Severe Hail Prediction with Spatiotemporal Relational Data Mining Techniques

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner
Tuesday, 4 February 2014: 9:30 AM
Room C204 (The Georgia World Congress Center )
David John Gagne II, National Center for Atmospheric Research, Boulder, CO; and A. McGovern, J. A. Brotzge, and M. Xue

Severe hail, or spherical ice precipitation over 1 inch in diameter, has caused billions of dollars in damage to crops, buildings, automobiles, and aircraft. Accurate predictions of severe hail with enough lead time can allow people to mitigate some hail damage by sheltering themselves and their vehicles and by rerouting their aircraft. Current pinpoint forecasts of severe hail rely on detection of hail in existing storms with radar-based methods. Predictions beyond an hour are limited to probabilistic predictions over larger areas based on expected environmental conditions. This presentation explains a technique that could increase the spatial precision and accuracy of severe hail forecasts by incorporating output from an ensemble of storm scale numerical weather prediction models into a spatiotemporal relational data mining model that would produce probabilistic predictions of severe hail. This study uses the Center for the Analysis and Prediction of Storms 2013 4 km Storm Scale Ensemble Forecast as well as a 500 m sub-ensemble nest. The spatiotemporal relational framework represents the ensemble output as a network of storm objects connected by spatial relationships. A spatiotemporal relational random forest is trained from the storm networks to produce a probability of severe hail for each storm in each ensemble member. Then, a spatial probability distribution based on Gaussian mixture models is calculated from the positions of the hail reports near each storm. Characterizations of the data are presented, and the performance of the spatiotemporal relational data mining model is compared with the performance of established machine learning methods. Initial results show positive skill for both types of models with the spatiotemporal model significantly outperforming the traditional machine learning model. Significant variables include the updraft speed, spatial distance between storms, and precipitable water.