RivUNet: An Attention-Gated Deep U-Net for Water Body Segmentation in Satellite Imagery to Enhance River Encroachment Monitoring

Effective water body mapping is fundamental to sustainable water resource management and river conservation in urban and semi-urban regions which face the risk of encroachment. In this chapter, we propose a customized Attention U-Net architecture called RivUNet that is specifically tailored for the semantic segmentation of water bodies in satellite imagery. The model is trained on the Satellite Images of Water Bodies dataset, comprising Sentinel-2 imagery accompanied by binary masks generated through the Normalized Difference Water Index (NDWI). TThese masks distinguish water from non-water regions, which enables a focused learning process for water body segmentation. The proposed model enhances the traditional U-Net architectures by incorporating attention gates within the skip connections, which allow the model to selectively emphasize spatially important features during the decoding process. The proposed model achieves a testing accuracy of 91% and an AUC score of 0.95, demonstrating its capability to generalize across diverse scenes and water body structures. This segmentation technique can be adopted to monitor how much river encroachment has occurred if analyzed, which accelerates river encroachment monitoring. This chapter lays a strong foundation for using semantic segmentation techniques in water management applications, including temporal river monitoring, hydrological analysis, and urban planning support.