Precise Surface Water Mapping from SAR Images in Kenya by a Multi-Scale Attention Network with Hybrid Dilated Convolution

Surface water is essential for ecosystem stability and socioeconomic development. In regions with high climate variability, such as Kenya, effective water management requires accurate and frequent monitoring. Cloud cover frequently limits optical remote sensing, whereas Synthetic Aperture Radar (SAR) provides an all‑weather alternative for surface water resource monitoring. However, identifying small water bodies and narrow rivers in SAR imagery is challenging due to background noise and similar backscatter characteristics between water and other land surfaces, such as smooth bare soil, wet ground, or shadowed terrain. The extracted water bodies often exhibit fragmented boundaries and incomplete shapes. To address these issues, a new network, CHA‑HRNet (Coordinate attention and Hybrid dilated convolution ASPP enhanced HRNet), is proposed for precise water mapping from Sentinel‑1 satellite imagery. A coordinate attention (CA) mechanism is introduced into the HRNet backbone to improve feature discrimination for small targets and complex boundaries. A Hybrid Dilated Convolution Atrous Spatial Pyramid Pooling (HDC-ASPP) module is integrated into the decoder, which aggregates multi-scale context information to enhance boundary continuity and connect linear water characteristics. The results show that CHA-HRNet effectively suppresses SAR image noise and handles different changes of water bodies well. Compared with other widely used deep learning models, the proposed model achieves the best performance, with the Precision, Recall, F1-score, and IoU of 98.38%, 96.56%, 97.46%, and 95.05%, respectively. Specifically, the high Precision indicates effective suppression of background noise and reduced false detections, and the high Recall reflects reliable detection of small and fragmented water bodies. The IoU further reflects the accurate morphological delineation of complex water features, and the F1-score confirms the overall robustness across complex environments. These advancements make CHA‑HRNet suitable for accurate water body monitoring in areas with complex climates, such as Kenya.