A Tornado Detection Algorithm using Deep Neural Networks, Full-Resolution Polarimetric Weather Radar Data, and Explainable AI

Weather radar is the primary tool used by forecasters to detect and warn for tornadoes in near-real time. While forecasters historically provide excellent warning performance, false alarm rates for tornado warnings remain relatively high. Particularly in cases of widespread convection, “data overload” can become an issue, leading to the desire to alert forecasters to potential storms of interest. This has led to multiple previous tornado detection algorithms in the literature and in operations, including the Weather Surveillance Radar – 1988 Doppler (WSR-88D) tornadic vortex signature (Brown and Wood 2012) and tornado detection algorithm (Mitchell et al. 1998), ProbSevere (Cintineo et al. 2020), and, most recently, the tornado probability algorithm (Sandmæl et al. 2023). While these approaches have continuously improved tornado detection capabilities, applications of image-based deep learning (DL) are yet to be extensively explored for the tornado detection problem.

In this study, a new database of full-resolution, polarimetric WSR-88D data spanning 2013-2022 is combined with the National Centers for Environmental Information (NCEI) tornado database, National Weather Service (NWS) warnings, and a plethora of metadata to train and evaluate a deep neural network (DNN) for tornado detection. The approach of using full-resolution Level-II polarimetric data is novel in this research space, and results have shown significant promise for improvement over previous methods. A baseline of performance is first evaluated with the improved dataset by comparing non-DL methods such as regression, random forests, and support vector machines, showing similar results to previous techniques. The dataset is then used to train a DNN using 12 modalities, including the lowest two tilts of horizontal reflectivity factor, radial velocity, spectrum width, differential reflectivity, co-polar correlation coefficient, and specific differential phase. Comparisons to the non-DL baselines are made and discussed, as well as the critical applicability to future tornado prediction methods using the same dataset. A specific focus is given to applications of eXplainable AI (XAI) in order to determine the specific areas of interest to the DNN. Using this approach, we can determine exactly which signatures across the 12 modalities are of most importance to the model, resulting in the ability to verify performance. Finally, the applicability of XAI to a future tornado prediction algorithm and discovery of tornadic precursors combined across multiple modalities is discussed.