Multisource SAR-Based Rural Flood and Partially Submerged Vegetation Mapping Using Fuzzy Logic and Machine Learning

Abstract

Flood mapping is critical for disaster response, yet hindered by cloud obstruction, vegetation interference, and limitations of single-source synthetic aperture radar (SAR) data. To address these challenges, this study proposes an innovative framework integrating fuzzy logic and machine learning for rural flood and partially submerged vegetation mapping. This study fuses multisource data, including pre/postflood dual-polarization SAR (VV/VH), optical-derived land-cover classifications, and digital elevation model-based hydrological features (height above nearest drainage, slope, and depressions). Then, five probabilistic maps are generated, including backscatter coefficient (FM1), hydraulic (FM2), reflection characteristics (FM3), water bodies, and partially submerged vegetation maps. A labeled dataset (500 million 10-m-resolution pixels) was constructed to train machine-learning models for defuzzification. Validated across five Asian river basins, the method achieved an overall accuracy of >0.95 and a Kappa coefficient of >0.84, with XGBoost achieving peak validation accuracy and LightGBM maintaining comparable accuracy with higher efficiency. Machine-learning-based defuzzification enables more effective capture of multiscale information, significantly outperforming fixed-threshold segmentation at the single-pixel level. This study, based on multitemporal dual-polarization SAR, integrates land-cover type and hydrological principles information, demonstrating the application potential for larger scale flood probability mapping and inundation extent mapping in rural areas.