Predicting Tornadoes: Deep Learning on ERA5 Proximity Soundings

Tornadoes are exceedingly rare events, and the environmental conditions that result in tornadogenesis are highly specific, involving a complex interplay between temperature, humidity, and wind within a 3-dimensional space. Right now, Deep Learning models are at the forefront of storm prediction, and this paper explores the predictive capabilities of a 1D Convolutional Neural Network with batch normalization and dropout layers on hourly vertical atmospheric profiles obtained from ERA5 reanalysis (2011-2020) across the Southeast region of the United States. To address the extreme imbalance, we apply oversampling and weighted binary cross-entropy during training. Model performance is evaluated using precision, recall, F1, and ROC-AUC against storm reports obtained from the NOAA Storm Prediction Center. Results show that the model achieves moderately high recall, but low precision, and struggles to identify rare tornadic environments.