. Preliminary results indicate that this passive sensing approach provides a scalable, high-density complement to traditional ionosonde data, offering a robust solution for regional space weather monitoring and wide-area atmospheric remote sensing without the need for additional spectrum allocation.

Back to poster list

P-10: Voxel Entropy Applications for Individual Tree Detection in High-definition Forest Point Clouds

Presenting author: Kasparas KARLAUSKAS

Authors: Kasparas KARLAUSKAS, Vilnius University, Institute of Data Science and Digital Technologies; Povilas TREIGYS, Vilnius University, Institute of Data Science and Digital Technologies

Keywords: LiDAR; Individual Tree Detection; precision forestry; entropy; TLS

High-definition 3D point clouds have enabled detailed analysis of forest areas via classical and learning-based algorithms. Individual Tree Detection, the task of finding the coordinates of all individual trees in a point cloud, is a pre-requisite for tree-level analysis such as inventorization or tree species classification. In Terrestrial and Mobile Laser Scanning, forest point clouds are captured from a low altitude below the canopy, making the stems of individual trees well-defined, providing a strong signal for detection. However, high-definition point clouds have high memory requirements, and due to the surface reflectivity and local geometry differences between stems and leaves, most laser pulse reflections and therefore most recorded points come from the canopy layer - data that is usually not necessary for Individual Tree Detection. Voxel Entropy, a measure of point distribution homogeneity within discrete cube volumes, can be used as a computationally efficient (linear time) coarse-scale discriminator between tree canopy and stem regions. This study showcases ways to seamlessly integrate Voxel Entropy into existing Individual Tree Detection workflows: pre-filtering trivial canopy points to reduce computational load of neighbourhood-based methods; automatically determining optimal regions of interest for methods based on narrow z-axis crops of the original forest point cloud. Finally, a competitive stand-alone method using Voxel Entropy as the only geometric feature for Individual Tree Detection is presented.

Back to poster list

P-12: Comparative Evaluation of Supervised Classification Algorithms for Land Use Mapping in Support of Asian Elephant Habitat Modeling

Presenting author: Zichen MIN

Authors: Zichen MIN, University of Tsukuba; Fatwa RAMDANI, University of Tsukuba

Keywords: Sentinel-2; Land Use and Land Cover; Land Use and Land Cover; Supervised Classification

Balancing economic development and wildlife conservation remains a major challenge in developing regions. In Yunnan Province, China, land use transformation has altered the habitat conditions of Asian elephants (Elephas maximus). Reliable land use and land cover (LULC) mapping is therefore essential for accurate habitat suitability modeling and subsequent spatial policy analysis. This study compares four supervised classification algorithms—Random Forest (RF), Support Vector Machine (SVM), Classification and Regression Tree (CART) and Gradient Boosted Decision Tree (GBDT) for high-resolution LULC mapping using Sentinel-2 imagery. Six spectral bands are used to classify forest, agriculture, urban, water, bare land, and greenhouse. Based on independent test samples, GBDT achieves the highest accuracy (overall accuracy = 0.744, Kappa = 0.693), closely followed by RF (overall accuracy = 0.738, Kappa = 0.686), while SVM (overall accuracy = 0.709, Kappa = 0.651) and CART (overall accuracy = 0.691, Kappa = 0.629) show relatively lower performance. Model agreement analysis reveals strong spatial consistency between RF and GBDT (94.6%) whereas CART exhibited greater variability relative to the ensemble methods. Classification discrepancies are mainly concentrated in urban–agricultural transition zones rather than reflecting systematic class-level bias. Overall, the results suggest that ensemble tree-based approaches provide more stable and reliable LULC inputs for subsequent habitat suitability modeling and spatial econometric analysis of human–elephant interactions.

Back to poster list

P-14: Wildfire Burned Area Mapping Using Early-fusion Deep Learning Change Detection Based on U-Net Model with Sentinel-2 Imagery

Presenting author: Amalia Gita AYUDYANTI

Authors: Amalia Gita AYUDYANTI, National Central University; Ying-Nong CHEN, Tamkang University; Hsuan REN, National Central University; Ilham JAMALUDDIN, National Central University

Keywords: wildfires; burned-area; Sentinel-2; U-Net; deep learning; change detection

Wildfires are a major environmental hazard that cause extensive ecological degradation and economic loss. Accurate and timely mapping of burned areas is essential for effective mitigation, recovery planning, and post-fire assessment. Remote sensing imagery is suitable data for burned area mapping due to its temporal resolution, large-scale observation, and ability to capture spectral changes associated with wildfire impacts. This study evaluates the performance of early-fusion deep learning change detection model based on U-Net architecture for mapping burned area affected by the wildfire in part of California in 2025. Sentinel-2 imagery acquired before and after the wildfire event was used as the primary input, supplemented with three related spectral indices to enhance feature representation. The U-Net architecture with a Residual Network (ResNet) backbone was used to perform semantic segmentation-based change detection. The U-Net model leverages encoder–decoder structures with skip connections to preserve spatial detail and improve segmentation accuracy. Experimental results demonstrate that the model achieves a mean Intersection over Union (mIoU) of 0.9854 and an IoU of 0.9830 for burned-area class. In addition, the trained model was tested on the independent testing site and shows the mean IoU of 0.9353. Overall, this study highlights the effectiveness of early-fusion deep learning change detection combined with Sentinel-2 imagery for rapid and reliable burned area mapping.

Back to poster list

P-16: Radiometrically Preserving Super-Resolution of Coarse-Resolution Satellite Imagery for Post-Disaster Damage Mapping

Presenting author: Angel MATOS

Authors: Angel MATOS, Institute of Science Tokyo; Masashi MATSUOKA, Institute of Science Tokyo; Miguel Luis R. LAGAHIT, Dynamic Map Platform Co., Ltd.

Keywords: Super-resolution; Sentinel-2; Radiometric consistency; Multispectral imagery; Cross-sensor domain adaptation; Disaster damage assessment

Rapid post-disaster damage assessment is essential for emergency response and recovery planning. While high-resolution (HR) satellite or aerial imagery enables accurate mapping of damaged buildings, such data are often limited by cost and global coverage immediately following an event. In contrast, Sentinel-2 imagery is globally available and frequently updated, but its native spatial resolution (10 m) is insufficient to reliably identify structural damage at the building scale. Super-resolution (SR) has been proposed as a potential solution to enhance low-resolution imagery. Nevertheless, most SR approaches for remote sensing focus primarily on visual quality, failing to preserve the radiometric consistency and spectral relationships required for physical interpretation and downstream analysis. For post-disaster damage mapping, maintaining physically meaningful band relationships is as important as improving spatial detail. In this study, we propose a physics-aware super-resolution framework designed to enhance Sentinel-2 imagery while maintaining radiometric consistency and inter-band relationships. The model increases spatial resolution by a factor of four and incorporates a degradation-consistency constraint to ensure the reconstructed high-resolution image remains anchored to the original Sentinel-2 observations. In addition, a cross-sensor radiometric alignment module is introduced to bridge the domain gap between multispectral satellite imagery and higher-resolution reference data. The proposed method is evaluated on a building damage segmentation task in a post-earthquake scenario. Performance is quantitatively compared with native Sentinel-2 imagery, bicubic upsampling, and available HR imagery to assess whether physics-aware SR improves damage detection while retaining interpretable reflectance information.

Back to poster list

P-18: Assessment of Damage Detection on the Metropolitan Expressway Caused by the 2011 Great East Japan Earthquake Using TerraSAR-X Imagery

Presenting author: Masaya KUBOTA

Authors: Masaya KUBOTA, Institute of Science Tokyo; Fumitaka OGUSHI, Terra Phase, Inc.; Takashi MIYAMOTO, Institute of Science Tokyo; Masashi MATSUOKA, Institute of Science Tokyo ; Mizuki OGAWA, Shutoko Technology Center ; Takanobu OSADA, Shutoko Technology Center ; Guangfeng ZHANG, Shutoko Technology Center ; Taishi NAKAMURA, Shutoko Engineering Co., Ltd.

Keywords: TerraSAR-X; Urban Infrastructure Monitoring; Expressway; Damage Detection; Coherence; Great East Japan Earthquake

The 2011 Great East Japan Earthquake inflicted damage not only on the Tohoku region but also on the extensive infrastructure networks of the metropolitan area. Particularly within the Metropolitan Expressway network, which supports transportation in the metropolitan area, damage was recorded primarily centered on structural components such as expansion joints, bridge piers, and bearings. In the immediate aftermath of a large-scale disaster, the rapid assessment of infrastructure damage is essential for achieving the early restoration of transportation functions. However, field investigations face challenges: they not only require significant time and resources but also expose inspectors to the risks of secondary disasters associated with aftershocks and structural instability. To address these issues, satellite SAR data serves as an effective tool. This study evaluated the feasibility of extracting localized damage sites on the Metropolitan Expressway using SAR imagery. The analysis utilized X-band images from the TerraSAR-X satellite, examining image pairs acquired in multiple periods before and after the earthquake. As indices for evaluating changes, the difference in backscattering intensity between pre- and post-disaster images and the coherence calculated from phase correlation were employed. A noteworthy challenge in the analysis of urban road surfaces is that scattering noise from vehicles traveling on the road during image acquisition obscures signals originating from structural damage. To address this issue, this study constructed a multi-temporal dataset consisting of images from multiple periods before and after the earthquake. By calculating the median backscattering intensity for each pixel, the study attempted to evaluate the change components inherent to the structures themselves while accounting for the effects of increased reflection intensity caused by moving objects such as vehicles. Visual inspection of the processed images revealed localized fluctuations in backscattering intensity and coherence in the vicinity of several sites where damage was reported by the Metropolitan Expressway Company Limited.

Back to poster list

P-20: Backscattering and Texture Characteristics of ALOS-2 PALSAR-2 Images for Buildings Damaged by the 2016 Kumamoto Earthquake

Presenting author: Yuriko ASAKA

Authors: Yuriko ASAKA, Institute of Science Tokyo; Masashi MATSUOKA, Institute of Science Tokyo

Keywords: 2016 Kumamoto Earthquake; ALOS-2; PALSAR-2; building damage assessment; polarimetric SAR analysis

Synthetic aperture radar (SAR) is capable of observations even at night and in bad weather, making it effective for quickly assessing building damage in areas affected by earthquakes. Damage analysis using images taken only after an earthquake is mainly based on backscattering intensity, but a single intensity index can make it difficult to identify the detailed extent of damage. In response, the use of texture indices using the Gray-Level Co-Occurrence Matrix (GLCM) is being promoted to capture the spatial distribution of brightness due to factors such as debris accumulation. However, most previous studies have been limited to the analysis of individual buildings, and there are few examples of wide-area analyses that cover the entire damaged area. This study aims to analyze wide-area data covering the entire area damaged by the earthquake and clarify the statistical differences between damage levels. This study focused on building damage caused by the 2016 Kumamoto earthquake, using L-band dual-polarization ALOS-2 PALSAR-2 data observed on April 21st after the disaster. The area analyzed was part of Mashiki Town, Kumamoto Prefecture, which suffered extensive damage, and targeted buildings that were visually classified into six levels of damage. The analysis involved quantifying the backscattering coefficient and GLCM texture index obtained from the dual-polarization SAR data using pixel values within the region and then conducting a statistical analysis. The results of the analysis revealed that the backscattering coefficients and GLCM texture indices of ALOS-2 PALSAR-2 dual polarization images showed distinct trends for each building damage classification. However, overlapping values for each index suggested that there are limitations in providing a definitive classification of damage levels using a single index. Correlations were found between multiple GLCM indices and damage levels, and significance tests revealed that each index was statistically effective in identifying specific damage levels.

Back to poster list

P-22: Spatial Extension of Tetracorder-Derived Mineral Signatures to Multispectral Sensor Coverage in Cuprite, Nevada

Presenting author: Hideki TSUBOMATSU

Authors: Hideki TSUBOMATSU; Satoru YAMAMOTO; Hideyuki TONOOKA

Keywords: Tetracorder; diagnostic output; absorption depth; hyperspectral; multispectral; AVIRIS; ASTER; random forest; Cuprite; mineral mapping

Tetracorder, an expert system developed by the U.S. Geological Survey (USGS), identifies surface minerals from hyperspectral (HS) imagery by evaluating the geometry of spectral absorption features. Beyond discrete class labels, it produces diverse continuous diagnostic outputs per pixel—such as absorption band depth—that capture multi-dimensional mineral expression with greater nuance. However, the spatial coverage of such HS-derived products is inherently limited by the narrow swath and sparse acquisition of HS sensors such as AVIRIS. This study investigates whether Tetracorder-derived diagnostic outputs and their spatially integrated derivatives can be extended to the broader coverage of multispectral (MS) sensors, specifically ASTER, at the Cuprite mining district, Nevada, USA. The processing pipeline proceeded as follows: (1) Tetracorder was applied to AVIRIS HS data to extract diverse per-mineral diagnostic maps across over 150 reference library entries; (2) these maps were consolidated into mineral-class images at multiple levels of aggregation through chemical grouping, weighted summation, and mixture normalization; and (3) Random Forest (RF) models were trained on co-located ASTER 14-band observations using the consolidated maps as prediction targets, and their performance was compared against conventional discrete mineral class-label prediction via 5-fold cross-validation. Results indicate that the conventional class-label approach consistently outperformed the continuous diagnostic targets across most mineral classes. This performance gap reflects a fundamental spectral resolution constraint: ASTER’s broad spectral bands cannot resolve the subtle absorption feature geometry that Tetracorder exploits, and these fine-scale signals are substantially degraded during spectral resampling to MS resolution. These findings clarify the physical limits of extending fine-grained HS diagnostic outputs to MS sensors. Future work will explore additional feature engineering, such as band math, and sensors with finer spectral resolution to improve signal recovery, and will assess transferability across additional sites and sensor combinations.

