Adapting the Segment Anything Model for SAR-Based Flood Detection Using Parameter-Efficient Fine-Tuning Techniques

Flooding, as a common natural disaster, poses severe threats to human life, property, and economic activities. To address these challenges, rapid, reliable, and robust flood extent detection plays a critical role in disaster prevention and mitigation. Recent advancements in computer vision, such as the Segment Anything Model (SAM), have introduced innovative approaches to flood detection by leveraging their strong feature extraction capabilities. However, their reliability in Synthetic Aperture Radar (SAR)-based flood detection tasks is limited due to the lack of relevant training samples. To address this limitation, this study fine-tunes SAM on SAR-based flood datasets using multiple Parameter-Efficient Fine-Tuning (PEFT) techniques to explore the feasibility of applying SAM for flood detection with SAR imagery. Five mainstream PEFT techniques—BitFit, Adapter Tuning, Prompt Tuning, Prefix Tuning, and LoRA—were employed. The experimental results demonstrate that all fine-tuned models significantly improved their performance in terms of Intersection over Union (IoU) and accuracy. Among them, the model fine-tuned with the LoRA technique achieved the best performance, with improvements of 34.88% and 44.33% in IoU and accuracy, respectively. This study highlights the potential of fine-tuning SAM for flood detection in SAR imagery and provides a novel approach to improving the accuracy and reliability of flood mapping.

Keywords: Flood Detection, SAR imagery, Segment Anything Model, PEFT