Real-Time Hail Prediction Using Machine Learning Algorithms and HREFv2

Severe hail is a weather hazard that, on average, results in more than a billion dollars of damage within the continental United States each year. Thus, timely and accurate operational hail forecasts are vital in allowing the public to take action.

Currently, multiple operational models and ensembles predict hail (and other severe weather hazards) using a variety of microphysical schemes and methods for parameterizing or simulating hail growth. This abundance of data can lead to cognitive overload for forecasters attempting to predict hail and other severe weather hazards. Machine learning (ML) can help to address this issue, synthesizing multiple datasets to obtain optimal predictions for severe weather hazards, including severe hail.

In this study, hail is predicted using classification ML models, with input from the High-Resolution Ensemble Forecast version 2 (HREFv2). HREFv2 is an operational version of the Storm Prediction Center (SPC) Storm-Scale Ensemble of Opportunity (SSEO), and consists of an eight-member ensemble that uses a mixture of WRF-ARW and NMMB members with varying configurations. Twenty-seven environmental and storm variables, determined via preliminary testing to be optimal for predicting severe and significant severe hail, are used from each ensemble member as inputs for a random forest machine learning algorithm for hail prediction. Maximum Estimated Size of Hail (MESH), a Multi-Radar Multi-Sensor (MRMS) product, was used as an observation dataset to train and verify the ML model.

Preliminary results suggest that the ML hail forecasts skillfully predict severe hail over spatially similar regions, as compared to observed hail from SPC hail reports. Results from analyses of ML model output to further investigate model performance and methods to further improve forecast skill will also be presented.