A Machine Learning Based High-Resolution Large Hail Climatology for the Contiguous United States

Over the past decade, hail has been responsible for most financial losses associated with severe convective storms, with costs steadily increasing. As the population grows and increasingly invests in vulnerable infrastructure, the risk of economic damage from large hail also rises. Accurate hazard assessment is therefore essential for effective mitigation.

Report-based hail climatologies are limited by observational biases, resulting in large uncertainties, particularly in sparsely populated areas. While recent machine learning approaches have enabled the development of hail climatologies for hazard assessment, many existing datasets remain limited by regional biases and coarse spatial and/or temporal resolution.

Here, we present a novel, high-resolution large hail (> 2.5 cm in diameter) hazard dataset on a 4 km × 4 km, half-hourly grid over the contiguous United States (CONUS), ranging from 2000 to 2022. This unprecedented spatiotemporal resolution is enabled by integrating hail reports with multiple high-resolution remotely sensed hail-proxy observations, including radar reflectivity, satellite-derived brightness temperature, and total ice water path, together with key hail-relevant environmental parameters from the ERA5 reanalysis. The large hail hazard model, which we trained to produce the dataset, is based on the gradient-boosted decision tree algorithm XGBoost and provides increased spatial and temporal detail relative to prior large hail climatologies.

When evaluated on held-out test data, the model accurately reproduces the interannual, seasonal, and diurnal cycles of large hail. It resolves fine-scale topographic influences and captures coherent hail tracks. Performance is strongest for typical events, while rarer, atypical cases present a trade-off between improved detection and increased false positives.

The resulting dataset provides a high-resolution basis for large hail risk assessment across the CONUS. Because it relies on globally available satellite observations and reanalysis, the framework is transferable to other regions and can be applied to kilometer-scale weather and climate model output.