Figure1

Figure 1. Overall framework of the proposed method. Historical radar frames are first encoded into hierarchical spatial features, then processed by a spatio-temporal module with CBAM refinement, and finally decoded into future precipitation frames. SimVP predicts five frames per forward pass, and four autoregressive rollouts generate the complete 20-frame forecast. Only representative frames are shown for clarity. During training, the predictions are jointly supervised by the MSE loss and the temporal consistency loss. CBAM: Convolutional block attention module; MSE: mean squared error.