Flood detection for satellite and UAV remote sensing based on fine-tuned SAM

This study presents a novel flood detection method using the fine - tuned Segment Anything Model (SAM) for satellite and UAV remote sensing data. Floods, as highly destructive natural disasters, demand accurate and timely detection. Traditional methods have limitations, while satellite and UAV remote sensing offer advantages in different aspects. The proposed method aims to address the challenges of existing deep - learning - based approaches. The dataset consists of Gaofen - 2 satellite data and UAV data from flood - affected areas in Anhui Province during a heavy rainstorm in 2024. After comprehensive pre - processing, the data was used to fine - tune SAM. The fine - tuning process adjusted the model's parameters with a combination of Dice Loss and Binary Cross - Entropy Loss over 75 epochs using the AdamW optimizer. The fine - tuned SAM was evaluated with metrics like IOU and OA. It achieved an IOU of 0.82 and OA of 0.88 on the Jilin - 1 satellite dataset, and an IOU of 0.85 and OA of 0.90 on the UAV dataset. It outperformed baseline models such as U - Net and Mask R - CNN, demonstrating its superiority in flood detection, thus contributing to more effective flood management.