Prior information assisted multi-scale network for precipitation nowcasting

Accurate precipitation nowcasting holds great significance for daily life. In recent years, deep learning networks have demonstrated excellent performance in the field of precipitation nowcasting. However, they did not fully harness important prior information such as the experience acquired from pre-trained model and the effects caused by terrain. In this paper, we propose a prior information assisted multi-scale network for precipitation nowcasting. Firstly, we employ a cross-attention mechanism to model the correlation between terrain elevation and radar echoes, enhancing the feature representation of the input. Subsequently, we introduce a teacher–student network, leveraging the pre-trained model’s capability in modeling echo movement as prior information to assist in the prediction. Finally, a multi-scale UNet network is proposed to cross-fuse large-scale and small-scale features so that the predicted images retain global information and more local details. We conduct precipitation nowcasting tests using real radar echo datasets within the 0–2 h range. Compared with the second best results(i.e., REMNet (Jing et al., 2022) for Probability of Detection (POD) and RainNet (Ayzel et al., 2020) for Critical Success Index (CSI)), our method improves the POD and CSI by 15.4% and 27.7%, respectively, demonstrating the superiority of our method.