Short-term rainfall prediction is typically defined as forecasting the spatial distribution of rainfall intensity over the next 0–6 h. Its accuracy plays a critical role in disaster prevention, urban management, agricultural activities, and numerous other domains. However, existing deep learning-based prediction models often rely on a single data source, which restricts their capacity to accurately capture the complex and dynamic evolution of rainfall processes. To address this limitation, we propose a novel prediction model called SDAUnet, which integrates satellite and radar data to enhance predictive performance. The SDAUnet incorporates static and dynamic attention mechanisms into the encoder of the U-Net framework, effectively capturing correlations among non-adjacent local information and relationships between consecutive frames. During multi-source data fusion, a channel adaptive fusion strategy is utilized, allowing the model to learn the adaptive importance weights of each data source. Extensive experiments conducted on the SEVIR dataset demonstrate the effectiveness and superiority of the proposed multi-source data fusion model in improving predictive accuracy.