An Overview of Hail Prediction Using Random Forests during the 2018 Hazardous Weather Testbed Spring Forecast Experiment

Each year, the Center for Analysis and Prediction of Storms (CAPS) at the University of Oklahoma (OU) performs a set of real-time ensemble forecast experiments in support of the NOAA Hazardous Weather Testbed (HWT) Spring Forecast Experiment (SFE). During the 2018 HWT SFE, CAPS produced an ensemble of 28 WRF-ARW forecasts covering the continental United States at 3 km grid spacing using a variety of multi-moment microphysical schemes and PBL parameterizations. These 28 members comprised 3 sub-ensembles: a single-physics ensemble, a mixed-physics ensemble, and a stochastic physics ensemble.

The 10 members of the mixed-physics sub-ensemble (also known as the core members), as well as data from the 8 members of the High Resolution Ensemble Forecast version 2 (HREFv2) ensemble, were used as the input for machine learning models designed for hail prediction – specifically an object-based system using random forests, using Hagelslag (a publically-available Python package for object tracking, classification, and machine learning). Machine learning forecast systems were trained separately for the CAPS SSEF and HREFv2 members, using a set of 2D (x-y) storm and environmental variable fields.

An overview of the machine learning system used will be presented, and the results of hail forecasts produced using SSEF and HREFv2 data during the 2018 HWT SFE will be summarized and verified using both subjective and objective methods. Preliminary results suggest that while both the SSEF and HREFv2 forecasts suffered from over-prediction of hail, forecasts using data from the CAPS SSEF system exhibited this error to a lesser extent. Preliminary results also suggest that both the SSEF and HREFv2-based forecasts exhibit ability to discriminate between hail and non-hail severe weather hazards.