A Spectral–Spatial Fusion Network for Hyperspectral Small Water Body Extraction

In hyperspectral imagery, small water bodies are often difficult to delineate due to their limited spatial extent, irregular shapes, and spectral similarity to surrounding land-cover types, which can lead to missed detections and fragmented boundaries. To address these challenges, this study proposes a spectral–spatial fusion network with parallel 3D CNN and ResNet-50 branches to extract complementary spectral–spatial and hierarchical spatial features. A progressive gated spectral–spatial fusion (PGSSF) module integrates these features across successive scales to help preserve weak water-body features and reduce background interference. The fused features are further processed by a scale-adaptive local–global (SALG) module, which combines local spatial details with broader contextual information to better characterize narrow watercourses and their continuity. Experiments were conducted using hyperspectral imagery from the Dongting Lake region, with comparisons against the normalized difference water index (NDWI), U-Net, Swin Transformer, and WaterFormer. The proposed method achieved a Water IoU of 80.68%, a recall of 90.33%, and an F1-score of 89.31%, with Water IoU exceeding that of WaterFormer by 4.76 percentage points. These results demonstrate the potential of the proposed method for regional small-water-body mapping and surface-water monitoring.