Answers are grounded in this paper and cite their sources.
As lead time increases, deterministic precipitation forecasts often underestimate localized rainfall peaks and become spatially smoother. We propose HM-SimVP, a precipitation nowcasting framework built on SimVPv2. The framework combines dual-scale spatial encoding, latent hyperedge interaction, and gated mean–contrast output refinement. The dual-scale encoder supplies local and contextual features for latent spatiotemporal prediction. In latent space, an adaptive hypergraph groups spatial locations by feature similarity. Neighboring hyperedges exchange information according to the overlap of their soft memberships, combining spatial adjacency with membership compatibility. After decoding, mean–contrast refinement adjusts patch-level means and within-patch contrast in the log-normalized precipitation field, with a learned gate controlling the correction magnitude. On CMPAS and MeteoNet, HM-SimVP achieves higher critical success index (CSI) values than eight deep learning methods and an optical-flow extrapolation baseline at all four evaluated rain-rate thresholds. At 20 mm h−1, CSI is 30.0% and 16.6% higher than that of the best-performing comparator on CMPAS and MeteoNet, respectively. Lead-time analysis shows that high-threshold CSI advantages persist at later forecast steps. Component and internal ablations support the contributions of latent hyperedge interaction and gated output refinement. These results show that combining latent hyperedge interaction with gated intensity refinement improves the retention of high-intensity precipitation in multistep forecasts.
No takes yet. Share an insight, caveat, or question.
Ma et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: