SAR Remote Sensing Flood Mapping Using A Multi-Feature Optimized Random Forest

Abstract. Flood disasters, with high suddenness, wide impact and severe consequences, severely threaten ecological security, human lives, property and socioeconomic development. Optical remote sensing, though advantageous in spatial coverage and revisit frequency for flood monitoring, is heavily constrained by rainy and harsh weather accompanying floods, while Synthetic Aperture Radar (SAR) enables all-weather earth observation and thus becomes a superior alternative. Conventional SAR-based flood extraction methods such as, thresholding, object-based and standard random forest models, however, face critical limitations of high false positive rates, inaccurate land cover discrimination and poor generalization ability. To address these issues, this study proposed a robust flood mapping approach based on Sentinel-1 SAR data, taking China’s Dongting Lake basin as the study area. First, Sentinel-1 data were preprocessed to extract polarization features optimal for water body identification and mapping precision. A Multiple Feature-Optimized Random Forest (MFORF) algorithm with a multi-level feature extraction framework was then developed to enhance the accuracy and reliability of flood-prone area delineation. Additionally, SAR-derived flood extents were fused with Sentinel-2-based land cover classification maps to accurately detect inundation dynamics. Quantitative and visual validations confirm that the MFORF method improves the Kappa coefficient of water extraction accuracy by an average of 3% compared with traditional SAR flood mapping techniques. This approach establishes a robust and efficient framework for rapid and accurate flood monitoring, and provides critical technical support for flood disaster response and mitigation practices.