WaveDSNet: A Wavelet-Enhanced Discrepancy Semantic Network for Water Change Detection in SAR Imagery
Abstract
Water change detection in synthetic aperture radar (SAR) imagery is important for inundation extent delineation and river course monitoring. However, existing methods still face three core challenges: inherent speckle noise interference, insufficient extraction of semantic-structural features, and the lack of dedicated benchmarks for evaluating high-resolution structural details. To address these issues, a wavelet-enhanced discrepancy semantic network (WaveDSNet) and the Xidian University SAR Water Change Detection (XDU-SWCD) dataset are proposed for water change detection in SAR imagery. WaveDSNet consists of two key components: first, a wavelet-based noise suppression (WNS)-embedded backbone performs wavelet-subband-aware denoising to suppress speckle while preserving structural cues; second, a semantic-boundary collaborative refinement (SBCR) module aggregates bi-temporal features into difference and semantic streams, and further coordinates them through bidirectional interaction and adaptive gating, with auxiliary boundary-aware supervision for fine-grained changed-water boundary localization. Furthermore, the XDU-SWCD dataset is constructed as a focused high-resolution benchmark for SAR water change detection. It combines fine-grained water-change structures with representative confusing nonwater regions, enabling a rigorous evaluation of boundary delineation and robustness to low-backscatter pseudo-changes. The experiments on XDU-SWCD show that WaveDSNet achieves stronger overall performance than nine representative change detection methods, with an overall IoU of 0.7796 and an overall ${F}1$ -score of 0.8590. Zero-shot cross-dataset evaluation on public benchmarks further demonstrates its promising transferability. These results suggest the effectiveness of both the proposed framework and the XDU-SWCD benchmark for SAR water change detection. Our code and XDU-SWCD dataset will be publicly available at https://github.com/gosling123456/WaveDSNet.git