A target-guided attention framework for medium-resolution glacial lake mapping in Southeastern Tibet

Accurate mapping of small and fragmented glacial lakes from medium-resolution imagery remains challenging because of mixed pixels, blurred shorelines, alpine background interference, and severe foreground–background imbalance. This study develops a target-guided framework for glacial lake mapping in Southeastern Tibet using 30 m Landsat 8 imagery. Landsat bands B3–B7, NDWI, NDSI, and terrain slope were integrated into an eight-channel spectral–index–topographic representation, while target-guided patch extraction increased informative samples and selective CBAM integration refined lake-related features in deep U-Net stages. Under a spatially independent evaluation protocol, CBAM-U-Net achieved a precision of 0.9158, recall of 0.9196, IoU of 0.8482, and F1 score of 0.9177. Relative to the U-Net baseline, IoU improved by 13.15% with only a marginal increase in parameter count. Lake-level mean IoU increased from 0.80 for 0.02–0.05 km2 lakes to 0.93 for lakes > 1.00 km2. Grad-CAM analysis further indicated more localized lake-related responses after CBAM integration. The framework provides a practical and interpretable approach for medium-resolution glacial lake inventory mapping and regional cryosphere monitoring.