Back to poster list

P-24: Initial Evaluation of a Cross‑Sensor Reflectance Transformation Algorithm Using Landsat 7 and 8 Time‑Series Data at Flux Sites

Presenting author: Yuma HORIKAWA

Authors: Yuma HORIKAWA, Aichi Pref. Univ.; Shoki FUNAYAMA, Aichi Pref. Univ.; Kenta OBATA, Aichi Pref. Univ.; Hiroki YOSHIOKA, Aichi Pref. Univ.

Keywords: Reflectance; Landsat 7; Landsat 8; Cross-Sensor Transformation; Flux Site

Long-term monitoring of terrestrial vegetation requires integration of earth observation data acquired by different generations of satellite sensors. Differences in specifications such as spectral responses among sensors lead to systematic differences in reflectances. Previous studies have applied regression and machine learning methods to reduce these systematic differences. Recently, an algorithm based on a linear mixture model was developed to harmonize reflectance data from multiple sensors (Obata, ISPRS J. P&RS, 2024). This algorithm was evaluated using green, red, and near-infrared (NIR) bands of the Landsat 4/5 Multispectral Scanner (MSS) and Thematic Mapper (TM), but was not evaluated for other wavelength regions because of band availability limitations. Thus, the objective of the present study was to broaden the spectral coverage of the algorithm and to evaluate the algorithm using visible (blue and red), NIR, and short-wave infrared (SWIR; centered at 1.6 μm) band time-series data from Landsat 7 Enhanced Thematic Mapper Plus (ETM+) and Landsat 8 Operational Land Imager (OLI). The pre-processing method for computing the parameter set (day-of-year dependent endmember spectrum) was improved by incorporating reflectance statistics of soil-dominated data, enabling its application across visible to SWIR bands. The evaluations of our algorithm were performed using ETM+ and OLI time-series data from a single pixel that includes three flux sites (OzFlux) comprised of eucalyptus forest, woodland savanna, and grassland, respectively. The results of the ETM+-to-OLI transformations showed that cross-sensor biases in reflectance and vegetation index time-series data tended to decrease, whereas slight over-corrections (negative biases in transformed reflectances) were sometimes observed in blue bands. The over-correction and other errors likely stem from uncertainties in endmembers used in the transformations as well as differences between the model and actual data. Future work should seek to improve the pre-processing method and evaluate the algorithm using data from a larger variety of sites.

Back to poster list

P-26: Classification Criteria for burned areas in Punjab, India, using Sentinel-2/MSI data

Presenting author: Nana MIZOGUCHI

Authors: Nana MIZOGUCHI; Kanako MURAMATSU

Keywords: Crop residue burning; Spectral reflectance; BAI; NBR; Flatness index

Crop residue burning in Punjab is a seasonal cause of air pollution in India. Rice and wheat cultivated under the double-cropping system, therefore, next crop must be planted within a short period. Despite being prohibited by law, the crop residue burning has continued due to low cost and reduced labor requirement. Accurate measurement of the burned area is crucial for assessing the impact of the burning on air pollution. Previous studies enabled identification of the burning even when satellite observations are out of sync with actual timing of the burning, by focusing on “burned area” rather than the active fires. However, moist soil was often misclassified as burned areas. This study aims to improve accuracy of identifying burned areas. Using Sentinel-2/MSI data (2022/4/1-2023/3/31) for a study area in Punjab. Burned areas were extracted via 1) identifying double-cropping fields using seasonal NDVI changes, and 2) a three-step process: (Step1) used ΔBAI and ΔNBR for candidate extraction; (Step2) removed misclassified areas, like moist soil via spectral characteristics; (Step3) applied pixel-based time-series analysis. Results: Step1 thresholds were set at ΔBAI>55 and ΔNBR<-0.15. While ΔBAI misclassified irrigated fields and ΔNBR misclassified pre-harvest and post-harvest rice fields, combining both indicators reduced these errors. However, misclassification persisted in “immature vegetation”, “pre-harvest wheat”, “bare soil”, and “moist soil”. While 2G_RBi, CIgreen, and TWI excluded most misclassifications, they could not exclude bare soil. Therefore, we developed a new index, “flatness”. This index focuses on the low and flat spectral shape of burned areas across the visible to near-infrared ranges. This index successfully distinguished burned areas from bare soil. Finally, based on the double-cropping system, only the first detection for each pixel was recorded per harvest season to avoid double counting. This approach enables accurate calculation of burned areas and facilitates trend analysis.

Back to poster list

P-28: Estimation of Mountain Slope Deformation Using Time-Series InSAR Analysis with ALOS-2/PALSAR-2

Presenting author: Masaya KOIKE

Authors: Masaya KOIKE, Kagawa Univ.; Atsuko NONOMURA, Kagawa Univ.; Eiji TAMURA, NAIBA Co., Ltd.

Keywords: ALOS-2/PALSAR-2, In SAR, Slope deformation, Time-Series Analysis, Hazard

In recent years, the risk of slope failures has increased due to torrential rains and earthquakes. To mitigate such damage, it is necessary to detect temporal trends in slope displacement and identify areas where displacement becomes apparent in advance. Although Interferometric SAR has been widely used for crustal deformation analysis, its application to the estimation and monitoring of slope movements is increasingly expected. However, in mountainous areas, geometric distortions (including foreshortening, layover and shadow) and low coherence might degrade the spatial and temporal accuracy of displacement estimates derived from single interferometric pairs. In contrast, interferometric SAR time-series analysis enables the identification of areas exhibiting temporally consistent displacement trends by integrating multi-temporal observations. This study focused on the Aruse district of Miyoshi City, Tokushima Prefecture, where large-scale landslide blocks became visibly displaced following the heavy rains of July 2018. Within the active landslide block, road surface steps, retaining wall damage, and loosening of multiple anchor heads were observed during field investigations. However, the spatial and temporal monitoring of displacement has not been sufficiently conducted. Therefore, time-series InSAR analysis was applied using ALOS-2/PALSAR-2 data, and the spatial and temporal characteristics were investigated by evaluating slope displacement detection methods. To improve the reliability of deformation estimates, filtering and coherence-based masking were incorporated into the processing workflow to reduce speckle noise in interferograms and suppress phase unwrapping errors. The analysis revealed long-term and significant displacement in the mid-slope section of the block. In addition, areas with smaller estimated displacement exhibited less severe surface deformation. Comparison with field survey results confirmed that the estimated displacement corresponds to ground surface deformation. Furthermore, the estimated displacement suggests the possibility of shallow sliding occurring within the landslide block. These results demonstrate that interferometric SAR time-series analysis is effective for evaluating the characteristics of landslide displacement in mountainous regions.

Back to poster list

P-30: TIME-SERIES ANALYSIS FOR URBAN CHANGE MONITORING BASED ON MULTI-TEMPORAL HIGH-RESOLUTION NANOSATELLITE IMAGERY

Presenting author: Taehong KWAK, Seoul National Univ.

Authors: Taehong KWAK, Seoul National Univ.; Hyeongseong CHOI, Seoul National Univ.; Yongil KIM, Seoul National Univ.

Keywords: Multi-Class Change Detection; Unsupervised Learning; Land Cover Change; Feature Representation

Time-series satellite imagery is a key data for long-term urban change monitoring. However, urban environments exhibit high spatial heterogeneity, abrupt and irregular changes, and weak seasonal periodicity, making their temporal characterization particularly challenging. Consequently, most previous studies have relied on coarse-resolution imagery and simplified complex urban change processes into a unidimensional metric, focusing primarily on changes in extent or spatial distribution. To address these limitations, this study leverages multi-temporal, high-resolution nanosatellite imagery with fine spatial and temporal resolution for urban time-series analysis. We employ piecewise harmonic regression to model urban change as a set of multi-layered temporal patterns, enabling the explicit representation of heterogeneous and non-stationary dynamics. Specifically, we construct stable temporal profiles by leveraging regression-derived variables, including coefficients, segment lengths, and periodic components, which serve as the basis for computing urban dynamics indicators. As a result, we derive three complementary urban dynamics indicators: change type, change timing, and change intensity. The proposed framework showed that the indicators successfully captured diverse urban change patterns across different stages of urban development. This study highlights the potential of high-resolution time-series analysis for detailed and scalable urban change monitoring at the city scale.

Back to poster list

P-32: Comparative Evaluation of Direct and Hierarchical Machine Learning Models for Forest Volume Estimation

Presenting author: Hiroyuki AOKI

Authors: Hiroyuki AOKI, Hokkaido Univ; Takeo TADONO,JAXA EORC; Funuju TANAKA, Hokkaido Univ

Keywords: Sentinel-2; DEM; XGBoost; Forest volume; Machine learning; REDD+

A regional-scale forest biomass estimation plays a critical role in both climate change mitigation and forest management.Consistent and spatially extensive biomass estimates enable the quantification of carbon emissions and removals associated with deforestation and forest degradation and provide a technical basis for greenhouse gas inventories and the measurement, reporting, and verification (MRV) framework under REDD+. Ground-based forest inventories alone are insufficient to provide spatially continuous and temporally consistent biomass information over large areas, highlighting the importance of remote sensing-based approaches. Satellite remote sensing enables objective, repeatable, and wide-area observation of forest conditions, which is essential for national to continental-scale biomass monitoring.The objective of this study is evaluating machine learning models for estimating the forest volume of areas predominantly composed of Japanese cedar and cypress, utilizing Sentinel-2 imagery and Digital Elevation Model (DEM) data. First, we confirmed that incorporating DEM-derived topographic features alongside satellite imagery improved the accuracy of forest height estimation.Next, we compared a direct approach with a hierarchical approach for forest volume estimation using the Extreme Gradient Boosting (XGBoost) algorithm.We initially tested a hierarchical method that calculates volume based on predicted tree species, forest height, and basal area. However, predicting basal area using optical imagery proved highly challenging for our dataset due to the dense canopies and complex topography, leading to substantial errors in the final volume calculation.Consequently, we refined the model to utilize only the reliably predicted forest height as an intermediate parameter.The results demonstrated that this refined approach improved the accuracy of forest volume estimation compared to the direct method.In future work, we plan to incorporate Synthetic Aperture Radar (SAR) data to further improve accuracy.

Back to poster list

P-34: Dual-Polarization SAR Analysis for Rice Growth Monitoring: Correlation Between Sentinel-1 Backscatter and UAV-Derived Plant Height

Presenting author: Yudai TANAKA

Authors: Yudai TANAKA; Masayuki MATSUOKA, Mie Univ.

Keywords: Sentinel-1; synthetic aperture radar; backscattering coefficient; rice plant height; unmanned aerial vehicle (UAV)

Monitoring rice plant growth is essential for efficient crop management, particularly in Japan where the agricultural workforce continues to decline. Synthetic aperture radar (SAR) offers the advantage of weather-independent, wide-area observation. To effectively utilize this advantage, this study analyzes the relationship between Sentinel-1 SAR backscattering coefficients and rice plant height derived from unmanned aerial vehicle (UAV) point cloud data, comparing VH and VV polarizations. We captured high-resolution aerial photos using a UAV and generated 3D point cloud data. Rice plant height was estimated as the difference between the point cloud elevation on each observation date and the ground surface elevation on May 14, prior to transplanting. Sentinel-1 C-band SAR data in Interferometric Wide Swath (IW) mode with VH and VV dual polarizations were acquired at 12-day intervals from June to October 2025, yielding a total of 12 scenes. Backscattering coefficients were extracted by applying radiometric calibration and terrain correction to the Sentinel-1 data. Using these data, the correlation between backscattering coefficients and rice plant height was analyzed for 12 paddy fields in Tsu City, Mie Prefecture. VH polarization exhibited moderate positive correlations with rice plant height across all fields, with correlation coefficients ranging from 0.50 to 0.71. In contrast, VV polarization showed weaker and more variable correlations, with coefficients ranging from 0.06 to 0.63. These results indicate that VH polarization, which captures volume scattering from rice leaves and panicles, is more suitable for monitoring rice plant height than VV polarization. The findings suggest that dual-polarization SAR analysis combined with UAV-derived height references provides a promising framework for large-scale rice growth monitoring. Ultimately, this approach will contribute to precise yield estimation and labor-saving management in smart agriculture.

Back to poster list

P-36: Survey of Solar Panel Distribution Using Sentinel-2/MSI

Presenting author: Shiori FUJII

Authors: Shiori FUJII, Nara Women’s Univ.; Kanako MURAMATSU, Nara Women’s Univ.

Keywords: Sentinel-2/MSI, Solar panel, Spectral reflectance, Distribution, Extraction

The use of renewable energy is promoting as a measure against global warming. In Japan, the solar power generation occupies the highest share in renewable energy. Therefore, it can be considered the solar power generation plays a important role in reducing carbon dioxide emissions from renewable energy in Japan. It is necessary to accurately understand the amount and distribution of solar power generation in order to accurately assess the impact of solar power generation on a measure against global warming. However, it is difficult to understand the distribution in detail, such as the installation locations and equipment area. Since satellite data can observe wide areas, it is an effective means of understanding the distribution of solar power generation. This study aims to extract solar panels using satellite data and clarify their distribution. Reflectance data obtained from Sentinel-2/MSI was used. The study area is Himeji City, Hyogo Prefecture. The spectral reflectance characteristics for solar panels, vegetation, water areas, and urban areas were compared. The spectral characteristics of each land cover were compared. The solar panels showed small changes in reflectance from band 1 to band 8A, an increase in reflectance at band 11, and a decrease in reflectance at band 12. From these characteristics, the solar panels were extracted under the conditions that the difference between band 8A and band 11 is more than 0.01 and the difference between band 11 and band 12 is more than 0.03. As a result, the solar panels were mostly extracted, but the fields were incorrectly extracted. This is because the reflectance change between band 11 and band 12 was used as the extraction condition for solar panels, but the fields showed the same characteristics. From now on, we will apply the conditions to remove incorrectly extracted section and will evaluate the extraction accuracy.

