CSW-SAM: a cross-scale algorithm for very-high-resolution water body segmentation based on segment anything model 2

Large-scale high-resolution water body (WB) extraction is one of the research hotspots in remote sensing image processing. However, accurate training labels for various WBs at Very-High-Resolution (VHR) are extremely scarce. Considering that low-resolution (LR) images and labels are more easily accessible, the challenge lies in fully leveraging LR data to guide high-precision WB extraction from VHR images. To address this issue, we propose a novel cross-scale CSW-SAM algorithm based on SAM2, which learns spectral information of WBs from easily accessible 10 m resolution LR images and maps it to 0.3 m resolution VHR remote sensing images for high-precision WB segmentation. In addition to fine-tuning the decoder, we enhance the encoder’s ability to effectively learn the mapping relationship between images of different resolutions by Adapter Tuning. We have designed the Automated Clustering Layer (ACL) based on the principle of feature similarity and local structure information clustering, to enhance the performance of SAM-based methods in cross-scale WB segmentation. To validate the robustness and generalization ability of the proposed CSW-SAM, we conducted extensive experiments on both a self-constructed cross-scale WB dataset and the publicly available GLH-Water dataset. The results confirm that CSW-SAM achieves strong performance across datasets with diverse WB conditions, demonstrating its potential for scalable and low-cost VHR WB mapping. Additionally, the model can be generalized with minimal cost, making it highly promising for large-scale global VHR WB mapping.