This study aims to improve the accuracy of high-resolution water body extraction in the Sanjiangyuan region. Using sub-meter Jilin-1 satellite imagery, a self-supervised pretraining approach based on the ViT(Vision Transformer) backbone is developed, followed by fine-tuning a water body extraction model with ViT-B as the encoder and UperNet as the decoder. The results demonstrate that the proposed method achieves an IoU( Intersection over Union) of 92. 93% , outperforming non-pretrained models by 11. 87% and surpassing other open-source pretrained weights. A total of 25 621 water body patches with an area of 1 544 km² are extracted. Comparative analysis with WorldCover 2021 reveals that the method effectively detects small water bodies missed by previous methods, providing a more accurate reference for water resource analysis in plateau wetlands.

high-resolution remote sensing images, pre-trained models, transfer learning, Sanjiangyuan region, water body extraction