SCFNet: A Swin-CNN Synergistic Fusion Network for Urban and Rural Water Extraction in Remote Sensing Images

Accurate monitoring of water resources is essential for disaster risk reduction and sustainable development amid global climate change. At present, various methods based on convolutional neural networks are widely used in the research field of remote sensing water extraction. However, traditional convolutional neural network-based methods, constrained by limited receptive fields, often yield discontinuous boundaries and fail to detect fragmented water regions, while Transformer-based models lack sensitivity to capture fine-grained spatial features. To address these above limitations, a novel Swin-CNN Synergistic Fusion Network is proposed in this study,which integrates three core modules: a Cross-Window Fusion Module to enhance spatial continuity, a Bidirectional Feature Bridge to connect local details with global semantics, and an Adaptive Multi-level Alignment and Fusion Module strategy to mitigate semantic inconsistency between deep and shallow layers. Experiments conducted on two benchmark remote sensing datasets demonstrate that the proposed method achieves 97.84% segmentation accuracy for coastline delineation using GF dataset, and 96.29% accuracy for inland water extraction using the GID-WD dataset. These results highlight the effectiveness of the proposed method in accurately extracting water boundaries under complex environmental conditions, enabling precise multi-scale water extraction—ranging from decameter-level coastal boundaries to meter-level lakes, rivers, and fragmented small water bodies—and underscore its potential applications in flood early warning, coastal zone protection, and urban water resource management.