Attention-Driven U-Net for Improved Water Body Segmentation from Satellite Imagery

Abstract

Accurate segmentation of water bodies from satellite imagery is essential for sustainable water resource management, flood monitoring, and hydrological modeling. In recent years, many deep convolutional neural network (DCNN) based techniques have been developed for the purpose of water bodies segmentation. These models often suffer from issues such as class imbalance, poor land-water contrast in urban environments, and loss of fine boundary details, although they have significantly improved segmentation accuracy over traditional methods. In order to improve spatial feature refinement, this study proposes a modified U-Net architecture that incorporates residual connections into the skip connections, followed by a Dual Attention Block (DAB). Furthermore, by combining features from several receptive fields, a Multi-Scale Fusion Block (MSFB) is incorporated at the bottleneck to improve contextual understanding. The model was trained on Sentinel- 2 satellite images and their corresponding segmentation masks obtained from the public dataset. Compared to traditional DCNN architectures, the proposed U-Net based model provides better qualitative and quantitative results and captures small and fragmented water features with greater robustness to spatial complexity and environmental variability.