Predicting 12-hour Storm Reports Using Random Forest Classification

Convective outlooks issued by the Storm Prediction Center (SPC) attempt to predict the probability that a given severe weather hazard (i.e., tornado, large hail, or damaging winds) will occur within 25 miles of a point location. Similarly, tornado and severe thunderstorm watches provide end users with probabilities that a severe hazard of a certain intensity will occur within the watch boundaries. While verification has proven the SPC probability forecasts to be reliable overall, there is currently a lack of probabilistic guidance that explicitly forecasts the occurrence of the local storm reports (LSRs) used to verify these severe weather products. It is the goal of this study to utilize historic sounding data to derive an empirical relationship between various atmospheric variables and the probability that an LSR will be received within a county warning area (CWA). Probabilities of LSR occurrence are generated by a random forest classification algorithm, and the model performance is evaluated using LSRs and severe warnings for verification. Finally, the most important variables in the model are broken down by season and analyzed.