The Prediction of Potential Tornado Damage Intensity Using Machine Learning Artificial Intelligence for the Earth Systems

This study uses nine classification machine learning algorithms to examine their skill in making short-fused, storm-based predictions of significant or nonsignificant tornado damage intensity, conditioned upon tornadogenesis, using pretornadic mesocyclone characteristics and the near-storm environment. Radar predictors are from approximately 30 min before tornadogenesis, while environmental predictors are from the model-analysis hour nearest but before the time of tornadogenesis. The most-skilled classifiers are logistic regression, random forests, and gradient boosting as measured by each model’s cross-validated accuracy (≈89%), precision (≈93%), and recall (≈73%) and other binary classification metrics. Learning curves indicate adequate training of models, and calibration curves reveal the reliability of predicted probabilities, with random forests being the most reliable. Also, permutation tests demonstrate the statistical significance of the cross-validated model accuracy. Out of the four radar predictors included in this study, radar-derived pretornadic mesocyclone width and differential velocity are the most important over convective mode and distance from the radar, followed by environmental vertical wind shear and composite parameters. Specifically, wider and stronger pretornadic mesocyclones in environments characterized by larger values of vertical wind shear and composite parameters increase the likelihood of significant tornadoes. The model results could build forecaster confidence in the anticipation of tornado damage intensity and aid forecasters in making informed impact-based warning tag decisions. This could better protect life and property by providing a summary of data relevant to potential tornado damage rating before tornado formation. Important future work includes the addition of other radar-based predictors and the development of a more diverse and realistic sample of tornadic events. Grant no. NA17OAR4590195