Abstract: In the current ground-based cloud image classification task,there are problems such as low recognition accuracy. In order to improve the accuracy of cloud classification,the DAR-CapsNet classification model for ground-based cloud images has been constructed by effectively integrating the features of depthwise separable convolution,attention mechanism and residual structure. Firstly,the ground-based cloud images were collected from the public database of the National New Energy Laboratory of the United States to build a cloud classification database;then,the proposed DAR-CapsNet classification model was trained and optimized;finally,experiments were conducted on different datasets to verify the performance of the proposed classification model. The experimental results show that the classification accuracy of the DAR-CapsNet model is as high as 95.50%,which is better than some published classification models,and the DAR-CapsNet model has better generalization performance on different datasets.