Convective Meteorology (Mesoscale Dynamics)

Day-Ahead Severe Hail Forecasting with Machine Learning and Convection-Allowing Model Ensembles

David John Gagne

School of Meteorology

30 October 2015, 3:00 PM

National Weather Center, Room 5600
120 David L. Boren Blvd.
University of Oklahoma
Norman, OK

Convection-Allowing Models (CAMs) have shown great value in predicting the location, evolution, and mode of severe thunderstorms; but they have a limited ability to infer the likelihood and intensity of severe hazards such as tornadoes and hail. Current displays of CAM output utilize storm proxies, such as updraft helicity, or simplified physical models, such as HAILCAST, to infer severe hazard potential. These methods have shown correlation with severe hazard intensity but are not calibrated to observations and are sensitive to model configuration. The goal of this seminar is to demonstrate how a track-based machine learning approach to severe hail forecasting can add value to current CAM ensemble forecasts by integrating spatiotemporal model output values into calibrated, probabilistic hail size forecasts.

Forecasts were made using the 2014-2015 CAPS Storm-Scale Ensemble Forecast system and the 2015 NCAR CAM Ensemble. Hail size observations were derived from the NOAA NSSL Multi-Radar Multi-Sensor radar mosaic Maximum Expected Size of Hail (MESH) product. Potential hailstorm tracks in each ensemble member are identified and matched with observed hailstorm tracks, and storm information is collected. Machine learning models are trained to predict hail occurrence and a bulk hail size distribution. Neighborhood ensemble probabilities are derived from the machine learning forecasts and compared with similar output from storm-proxy variables and physics-based approaches. The machine learning methods are able to reduce the False Alarm Ratio with minimal impact to Probability of Detection. The calibrated hail size distributions can also infer the probability of extreme hail events. In addition to hail forecasting, the hail track information provides insights into the impact of physics parameterizations on modeled storms. Applications of similar machine learning techniques to other severe weather prediction problems will also be discussed.

Convective Meteorology (Mesoscale Dynamics) Seminar Series website