Flood Mapping employing U-Net: A Case Study of Teesta Dam

Abstract

River Teesta is prone to weather fluctuations that leads to cloud burst or heavy rainfall resulting in overflowing of Teesta dam (27.0018057°N, 88.4404352°E) situated across the Teesta Basin region. Dam overflow is one of the major causes of sudden and uncontrolled release of water, which often leads to flooding in nearby downstream areas. Accurate flood mapping is important to reduce the damage caused by floods in low-lying areas. This study uses a U-Net deep learning model to identify flooded regions near Kalimpong in the Teesta River area using satellite images. The proposed method identifies flood-affected areas accurately and provides useful support for early warning and water management systems. The model was compared with existing methods such as Otsu thresholding, Random Forest, SegNet, and PSPNet using IoU, F1-score, precision, and recall values. The proposed U-Net model achieved 98% IoU, 98% F1-score, 99% precision, and 97% recall, showing better results than the other methods. The IoU value was 26% higher than Otsu thresholding, 18% higher than Random Forest, 10% higher than SegNet, and 8% higher than PSPNet. The training and validation graphs show stable learning with very little overfitting and steady performance during all 100 epochs. In addition, NDVI time-series analysis for Lachung city and the Lachung–Teesta meeting point showed important changes in vegetation that are useful for water resource management. These results show that the proposed U-Net model can identify flood areas effectively and can help improve early warning systems and safe dam management. However, the study is limited to Sentinel-2 image-based flood segmentation for the Teesta basin and does not include real-time hydrological inputs.