M4Caster: Multi-source, multi-spatial, multi-temporal modeling for precipitation nowcasting

Dan Niu, Chunlei Shi, Tianbao Zhang, Hongbin Wang, Zengliang Zang, Mingbo Jiang, Jun Yang

Abstract: Highlights•We propose M4Caster, a unified framework that jointly models radar and meteorological-satellite observations for 0–1 h precipitation nowcasting, thereby overcoming the limitations of single-source extrapolation.•We develop a MMA that extracts hierarchical spatiotemporal features via multi-scale patch embedding, adaptive cross-scale perception refinement, and multi-temporal self-attention, enabling balanced integration of local details and global context.•We introduce a bidirectional bridging fusion mechanism in which dual cross-attention pathways align and mutually enhance radar and satellite representations through residual connections.•Extensive evaluation on the Yangtze River Delta dataset demonstrates that M4Caster significantly outperforms some excellent ConvRNN and transformer baselines, achieving substantial gains in heavy-rain nowcasting accuracy and convective-initiation detection.