Efficient Multi-Scale Water Body Extraction Based on Synergistic Optimization of Shunted Transformer and Dual-Path Feature Fusion

Remote sensing image-based water body extraction faces several challenges, including background interference, sparse target distribution, and high similarity between water bodies and background classes. Existing methods struggle with complex edges and small targets, while Transformer models are often constrained by computational complexity. To address these issues, this study proposes a synergistically optimized Shunted Transformer and Dual-Path Feature Fusion Network (SDA-Net). The proposed method improves the Shunted Transformer module by introducing a multi-scale token aggregation and attention head grouping strategy, which significantly reduces computational costs while achieving efficient multi-scale global dependency modeling. Additionally, a Convolutional Feature Fusion (CFF) module is incorporated, utilizing wavelet convolution and an adaptive channel selection mechanism, which enhances the extraction of crucial local details, such as water body edges, without compromising model lightweightness. Furthermore, a Dual-Skip Connection (DSC) mechanism is designed to simultaneously transmit the complete self-attention blocks and their internal attention maps from the encoder to the decoder, overcoming the limitations of single-skip connections and deeply fusing local details with global contextual information. An Adaptive Learning Module (ALM) is also introduced, which dynamically adjusts feature weights through cross-channel correlation learning to effectively suppress noise and enhance key semantics, ensuring the fine structure of water bodies is retained. Experiments on the LoveDA and GID datasets demonstrate that SDA-Net significantly outperforms existing methods, achieving IoU scores of 80.39% and 91.03%, respectively, and F1 scores of 89.13% and 95.53%. The model requires only 21.04M parameters and 30.41G FLOPs, maintaining lightweight properties while significantly improving the handling of complex edges and small target water bodies, thereby validating its superior accuracy and efficiency.