Foundation model for waterbody segmentation based on ground camera images

초록

"With the increasing flash floods due to climate change, real-time flood detection methods have gained significant attention. For real-time flood detection, ground-based camera images are primarily used. Unlike satellite images, deep learning-based waterbody segmentation techniques are applied to handle the high variability of waterbody segmentation. To address the challenge of constructing a large volume of ground truth images, Segment Anything Model 2(SAM 2), based on the foundation model and the concept of prompts, has been proposed. Applying this model to flood detection is expected to enable high-accuracy waterbody segmentation with relatively small training datasets through fine-tuning. Therefore, this study proposes an advanced waterbody segmentation model based on SAM 2 and compares their performance with the existing model (V-FloodNet). The improved model, which selectively applies prompts, achieved a 10% improvement in performance compared to the existing model, demonstrating higher accuracy and greater efficiency in the segmentation process. © 2025 Korea Water Resources Association.

키워드