Super-resolution enhanced Segment Anything Model 2 for accurate inundation scenario mapping

Flood disasters are escalating in severity, underscoring the importance of timely and accurate inundation scenario monitoring using remote sensing imagery for effective disaster prevention and management. However, the scarcity of high-quality imagery poses significant challenges for precisely delineating flooded areas. We proposed the Super-resolution Segment Anything Model 2 (SSAM2), a framework that integrated a super-resolution module into the SAM2 segmentation framework for mapping inundations from low-quality remote sensing imagery. Furthermore, we introduced a flood prompt generation mechanism using the Normalized Difference Water Index to prompt the pre-trained SAM2 model to learn flood-related information. To evaluate our approach, we created a dataset based on FloodNet, comprising 8273 pairs of low- and high-quality images with corresponding inundation labels. Experimental results demonstrated that it significantly improves inundation segmentation accuracy in the absence of high-quality remote sensing imagery. Then, we applied the trained model to flood dynamic monitoring using Landsat-8 imagery of Taihu Lake, observing that the area of Taihu Lake changed, which has demonstrated its efficacy in accurately identifying inundated regions. We showcase the substantial potential of SSAM2 in satellite remote sensing-based flood mitigation applications, offering a technological solution for the dynamic monitoring of large-scale flood disasters when lacking high-quality images.