3.3 Short-Term Tornado Prediction via Deep Learning on 3D Multiscale Data

Tuesday, 14 January 2020: 2:00 PM
258B (Boston Convention and Exhibition Center)
Ryan A. Lagerquist, CIMMS, Norman, OK; and A. McGovern, C. R. Homeyer, D. J. Gagne II, and T. M. Smith

Machine learning (ML) has been used successfully in the atmospheric sciences for decades, including tornado prediction (Marzban and Stumpf, 1996; Cintineo et al., 2014, 2018). Deep learning is a subset of ML with the ability to learn directly from spatiotemporal grids, such as radar images, without the need to pre-process into scalar features such as principal components. We use convolutional neural networks (CNN), a flavour of deep-learning model, to predict next-hour tornado occurrence on a storm-by-storm basis. Thus, each input to the model is a pre-existing storm; the output is a probability. Predictors include a storm-centered radar image and proximity sounding. The proximity sounding comes from the Rapid Refresh model (or Rapid Update Cycle for times before May 2012), while the radar image comes from either the Multi-year Reanalysis of Remotely Sensed Storms (Ortega et al., 2012) or GridRad (Homeyer and Bowman, 2017). Both are quality-controlled composites of WSR-88D data over the continental United States. We train separate models on the MYRORSS and GridRad data.
We have achieved considerable skill on testing data (independent of data used for training and model selection): an AUC (area under the ROC curve) of 0.96 and CSI (critical success index) of 0.08. CSI is low due to the extreme rarity of the event (only 0.1% of storms are tornadic in the next hour). We have made efforts to address the rare-event issue by eliminating "trivial null cases" before machine learning, and we will share insights from these experiments. Also, we will compare performance of the MYRORSS and GridRad models on a single testing year while highlighting the failure modes (storm structures, evolutions, and surrounding environments that lead to egregious misses and false alarms).
Cintineo, J., M. Pavolonis, J. Sieglaff, and D. Lindsey, 2014: "An empirical model for assessing the severe weather potential of developing convection." Weather and Forecasting, 29 (3), 639–653.
Cintineo, J., and Coauthors, 2018: "An empirical model for assessing the severe weather potential of developing convection." Weather and Forecasting, 33 (1), 331–345.
Homeyer, C., and K. Bowman, 2017: "Algorithm Description Document for Version 3.1 of the Three-Dimensional Gridded NEXRAD WSR-88D Radar (GridRad) Dataset." Tech. rep., University of Oklahoma. URL http://gridrad.org/pdf/GridRad-v3.1-Algorithm-Description.pdf.
Marzban, C., and G. Stumpf, 1996: "A neural network for tornado prediction based on Doppler radar-derived attributes." Journal of Applied Meteorology, 35 (5), 617–626.
Ortega, K., T. Smith, J. Zhang, C. Langston, Y. Qi, S. Stevens, and J. Tate, 2012: "The Multi-year Reanalysis of Remotely Sensed Storms (MYRORSS) project." Conference on Severe Local Storms, Nashville, Tennessee, American Meteorological Society.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner