SDNet: A General Framework for SWOT Denoising and Water Body Extraction

Dynamic water resource monitoring is essential for global sustainable development. Although the Surface Water and Ocean Topography (SWOT) satellite enables global water body observation, its Ka-band radar interferometer introduces noise and systematic errors, resulting in stripe noise and spatial heterogeneity. This study proposes SDNet, a high-precision water extraction framework based on the Transformer architecture. SDNet was validated in the Longyangxia Reservoir, Khukh Lake, and Kherlen River, showing strong performance in mitigating stripe noise, voids, and edge gaps. Prediction results were further used to analyze seasonal variations and driving factors of inland water bodies. Experimental findings show: (1) SDNet achieved high accuracy across different water body scales. For large-scale bodies, the Water Intersection Ratio (WIR) and Background Intersection Ratio (BIR) improved by 23.68% and 45.96%, respectively. For medium-scale bodies, WIR rose by 2.83%, and for small-scale bodies, by 3.57 times. (2) Altimetry error correlated positively with cross-track distance. The 250 m resolution data had a lower error (0.038 m) than the 100 m data (0.052 m), with median relative error between 10-5% and 10-4%. (3) Precipitation primarily drove water level changes in Longyangxia Reservoir, while ice-jam floods and topography influenced Khukh Lake and the Kherlen River.