Back to poster list

P-38: Deep Learning–Based Landslide Mapping Using Sentinel-1 SAR Time Series and DEM-Derived Features

Presenting author: Byunggun KIM

Authors: Byunggun KIM, Kangwon National Univ.; Hyangsun HAN, Kangwon National Univ.

Keywords: Landslide; Sentinel-1; SAR; DEM; Deep learning

Rapid landslides, including debris flows, can travel downslope with little warning, and result in severe casualties and infrastructure damage. Synthetic Aperture Radar (SAR) is well suited for rapid landslide mapping because it is less affected by weather conditions and illumination than optical imagery. In this study, we proposes a deep learning-based framework to detect landslide-affected areas by integrating Sentinel-1 SAR time-series data with DEM-derived topographic and hydrologic features. However, reliable identification of landslide-affected areas from SAR remains challenging due to variations in scattering behavior over vegetated and rain-wetted surfaces. To address this limitation, we construct a multi-source features composed of Sentinel-1 backscattering time-series metrics and DEM-derived variables that represent the topographic, geographic, and hydrologic characteristics of landslide-damaged areas. A deep learning-based segmentation model is trained to discriminate landslide from non-landslide areas. The results demonstrate that the integration of time-series SAR features and DEM-derived variables with a deep learning segmentation approach has a strong potential for accurate landslide mapping. The proposed framework can support rapid landslide delineation of landslide-affected areas for emergency response and hazard management.

Back to poster list

P-40: Applicability of Idle Farmland Category Discrimination Using Random Forest with Satellite Imagery and GIS Data

Presenting author: Konomi OTOMO

Authors: Konomi Otomo, Ibaraki Univ.; Hideyuki Tonooka, Ibaraki Univ.

Keywords: random forest, sentinel-2, GIS data, NDVI, idle farmland

The survey of idle farmland conducted every year by municipal agricultural committees in Japan requires significant labor and time. To alleviate this burden, in our previous study, we developed a random forest model using satellite imagery and GIS data to distinguish between non-idle farmland, mildly-degraded idle farmland (green class), and highly-degraded idle farmland (yellow class). In the cross-validation study, the method showed a consistent performance with a F1-score of approximately 75%, but it was also indicated that discovering more explanatory features is crucial for further improving accuracy. On the other hand, the idle farmland survey results may involve subjective differences in evaluation criteria, since these surveys are based on field inspections. Such inherent discrepancies between qualitative field assessments and quantitative satellite-based features may limit the achievable classification performance. Therefore, in this study, we analyzed the characteristics of misclassified fields and the relationship between prediction probability and performance to evaluate the applicability of the method. Since the previous analysis included excessively small farmland and cloud effects, we first performed classification targeting paddy fields with no cloud effects and an area of 300 square meters or larger, resulting that the misclassified farmlands revealed high rate of misclassification between the green class and the yellow class. Particularly, yellow-class farmland tended to be misclassified when NDVI values were relatively low, which indicates that it is difficult to precisely distinguish the degree of degradation using vegetation features used, but the method performed effectively for the farmlands that have consistently shown high NDVI values over a long period. Furthermore, using only data with higher maximum prediction probabilities consistently improved the F1-score. These results suggest that by focusing on fields where the model exhibits high confidence, the method can improve accuracy and serve as an effective supporting tool for field surveys.

Back to poster list

P-42: Observation of Typhoon Cold Wake Recovery Time in the Waters Surrounding Taiwan Using Himawari 8/9

Presenting author: Yung-Szu LIANG

Authors: Yung-Szu LIANG, National Taiwan Ocean University; Chung-Ru HO, National Taiwan Ocean University

Keywords: Himawari-8/9; sea surface temperature; typhoon; cold wake; recovery time; Taiwan waters

When a typhoon moves across the ocean, its strong winds cause vertical mixing by entraining subsurface seawater and induce upwelling through Ekman pumping. These processes cause a significant drop in sea surface temperature (SST) and create a “cold wake” along the typhoon’s path. After a typhoon passes, SST within the cold wake gradually returns to background levels; the associated recovery time reflects the upper ocean’s thermal resilience to atmospheric disturbances and varies across different typhoon events. This study focuses on the waters surrounding Taiwan, using high-spatiotemporal-resolution SST data from the Himawari 8/9 geostationary satellites and analyzing representative typhoon events from recent years. We quantitatively assess the SST recovery time within the cold wake region and compare differences across various typhoon events. The average SST measured several days before the typhoon’s passage is defined as the background field, while SST anomalies delineate the extent of the cold wake. Additionally, we employ the e-folding method to determine the recovery time of SST. The results indicate that the Kuroshio region experiences a shorter SST recovery time due to replenishment from horizontal thermal advection and higher ocean heat content. In contrast, the SST in the shallow waters of the Taiwan Strait takes longer to recover, as it primarily depends on solar radiation, leading to a significant “thermal lag” phenomenon.

Back to poster list

P-44: Statistical Analysis of Anomalous Transitions in the Kuroshio Surface Currents

Presenting author: Hsiang LEE

Authors: Hsiang LEE,National Taiwan Ocean University; Chung-Ru HO,National Taiwan Ocean University

Keywords: Kuroshio; anomalous; probability transition matrix; typhoon; surface currents

The Kuroshio is a major western boundary current flowing along the eastern coast of Taiwan and plays a critical role in regional climate, fisheries, and marine ecosystems. The Kuroshio varies under different weather conditions (e.g., typhoons) or ocean dynamics (e.g., oceanic eddies). Therefore, understanding the variations of Kuroshio is of considerable research interest. This study used satellite surface current data from the Copernicus Marine Environment Monitoring Service (CMEMS) to construct probability transition matrices for the current speed and direction in the Kuroshio region for the period 2023–2025. Statistical deviation measures are applied to identify anomalous transition behaviors in the surface current field, with a particular focus on flow features characterized by significantly reduced transition probabilities. The causes and variability of these anomalous behaviors are further examined to investigate the temporal and spatial variability of the Kuroshio and its recovery characteristics following disturbances. The results show distinct temporal and spatial patterns in the surface current under the influence of typhoons.

Back to poster list

P-46: A pixel-Level Comparison of Satellite and Commercial Aircraft Observation Frequencies for Coral Reef Monitoring in the Nansei Islands, Japan

Presenting author: Sora TAKAMIYA

Authors: Sora TAKAMIYA, NIT, Kisarazu College, NIES; Asahi SAKUMA, NIT, Kisarazu College, NIES; Hiroyuki OGUMA, NIES, Shinshu University; Ryotaro OKAMOTO, NIES; Masayuki KOJIMA, JAL; Daigo KAWAMURA, JAL; Shohei YAMAUCHI, JAL; Katsuto HARADA, JAL; Hiroya YAMANO, NIES, UTokyo

Keywords: Landsat; Sentinel-2; civil aircraft; flight data; simulation

Earth observation satellites, such as Landsat and Sentinel-2, are primary tools for coral reef monitoring because of their broad spatial coverage and regular revisit cycles. However, persistent cloud cover in Japan’s Nansei Islands poses a significant challenge for sustained, high-frequency observations. Cloud-induced data gaps, combined with spatial resolution limitations of satellite sensors, often impede the detection of fine-scale and short-term changes in reef environments. To address these limitations, commercial aircraft have been proposed as a complementary observation platform that leverages existing aviation infrastructure (Yamano et al., 2025). By utilizing routine inter-island routes, sensors can achieve very high temporal sampling and superior spatial resolution due to low flight altitudes (Takamiya et al., 2025). The feasibility of satellite observation in this region has been evaluated based on scene-level metadata (e.g., Akiyama and Kawamura, 2003). Scene-based metrics, however, often fail to capture localized cloud dynamics over small islands and coastal areas, which obscures the spatial bias between high-relief islands, prone to orographic cumulus, and low-lying islands. Moreover, no study to date has performed a direct, pixel-level comparison of effective observation capability between cloud-affected satellite imagery and commercial aircraft-based observations without cloud contamination. In this study, we shift the analytical unit from scenes to pixels to quantify the effective monitoring capabilities of satellites and commercial aircraft under various cloud conditions. Specifically, we (1) calculate and compare effective observation frequency by excluding cloud-contaminated pixels; (2) analyze the influence of topography and land cover on observation success to identify optimal locations and seasons for aircraft-based monitoring; and (3) propose a regular, multi-platform monitoring framework for episodic events such as sediment runoff and coral bleaching. This pixel-level approach provides a robust basis for integrating satellite and aircraft observations toward effective regional/local reef management.

Back to poster list

P-48: Influence of Southern Landfast Sea Ice on the Flow Variability of Drygalski Ice Tongue in East Antarctica Inferred from Sentinel-1 SAR Imagery

Presenting author: Junbeom PARK

Authors: Junbeom PARK; Hyangsun HAN

Keywords: Landfast sea ice, Drygalski Ice Tongue, Ice flow variation, Sentinel-1, SAR

This study investigates the factors controlling landfast sea ice variability along the southern side of Drygalski Ice Tongue (DIT), East Antarctica, and examines its relationship with DIT flow variability. Landfast sea ice was delineated using the correlation coefficient and interferometric coherence derived from Sentinel-1 SAR time-series imagery between February 2017 and February 2021. The derived landfast ice extent was validated against U.S. National Ice Center ice charts and showed strong agreement. Analysis of ERA5 and ORAS5 reanalysis indicated that landfast sea ice extent is more strongly influenced by atmospheric and oceanic temperatures than by wind conditions. Two-dimensional surface velocities of DIT derived from Sentinel-1 offset tracking revealed that transverse velocity variations are associated with changes in landfast sea ice extent. When landfast sea ice is present along the southern margin of DIT, the ice tongue exhibits enhanced southward bending. Greater landfast sea ice extent along the southern side of DIT corresponds to stronger southward bending of the ice tongue. These results suggest that landfast sea ice along the southern side of DIT plays an important role in modulating the ice flow and deformation of the ice tongue.

Back to poster list

P-50: Classification of Landfast Sea Ice in Terra Nova Bay, East Antarctica, Using Sentinel-1 SAR Backscattering and Interferometric Coherence

Presenting author: Hyewon SONG

Authors: Hyewon SONG, Kangwon National Univ.; Hyanson HAN, Kangwon National Univ.

Keywords: Landfast sea ice, SAR, Backscattering, Coherence, Machine learning

Landfast sea ice is anchored to the coast or an ice shelf and remains largely immobile for at least two weeks. Based on its formation mechanism, landfast sea ice can be classified into level fast ice (LFI), which forms through progressive coastal freezing, and rough fast ice (RFI), which develops through junction and consolidation processes involving drifting pack ice. The relative proportion of these landfast sea ice types can influence the persistence and breakup timing of the landfast fast ice region, influencing local oceanic environment. In this study, we compiled a time series of Sentinel-1 SAR images acquired in IW mode with HH polarization over Terra Nova Bay, East Antarctica, from 2017 to 2024. We derived backscattering coefficients, GLCM texture features, and two 12-day interferometric coherence from three consecutive SAR acquisitions. Using these features, we developed a machine learning-based landfast sea ice classification model to discriminate LFI and RFI. Model performance was evaluated against reference labels produced by manual interpretation of ICESat-2 sea ice freeboard, Sentinel-1 backscattering coefficients, coherence, and interferometric phases, and optical imagery. The proposed classification model achieved high performance in separating LFI and RFI, demonstrating the potential of combined SAR backscattering- and coherence-based features for monitoring landfast sea ice type variability.

Back to poster list

P-52: Measurements of PM10 Spatial Distributions in Seoul using High Spatial Resolution Aerosol Scanning LiDAR

Presenting author: Kwanuk Kye

Authors: Kwanuk KYE; Wonei CHOI; Hanlim Lee

Keywords: PM10, Aerosol Lidar, Horizontal Spatial Distribution, Metropolitan, Scanning

To establish an effective PM10 reduction strategy in a metropolitan city, highly resolved spatial distributions of PM10, known to have adverse health effects, are essential. Conventionally, in-situ instruments have been used to monitor PM10 concentrations in areas with multiple aerosol sources. While these instruments provide good accuracy, they have limitations in achieving high spatial resolution. In this study, we present a novel LiDAR (Light Detection and Ranging) technique combined with a machine learning–based PM10 retrieval algorithm, enabling measurements with high spatial resolution (7.5 m) and reliable accuracy. LiDAR observations were conducted at the Seoul Institute and Korea University, measuring PM10 distributions at one-hour intervals over half-circular scans with a 5 km radius covering Seocho-gu, Gangnam-gu, and Seongbuk-gu, where several in-situ monitoring sites are located. Elevated PM10 concentrations were detected in Seocho-gu and Gangnam-gu, including construction sites near Seoul Seonjeongneung, Eonju-ro, and densely populated residential and school areas near Seolleung Station, which were not clearly identified by conventional in-situ monitoring. In Seongbuk-gu, higher PM10 concentrations were observed at the westbound intersection of Dongdaemun History & Culture Park, the southbound intersection of Wangsimni Culture Park, and a commercial area near Myeongdong Cathedral. LiDAR-derived PM10 concentrations show strong agreement with in-situ observations. For S-DoT data, the correlation coefficient and mean bias are 0.96 and -0.73±12.25 μg/m³, respectively, with an RMSE of 12.27 μg/m³ and a percentile difference of 13.41±11.28%. For AirKorea data, the correlation coefficient and mean bias are 0.98 and -0.123±10.49 μg/m³, respectively, with an RMSE of 10.49 μg/m³ and a percentile difference of 11.40±9.28%. This study demonstrates that LiDAR is effective in identifying previously unresolved PM10 source locations and provides valuable aerosol information for satellite validation.

