A Deep Learning Approach to Severe Weather Subseasonal Forecasting over the United States

Severe weather is a prevalent phenomena in the United States, with hazards spanning droughts, cold air outbreaks, floods, strong winds, wildfires, hurricanes, and tornadoes.Various factors contribute to the heightened frequency of severe weather, some of which include the mid-latitude location of the U.S., topography, and both mesoscale and synoptic-scale processes. Fortunately, there have been huge strides in severe weather forecasting over the last 20 years. This is largely attributed to technological advancements, which enable forecasters to access new and higher quality weather data, while also accelerating data processing and analysis. However, even with recent technological advances, extending predictability for severe weather beyond a few days remains a persistent challenge. Previous work has shown promise for prediction at longer timescales through consideration of various teleconnections like the Madden Julian Oscillation and the El Niño-Southern Oscillation. In this work, we utilize deep learning to better understand and extend predictability of severe weather, specifically focusing on tornado and hail hazards at subseasonal timescales. Using a convolutional neural network (CNN), large-scale predictors like the mid-latitude jet stream, geopotential heights, and low-level temperature from ERA-5 reanalysis data, as well as tropical predictors like outgoing longwave radiation, are used to predict severe weather likelihood over the U.S. from Practically Perfect hindcast data. Subsequently, we apply explainable AI (XAI) to lend insight into the integral processes underpinning extended predictability. The integration of these deep learning techniques in a scientific discovery framework enhances our understanding of impactful teleconnections and opportunities for refined subseasonal forecasts.