Abstract: The environment of the Nanji Wetland in the Poyang Lake is complex and the wetland types are diverse. It is of great significance to realize the dynamic monitoring and automatic extraction of wetland information for the protection of wetland ecosystem. In this study, we used the Sentinel-2 image as the data source, the DEM data and the Google Earth high-definition image as the auxiliary data, and selected the image pure pixels as the training samples and the verification samples. According to the spectral and spatial characteristics of samples, the classification and regression tree(CART) model was constructed to realize the classification of the Nanji Wetland, and to study the types transfer changes of the Nanji Wetland from 2018 to 2021. The results showed that the combination of object-oriented CART decision tree classification method could get the classification results with high accuracy. Its overall classification accuracy could reach 86.36%, with a Kappa coefficient of 0.83, while the overall accuracy of the traditional supervised classification method was only 80.07%, the Kappa coefficient was 0.76, which was an improvement of the overall accuracy of 6.29%. The wetland types in the Nanji Wetland had obvious seasonal characteristics, with the area of grassland widely distributed in spring and winter, and most of the mudflat and grassland being flooded in summer and fall. The wetland types in the Nanji Wetland were shifted and changed mainly due to the influence of the water level and the topography factors, which mainly reflect the mutual transformation between the water area, mudflat and grassland. By the influence of water level and topographic factors, the wetland types in the Nanji Wetland had been shifted mainly in the mutual transformation among water, mudflat and grassland, and the distribution of cultivated land, forest land and building land were relatively stable. Compared with the traditional classification method, the object-oriented CART decision tree classification results were better, and the wetland information extraction could be realized quickly and efficiently.