Back to poster list

P-54: Automated Traffic Sign Recognition and Photogrammetric Localization for Large-Scale Surveys

Presenting author: Hsin-Hung HU

Authors: Hsin-Hung HU; Yi-Tang LI; Jyun-Ping JHAN

Keywords: Mobile Mapping System (MMS); RTK-GNSS Georeferencing; YOLO Instance Segmentation; SuperPoint; LightGlue; Photogrammetric Forward Intersection; Agisoft Metashape; GIS Layer Publishing

Traffic sign maintenance in Taiwan predominantly relies on manual field inspections, resulting in high labor costs and limited digital traceability. This research presents a Mobile Mapping System (MMS) workflow designed to acquire continuous street-level imagery synchronized with high-precision positioning data. The system integrates wide-angle cameras and a roof-mounted RTK-GNSS for precise georeferencing, while a rear-wheel encoder triggers distance-based image capture at 1 m intervals to ensure spatiotemporal consistency between imagery and spatial coordinates. A pipeline comprising Detection, Matching, Localization, and Interactive Visualization is proposed. First, traffic signs are extracted using YOLO-based instance segmentation to generate precise masks and regions of interest (ROIs). Subsequently, SuperPoint and LightGlue are employed for inter-frame feature matching to establish consistent object IDs across sequential frames, thereby constructing multi-view observation tracks. These observations are then processed in Agisoft Metashape, where 3D coordinates are derived via photogrammetric forward intersection. Finally, the georeferenced outputs are published as queryable GIS layers to support downstream asset inventory and maintenance management. Preliminary validation across 15 sign categories and 22 sites achieve a Mask mAP@50 of 0.978. While the initial cross-image association success rate is 73%, this serves as a baseline for ongoing algorithmic refinements aimed at improving robustness under diverse real-world conditions. Overall, the proposed procedure provides an automated solution for rapid traffic sign recognition and localization, significantly enhancing the efficiency of large-scale field surveys.

Back to poster list

P-56: A method for estimating levee crest heights using AW3D enhanced DEM products

Presenting author: Takeru FURUICHI

Authors: Takeru FURUICHI, Ibaraki Univ.; Susumu TAKAGISHI, PASCO Corp.; Hiromune YOKOKI, Ibaraki Univ.; Hideyuki TONOOKA, Ibaraki Univ.

Keywords: AW3D; DEM; levee; flood simulation; RTK-GNSS

Accurate GIS data on levee locations and heights is critical input for flood simulation analyses of storm surges and floods, significantly influencing simulation accuracy. However, existing levee GIS data often contains inaccuracies due to discrepancies in data creation timelines or human errors during compilation. Therefore, this study developed a method to estimate levee crest heights using the AW3D Enhanced DEM product, a representative DEM generated from high-resolution optical satellite observation data. However, when the crest width is narrow or structures exist nearby, pixel blending or smoothing associated with image matching may prevent accurate crest height estimation. Therefore, our developed method estimates crest height through the following steps: First, a pseudo-centerline is generated from the levee polygon. At each point along this centerline, an initial estimate of the crest elevation is calculated using the elevation values of surrounding pixels. Next, by comparing the elevation of structures near each point and analyzing cross-section profile trends, areas affected by structures are detected and treated as missing data. Subsequently, remaining outliers are removed using statistical methods. Finally, the missing sections are interpolated using the surrounding levee heights as references. To evaluate this method, coastal and river levees in the central to northern regions of Ibaraki Prefecture, Japan, were targeted. The estimated values obtained by this method were evaluated against the ground truth values of levee crest heights measured on-site using PASCO Corporation’s SmartSOKURYO POLE, an RTK GNSS surveying system. As a result, the estimates obtained using this method tended to underestimate overall due to the spatial sampling limitations of the AW3D product, but the root mean square error for the crest elevation was less than 1 meter. Future plans include verifying the generalizability of this method to other regions and implementing improvements to further enhance its accuracy.

Back to poster list

P-58: Robust Phase Linking for SAR Interferometric Phase Estimation

Presenting author: Youngjae CHAE, Sejong Univ.

Authors: Taeseok LIM, Sejong Univ.; Youngjae CHAE, Sejong Univ.; Hana YOO, Sejong Univ.; Sang-Wan KIM, Sejong Univ.

Advances in Synthetic Aperture Radar (SAR) have expanded Interferometric SAR (InSAR) deformation monitoring. Persistent Scatterer InSAR (PSInSAR) enables millimeter level deformation analysis using time series phase from stable PS, However it is limited where PSs are sparse, so Distributed scatterer (DS) use is required. Phase Linking builds a Sample Covariance Matrix (SCM) from Statistically Homogeneous Pixels (SHP) around each DS and estimates an optimal phase vector with a stable scattering mechanism. Because the phase vector is estimated from the SCM, performance is sensitive to SCM quality. Statistical test errors during SHP extraction can introduce outliers and degrade phase estimation. Conventional Phase Linking (CPL) forms the SCM in a full stack structure using all acquisitions at once, so temporal decorrelation driven by acquisition time differences strongly affects it. To mitigate this, Sequential Phase Linking (SPL) uses mini stacks, and Non-Local Phase Linking (NPL) applies Non-Local weights. But SPL reduces the sample count per stack, which increases statistical test errors and can allow outliers to enter. NPL requires full stack processing plus weighting, which increases computational burden and limits long time series analysis. To address these limitations, the SCM is stabilized prior to Phase Linking using Oracle Approximating Shrinkage, which applies shrinkage to the SCM eigenvalue distribution and improves robustness under temporal decorrelation. Mahalanobis distance based weights reflecting the data structure are also applied to suppress outlier influence. The SPL mini stack structure is then combined with NPL Non-Local weights to improve both phase estimation performance and computational efficiency. In 1,000 Monte Carlo simulations, RMSE decreases by about 46% versus CPL, 32% versus SPL, and 9% versus NPL. Relative to NPL using the same Non-Local weighting, memory usage and processing time decrease by about 85% and 74%, confirming reduced computational burden while maintaining phase estimation performance.

Back to poster list

P-60: Cross-Regional Generalization of a Deep Learning Model for Short-Term Spatiotemporal Population Distribution Forecasting Using Mobile Spatial Statistics

Presenting author: Ryuzo TAKAHASHI

Authors: Ryuzo TAKAHASHI, Tohoku Univ.; Bruno ADRIANO, Tohoku Univ.; Erick MAS, Tohoku Univ.; Shunichi KOSHIMURA, Tohoku Univ.

Keywords: Mobile Spatial Statistics; Deep learning; Human mobility nowcasting

Rapid disaster response requires accurate estimates of populations exposed to hazards. Mobile phone network-derived data, such as Mobile Spatial Statistics (MSS), functions as a near real-time, high-resolution human population surface sensing system, capturing real-time population surfaces across regular grids. However, operational aggregation causes multi-hour data latency, forcing disaster management to rely on simple persistence or historical temporal averages. These operational heuristics introduce severe errors in dense urban environments undergoing rapid population shifts. To overcome this latency, we frames short-horizon population nowcasting as a spatio-temporal prediction task on grid-based remote sensing data, proposing ST-Swin. ST-Swin is a transformer-based model specifically designed for cross-region generalization without the need for additional retraining. The architecture employs window-based self-attention to capture local spatial interactions on regular grids, combined with explicit absolute time encoding (Time2Vec) to represent diurnal and weekly mobility patterns. To ensure robustness against large inter-region scale disparities and non-stationarity ST-Swin integrates reversible instance normalization with data-dependent non-stationary parameters. The model was evaluated across multiple first-level mesh regions in Japan using MSS data with a spatial resolution of approximately 500 m. ST-Swin demonstrated consistent superiority over operational baselines and LSTM models in cross-region transfer scenarios, particularly regarding enhancing accuracy in high-density urban cells. Furthermore, the model was validated using a real-world disaster scenario: the heavy rainfall in Omuta City in July 2020. ST-Swin successfully reduced exposed-population estimation errors to approximately one-third of the baseline. The results demonstrate the operational viability of the model as a reliable, rapidly deployable component for near real-time exposed-population estimation in disaster response workflows.

Back to poster list

P-62: Moving Target Velocity Estimation Using Single-Channel SAR Imagery

Presenting author: Juyoung SONG

Authors: Juyoung SONG, Seoul National University; Duk-jin KIM, Seoul National University

Keywords: SAR; MTI; Velocity Estimation; Refocusing; AIS; ADS-B

Synthetic Aperture Radar (SAR) is an essential tool for all-weather remote sensing, yet traditional velocity estimation via multi-channel techniques like ATI or DPCA is often limited by the scarcity of multi-channel hardware on satellite platforms. To overcome this, the study proposes a robust Target Motion SAR Phase Refocusing algorithm designed for single-channel imagery. The method derives a refocusing function to model phase distortion from azimuth velocity, utilizing a minimum-entropy optimization strategy to identify the velocity that minimizes image defocusing. The algorithm’s accuracy was validated using high-resolution SAR data from ICEYE and Umbra satellites alongside the UAVSAR airborne system. Experiments involving 42 AIS-equipped vessels and 20 ADS-B-equipped aircraft yielded average velocity errors of 0.59 m/s and 2.22 m/s, respectively. Results confirm the technique is highly effective for diverse moving targets, including ships, aircraft, vehicles, and trains. With a processing time of under 5 seconds for each target, the algorithm is suitable for real-time applications. Beyond velocity estimation, the process enhances moving target focusing performance, significantly aiding in accurate identification.

Back to poster list

P-64: Automated Measurement of Intensity Quality in MMS LiDAR Data

Presenting author: Seungmin NOH

Authors: Seungmin NOH, KRIHS; Kyoungah CHOI, KRIHS

Keywords: MMS LiDAR; Intensity Measurement; Quality Assessment; Automated Method; Analysis of Similarities(ANOSIM)

MMS LiDAR data have become an essential source of high-resolution 3D spatial information for road infrastructure management and High-Definition map production. In current HD road map construction guidelines, intensity is designated as a major quality indicator. however, it is commonly assessed subjectively based on whether human inspectors can visually distinguish road markings. Such inspection-based evaluation is labor-intensive and lacks reproducibility and scalability for large-scale datasets. This study proposes an automated and quantitative framework for evaluating intensity quality from a purpose-driven usability perspective. Focusing on crosswalk regions, intensity quality is defined by the distinguishability of road markings from asphalt surfaces. In crosswalk areas, intensity values typically exhibit a bimodal distribution representing asphalt surfaces and road markings. We assume that reliable identification of road markings requires sufficient mean separation between the two groups, limited distribution overlap, and preservation of the bimodal structure. To evaluate these structural properties, two complementary approaches were adopted. First, a Gaussian Mixture Model was applied to partition intensity values into high- and low-reflectance groups, and structural separability was quantified using ANOSIM (Analysis of Similarities). Second, relative and absolute dynamic range metrics were analyzed to assess range-based sensitivity. To validate the proposed metrics, multiple degradation scenarios, including increased incidence angle, contrast reduction, unimodal collapse, and range compression, were synthetically generated from high-quality datasets. Segmentation accuracy was computed using the original dataset as ground truth, and correlations between segmentation accuracy, range-based indices, and ANOSIM-based separability were examined. The results indicate that range-based metrics reflect visual distinguishability but do not adequately explain segmentation accuracy or noise effects. Structure-based separability metrics show limitations in capturing perceptual visibility. This study establishes a reproducible and automated framework for intensity quality assessment and contributes to a systematic and quantitative quality control scheme for LiDAR intensity data.

Back to poster list

P-66: Evaluation of the Range–Doppler Model for Capella Spotlight Imagery Across Different Imaging Formations

Presenting author: Tee-Ann TEO

Authors: Tee-Ann TEO, National Yang Ming Chiao Tung University; Tsung-Han WEN, National Yang Ming Chiao Tung University; M.C. WU, Taiwan Space Agency; Y.S. LEE, Taiwan Space Agency

Keywords: Synthetic Aperture Radar (SAR), Range–Doppler Model, Capella Space, Spotlight Mode, Back-Projection Algorithm (BPA), Polar Format Algorithm (PFA), Geometric Correction.

Geometric correction is fundamental for maintaining the spatial integrity of Synthetic Aperture Radar (SAR) data. The Range–Doppler (RD) model, which incorporates backward and forward geocoding, serves as the primary sensor model for establishing the relationship between 3D ground coordinates and 2D image coordinates. This study evaluates the RD model’s performance for Capella Space Spotlight imagery, specifically comparing two distinct imaging formations: the Back-Projection Algorithm (BPA) and the Polar Format Algorithm (PFA). Unlike Stripmap mode, which uses a fixed antenna pointing, Spotlight mode steers the beam toward a specific target to enhance azimuthal resolution, necessitating precise geometric modeling. While BPA operates in the time domain to handle non-linear flight paths, PFA utilizes frequency-domain processing for computational efficiency. In this research, RD models were developed for both formations and validated against the SICD SAR Toolbox benchmark. Experimental results using BPA data from Singapore and PFA data from Los Angeles demonstrate sub-meter positioning accuracy. For backward geocoding, the PFA model yielded mean range and azimuth differences of 0.004 and -0.014 pixels, respectively, while the BPA model showed -0.103 and -1.002 pixels. In forward geocoding, mean errors in WGS84 (X, Y, Z) were -0.310 m, -0.070 m, and -0.271 m for BPA, and -0.334 m, 0.227 m, and -0.738 m for PFA. The findings indicate that while both formations support high-precision applications, the PFA-based RD model exhibits superior mathematical alignment with standard SICD projection models. This study provides a critical technical basis for the geometric calibration of high-resolution SAR systems.

