RadarEchoMamba: A Fast, High-Fidelity Pyramidal Bidirectional Mamba Model for Radar Echo Extrapolation

Radar echo extrapolation is a fundamental task in precipitation nowcasting. However, existing models based on RNNs, CNNs, and Transformers are often constrained by error accumulation, the loss of temporal information, and quadratic computational complexity. Consequently, these limitations can lead to prediction blurring and the distortion of fine-grained details in extrapolated radar images. To address these challenges, we propose RadarEchoMamba, a novel extrapolation framework built on the efficient Mamba model. RadarEchoMamba uses a pyramidal architecture to capture multi-scale features, and its core is a Bidirectional Spatiotemporal Mamba backbone that models dependencies within the observed historical input window. Furthermore, we introduce a spatiotemporal prior module driven by the input sequence to guide the reconstruction of high-resolution features during the decoding phase. This design enables the model to generate detail-rich predictions, improving upon the blurring issues prevalent in traditional extrapolation models. Experimental results on the South China and Shanghai-2020 datasets show the effectiveness of our method. Compared with strong Transformer-based baselines, the lightweight variant uses fewer parameters and achieves lower measured inference latency, while maintaining competitive prediction quality. Its main advantages are observed in visual-fidelity metrics, with improved SSIM and reduced LPIPS on the South China dataset.