Abstract: To improve the accuracy of rainfall prediction and to solve the problem of high computational complexity, a DA-UNet with multimodal fusion for rainfall prediction based on convolutional block attention mechanism and depthwise separable convolution is proposed. The DA-UNet integrated multimodal data, captured the feature dependence between each part of the image through the attention mechanism to improve the accuracy of precipitation forecast, and reduced the network parameters by varying the feature extraction process. Comparative experiments and ablation experiments were carried out on MeteoNet dataset. The results show that the multimodal fusion of DA-UNet improves the overall rainfall forecasting performance compared with other algorithms, while reducing the parameter volume by three-quarters.