Back to poster list

P-68: Pre-launch Wavelength and Radiometric Calibration of Hyperspectral Sensors Using an Offner Spectrometer

Presenting author: Yuya KOAKUTSU

Authors: Yuya KOAKUTSU,Tokyo Univ.; Daichi MATSUOKA,Tokyo Univ.; Takahiro KAWASHIMA,Tokyo Univ.; Akira IWASAKI,Tokyo Univ.

Keywords: Hyperspectral remote sensing, Pre-launch calibration, Wavelength assignment, Radiometric response, Planck’s Law

In recent years, advancements in detectors and diffraction gratings have led to an increase in sensors with broader observation wavelength ranges. For instance, the Hyperspectral Imager SUIte (HISUI), a Japanese sensor onboard the International Space Station (ISS), acquires wide-range spectral data from 400 nm to 2500 nm. These data are utilized in various fields, such as precision agriculture, geological exploration, and water environment monitoring. Since these wide-band observations capture numerous atmospheric absorption features (e.g., water vapor) for use in retrieval analysis, precise pre-launch calibration of both wavelength and radiometry is essential. In this work, we performed multi-point wavelength calibration using both first-order emission lines and their higher-order diffraction peaks from a xenon lamp. This approach improved the accuracy of wavelength assignment for each detector element. Validation using band-pass filters demonstrated a wavelength estimation accuracy of 0.379 nm (0.3 pixels). Regarding radiometric calibration, we derived response characteristics using a planar blackbody. We then reduced calibration uncertainty by comparing these results with theoretical simulations that modeled the relationship between input radiance (based on Planck’s Law) and the detector’s digital numbers (DNs). These methods establish a robust framework for converting raw data into physical quantities, enabling advanced image analysis and retrieval using wide-band hyperspectral data.

Back to poster list

P-70: Automated method for assessing cyanobacterial levels using image classification based on deep learning model with RGB images

Presenting author: Manabu WATANABE

Authors: Manabu WATANABE, blue and tech, Ltd.; Hazuki NEMOTO, blue and tech, Ltd.; Masayuki FUJITA, blue and tech, Ltd.; Akemi TANAKA, blue and tech, Ltd..; Natsuko KAWAKAMI, Hiyoshi Corporation.; Tomoharu SHIMONO, Hiyoshi Corporation.

Keywords: cyanobacteria; blue-green algae; Appearance Algal Bloom Index; deep learning; YOLO; RGB images

Since around 2016, Lake Nishinoko in Omihachiman city, Shiga prefecture, Japan, has experienced deteriorating water quality, resulting in recurrent massive phytoplankton blooms. Cyanobacteria (blue-green algae) have appeared almost annually, causing musty odors at nearby drinking water treatment plants. Given the projected progression of global warming, further deterioration of water quality is a growing concern. In response, Shiga Prefecture has implemented measures to improve the lake’s water environment and has monitored algal occurrence at fixed monitoring sites using “Appearance Algal Bloom Index.” This index is a practical method for visually evaluating algal activity based on predefined criteria. However, when humans conduct visual assessments, it is necessary to visit monitoring sites regularly. Consequently, while algae activity is known to change hourly, multiple monitoring within a single day are rarely conducted. To address these limitations, this study develops a automated method for assessing algal levels using image classification based on a deep learning model with RGB images as input. Hourly RGB images of the lake surface were acquired from August 2 to November 30, 2024, using monitoring equipment installed at Lake Nishinoko. Appearance Algal Bloom Index was then automatically estimated using YOLO Classify v8, a convolutional neural network–based image classification model provided by Ultralytics. A total of approximately 1,400 images captured during daytime were used for model training and validation. Model performance was evaluated under three training/validation split ratios: 85%/15%, 75%/25%, and 50%/50%. The results demonstrated the feasibility of automatically estimating Appearance Algal Bloom Index. Among the data classified as algal levels level 1 or higher, it was found that levels 4 and 5—both representing high algal activity—could be identified with relatively high accuracy, with average accuracies of approximately 80% for each level. In contrast, the accuracy for level 1 classification was approximately 50%, indicating the need for further refinement.

Back to poster list

P-72: Benchmarking FAST-LIO2 and LIO-SAM SLAM Systems in Complex Urban Environments

Presenting author: Yen-Ru Lai

Authors: Chi-An-Chen, National Chung Hsing University; Yen-Ru Lai, National Chung Hsing University; Ching-Ming-Lai, National Chung Hsing University; Wen-Han-Chiang，National Chung Hsing University；Hao-Ting-Lin，National Chung Hsing University

Keywords: SLAM, FAST-LIO2, LIO-SAM

Complex urban environments present substantial challenges to LiDAR–IMU-based Simultaneous Localization and Mapping (SLAM) systems due to structural occlusions from high-rise buildings, limited geometric diversity, dynamic traffic participants, and frequent viewpoint changes. Such urban canyon characteristics can reduce feature observability, introduce scan-to-map ambiguities, and increase accumulated estimation drift, thereby affecting localization stability and map consistency. Although numerous open-source SLAM frameworks have been proposed, quantitative validation under real-world urban driving conditions remains limited. This study benchmarks two representative tightly coupled LiDAR–Inertial Odometry (LIO) frameworks, FAST-LIO2 and LIO-SAM, in urban environments. FAST-LIO2 adopts a filter-based architecture that uses an iterated Extended Kalman Filter (iEKF), direct raw-point-cloud registration, and an incremental k-d tree map structure, emphasizing computational efficiency and real-time state estimation. In contrast, LIO-SAM integrates feature-based front-end odometry with factor-graph optimization and loop-closure to enhance global map consistency. These systems represent two distinct design philosophies: local filtering-based state estimation versus graph-based global optimization. A multi-sensor platform integrating a multi-beam LiDAR and a high-precision IMU was deployed for field experiments in dense urban road networks. A NovAtel tactical-grade GNSS/INS navigation system was employed solely as the reference trajectory to quantitatively evaluate SLAM navigation accuracy. Performance metrics include Absolute Trajectory Error (ATE), Relative Pose Error (RPE), and qualitative assessment of map structural integrity under challenging conditions such as high-speed motion and sharp turns. By benchmarking these SLAM architectures against a tactical-grade reference in complex urban scenarios, this study provides insights into their performance characteristics and architectural trade-offs for urban autonomous navigation applications.

Back to poster list

P-74: Segmentation-Based NN-PSI with Change Point Detection for Nonlinear Deformation Reconstruction

Presenting author: Ho-Yeong YOU, Sejong Univ.

Authors: Ho-Yeong YOU, Sejong Univ; Gyeongmin KIM, Sejong Univ.; Boram LEE, Sejong Univ.; Sang-Wan KIM, Sejong Univ.

Persistent Scatterer Interferometry (PSInSAR) is a time-series synthetic aperture radar (SAR) interferometric technique capable of estimating ground deformation with millimeter-level precision by analyzing signals from stable scatterers over long periods. However, conventional PSInSAR generally assumes a linear deformation model, which can lead to estimation errors when the actual ground motion exhibits nonlinear temporal behavior. In such cases, phase ambiguity and model mismatch may degrade the reliability of deformation retrieval. To address this limitation, Ogushi et al. (2019) proposed Non-linear Non-parametric PSInSAR (NN-PSI), which reconstructs deformation by analyzing a velocity spectrum derived from the observed interferometric phase. This approach enables the retrieval of nonlinear deformation without assuming a parametric displacement model. Nevertheless, when a single velocity spectrum is applied to the entire observation period, the reconstruction accuracy may decrease if deformation occurs only during specific time intervals. In such cases, stationary periods tend to dominate the spectral energy, making it difficult to accurately recover localized deformation signals. In this study, we propose a segmentation-based NN-PSI framework that incorporates change point detection (CPD) to identify temporal transitions in deformation behavior. Specifically, change points are detected by minimizing a cost function based on the circular variance of the wrapped phase time series, partitioning the observation period into segments with statistically consistent deformation characteristics. NN-PSI reconstruction is then performed independently for each segment, allowing the velocity spectrum to better capture the deformation dynamics within each interval. To evaluate the proposed approach, simulated deformation phases representing bilinear, step-like, and exponential patterns were generated, and Monte Carlo simulations were conducted to compare the conventional and segmented NN-PSI. The results demonstrate that the proposed method reduces mean RMSE, particularly when deformation occurs intermittently within the observation period. Future work will focus on validating the framework using real SAR datasets over areas with nonlinear deformation.

Back to poster list

P-76: Integrating Multi-temporal SAR Features and Ensemble Feature Selection for High-Precision Forest Type Classification

Presenting author: Geng Gui WANG, National Chung Hsing University

Authors: Geng Gui WANG, National Chung Hsing University; Hui Ping TSAI, National Chung Hsing University

Keywords: Forest Type Classification; Sentinel-1; Ensemble Feature Selection; XGBoost; LightGBM

Forest ecosystems play a critical role in global carbon cycling and climate regulation. However, climate change and anthropogenic disturbances threaten forest stability, making the precise monitoring of forest type distribution a primary task for assessing biomass and carbon sink functions. Although remote sensing has become the mainstream for large-scale surveys, frequent cloud cover and complex terrain in alpine regions have long limited the effectiveness of traditional optical sensors. To overcome these technical bottlenecks, this study focuses on Taiwan’s Shei-Pa National Park to develop a high-precision forest type classification framework integrating multi-temporal Sentinel-1 Synthetic Aperture Radar (SAR) imagery and topographic factors. Using 2016 Sentinel-1 data, we extracted backscattering coefficients, Gray-Level Co-occurrence Matrix (GLCM) texture features, polarimetric decomposition parameters (entropy, anisotropy, alpha angle), and the Radar Vegetation Index (RVI), and constructed a multidimensional feature space using Digital Elevation Model (DEM) data. To address high-dimensional redundancy, a Heterogeneous Centralized Ensemble Strategy integrating ReliefF, mRMR, and Recursive Feature Elimination (RFE) was employed to optimize features. We evaluated the classification performance of XGBoost and LightGBM across coniferous, broad-leaved, and mixed forests. The results show that both models effectively reduced parameter counts, with XGBoost achieving peak performance when combined with 56 features selected via RFE. On a temporal scale, seasonal features incorporating phenological characteristics reached an Overall Accuracy (OA) of 0.958. Notably, the F1-Score for mixed forests increased significantly from 0.779 to 0.934 after introducing seasonal features, effectively breaking the bottleneck of frequent misclassification in traditional research. Furthermore, GLCM texture features exhibited the highest stability and retention rates, reflecting significant differences in forest structure. This study demonstrates that multi-temporal SAR combined with machine learning can precisely capture vegetation dynamics, providing a vital digital monitoring foundation for future alpine forest carbon sink assessments.

Back to poster list

P-78: Autonomous relay based navigation and high speed data relay system for satellite constellations

Presenting author: Sang-Youn SHIN

Authors: Sang-Youn SHIN, Korea Aerospace Research Institute; Yee-Jin CHEON, Korea Aerospace Research Institute

Keywords: satellite constellation, high speed data link, x-band

This study proposes an autonomous relay architecture for satellite constellations where a relay satellite collects payload data via inter-satellite links and performs high-speed X-band downlink. Centralized navigation and timing improve efficiency, reduce onboard complexity, and support scalable and flexible autonomous operation.

Back to poster list

P-80: On-board Edge Computing Framework for Real-time Rice Lodging Detection via Multi-task Learning and Multispectral Imagery

Presenting author: Hsin-Hung TSENG

Authors: Hsin-Hung TSENG, National Chung Hsing Univ.; Ming-Der YANG, National Chung Hsing Univ.

Keywords: Edge Computing; UAV; Multispectral Imagery; Rice Lodging; Multi-task Learning

The rapid expansion of Unmanned Aerial Vehicles (UAVs) in remote sensing is often hindered by the latency of traditional post-flight image processing. This delay is critical in time-sensitive tasks like crop monitoring and disaster assessment, where missing immediate data can lead to mission failure. Furthermore, developing robust AI models for these tasks is frequently challenged by data scarcity and fragmented landscapes. To address these limitations, this study proposes an on-board edge computing framework designed for real-time image processing during UAV missions. Using rice lodging detection from multispectral images as a validation case, we developed the RBMTL model based on the EDANet architecture. This high-performance multitask learning model incorporates attention mechanisms and feature fusion to enhance recognition in complex scenarios. Concurrently, a weight fine-tuning transfer learning strategy is introduced to maximize performance using limited annotated samples. Experimental results indicate that RBMTL outperforms baseline models, improving F1-scores by 3% to 10%. The model achieved an F1-score of 97.65% for the background class and 85.52% for the rice lodging class. For deployment, the system was optimized on the NVIDIA Jetson AGX Orin platform, achieving inference speeds exceeding 30 frames per second (FPS) on 1536 × 2064 images. This research successfully integrates advanced computational architecture with efficient AI to create an on-board, real-time remote sensing solution. The system significantly enhances efficiency and provides key technological support for real-time precision agriculture and rapid disaster assessment.

Back to poster list

P-82: Benchmarking SpectralGPT Against Traditional OBIA-SVM for Solar Site Selection in Fragmented Indigenous Terrains.

Presenting author: Hsuan-Yin HUANG

Authors: Hsuan-Yin HUANG, National Chengchi Univ.; Fang-Shii NING, National Chengchi Univ.

