Flood Index-Enhanced deep learning model for coastal inundation mapping in SAR imagery

• Coastal inundation mapping with dual-polarized and bitemporal Sentinel-1 images. • Novel CNN integrated SAR-derived flood indices to enhance performance. • Four specialized modules facilitate the feature extraction and interaction. • Successful applications highlight the method’s potential for disaster response. This study introduces a flood index-enhanced deep learning model for coastal inundation mapping leveraging dual-polarization and bitemporal Sentinel-1 synthetic aperture radar (SAR) imagery. The proposed model, FIE-Net, introduces two key innovations to improve flood mapping accuracy. First, it integrates the Ratio Image (RI), a primary flood index derived from bitemporal SAR data, which provides a reliable reference for precise inundation area delineation. The index offers a clear, flood-specific signal that enhances segmentation precision. Second, the model employs a dual-branch U-Net framework, augmented with specialized modules to enhance feature extraction and better integrate diverse data sources. This combination enables the model to handle complex flood scenarios more effectively, thereby boosting overall performance. Using data from the Copernicus Emergency Management Service, 16 pairs of representative SAR images from Madagascar’s Tropical Cyclones 2018 AVA and 2023 Cheneso were collected under various terrain conditions. After semi-automatic labeling and cropping, 4350 sample pairs were processed, with 2784/696/870 used for model training/validation/testing. The proposed model achieved the highest Intersection over Union (IOU) of 79.44%, outperforming the state-of-the-art models across all evaluation metrics. The experiments also demonstrate that each introduced innovation contributes to improved accuracy, with the flood index making the most significant impact. Furthermore, the model’s applicability was further confirmed with three independent flooded areas (689 samples, not included in training) from the event Cheneso. IOUs in all three scenes consistently exceeded 75%, underscoring the model’s reliability and robustness in real-world scenarios.