Abstract: In order to precisely classify water bodies and further enhance the accuracy and efficiency of water body extraction in the Changjiang River, we took the water bodies of Hukou Hydrographic Station of the Lower Changjiang River as the research object, the GF1B image water data of Zhangjiazhou Reach and Poyang Lake exit of the lower Changjiang River was selected, the water extraction data set of ZP-River semantic segmentation was constructed by manually annotating the water data set. The water extraction performance of U-Net, BiSeNet, U-NET ++, DeepLabv3+ and MAResU-Net methods in ZP-River dataset (at the exit of Changjiang River and Poyang Lake) were compared and analyzed. Experiments showed that compared with the other methods, the MAResU-Net model was superior to the deep learning method in many evaluation indexes, and its average F1 score was 98.74%. The model had shown excellent performance in accurately distinguishing between small water bodies (such as reservoirs, small ponds, pits and man-made ponds) and confusing features around the Changjiang River (including ships, dam structures and so on). Compared with U-Net method, the recognition efficiency of this model was improved by about 40%. The research results can provide a reference for the construction of digital twin basin intelligent model, and help build a monitoring and defense system for water and drought disasters in the Changjiang River.