WASNet: Water Area Segmentation Network with Enhanced Positional Encoding and Large-Kernel Attention for SAR-Based Hydrological Monitoring

It is an important task to automatically extract water areas from various synthetic aperture radar (SAR) images for hydrological monitoring activities, such as water resource management, flood risk assessment, and drought detection. However, it is challenging to achieve accurate water area extraction, due to image noise, mountain shadows, and land cover interference. To address these limitations, this paper proposes an accurate segmentation network for water area extraction based on the U-Net framework called WASNet. It first integrates Transformer and spatial convolutional networks to cooperatively learn features from both global and local perspectives, which enables effective combination of multiscale features to achieve accurate segmentation of water areas in SAR images. Subsequently, we propose large-kernel generalized attention (LKGA) in each Transformer block of its decoder, to expand its receptive field and enhance global context understanding, for capturing a wide range of feature-associated regions. Furthermore, we design extended local positional encoding (ELPE) to enhance the effectiveness of positional embedding, and add it into the computation of the attention mechanism for each Transformer block, which reinforces the ability to capture water area edge details. Additionally, we devise a multiscale enhanced feed-forward network (MSEFF) to dynamically balance the importance of feature information across different branches, strengthening the multiscale contour details of water areas. Eventually, we verify the performance of our proposed WASNet, by quantitative and qualitative experiments on a representative subset with 1,600 Sentinel-1 SAR image pairs extracted from the publicly available GID-15 data set. Extensive experimental results demonstrate that the WASNet significantly outperforms state-of-the-art water area segmentation networks, particularly in the identification of small branches and edge details, enabling accurate extraction of water areas for understanding climate-driven migration.