Expanded Training Data Extends Hail Nowcast Lead Times by 25 Minutes in Spatiotemporal Neural Networks

The arXiv preprint by George Pacey and colleagues tested how data volume, augmentation, and input timesteps affect a convolutional-recurrent model for 0-60 minute hail nowcasts without changing network weights or layers. Experiments scaled training from a few years to over a decade of pan-European composites while holding architecture constant, revealing monotonic gains in critical success index that grew with lead time. Data augmentation via mirroring and rotation boosted performance only when the base dataset exceeded roughly eight years; smaller sets showed degradation, suggesting augmentation amplifies noise when examples are scarce. Sensitivity to the number of input radar frames proved weaker than sensitivity to total training years, indicating that sheer volume of diverse storm examples outweighs marginal gains from longer temporal context once a minimum sequence length is met. These results align with earlier findings in precipitation nowcasting where dataset scale dominated architectural tweaks, yet diverge by isolating hail, a rarer and more localized hazard. The work implies that national meteorological services can extract substantial forecast gains by pooling historical archives and applying careful augmentation thresholds rather than immediately pursuing larger models. Future operational pipelines should therefore prioritize multi-decadal, quality-controlled radar reanalyses before investing in parameter scaling.