Deep learning-based flood mapping in monsoon climates using sentinel-1 SAR

Abstract

Monsoon floods pose significant disaster management challenges due to persistent cloud cover, which renders optical remote sensing ineffective for rapid response. To overcome this limitation, this study proposes a computationally efficient, all-weather flood mapping framework that integrates Sentinel-1 Synthetic Aperture Radar (SAR) data with Height Above Nearest Drainage (HAND) topographic constraints within a deep learning pipeline. Three model architectures were compared systematically: a baseline U-Net, a U-Net with spatial augmentation (UNet_Aug), and a U-Net++ with Focal-Tversky loss (UNetPP_TMA). Bi-temporal Sentinel-1 scenes (VV and VH polarisations) from the 2024 Nawalpur flood event in Nepal were combined with a hydrologically corrected 30 m SRTM DEM, utilizing Sentinel-2 imagery and field data as high-fidelity ground truth. Models were trained using patch-based sampling to ensure computational efficiency. Quantitative results demonstrated that while UNet_Aug achieved the highest overall balance of metrics (IoU of 0.6101, F1-Score of 0.7578, Precision of 0.7496, and AUC-ROC of 0.9901), UNetPP_TMA achieved the highest recall of 0.7704. Overall, UNet_Aug is recommended as the most suitable operational model because it provided the best balance across IoU, F1-score, Kappa, accuracy, and AUC, while UNetPP_TMA remains useful as a high-recall option for safety-critical screening. This study contributes a reliable flood delineation approach for cloud-obscured monsoon conditions and provides an open-source pipeline and dataset that can serve as a reproducible benchmark for future disaster management in data-scarce regions.