Using Machine Learning to Predict Tornadoes Based on Sounding and Reanalysis Data

Severe weather events such as tornadoes and high-wind thunderstorms pose a significant threat to human life and well-being. Despite these impacts, tornado forecasting remains a challenge for meteorologists. Most forecast methods rely on an analysis of the instability of an air column, often through indices such as convective available potential energy (CAPE), convective inhibition (CIN), and bulk Richardson number (BRN), as well as their derivatives (ERH, STP, etc). However, few studies have analyzed the long-term efficacy of these indices for severe weather and tornado forecasting. We analyze the predictive capacities of CAPE, CIN, and BRN for severe weather and tornado occurrence between 1979 and 2014. Index values are derived from 37 sounding stations across the CONUS with continuous data collection for the time period and, independently, from the North American Regional Reanalysis (NARR). Severe weather and tornado records are taken from NOAA’s Storm Events dataset. We find that while both CAPE and CIN are significant predictors for the occurrence of all severe weather event types, only CAPE is significant for tornado occurrence. The use of NARR data further in our investigation is validated through a comparison with sounding data. We present the results of the development of a tornado prediction index derived empirically through a machine learning algorithm trained on NARR historical data. Using NARR allows us to consider more accurate locations and timing, as well as a wider range of atmospheric parameters that could impact tornado occurrence. Our analysis suggests that such an index could present an improvement in the field of tornado prediction. Preliminary estimates using a logistic regression show significant improvement in model accuracy when sounding data is replaced by NARR, allowing for true positive prediction rates up to at least 85%. We investigate the application of traditional neural networks.