Abstract: Aiming at the problem of insufficient accuracy of automatic extraction of multi-feature inland water bodies by satellite remote sensing,this paper takes Zhejiang province as the research scope to explore the extraction accuracy of the Vision Transformer (ViT)largevision model for inland water bodies with different features. Through the results of historical geographical national conditions monitoring,large scale samples are obtained to get a pre-trained model.Combined with the multi-level perception characteristics of inland water bodies in Zhejiang province,the UPerNet network is used to optimize the output layer of the ViT model from different aspects such as scenes,objects,parts,materials and textures,further increasing the perception ability of the ViT model for multi-scale and multi-feature water bodies. The accuracy and recall rate of the algorithm in this paper can reach 90%,which is 15% higher than the traditional exponential threshold method and 10% higher than the pre-trained model.Through post-processing,it can meet the accuracy requirements of the water surface area survey and monitoring in Zhejiang province.The results show that the feature-optimized visual large model can be well applied to the extraction of multi-feature inland water bodies and serve the national water resources survey.

Key words: vision large model, inland water bodies, satellite remote sensing, multi-level perception