Water bodies serve as one of the three major elements in maintaining wetlands. Their dynamic monitoring can effectively protect wetland ecosystems. Conventional methods for monitoring water bodies in wetlands employ field surveys or manual interpretation of remote sensing images, which are costly and inefficient, and inapplicable to continuous dynamic monitoring. In recent years, using methods like machine and deep learning to extract water body features from satellite remote sensing images has developed into an effective means for monitoring water bodies in wetlands. Based on the hyperspectral images from the ZY1-02D satellite, this study classified the water bodies in the Baiyangdian wetland using machine learning, convolutional and transformer neural networks. The accuracy and computational efficiency of water body classification under different spectral preprocessing methods and different image neighborhood sizes in training were compared to explore the optimal data preprocessing method and classification model for water bodies in wetlands. The results indicate that deep learning significantly outperformed machine learning in classification accuracy and computational efficiency. In particular, the spectral-spatial residual network (SSRN) model based on the convolutional neural network achieved the highest classification accuracy (OA: 99.09 %, Recall: 99.62 %, F1-score: 0.99) under conditions of all spectral bands and a 9×9 neighborhood size. Besides, despite a low signal-to-noise ratio, the atmospheric water vapor absorption band contained significant information, assisting in improving the classification accuracy of water bodies in the wetland during model training and prediction. The results of this study are expected to provide methodological support for the business operation of water body classification of wetlands.