Deep Learning Models for Flood Detection in Nepal: Challenges and Insights

Abstract

Climate-driven increases in flood frequency and intensity have made rapid and reliable flood monitoring essential for effective emergency response. Synthetic Aperture Radar (SAR) imagery from Sentinel-1 is particularly valuable for this purpose because it operates independently of cloud cover and illumination conditions, enabling flood observation during extreme weather events. In this study, a 5-channel U-Net was developed for flood mapping in Nepal, using multi-temporal Sentinel-1 SAR data and hydrologically relevant topographic information. The input configuration combined pre-flood VV and VH backscatter, post-flood VV and VH backscatter, and Height Above the Nearest Drainage (HAND) data to improve the discrimination of flooded and non-flooded areas. The model demonstrated stable convergence during training, with continuously decreasing loss and improving Dice and IoU scores on both training and validation datasets. Quantitative evaluation showed that the proposed model achieved an Intersection over Union (IoU) of approximately 0.51, an F1-score of 0.67, precision of 0.68, recall of 0.66, near-perfect specificity, and an AUC of 0.996, indicating excellent class separability and strong suppression of false positives. Qualitative assessment further confirmed that the predicted inundation patterns were spatially coherent and closely matched the reference flood masks, although some local boundary overestimation and omission of small fragmented flood patches remained. These results demonstrate that integrating multi-temporal SAR backscatter with HAND in a U-Net framework provides a robust and operationally relevant approach for flood extent mapping in cloud-prone regions.