Keywords: SpectralGPT; Foundation Model; Solar Site Selection; Indigenous Resilience; Kompsat-3A; Comparative Study

Due to terrain-related damage to infrastructure, indigenous communities in Taiwan which in mountain areas often experience power outages during natural disasters. To address this challenge, establishing self-sufficient solar power systems is an essential strategy to building community resilience. However, achieving efficient land-use classification and site selection in these fragmented and complex terrains is technically challenging. Traditional methods, such as Object-Based Image Analysis (OBIA) combined with Support Vector Machines (SVM), have been widely applied but are restricted by the heavy dependency on manual parameter adjustments, including factors like scale, shape, and so on. These limitations reduce the scalability and cross-scene adaptability in diverse landscapes. This study explores an alternative by comparing traditional OBIA-SVM workflows with SpectralGPT, an advanced spectral remote sensing foundation model. The research evaluates SpectralGPT’s pre-trained encoder for its ability to automatically extract detailed spatial-spectral features, using high-resolution Kompsat-3A multispectral imagery (0.55m) of the Anpo Tribe in Pingtung, Taiwan. A comparative analysis is conducted through linear probing to test classification accuracy under small-sample conditions for identifying “available land” under scenarios with limited training data. The preliminary assessments indicate that the SpectralGPT-based framework delivers results suitable for practical land-use mapping, effectively capturing textural details in complex terrains with minimal manual adjustments. By transitioning from subjective parameter settings to automated feature learning, it emphasizes a more efficient approach for disaster-resilient planning by employing foundation models for geospatial decision-making at the community level in vulnerable and isolated regions.

Back to poster list

P-84: Bridging Taxonomy Gaps: A Label-Invariant Fusion Strategy for Multi-Dataset High-Resolution Aerial Segmentation

Presenting author: Jae Young CHANG

Authors: Jae Young CHANG, Korea Aerospace Research Institute; Kwang-Jae LEE, Korea Aerospace Research Institute; Yoon-Soo KIM, Korea Aerospace Research Institute

Keywords: Deep learning, Land cover, Semantic segmentation, High-resolution satellite imagery, Multi-dataset fusion

The primary limitation of existing deep-learning-based land cover segmentation models for high-resolution aerial images is their inconsistent performance. They frequently struggle with performance when presented with new, untrained images. This issue arose from two main reasons: the limited spatial and temporal ranges of datasets, along with poor-quality labels, which led to model overfitting. There were various approaches to addressing this issue. Foundation models address the challenge by enhancing the intrinsic strengths of backbones, allowing for minimal transfer learning to yield more robust performance. Unsupervised domain adaptation methods aim to help a model generalize to unlabeled target data. The OpenEarthMap dataset aimed to integrate various land-cover datasets derived from images taken by different sensors into a single unified label space. While these methods showed significant performance, they are all indirect approaches to achieving a specific objective and therefore require considerable additional work, such as transfer learning. This study introduces a new method for integrating heterogeneous datasets that have different class compositions. For each dataset, redundant models predict the class probabilities for every pixel location. These probabilities are then combined to create a unified label-space probability. By training a single model on this unified dataset, we achieve robustness across a variety of images while also catering to specific target customizations. We applied this method and validated its effectiveness using three datasets: the FLAIR dataset, the OpenEarthMap dataset, and the Cloud Dataset from AIHub in Korea. The results showed improved robustness in customized target classification and allowed the model to identify regions obscured by clouds.

Back to poster list

P-86: Case study of the Influence of upstream oil palm expansion on flooding in an Aceh catchment, Indonesia

Presenting author: Chihiro NAITO

Authors: Chihiro NAITO, The University of Tokyo; Wataru TAKEUCHI, The University of Tokyo

Keywords: Sentinel-1, Watershed, Cumulative precipitation, Topographic features, Flow accumulation, Hierarchical Bayesian analysis, Land cover, Spatial correlation

A massive flood disaster hit northern Sumatra, Indonesia, from late November to December 2025. Allegations suggest that deforestation may have caused the severe flooding. Oil palm, which is a vital crop regionally and globally, is often one of deforestation drivers in Indonesia. This study aimed to investigate the influence of upstream oil palm expansion on downstream flood occurrence within a catchment, to provide scientific evidence for sustainable oil palm production planning. Sentinel-1 descending orbit data in 2016, 2020, and 2025 were utilized to detect flooded areas using a Random Forest classifier on Google Earth Engine, focusing on events following precipitation exceeding the 95th percentile. The dependent variable was defined as the increase in flood frequency per pixel between two periods. Explanatory variables included changes in water flow accumulation weighted by upstream oil palm presence, differences in maximum cumulative precipitation, topographic features, and distance from water bodies. One thousand random points were sampled from pixels where flow accumulation changed, and a hierarchical Bayesian correlation analysis was performed. The results indicated that flood frequency increased across a wide range of the study site, particularly between 2016 and 2020. The Bayesian analysis revealed a positive correlation for the increase in oil-palm-weighted flow accumulation, with a slightly stronger influence on flooding than changes in cumulative precipitation, especially between 2016 and 2020. These findings suggest that oil palm expansion may exacerbate flood frequency by affecting surface water flow. While further hydrological investigation is needed, this study provides statistical evidence linking upstream oil palm expansion to downstream flood events, highlighting the need for careful spatial planning and zoning for future plantations.

Back to poster list

P-88: Quantifying Multi-Temporal Domain Shift in UAV-Based Sweet Potato Plant Localization Using YOLO

Presenting author: Yu-Chun HSU

Authors: Po-Hsien LU, Department of agronomy, National Chung Hsing University (NCHU); 2. Chiayi Branch, Taiwan Agricultural Research Institute, Ministry of Agriculture; Shu-Yun CHEN, Department of agronomy, National Chung Hsing University (NCHU); Yu-Chun HSU, Department of Forestry, National Chung Hsing University (NCHU)

Keywords: Multi-temporal imagery; Domain shift; Object detection; Plant localization; UAV;

Multi-temporal unmanned aerial vehicle (UAV) imagery plays an increasingly important role in high-resolution crop monitoring. However, rapid canopy development across growth stages introduces substantial temporal domain shift, which significantly affects object detection performance. This study investigates multi-temporal variability in UAV-based plant localization using a heterogeneous sweet potato hybrid population. Time-series RGB imagery was collected over an experimental field containing diverse hybrid progenies, representing substantial variability in canopy architecture and growth dynamics from early sparse stages to dense canopy closure. Orthomosaic images were generated for each acquisition, and 785 individual plants were manually annotated to construct a cross-stage object detection dataset. A lightweight YOLOv11-nano detector initialized with COCO pre-trained weights was fine-tuned using stage-specific UAV imagery. Three temporal training strategies, Single-to-Single, Single-to-Multiple, and Multiple-to-Multiple, were systematically evaluated under identical model settings to isolate the effect of temporal data configuration. Results demonstrate pronounced temporal domain shift across growth stages, particularly under heterogeneous canopy conditions. Single-stage training led to up to 20% reduction in F1-score when applied to different stages. In contrast, multi-temporal training improved cross-stage robustness, increasing mean F1-score by over 15% and reducing center-point localization error by approximately 2 cm. Detection precision remained above 0.90 across representative canopy structures. These findings highlight the importance of temporal representativeness in UAV-based agricultural object detection and provide quantitative evidence that multi-temporal training mitigates domain shift in phenotypically diverse crop populations.

Back to poster list

P-90: Building Damage Mapping Using Dual-Orbit DInSAR and SAR–Optical Data for the 2024 Noto Earthquake Case Study

Presenting author: Yifan YANG

Authors: Yifan YANG, Japan Advanced Institute of Science and Technology; Arliandy P. Arbad, Institute of Industrial Science, The University of Tokyo; Khin Myat Kyaw, Institute of Industrial Science, The University of Tokyo; Hideomi GOKON, Japan Advanced Institute of Science and Technology; Wataru TAKEUCHI, Institute of Industrial Science, The University of Tokyo

Keywords: DInSAR; building damage assessment; SAR–dual-orbit; Random Forest; Noto earthquake

On 1 January 2024, a magnitude 7.6 earthquake struck the Noto Peninsula, Japan, causing catastrophic building damage in Suzu City. Rapid building-scale damage assessment is vital for effective emergency response and recovery planning. This study presents a robust screening approach by integrating dual-orbit Differential Interferometric SAR (DInSAR) displacement with multi-source remote sensing indicators. Using ALOS-2 (ascending/descending) and Sentinel-1 SAR data alongside Sentinel-2 optical imagery, building-level features were extracted for 925 validated structures. Three Random Forest classification models were evaluated to identify the optimal feature combination. The model integrating dual-orbit DInSAR displacement achieved the highest performance (AUC = 0.702), significantly outperforming single-orbit alternatives. Feature importance analysis reveals that while 3D-sensitive displacement is the primary driver for damage identification, the fusion of radar backscatter and optical indices provides critical complementary information. These findings demonstrate that multi-geometry SAR observations, when synergized with optical data, can effectively prioritize field surveys and enhance large-scale post-earthquake damage screening.

Back to poster list

P-92: Long-Term Precision Assessment of ICESat-2 Laser Altimetry over the Greenland Ice Sheet (2018–2025)

Presenting author: An-Ting Chang

Authors: An-Ting Chang, National Chung Hsing University; Yen-Ru Lai, National Chung Hsing University

Keywords: ICESat-2; Laser Altimeter System; Greenland ice sheet; Crossover Analysis; Elevation Precision Assessment

The accelerating mass loss of the Greenland Ice Sheet (GrIS) is a major contributor to global sea-level rise, underscoring the need for stable, high-precision elevation measurements to quantify long-term ice mass balance. Satellite laser altimetry provides a critical observational framework for detecting subtle surface elevation changes across diverse topographic and climatic regimes. ICESat-2, equipped with the Advanced Topographic Laser Altimeter System (ATLAS), delivers centimeter-level elevation measurements through photon-counting technology and a six-beam configuration with dense along-track sampling. Its high spatial resolution and sensitivity to surface slope enable detailed characterization of elevation variability across both the interior and marginal regions of the GrIS. In this study, we conduct a long-term assessment of ICESat-2 intra-mission precision using ATL06 Version 7 land-ice elevation data spanning October 2018 to August 2025. Precision is evaluated using a near-time (≤30 days) ascending–descending crossover analysis, with crossover elevations derived from locally fitted polynomial representations of satellite tracks, followed by a 3σ iterative outlier-removal procedure. The standard deviation of the crossover residuals serves as the primary precision metric. Beyond ice-sheet-wide statistics, we further characterize spatial variability by dividing GrIS into grid cells and estimating crossover-based precision within each cell. Results indicate that ICESat-2 maintains stable centimeter-level precision over the seven-year observation period. Over relatively flat interior regions, precision is typically within several centimeters, whereas in steeper terrain (~1° surface slope), the crossover dispersion increases to approximately 10-15 cm. Beam-combination statistics (strong–strong, weak–weak, and strong–weak crossovers) show no systematic precision differences among beam types, confirming the internal consistency of the six-beam configuration. Overall, this long-term analysis demonstrates the robustness and spatial stability of ICESat-2 elevation measurements for monitoring changes in the Greenland ice sheet.

Back to poster list

P-94: Wildfire Burn Severity and Scarring in Protected Areas Using Sentinel-2 dNBR for Rapid Post-Fire Mapping: Evidence from the 2024 Belize Wildfires

Presenting author: Rosalia DOMINGUEZ

Authors: Rosalia DOMINGUEZ; Tang-Huang LIN, National Central University

Keywords: Wildfire burn severity; dNBR; Sentinel-2; Post-fire mapping; Protected areas; Belize; USGS classification

Wildfire impacts during Belize’s 2024 fire season were quantified using a differenced Normalized Burn Ratio (dNBR) framework to map burn severity across the combined extent of Chiquibul Forest Reserve and Mountain Pine Ridge Forest Reserve. Sentinel-2 Surface Reflectance (HARMONIZED) imagery was processed in Google Earth Engine using a pixel-level cloud and shadow masking workflow (s2cloudless probability with shadow projection and buffering) to generate robust pre-fire (1 December 2023–14 February 2024) and post-fire (15 June 2024–30 September 2024) composites. The Normalized Burn Ratio (NBR) was calculated from near-infrared (B8) and shortwave infrared (B12) bands, and dNBR was derived as NBR_pre-NBR_post. To ensure analytical consistency, the final dNBR surface was restricted to pixels valid in both temporal composites, minimizing bias from cloud contamination or incomplete coverage. Burn severity was classified directly from the dNBR surface using USGS threshold conventions and burned area reporting followed the USGS standard definition of severity classes 2–5 (low to high severity). Area statistics were computed at 20 m spatial resolution. Across the combined protected-area landscape (209,907.48 ha), total burned area reached 29,304.52 ha, representing 14.0% of the extent. Fire impacts were highly heterogeneous: Mountain Pine Ridge exhibited 18,799.92 ha burned (43.4% of its area), whereas Chiquibul recorded 10,518.12 ha (6.3%). Severity distribution derived from dNBR was dominated by low-severity burns (18,541.36 ha), with smaller extents in moderate–low (5,582.56 ha), moderate–high (3,692.48 ha), and high severity (1,488.12 ha) classes. MODIS C6.1 active-fire detections provided qualitative corroboration of observed burn patterns. These findings demonstrate the utility of dNBR-based mapping for operational wildfire assessment and regional fire-impact reporting in Belize.

