MYRORSS Hail Detection and Size Classification Models

This repository contains the trained machine-learning models, model-generated outputs, and supporting materials developed for an object-based hail classification framework using the Multi-Year Reanalysis of Remotely Sensed Storms (MYRORSS) storm-cluster dataset for 2005–2011. The framework consists of two stages. Stage 1 uses an XGBoost classifier to identify hail-producing storm-cluster snapshots. Stage 2 uses a Random Forest classifier to classify detected hail-producing clusters into non-severe, severe, and significant-severe hail categories. The repository includes: trained model files for Stage 1 and Stage 2; model configuration files and parameter settings; the final input-feature lists required by each model; model-generated hail detection and hail-size classification outputs; and a README with instructions for loading, running, and applying the models. The complete MYRORSS predictor namelist and the underlying storm-cluster dataset are provided through the companion tornado-model repository and are not duplicated here: MYRORSS Storm Cluster Dataset for Tornado Detection and Climatology (2005–2011) DOI: 10.5281/zenodo.19644585 Users seeking to reproduce or rerun the hail models should obtain the storm-cluster data and full variable documentation from the companion repository and prepare the input data according to the feature names, ordering, preprocessing requirements, and file structure described in the README. The models use radar-derived storm characteristics and near-storm environmental variables from the Rapid Update Cycle analysis. Geographic coordinates and population density were excluded from model training to reduce direct dependence on spatial reporting biases. The model outputs provide object-based hail predictions for MYRORSS storm-cluster snapshots, including Stage 1 hail-detection results and Stage 2 hail-size classifications. These outputs are intended to support reproducibility, hail climatology research, storm-scale hazard analysis, model evaluation, and comparison with report-based or radar-derived hail datasets. Users should consider several limitations. The training labels are based on reported hail events and may be affected by population and reporting biases. Radar-derived predictors may be influenced by radar range, beam geometry, terrain blockage, sampling resolution, and changes in radar processing or observing systems. Near-storm environmental predictors are also subject to uncertainty in the underlying numerical analyses. Individual 5-minute storm-cluster snapshots are treated independently, so the framework does not explicitly represent complete storm lifecycles or hail-swath evolution. Rare significant-severe hail events are less well represented and are more difficult to classify reliably than more common hail categories. Application of the models to datasets outside the original MYRORSS period, domain, or processing framework should be approached cautiously because differences in radar processing, predictor definitions, observing systems, environmental analyses, and input-data distributions may affect model performance. Please cite the associated hail manuscript, this Zenodo record, and the companion tornado-model repository when using the models, outputs, or supporting data products.