SAR-FM: SAR-based flood rapid mapping using multi-modal data fusion

Adverse climatic changes have increased the frequency of natural hazards, emphasizing the need for timely detection to support effective disaster management and mitigation efforts. Floods are among the most common weather-related natural disasters, causing significant damage to human settlements and the environment. While both Synthetic Aperture Radar (SAR) and Electro-Optical (EO) remote sensing data have been used to detect flooded areas, SAR is preferred due to its independence from weather and illumination conditions. However, detecting flood-inundated areas with SAR is challenging due to volume and double-bounce scattering, unlike permanent water bodies that are easily detected. To address this issue, we fuse SAR imagery with other multimodal data and introduce three novel derived features: Water-Proximity Elevation Map (PEM), Water-Terrain Binary Map (WTM), and Dual-Pol SAR Map (DPM). We also present three deep learning architectures for flood mapping that effectively fuse multi-modal data. Extensive experiments and comparisons with state-of-the-art methods demonstrate that our proposed approach achieves improved performance with significant reduction in model complexity.