SAR coherence-guided machine learning framework for flood-induced damage mapping using a single fully polarimetric SAR scene
- 1Department of Civil Engineering, Indian Institute of Technology Roorkee, India
- 2Corresponding author.
Abstract
Flood-induced damage assessment is challenging due to lack of standardized damage labels for flood events. Flood inundation maps are often used as proxy for damage, although inundation does not necessarily represent actual structural or surface damage. In this study, a novel framework is proposed that utilizes Synthetic Aperture Radar (SAR) coherence-based labels and a single fully polarimetric ALOS-2 scene for flood induced damage assessment. The proposed framework is used to detect damage induced by the 2019 Typhoon Hagibis flood event in Japan. Multiple polarimetric descriptors are selected using a combined Jeffries-Matusita distance and Total Decrease-based importance to train machine learning model of Random Forest to identify flood induced damage pixels. The damage maps from the trained machine learning model are validated using flood water depth estimated from the Floodwater Depth Estimation Tool (FwDET) with the 5 m GSI DEM5A dataset. Results show most of the damage pixels are in agricultural lands, low-lying vegetation, and river floodplains. Fewer damage pixels are found in dense vegetation, compact urban areas, and river boundaries. The framework is limited by reduced coherence sensitivity in dense vegetation and complex scattering in urban and water-land boundary areas, as well as the limited availability of suitable multi-sensor observations. Further, cross-event and cross-region validation is required to assess its robustness and transferability in future work. Overall, the proposed framework generates spatially consistent damage patterns with fewer false detections over highland vegetated regions, demonstrating the effectiveness of integrating SAR coherence, polarimetric scattering information, and flood depth analysis for large-scale flood damage assessment.
Related
- Hybrid machine learning framework for drought forecasting: integration of LSBoost, RQNN, CNN-RNN, CNN-LSTM and advanced data decomposition
- Integrated machine learning and AHP-based framework for spatial delineation of rainwater harvesting potential
- SeisBench DAS: A machine learning framework for Distributed Acoustic Sensing
- First Demonstration of an Onboard Polarimetric Calibration and Ice-Cloud-Based Validation Framework for the Particulate Observing Scanning Polarimeter (POSP)
- A geospatial machine-learning framework for regional electrification estimation mapping and trend analysis using demographic and infrastructure features
BibTeX
@article{adhikari2026coherenceguided,
title = {{SAR coherence-guided machine learning framework for flood-induced damage mapping using a single fully polarimetric SAR scene}},
author = {Ruma Adhikari and Alok Bhardwaj},
journal = {Science of Remote Sensing},
year = {2026},
month = oct,
doi = {10.1016/j.srs.2026.100527},
url = {https://doi.org/10.1016/j.srs.2026.100527},
note = {CC BY},
}
RIS
TY - JOUR
TI - SAR coherence-guided machine learning framework for flood-induced damage mapping using a single fully polarimetric SAR scene
AU - Ruma Adhikari
AU - Alok Bhardwaj
PY - 2026
DA - 2026/10/30
JO - Science of Remote Sensing
DO - 10.1016/j.srs.2026.100527
UR - https://doi.org/10.1016/j.srs.2026.100527
AB - Flood-induced damage assessment is challenging due to lack of standardized damage labels for flood events. Flood inundation maps are often used as proxy for damage, although inundation does not necessarily represent actual structural or surface damage. In this study, a novel framework is proposed that utilizes Synthetic Aperture Radar (SAR) coherence-based labels and a single fully polarimetric ALOS-2 scene for flood induced damage assessment. The proposed framework is used to detect damage induced by the 2019 Typhoon Hagibis flood event in Japan. Multiple polarimetric descriptors are selected using a combined Jeffries-Matusita distance and Total Decrease-based importance to train machine learning model of Random Forest to identify flood induced damage pixels. The damage maps from the trained machine learning model are validated using flood water depth estimated from the Floodwater Depth Estimation Tool (FwDET) with the 5 m GSI DEM5A dataset. Results show most of the damage pixels are in agricultural lands, low-lying vegetation, and river floodplains. Fewer damage pixels are found in dense vegetation, compact urban areas, and river boundaries. The framework is limited by reduced coherence sensitivity in dense vegetation and complex scattering in urban and water-land boundary areas, as well as the limited availability of suitable multi-sensor observations. Further, cross-event and cross-region validation is required to assess its robustness and transferability in future work. Overall, the proposed framework generates spatially consistent damage patterns with fewer false detections over highland vegetated regions, demonstrating the effectiveness of integrating SAR coherence, polarimetric scattering information, and flood depth analysis for large-scale flood damage assessment.
SN - 2666-0172
ER -