Integrity Extraction of Water Bodies in Complex Scenes Based on the Adaptive Watermamba Framework

Water body integrity extraction is crucial for watershed management, flood monitoring, and other related tasks. Existing methods are prone to noise interference in water scenes that include ancillary facilities, leading to inaccurate boundary detection and disconnected end water bodies. To address these issues, we propose the adaptive watermamba framework (AWF), consisting of the watermamba network (WM-Net) and a skeleton-point-based small-water-bodies post-processing (SSP) method. Firstly, the WM-Net enhances the adaptability of water bodies to complex background interference by integrating the multiple equalization-channel vision mamba (MEVM) module and the mamba boundary attention (MBA) module. The MEVM uses a parallel mamba architecture to perform linear scanning across different channel dimensions, capturing richer global-local features and effectively mitigating interference from background noise. The MBA module improves the integrity of water body boundary extraction by capturing global boundary features and enhancing sensitivity. Secondly, to address the issue of disconnected small water bodies, this paper designs the SSP method. By extracting skeleton points from the boundary of water body images, comparing the omitted parts in the initial water body results, and performing selective connection, the integrity of disconnected small water bodies is improved. In addition, we have constructed and publicly released two high-quality annotated datasets: the Minjiang river basin (MRB) and Dongting lake (DTL). Experimental comparative analysis was conducted on MRB, DTL, and the public dataset GID. The results showed that the intersection over union (IoU) of AWF reached 95.34%, 95.25%, and 93.74%, respectively, outperforming the state-of-the-art methods. At the same time, it further verified the advantages of the WM-Net and the SSP in enhancing boundary integrity and flow continuity. Both boundary IoU (BIoU) and structural similarity (SSIM) of the proposed method are superior to those of the SOTA method. The AWF code can be obtained at https://github.com/329719386/Adaptive-WaterMamb-.git.