MSA-TCNet: a multi-scale attentional transformer-CNN network for detail-preserving water extraction in remote sensing images

High-precision water extraction is essential for water resource management, disaster mitigation, and engineering applications. Although deep learning has greatly advanced water body extraction, achieving accurate delineation in complex environments remains challenging, especially for fine-scale structures, boundary continuity, and fragmentation caused by diverse morphological patterns. To address these issues, we propose MSA-TCNet, a multi-scale attentional transformer-CNN network In the global pathway, we design a Multi-Scale Multi-Head Window Linear Self-Attention (MSMH-WLSA) and a Global Convolutional MLP (GC-MLP). The former enhances the acquisition of global and multi-scale contextual information with high computational efficiency, while the latter preserves local spatial topology and strengthens cross-window feature interaction in the MSMH WLSA module. To mitigate spatial information loss, an additional spatial pathway is introduced to emphasize fine-grained details. In the decoding stage, a Dynamic Fusion Attention module progressively restores spatial resolution and reduces semantic inconsistencies across hierarchical features, producing high-fidelity water extraction results. Extensive experiments conducted on one Self-made dataset and two public datasets demonstrate that MSA-TCNet achieves superior performance in both extraction accuracy and boundary coherence compared with state-of-the-art methods.