Utilizing Deep Neural Networks for Short-Term Tornado Prediction and Detection using Multi-Modal Datasets

Weather radar remains the primary tool for real-time tornado detection and warnings. While tornado warning performance is much-improved in recent decades, forecasters continue to grapple with high false alarm rates. The recent surge in machine learning and artificial intelligence research, both within the meteorological community and in general, has led to the development of novel algorithms for tornado detection and prediction, primarily leveraging radar data. Yet, the application of deep learning techniques in this domain remains underexplored. This research employs an enhanced version of the TorNet (https://github.com/mit-ll/tornet) database, incorporating high-resolution polarimetric Level-II WSR-88D radar data, GOES satellite observations, numerical weather prediction model outputs, lightning data, the National Centers for Environmental Information storm events database, and National Weather Service warnings. These comprehensive data sources are utilized to train and validate deep neural networks (DNNs) aimed at improving tornado detection and prediction. This study both expands the spectrum of data modalities for tornado detection and introduces prediction capabilities extending up to 15 minutes prior to tornadogenesis. An overview of the expanded TorNet dataset is presented, detailing the data curation and holdout strategies employed for evaluation. Comparative analyses between traditional methods and the DNN-based approach are discussed, with results showcased through standard performance metrics and selected case studies. Finally, a summary of ongoing research efforts and potential future applications for operational deployment is presented.