SDNet: Sandwich Decoder Network for Waterbody Segmentation in Remote Sensing Imagery

Waterbody extraction is essential for monitoring surface changes and supporting disaster response. However, differences in morphology, dimensions, and spectral reflectance, make it problematic to segregate waterbodies accurately in remote sensing (RS) photographs. Although self-attention (SA) and convolutional neural networks (CNN) have demonstrated strong modeling capabilities on RS image recognition, their efficiency and expressiveness still deserve further optimization. In this paper, we propose SDNet, an innovative deep learning structure designed to improve segmentation accuracy and scalability in waterbody extraction while minimizing computational costs. SDNet introduces a Two-Step Self-Attention (TSSA) strategy to selectively aggregate spatial features and a Sparse-Tight Pattern Fusion (STPF) module to capture context across various manipulation spaces, enhancing axial and space information fusion. The network's sandwich-shaped decoder structure allows for easy integration with mainstream encoders. Compared to previous influential methods, SDNet exhibits superior sensitivity in detecting small waterbodies and delivers better overall performance. Experiments performed on public benchmark datasets, like MSRWD, UAVid, and LoveDA, indicate that SDNet exceeds prior state-of-the-art techniques, with a 2%-10% increase in mIoU and a 3%-9% improvement in F1-score. On the MSRWD, SDNet attained an IoU of 94.21% for the water class, with an F1-score of 97.03% and an overall accuracy of 98.87%.