Precipitation nowcasting plays an important role in mitigating the impacts of extreme weather events on social production and daily life. However, existing methods still face two major limitations. (1) Convolutional neural network-based methods are insufficient in modeling the temporal dependencies of radar echo sequences, which may lead to information loss in the prediction results. (2) Most existing methods enlarge the receptive field by stacking convolutional layers. This strategy makes it difficult to obtain a truly global receptive field and effectively model global dependencies, resulting in limited accuracy in heavy rainfall prediction. In addition, spatiotemporal information at different time steps is not fully integrated, and the multi-scale directional features of rainbands are often ignored. To address these issues, this paper proposes ST-TriMambaUNet, which consists of an encoder, a decoder, and a feature enhancement module. First, a Spatiotemporal Fusion Attention (STFA) was designed, including global spatial attention and temporal attention. It can effectively learn long-range spatial correlations and capture the temporal dependencies of radar echo sequences in a parallel manner. Second, a Multi-Scale Interaction Mamba (MSIM) module was developed with three branches. The first branch leverages Mamba to model global spatiotemporal dependencies with linear complexity. The second branch promotes spatiotemporal information interaction through channel shuffle and further combines Mamba to model global spatiotemporal dependencies. The third branch designs Multi-Scale Directional Convolution (MSDC) to learn the multi-scale directional features of rainbands. Finally, the features from the three branches are dynamically fused through the designed adaptive gated fusion mechanism. This enhances the model’s representation capability for strong-echo core regions and multi-scale precipitation band structures. Experimental results on two public datasets, SEVIR and CIKM, demonstrated that the proposed ST-TriMambaUNet achieved clear advantages in both overall prediction accuracy and heavy rainfall scenarios. In particular, under the high-threshold precipitation scenarios of SEVIR (160, 181, and 219), the CSI was improved by up to 10.19%. In the heavy rainfall scenario of CIKM at 40 dBZ, CSI, POD, and HSS were improved by 5.29%, 8.07%, and 4.65%, respectively.
Wang et al. (Tue,) studied this question.