A Margin-Based Decomposition Framework via Intensity Flow Matching for Precipitation Nowcasting

Precipitation nowcasting is an important and challenging spatiotemporal prediction task. Most of the previous studies are hindered by the gradual fading of high-value echoes and positional inaccuracies, which stem from the inherent aggregation and dispersion phenomena in the evolution of the precipitation system. To address this problem, we propose a novel margin-based decomposition framework that effectively decouples the precipitation system into advection and intensity, where the advection branch models the overall motion trend and the intensity perception network provides an initial estimate of intensity-related variations. To address the branch degeneracy issue of conventional decomposition methods, an effective margin-based decomposition strategy is introduced. Additionally, to better characterize intensity variations and preserve local details, an intensity flow matching network is developed to refine the intensity component. Extensive experimental results on four publicly available radar datasets, namely Storm EVent ImagRy (SEVIR), MeteoNet, Shanghai Radar, and CIKM, demonstrate the effectiveness and superiority of the proposed framework. Across these benchmarks, the proposed method consistently improves multiple forecasting backbones, achieving relative gains of up to 15.7% in critical success index (CSI) and 14.6% in Heidke skill score (HSS).