Abstract

Accurate precipitation nowcasting for extended horizons (0-3 hours) remains a critical challenge in meteorology, as existing methods typically focus on shorter periods (0-1 or 0-2 hours) and suffer from rapid performance degradation beyond 90 minutes when relying solely on radar observations. While generative models show promise, their slow inference speeds limit operational deployment. To address these limitations, we propose MambaRain, a novel multi-scale encoder-decoder framework that synergistically combines Mamba's long-range temporal memory capabilities with self-attention mechanisms for extended precipitation nowcasting. Our core contribution lies in designing a hybrid architecture where Mamba blocks capture global spatiotemporal dependencies across extended sequences, while self-attention mechanisms complement Mamba's sequential processing by capturing explicit spatial dependencies and enabling parallel global context aggregation. Furthermore, we propose a spectral loss to mitigate the averaging blur effect commonly observed in chaotic precipitation systems, thereby preserving the clarity of fine-scale motion patterns over the 0-3 hour horizon. Unlike previous approaches that struggle with balancing global context and local details or rely on computationally expensive generative models, our deterministic framework maintains efficiency while extending the effective forecasting window. Extensive experiments on the Xinjiang and Southeast China SWAN datasets demonstrate that MambaRain achieves superior performance in 0-3 hour precipitation nowcasting, significantly outperforming existing methods in both accuracy and computational efficiency.