Hail Size Prediction Using Machine Learning Techniques Applied to Storm Scale Ensembles

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner
Tuesday, 4 November 2014
Capitol Ballroom AB (Madison Concourse Hotel)
David John Gagne II, CAPS/Univ. of Oklahoma, Norman, OK; and A. McGovern, J. Brotzge, J. Correia Jr., M. C. Coniglio, and M. Xue

Handout (1.5 MB)

Ensembles of convection-allowing numerical weather prediction models have enabled forecasters to predict the likelihood and characteristics of severe thunderstorms hours and days in advance. While these convection-allowing models can partially resolve individual storms, they cannot resolve the major hazards from severe storms, which include tornadoes and large hail. Although these severe hazards cannot be directly resolved, their presence and strength can be inferred using techniques that associate model output with observed severe events. This study examines machine-learning-based techniques to estimate hail size and compares their skill with a physics-based column hail growth model. The study utilizes model output from the 2013 and 2014 Center for Analysis and Prediction of Storms Storm-Scale Ensemble Forecast system for training, evaluating, and testing the different hail forecasting approaches. Gridded radar-estimated Maximum Expected Size of Hail comes from the NOAA National Severe Storms Laboratory Multi-Radar Multi-Sensor System. For each forecast hour, local maxima in the 1-hour maximum column graupel field are identified using the enhanced watershed image segmentation technique. Statistics describing the storm strength and environmental conditions are extracted from each identified region. Each region is then matched with nearby hail size maxima. Gradient boosting regression trees, random forest, and ridge regression are trained to associate the model variables from each ensemble member with the nearby observed hail size. The size and spatial skill of those algorithms are compared with the HAILCAST column model. Both the gradient boosting regression trees and HAILCAST were run in real-time with the 2014 NOAA Hazardous Weather Testbed Experimental Forecast Program. Verification statistics show that the forecast size errors from each machine learning algorithm are similar, and that all approaches outperformed the HAILCAST algorithm. Comparisons of neighborhood ensemble probability from each approach showed that the machine learning algorithms provided reliable probability forecasts while HAILCAST tended to overforecast the probability of severe hail. The machine learning approaches generally produced a lower false alarm ratio for a given probability of detection. These results showcase the potential improvements to severe hail forecasts that machine-learning-based approaches can provide.