Improving the Anticipation of Potential Tornado Intensity Across Different Time Scales

Previous studies of tornadoes have focused on understanding and predicting tornadogenesis or diagnosing the intensity of an ongoing tornado. Given that the majority of damage and fatalities are caused by strong to violent tornadoes, there is a need for robust operational tools that focus on anticipating tornado intensity rather than simply on tornadogenesis or ongoing tornadoes. The anticipation of tornado intensity is operationally relevant across different time scales including during the pre-tornadic period of ongoing thunderstorms as well as during the tornado watch time scale several hours before storms develop. The first part of this work utilized the previously demonstrated robust relationship between the intensity of a tornado and its pre-tornadic mesocyclones characteristics as well as its near-storm environment to explore tornado-intensity prediction approaches through machine learning applications using a dataset of 300 tornadic events. Several classification machine learning algorithms were implemented and used to examine their skill in predicting significant or non-significant tornado intensity for a given storm. Logistic regression, random forests, and gradient boosting were found to be the most skilled classifiers as measured by several cross-validated binary classification metrics as well as performance diagrams. Feature importance and the decision-making process within each model was also explored to help reveal a more physical understanding of the model performance and results, as well as relationships between the predictors and tornado intensity including the establishment of critical thresholds of predictors. The predictors of radar-derived pre-tornadic mesocyclone width and differential velocity were the most important, followed by environmental vertical wind shear and composite parameters. The results demonstrate a skilled and reliable binary prediction of tornado intensity, conditioned upon tornadogenesis. Accordingly, they demonstrate the potential for these machine learning applications to become a helpful resource in an operational setting, allowing an operational forecaster to be aware of and communicate information about potential tornado intensity before a tornado forms to better protect life and property during ongoing thunderstorms. The second part of this work uses High-Resolution Rapid Refresh (HRRR) model forecasts of storm-scale diagnostics such as updraft helicity (UH), vertical vorticity, and the Okubo-Weiss number (OW) as well as other environmental parameters such as the significant tornado parameter (STP) prior to convection initiation to examine their skill in predicting whether a severe weather event will be associated with no tornadoes, non-significant tornadoes (EF0-1), or significant tornadoes (EF2+). Each event and the associated HRRR forecasts are sampled within Storm Prediction Center outlook regions while exploring different thresholds of surrogate diagnostics in combination with the large-scale environment. The presence of any skilled separation of surrogate diagnostics between non-tornadic events and both non-significantly and significantly tornadic events is explored. This separation is expected to be dependent upon the inclusion of environmental parameters as well as the thresholds of storm-scale diagnostics used. The diagnostics of 0-2 km UH, 0-1 km vertical vorticity, and near-surface OW are expected to exhibit the greatest skill. The ultimate goal of this collective work is to improve the understanding of and provide new tools for the anticipation of tornado intensity before tornadoes form, from the tornado watch time scale to the pre-tornadic stages of ongoing thunderstorms.