Back to poster list

P-96: Assessing Urban Thermal Environments through LCZ Mapping and Satellite-Derived LST Validation of Microclimate Models: A Case Study of Chiayi City

Presenting author: Liang-Yun KU

Authors: Liang-Yun KU, National Chung Hsing University; Jou-Man HUANG, National Chiayi University; Hui-Ping TSAI, National Chung Hsing University

Keywords: Local Climate Zone (LCZ); ENVI-met; Land Surface Temperature (LST); Urban Heat Island (UHI); Model Validation; Multi-source Remote Sensing

This study aims to assess the Urban Heat Island (UHI) effect in Chiayi City, a typical subtropical medium-sized city in Taiwan, which exhibits a UHI intensity of 4.1°C, a magnitude comparable to that of large metropolitan areas. To systematically evaluate the urban thermal environment, this research delineates the city’s Local Climate Zones (LCZs) and validates microclimate simulation outputs using multi-source satellite imagery. Utilizing Google Earth Engine and the open-source LCZ Generator, an LCZ map of Chiayi City at a 100-meter resolution was generated, achieving a high weighted classification accuracy of 0.94. The results indicate that the city primarily comprises built types such as LCZ 2, LCZ 4, LCZ 6, and LCZ 8, alongside natural classes including LCZ A, LCZ B, LCZ D, and LCZ G. Subsequently, the ENVI-met microclimate model was employed to simulate the surface temperatures of representative LCZ types. By integrating the LCZ classification with the microclimate simulation results, an urban climatic map was developed to characterize the city’s thermal patterns. To evaluate the reliability of the simulated surface temperatures, this study proposed a rigorous validation framework integrating Land Surface Temperature (LST) products derived from Landsat and MODIS. The validation workflow involved critical geospatial preprocessing: coordinate transformation was executed to accurately align the satellite data with the local spatial reference system of the microclimate model. Furthermore, spatial resampling techniques were applied to harmonize the varying spatial resolutions of Landsat and MODIS imagery with the ENVI-met simulation grid. Preliminary cross-validation revealed varying degrees of temperature discrepancies between the satellite-derived LSTs and simulated surface temperatures across different LCZ classes. Currently, iterative calibration and sensitivity analyses of material properties and boundary conditions are being conducted to optimize model performance and minimize simulation uncertainties. Ultimately, this comprehensive validation methodology aims to establish a robust, evidence-based foundation for accurate urban climate mapping and UHI assessment.

Back to poster list

P-98: Integrating MAE-SegFormer and stacking ensemble regression for precision tea tree physiological monitoring based on UAV multispectral imagery

Presenting author: Zhong-Han ZHUANG

Authors: Zhong-Han ZHUANG, National Chung Hsing Univ.; Hui Ping TSAI, National Chung Hsing Univ.

Keywords: Tea; Unmanned aerial vehicle; Multispectral; Ensemble learning; Deep learning

Tea is among the most ancient and globally consumed beverages. In Taiwan, the tea industry’s annual production value exceeds US$ 240 million, with 65% from Nantou County. The physiological status of tea trees is a primary determinant of both yield and quality. Specifically, leaf area index (LAI) and chlorophyll content (Chl) serve as essential biomarkers for characterizing canopy structure, photosynthetic efficiency. This study employed a five-band multispectral sensor mounted on an unmanned aerial vehicle (UAV) to conduct monitoring from January 2021 to July 2024. Data were collected across 20 sites in Nantou, spanning low, middle, and high altitudes and comprising both conventional and agroecological plantations. The research framework consists of two stages. The first stage focusing on enhanceing spectral feature extraction using a SegFormer encoder with Transformer-based Masked Autoencoder (MAE) self-supervised pre-training. By evaluating three decoders—FCN, DeepLab V3+, and SegFormer Decoder—the SegFormer-DeepLab V3+ configuration was optimized for its superior ability to fuse multi-scale features and reconstruct precise object boundaries. This model achieved an overall accuracy of 0.85, with a precision of 0.88 and a recall of 0.87 for the tea tree class. In the second stage, a stacking ensemble regression, integrating optimized Random Forest, XGBoost, Support Vector Machine, and Multi-Layer Perceptron as base learners, was implemented with an ElasticNet meta-model. The framework achieved robust prediction performance for LAI (RMSE: 1.11, MAPE: 24.70%) and high precision for Chl (RMSE: 5.33, MAPE: 5.32%). The correlation coefficients between predicted and observed values reached 0.76 and 0.75 for LAI and Chl, respectively. This research demonstrated that integrating advanced deep learning architectures with ensemble learning effectively captures the complex nonlinear relationships between multispectral features and physiological traits, providing a robust scientific foundation and demonstrates significant potential for the digital transformation and precision management of the tea industry.

Back to poster list

P-100: Evaluating Arctic Tundra Vegetation Properties Using in-situ Remote Sensing Observation in Svalbard

Presenting author: Hibiki NODA

Authors: Hibiki NODA, National Institute for Environmental Studies; Soshi OSAKI, Kitakyushu Museum of Natural History and Human History; Masaki UCHIDA, National Institute of Polar Research; Takayuki NAKATSUBO, Hiroshima University

Keywords: Arctic, tundra, terrestrial ecosystem, climate change, ground observation

Recent climate change impacts are particularly pronounced in high‑latitude regions. In Arctic tundra ecosystems, newly exposed land surfaces resulting from rapid glacier retreat and ongoing vegetation succession have been observed over the past few decades. These terrestrial ecosystem changes may significantly affect the Arctic carbon cycle. Therefore, it is necessary to monitor the changes in tundra vegetation to evaluate the impacts of climate change on the Arctic region. Satellite remote sensing provides a powerful tool for monitoring long‑term changes in vegetation distribution and structure. However, to translate remotely sensed optical data into ecosystem structure, it is necessary to clarify the relationships between vegetation indices and vegetation traits, e.g., vegetation coverage, biomass, and leaf area index, using in‑situ observed data. In the present study, we conducted in‑situ observations in the tundra around Longyearbyen (78°13′N, 15°38′E) and Ny‑Ålesund (78°55′N, 11°55′E), Svalbard, in July 2025. We selected 80 points in Longyearbyen and 44 points in Ny‑Ålesund, representing typical tundra landscapes with varying degrees of vegetation cover. Reflectance spectra at each point were measured using an MS‑720 hand‑held spectroradiometer (Eko Instruments Co., Ltd., Japan), and vegetation indices, such as NDVI (normalized difference vegetation index), EVI (enhanced vegetation index), and GRVI (green‑red vegetation index), were calculated. Vegetation coverage ratios within the MS‑720 ground footprint were derived from images taken with a digital camera equipped with a fisheye lens. Plants and mosses within the footprint were harvested and the leaf area of vascular plants and vegetation biomass were also measured. In this presentation, we will show the relationships between vegetation indices and vegetation biomass, coverage, and leaf area index, along with analyses of satellite‑data derived from these relationships.

Back to poster list

P-102: Leveraging SWOT Observations for Monitoring Inland Water Dynamics in Taiwan

Presenting author: Kuo-Hsin TSENG

Authors: Kuo-Hsin TSENG, National Central University; Chia-Hao WANG, National Central University

Keywords: Water Level, SAR Altimetry, Water Resources, SWOT, PIXC

The Surface Water and Ocean Topography (SWOT) mission, launched in late 2022, delivers unprecedented spatial resolution and vertical precision for characterizing subtle water-surface gradients. Its wide-swath interferometric altimetry enables systematic and repeat monitoring of ungauged rivers and small inland water bodies. In this study, we exploit two Level-2 high-rate products, Lake Single Pass Observations (L2_HR_LakeSP_Obs) and Pixel Cloud (L2_HR_PIXC), to retrieve water surface elevations for 14 reservoirs and 14 ponds across Taiwan, and validate them against in situ gauge records. Preliminary analyses demonstrate that both absolute water levels and their temporal variability are effectively captured. For reservoirs, the standard deviation (STD) decreases from 1.02 m to 0.31 m and the RMSE from 1.14 m to 0.53 m after applying LakeSP and PIXC refinements. For the 14 Taoyuan ponds, the STD is reduced from 0.39 m to 0.16 m and the RMSE from 0.41 m to 0.16 m. These results confirm that SWOT provides a robust, independent perspective on hydrological dynamics, supporting improved water resource planning and operational management.

Back to poster list

P-104: Discriminating Snowfall Effects in Post-Earthquake Damage Detection Using Polarimetric Coherence: A Case Study of the 2025 Aomori Prefecture Eastern Offshore Earthquake

Presenting author: Fumitaka OGUSHI

Authors: Fumitaka OGUSHI, Terra Phase Inc.; Masashi MATSUOKA, Institute of Science Tokyo; Makoto YAMAGUCHI, General Insurance Rating Organization of Japan; Kazuma NISHIMURA, General Insurance Rating Organization of Japan

Keywords: Cross-POL; Coherence; Snowfall; Damage Detection; Earthquake; Sentinel-1

Snowfall can significantly affect synthetic aperture radar (SAR)-based earthquake damage detection when coherence is used as an indicator of structural changes. In built-up areas, snow accumulation alters surface scattering characteristics and may induce decorrelation unrelated to structural damage. In particular, damage assessment based solely on VV-polarized coherence can lead to increased false detections in the presence of snowfall. This study investigates snowfall-induced false detections in post-earthquake damage assessment and demonstrates a polarimetric coherence-based discrimination approach. The 2025 Aomori Prefecture Eastern Offshore Earthquake is analyzed as a case study using Sentinel-1 SAR data. Coherence maps derived from VV and VH polarizations were examined over built-up districts affected by both the earthquake and subsequent snowfall. The results indicate that snowfall causes significant decorrelation in VV coherence, leading to overestimation of damaged areas. In contrast, VH coherence exhibits smaller decorrelation under snowfall conditions. This differential polarimetric response enables the discrimination of areas where coherence loss is primarily caused by snow cover rather than earthquake-induced structural damage. Analyzing both VV and VH coherence, snowfall-induced false damage detections can be effectively discriminated. The proposed approach improves the robustness of SAR-based rapid damage assessment in snowy urban environments and highlights the importance of incorporating cross-polarimetric information when seasonal surface changes are present.

Back to poster list

P-106: Toward Reliable Bridge Infrastructure Monitoring: Demonstrating the Effectiveness of Complex-Valued SAR Simulation

Presenting author: Fumitaka OGUSHI

Authors: Fumitaka OGUSHI, Terra Phase Inc.; Takashi MIYAMOTO, Institute of Science Tokyo

Keywords: SAR; Simulation; Digital-Twin; Bridge Monitoring; Displacement

Synthetic Aperture Radar (SAR) simulation has been widely used to interpret complex backscattering mechanisms; however, many existing approaches primarily emphasize amplitude reproduction, while comprehensive modeling of both amplitude and phase remains limited. This restriction reduces their effectiveness in operational interferometric SAR (InSAR)-based infrastructure monitoring, where deformation assessment relies on the complex-valued nature of SAR signals. Bridge infrastructures present a particular challenge because the strong contrast between water surfaces and artificial structures produces highly complex SAR geometry. Specular reflections from water bodies, double-bounce interactions between bridge components and the underlying surface, and multipath effects generate intricate scattering behavior and phase patterns. These phenomena make deformation estimation difficult, especially under large structural displacements where phase unwrapping errors can severely degrade estimation accuracy. In this study, we develop a complex-valued SAR simulation framework capable of reproducing both amplitude and phase components of the scattered field under realistic acquisition conditions. A three-dimensional environment is constructed by integrating a digital elevation model with a 3D bridge structure model whose deformation is reproduced through structural simulation. Sensor parameters, such as wavelength, flight direction, incidence angle, and slant-range distance, are explicitly incorporated. By modeling single- and multiple-bounce scattering mechanisms, complex signals are generated to analyze amplitude variations and interferometric phase signatures. Through comparison with observed Sentinel-1 and ALOS-2 SAR data, we conduct numerical evaluation and visual inspection of amplitude and phase behavior. The results demonstrate the effectiveness of complex-valued SAR simulation for reliable bridge infrastructure monitoring, providing a practical link between structural modeling and SAR-based deformation analysis within a digital twin–oriented framework.

Back to poster list

P-108: Retrieval-Optimized Matched Filtering of Sentinel-2 SWIR Observations for Single-Acquisition Methane Leak Detection and Quantification in Biogas Facilities

Presenting author: Harika NULU

Authors: Harika NULU, Origin Tech Ltd

Keywords: Sentinel-2, ERA5, Background Estimation, Matched Filter, Monte Carlo Simulation

Methane (CH₄) emissions originating from biogas facilities constitute a considerable yet often underestimated source of greenhouse gases. This research introduces a singular methane leak detection framework utilizing shortwave infrared (SWIR) observations from Sentinel-2, aimed at reliable plume detection and emission measurement over active biogas plants. The approach incorporates a retrieval-optimized matched filtering (MF), where methane absorption cross-sections obtained from the HITRAN database are blended with Sentinel-2 spectral response functions to develop a precise target spectrum for chosen SWIR bands. Atmospheric parameters specific to the scene, such as surface pressure, temperature, dew point temperature, wind data, and boundary layer height, are obtained from ERA5 reanalysis data to enhance the retrieval of methane column enhancements. After calculating the MF, methane enhancements (in parts per billion by volume) are derived across the entire scene. Background characterization is accomplished using kernel-based SWIR detrending and robust covariance estimation for whitening the MF. Plume detection is then carried out to distinguish methane emissions from biogas infrastructure. Due to the common occurrence of industrial buildings and varying vegetation textures that may cause spectral interference and noise, quality controls are put in place. These include analyzing plume coherence, confirming transport consistency with wind patterns, and assessing background stability with a noise-resistant ring filter. This comprehensive screening greatly enhances the reliability of plume identification and the stability of extraction. The calculations for Integrated Methane Enhancement (IME) and emission flux are further improved through Monte Carlo simulations, which adjust atmospheric conditions and measurement noise to establish 95% confidence intervals. This framework effectively identifies methane plumes originating from active biogas facilities with emission rates ranging from 1 to 3 t h⁻¹, while decreasing scene noise by roughly 30 to 50%. The approach adapts to multiple time periods and seasons, enabling detailed analysis of methane leakage patterns specific to each location.

