Spatiotemporal Tornadic Prediction With Recurrent Neural Networks

Tornadoes are a dangerous weather phenomenon that causes an average of 80 fatalities and millions of dollars in damages yearly in the United States (NCEI, 2024; III, 2023). Forecasting these events is extremely challenging due to the complex atmospheric interactions that lead to the rapid development and dissipation of tornadoes. Due to the speed at which tornadoes can form and decay, operational meteorologists often rely on real-time weather radar analysis to predict the location and timing of tornadoes. To aid forecasters, the National Weather Service (NWS) utilizes radar-based algorithms such as the Mesocyclone Detection Algorithm (MDA) (Stumpf et al. 1995; Stumpf et al. 1998) and Tornado Detection Algorithm (TDA) (Mitchell et al., 1998) to determine if a thunderstorm has the potential to produce a tornado based on predefined threshold values. With the aid of these algorithms, the NWS typically provides around nine minutes of advance warning before a tornado occurs and achieves a probability of detection (POD) of approximately 62% (0.62) (Bentley et al., 2021; Rice and Woolverton, 2021). However, these algorithms are also known for their high false alarm ratios (FAR), averaging 70-75% (0.70-0.75) (Lim et al., 2019; Rice and Woolverton, 2021).

Researchers are exploring prediction with machine learning and deep learning models to improve forecasting capabilities. Marzban et al. (1996) demonstrated that a simple feed-forward neural network outperformed the MDA for predicting tornado occurrences, while Lagerquist et al. (2020) developed a Convolutional Neural Network (CNN) model that produced superior results in terms of FAR when compared to the NWS's average FAR. However, research gaps exist in leveraging Recurrent Neural Networks (RNN) and attention-based models for tornado prediction. This research addresses this gap by developing attention and non-attention-based RNN variants for predicting tornadic events. These deep learning models ingest a time series of 0.5-degree base-reflectivity and spectrum width images and produce the probability of tornadic occurrence for each pixel in a time series of images. The radar data is derived from the GridRad-Severe dataset (University of Oklahoma, 2021), and geolocated to tornadic events found in the GridRad-Severe dataset and historical tornadic events from the Storm Prediction Center (2023). Preliminary modeling indicates that these models can predict the location of tornadic events 10 to 25 minutes in advance at a 0.0208-degree resolution and achieve FAR values comparable to the NWS. Ongoing model tuning will likely lead to lower FARs, suggesting that the models could aid forecasters in making more precise forecasts with longer lead times, increasing their effectiveness in providing warnings to the public of imminent dangers.

Bentley, E. S., R. L. Thompson, B. R. Bowers, J. G. Gibbs, and S. E. Nelson, 2021: An analysis of 2016–18 tornadoes and National Weather Service tornado warnings across the contiguous United States. Wea. Forecasting, 36, 1909–1924.

Insurance Information Institute (III), 2023: Facts + Statistics: Tornadoes and thunderstorms. Available online at www.iii.org/fact-statistic/facts-statistics-tornadoes-and-thunderstorms. Accessed 18 May 2023.

Lagerquist, R., A. McGovern, C. R. Homeyer, D. J. Gagne II, and T. Smith, 2020: Deep learning on three-dimensional multiscale data for next-hour tornado prediction. Mon. Wea. Rev., 148, 2837–2861.

Lim, J. R., B. F. Liu, and M. Egnoto, 2019: Cry wolf effect? Evaluating the impact of false alarms on public responses to tornado alerts in the southeastern United States. Wea. Climate Soc., 11, 549–563.

Marzban, C., and G. J. Stumpf, 1996: A neural network for tornado prediction based on Doppler radar-derived attributes. J. Appl. Meteor. Climatol., 35, 617–626.

Mitchell, E. D., S. V. Vasiloff, G. J. Stumpf, A. Witt, M. D. Eilts, J. T. Johnson, and K. W. Thomas, 1998: The national severe storms laboratory tornado detection algorithm. Wea. Forecasting, 13, 352–366.

National Center for Environmental Intelligence (NCEI), 2023: U.S. tornadoes. Available online at https://www.ncei.noaa.gov/access/monitoring/tornadoes/12/1?fatalities=true. Accessed 18 May 2023.

Rice, D., and P. Woolverton, 2021: Our tornado warning system has come a long way, but it could be better. USA Today. Available online at https://www.usatoday.com/in-depth/news/nation/2021/06/16/tornado-warning-system-has-come-long-way-but-could-better/5152104001/. Accessed Jun 2021.

School of Meteorology, University of Oklahoma, 2021: GridRad-Severe - Three-Dimensional Gridded NEXRAD WSR-88D Radar Data for Severe Events. Research Data Archive at the National Center for Atmospheric Research, Computational and Information Systems Laboratory, Boulder, CO. Available online at https://doi.org/10.5065/2B46-1A97.

Storm Prediction Center, 2023: Storm prediction center maps, graphics, and Data Page. National Oceanic and Atmospheric Administration. Available online at https://www.spc.noaa.gov/wcm/data. Accessed Apr 2023.

Stumpf, G. J., C. Marzban, and E. N. Rasmussen, 1995: The new NSSL Mesocyclone Detection Algorithm- A paradigm shift in the understanding of storm-scale circulation detection. 27th Conf. on Radar Meteorology, Vail, CO, Amer. Meteor. Soc., 208–210.

Stumpf, G. J., A. Witt, E. D. Mitchell, P. L. Spencer, J. T. Johnson, M. D. Eilts, K. W. Thomas, and D. W. Burgess, 1998: The National Severe Storms Laboratory mesocyclone detection algorithm for the WSR-88D. Wea. Forecasting, 13, 304–326.