Answers are grounded in this paper and cite their sources.
Flood mapping in urban and vegetated environments remains challenging when it relies solely on polarization-based synthetic aperture radar (SAR) signals, as different surface types may exhibit similar backscatter intensities. This study introduces a polarimetric SAR (PolSAR) decomposition-based deep learning framework that exploits scattering information from C-band dual-polarization Sentinel-1 and quad-polarization RADARSAT-2 data. Decomposition features derived from multiple decomposition techniques were refined using variance inflation factor (VIF)–based multicollinearity screening and gravitational search algorithm (GSA) optimization to reduce redundancy and retain informative predictors. Flood and non-flood samples were delineated using normalized difference water index differences and false-color composites, followed by a 70:30 division into training and testing datasets. Using the selected decomposition features, the disjunctive normal form network (DNF-Net) and tabular network (TabNet) achieved average accuracies of 0.9689 for Sentinel-1 and 0.9680 for RADARSAT-2, generally outperforming polarization-only inputs across the evaluation metrics. Additional validation using Sentinel-1 imagery from an adjacent Red River region and Kauai, Hawaii, confirmed the practical transferability of the framework, while also indicating sensitivity to regional land-cover variability and acquisition conditions. These results demonstrate that PolSAR decomposition features provide physically meaningful and effective predictors for SAR-based flood mapping in complex flood-prone landscapes.
No takes yet. Share an insight, caveat, or question.
Kim et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: