DFAMamba - a dual-channel fusion precipitation nowcasting network based on attention and Mamba model

The extrapolation of short-term precipitation forecasting from radar echo maps is of great significance for disaster prevention and control, but the existing methods mostly rely on a single radar data source, which is difficult to capture complex precipitation patterns, and the prediction accuracy is insufficient and the computational complexity is high. In this paper, a multi-source fusion model based on Mamba, DFAMamba is proposed, which uses the U-Net architecture and Vision Mamba encoding and decoding to fuse multi-dimensional radar features through the spatio-temporal attention module, and introduces satellite data to enhance feature complementarity. Experiments on the Yangtze River Delta dataset show that compared with ROVER, TrajGRU, ConvLSTM, SimVP, SwinTransformerUNet and other methods, the proposed model has better precipitation accuracy indexes, reduced computational complexity, and significantly improves the prediction ability of the initial stage of convection, providing an efficient and accurate solution for short-term precipitation forecasting.