Back to poster list

P-110: Integrating SAR and Optical Satellite Imagery for Monitoring Land Reclamation and Surface Deformation at a Commercial Port

Presenting author: Chia-Che CHANG

Authors: Chia-Che CHANG, CECI Engineering Consultants, Inc., Taiwan; Tzu-Min HUANG, CECI Engineering Consultants, Inc., Taiwan; Hsi-Hsien WU, CECI Engineering Consultants, Inc., Taiwan; Kuo-Hsin TSENG, Center for Space and Remote Sensing Research, National Central University, Taiwan; Fu-An TSAI, Center for Space and Remote Sensing Research, National Central University, Taiwan

Keywords: Land Reclamation; InSAR; Sentinel-1; Sentinel-2; Surface Deformation

This study explores the integration of synthetic aperture radar (SAR) and optical satellite imagery to monitor land reclamation and surface deformation at a commercial port in Taiwan. The port serves as an important international freight hub, and large areas were developed through land reclamation on soft sedimentary deposits, making the newly reclaimed ground susceptible to subsidence. Because traditional ground-based monitoring methods require substantial installation and maintenance efforts in harsh coastal environments, satellite remote sensing provides an efficient approach for wide-area observation. To assess long-term ground stability, multi-temporal Sentinel-1 radar imagery is analyzed using the Persistent Scatterer Interferometric Synthetic Aperture Radar (PSInSAR) technique. The results indicate that most completed reclamation areas remain generally stable, while localized deformation zones are identified in limited locations. In addition, multispectral Sentinel-2 optical imagery is used to detect land surface changes and to quantify reclamation development in areas that remain under construction. The optical observations show a continuous expansion of reclaimed land, suggesting that the reclamation is approaching completion. By integrating radar and optical satellite observations, both ground conditions and reclamation progress can be examined over large areas and extended periods, providing useful information for coastal engineering development and infrastructure management.

Back to poster list

P-112: Automatic flood detection using ALOS-2 and ALOS-4: algorithm updates and operation results

Presenting author: Masato OHKI

Authors: Masato OHKI, Japan Aerospace Exploration Agency; Tamotsu MOTOKI, Japan Aerospace Exploration Agency; Shiro KAWAKITA, Japan Aerospace Exploration Agency; Takeo TADONO, Japan Aerospace Exploration Agency

Keywords: disaster monitoring; inundation; flood mapping; ALOS-2; PALSAR-2; ALOS-4; PALSAR-3

We have developed an automated flood detection algorithm using the Advanced Land Observing Satellite-2 (ALOS 2). This algorithm has been operational since 2022 and provides flood extent polygons to end-users within approximately one hour after the data input. For robust and rapid computation, our algorithm uses high resolution SAR data from ALOS 2 together with quasi-real-time flood simulation data from Today’s Earth (TE). In addition, the algorithm can experimentally estimate floodwater depth using a high resolution digital elevation model. We validated the accuracy of our algorithm by comparing its outputs with reference flood maps based on arial photographs and social media information. The results indicate that our algorithm can detect flood areas with acceptable accuracy across incidence angles ranging from 15 to 59 degrees. This broad incidence angle range enables high frequency monitoring (more than once per day) even with a single satellite. In 2024, the successor satellite, ALOS 4 was successfully launched. ALOS 4 is currently operated primarily in the fixed PRF mode to improve image quality, resulting in the presence of blind (invisible) regions within the imaging swath. We modified our algorithm to accommodate ALOS-4 data and automatically mask the blind areas. In this presentation, we also report recent flood detection results using both ALOS 2 and ALOS-4 data for the latest flood events.

Back to poster list

P-114: Landslide detection using Sentinel-1 and ALOS-2, a case study of Noto Earthquake, Japan

Presenting author: Shouhao CHIANG

Authors: Shou Hao CHIANG; Center for Space and Remote Sensing Research, National Central University

Keywords: ALOS-2; Sentinel-1; SAR2OPT; Noto Earthquake

This research presents a framework for rapid landslide detection following the 2024 Noto Earthquake in Japan. Because cloudy and rainy weather often limits the utility of traditional optical sensors for emergency monitoring, this study evaluates the application of SAR-to-Optical Image Translation (SAR2OPT). The methodology leverages Synthetic Aperture Radar (SAR) data from Sentinel-1 (C-band) and ALOS-2/PALSAR-2 (L-band) to simulate optical-like imagery using deep learning techniques. Models such as Latent Diffusion Models (LDM) and Generative Adversarial Networks (GANs) were trained on paired radar and optical data to reconstruct RGB true-color images and NDVI proxies from SAR input. The study compares various analytical methods, including backscattering intensity, polarimetric correlation, and decompositions like Pauli and Freeman-Durden. Key findings indicate that the SAR2OPT framework successfully simulated post-event NDVI, showing strong potential as a robust tool for emergency landslide interpretation in all-weather conditions. Overall, the integration of multi-source SAR products into the SAR2OPT model offers a promising approach to improve the performance and reliability of disaster monitoring systems.

Back to poster list

P-116: Establishment of an Intelligent Data Platform for Detecting Nuclear Activities: A Multilingual Ontology and OSINT-GEOINT Fusion Framework

Presenting author: Yonhong JEONG

Authors: Yonhong JEONG, KINAC; Gayeon HA, KINAC; Sueyeon LEE, KINAC

Keywords: North Korea, Remote Monitoring, Data Platform, OSINT, GEOINT, Ontology database

The surveillance and early detection of covert nuclear activities, particularly those of North Korea, represent a critical challenge extending beyond Northeast Asian security to international security. Conventionally, such surveillance has relied on highly classified intelligence collection methods, including HUMINT (Human Intelligence) and SIGINT (Signals Intelligence). Nonetheless, the proliferation of commercial satellite technology and digital information has established open-source intelligence (OSINT) and geospatial intelligence (GEOINT) as essential complementary tools. The assessment of North Korea’s nuclear activities has been historically undertaken by specialist outlets such as 38 North and Beyond Parallel, which have utilized commercial imagery in their analysis. However, this process continues to be constrained by an excessive reliance on limited human expertise. Recent technological advances in commercial satellites have led to substantial improvements in the spatial and temporal resolution of imagery, encompassing regions of interest. This trend has established the foundation for expanding the analysis of nuclear activity indicators—previously dependent exclusively on the interpretation of specialists—into the domain of data and systems. Accordingly, the Korea Institute of Nuclear Safety (KINAC) has developed a data platform that automatically collects geospatial intelligence (GEOINT) and open-source intelligence (OSINT), fusing them within an ontology-based knowledge framework to enhance the effectiveness and efficiency of nuclear activity interpretation tasks.

Back to poster list

P-118: Spatiotemporal Analysis of Forest Cover Dynamics in Darkhan-Uul Province, Mongolia (2018–2021)

Presenting author: Ochirkhuyag LKHAMJAV, National Central University

Authors: Battuya SANJAAKHAND, Mongolian University of Life Science; Bayanmunkh NOROVSUREN, Centre for Policy Research and Analysis of Ulaanbaatar Municipality; Ulziisaikhan Ganbold, School of Geology and Mining Engineering, Mongolian University of Science and Technology; Ochirkhuyag LKHAMJAV, National Central University; Fuan Tsai, National Central University; Tsolmon Altanchimeg, Institute of Geography and Geoecology, Mongolian Academy of Sciences

Keywords: Forest Monitoring, Random Forest, SPOT, Mongolia, Darkhan-Uul, Land Cover Change

The forest-steppe ecotone of northern Mongolia constitutes a critical transition zone increasingly vulnerable to synergistic pressures from climate variability and anthropogenic disturbance. This study quantifies forest cover change in Darkhan-Uul Province over a three-year period (2018–2021) employing high-resolution SPOT 6/7 satellite imagery analyzed through Random Forest classification algorithms. The methodological framework incorporates rigorous radiometric normalization protocols and temporally consistent training strategies to minimize classification artifacts and ensure inter-epoch comparability. The Random Forest classifier demonstrated robust performance across all temporal epochs, with overall accuracies ranging from 92.8% to 94.5% and Kappa coefficients between 0.89 and 0.92, thereby validating the reliability of the classification framework for regional-scale monitoring applications. The analysis reveals a statistically significant declining trajectory in forest resources. Total forest cover declined from 125,450 hectares (28.5% of provincial area) in 2018 to 120,890 hectares (27.5%) in 2021, representing a net loss of 4,560 hectares and an average annual deforestation rate of 1.2%. Spatial pattern analysis indicates heterogeneous forest loss distribution, with the most pronounced degradation concentrated along forest patch edges and in zones proximate to transportation corridors and the urban center of Darkhan city. These spatial patterns provide strong empirical evidence implicating anthropogenic drivers—including fuelwood extraction, urban encroachment, and pastoral pressure—as principal deforestation agents. While the binary classification schema inherently limits the detection of subtle degradation gradients, this research establishes a high-resolution baseline dataset essential for evidence-based forest management in Mongolia. The study demonstrates a replicable methodological framework for continuous monitoring of vulnerable boreal-steppe ecotones under environmental change.

Back to poster list

P-120: Accuracy Validation of ALOS-4 InSAR for Landslide Deformation Monitoring: A Case Study of the Koshio Landslide, Japan

Presenting author: Seiya USAMI

Authors: Seiya USAMI, Hokkaido Research Oeganization; Takeo TADONO, JAXA EORC

Keywords: Advanced Land Observing Satellite-4; PALSAR-3; Synthetic Aperture Radar; Interferometry

This study validates the displacement measurement accuracy of ALOS-4/PALSAR-3 InSAR by comparison with GNSS observations at the Koshio landslide, Japan. Seven GNSS stations installed within the landslide area were used for validation. The root-mean-square error (RMSE) between GNSS- and InSAR-derived line-of-sight (LoS) displacements was 16.4 mm, which is comparable to that reported for ALOS-2 in previous studies. The results demonstrate that ALOS-4 InSAR achieves a displacement accuracy suitable for landslide deformation monitoring.

Back to poster list

P-122: Modeling Land Use and Land Cover Change Dynamics and Prediction Using Machine Learning and QGIS-MOLUSCE

Presenting author: Enkhzaya ENKHTAIVAN

Authors: Oyunbileg ERDENEBADRAKH, Centre for Policy Research and Analysis; Gandoljin NERGUI, Centre for Policy Research and Analysis; Enkhjargal NATSAGDORJ, Centre for Policy Research and Analysis; Enkhzaya ENKHTAIVAN, Centre for Policy Research and Analysis, Batbileg BAYARAA, Mongolian University of Life Sciences, Bolormaa BATSUURI, National University of Mongolia

Keywords: LULC, Machine Learning, QGIS-MOLUSCE, Support Vector Machine, Classification and Regression Trees

Globally, agricultural land covers approximately 5 billion hectares, accounting for nearly 38% of the Earth’s total terrestrial surface. About one-third of this area is utilized for cropland, while the remaining two-thirds consist of pasture and rangelands. Global warming results have intensified challenges to agricultural production by facilitating the spread of new pests and diseases, shortening growing seasons, disrupting flowering periods, and degrading soil fertility. These impacts have heightened concerns regarding food security and the sustainability of agricultural systems. In this study, LULC changes were investigated for the Erdeneburen agricultural area in Khovd Province and the Khuren Tal area in Zavkhan Province, Mongolia, using Landsat satellite imagery and ML algorithms. High-resolution LULC maps for the period 1995–2025 were produced using Random Forest (RF), Support Vector Machine (SVM), k-Nearest Neighbours (k-NN), and Classification and Regression Trees (CART) algorithms. Among these methods, the Random Forest algorithm demonstrated strong capability in distinguishing complex land cover classes and was particularly effective for detailed classification. Future LULC scenarios for the years 2030 and 2050 were simulated using a Cellular Automata–Artificial Neural Network (CA–ANN) model implemented through the QGIS MOLUSCE plugin. The results indicate that cropland areas have expanded by approximately 3%–10% over the study period. However, if soil conservation measures are not effectively implemented, average forage yields are projected to decline by 20%–30%, while potato yields may decrease by 30%–40% across the study regions. These findings highlight the growing necessity for long-term, sustainable assessment of agricultural land quality and condition through remote sensing–based monitoring. Furthermore, there is an increasing demand to identify future development trends in agricultural land use using quantitative analysis and spatial modeling approaches to support evidence-based land management and policy decision making.

Back to poster list

P-124: Comparison of multiple strategies for detecting landslide affected areas with optical remote sensing and object-based image analysis

Presenting author: Jhe-Syuan LAI

Authors: Jhe-Syuan LAI, Feng Chia Univ.; Sheng-Wei WU, Feng Chia Univ.; Yun-Chen TU, Feng Chia Univ.

Keywords: Landslide; Object-based Image Analysis; Optical Remote Sensing; Support Vector Machine; Typhoon Morakot

Pixel-based image analysis is a critical approach to extract landslide features and affected areas. However, these achievements might be discontinuously detected because of the salt and pepper effect. To address this issue, this study applied the object-based image analysis and the support vector machine (SVM) algorithm with optical remote sensing images to extract the relatively complete landslide affected regions. Landslide inventory can be further produced by the remotely sensed landslide detections after separating landslide source and runout areas. Multiple strategies of detecting landslide affected areas were further proposed and compared in this study; for example, binary (i.e., landslide and non-landslide classes for the post-event image) and multi-label (i.e., land-cover classes for the post-event image and to merge as landslide and non-landslide classes) classification, and change detection (i.e., compared labels from pre- and post-event images to identify landslide and non-landslide classes) based schemes were used to detect landslide affected areas individually. Typhoon Morakot was selected as the target case. Reference data was produced manually based on the stereoscopic aerial photographs and other auxiliary data. The detection results were then verified against the reference data to evaluate the effectiveness of the constructed SVM models. The best solution derived from all proposed strategies reached 91 % of overall accuracy. This result indicated very high consistency between landslide detections and the reference data, reducing the cost of production. The strategy of land-cover classification with the post-event image and to merge the multi-labels is suggested for detection of landslide affected areas in Typhoon Morakot event.

Back to poster list