SAREdgeNet: Lightweight Parallel Deep Learning via Task Decoupling for Large-Scale SAR Lake Segmentation

Monitoring lake extent is essential for understanding hydrological processes and climate change. Synthetic aperture radar (SAR) provides all-weather observations, but speckle, ambiguous backscatter, and limited spatial context complicate large-area lake extraction. We present SAREdgeNet, a lightweight framework based on Open Admission, Strict Graduation (OASG), which separates candidate generation from contour classification. Full-image edge detection generates potential shoreline contours. A parallel dual-path classifier combines contour masks with SAR-intensity, digital elevation model (DEM), and geometric descriptors to reject non-lake candidates before lake-mask reconstruction. The classifier was trained and validated on Kokoxili contours. Regional SAREdgeNet scores were computed for lakes ≥ 1 km2 against a 2022 Landsat-derived reference in other covered areas of the Qinghai–Tibet Plateau. Across ten training seeds, precision, recall, and F1-score were 96.23 ± 0.37%, 94.19 ± 0.27%, and 95.20 ± 0.32%, respectively (mean ± sample standard deviation). These results support task decoupling for extensive lakes; recall for lakes < 1 km2 was 12.39%, and weak shoreline candidates remain a limitation.