Refining Water Body Extraction by Remote Sensing With Deep Learning Models: Exploring Different Band Combinations

Satellite-based remote sensing is essential for monitoring water resources and supporting ecological conservation and sustainable development. However, complex scenarios, such as cloud shadows and ice–snow misclassification, challenge traditional methods, such as spectral indices and thresholding. This study of the arid Xinjiang region in northwest China introduced a Transformer-based deep learning model to improve lake extraction. Seven training datasets were constructed to ensure the quality and diversity of the training samples. Cloud-shadow negative samples were incorporated to enhance robustness. Different three-band combinations from Sentinel-2 imagery were composed and evaluated, including visible bands (B2–B4), the near-infrared band (B8), and shortwave infrared bands (B11 and B12). The model trained with the B4, B8, and B11 combination delivered the best performance (intersection over union: 0.9715, F1: 0.9855). Validation with Dynamic World and CLCD datasets demonstrated strong generalization capabilities in arid regions, confirming the model’s effectiveness for large-scale water body extraction in complex environments. Model portability was assessed in selected regions with diverse geographical conditions in the Tibetan Plateau, North America, and South America. Refining model training and validation can resolve misclassification issues identified in other regions. The findings provide a robust water resource monitoring and management tool with promising applications in similar complex environments.