Sunday, 28 January 2024
Hall E (The Baltimore Convention Center)
The ability to predict severe weather events over the United States diminishes beyond a three-day lead time. However, increasing lead times for severe weather hazards is vital for those such as emergency managers and stakeholders, as it enables them to better prepare for potential emergencies. In this study, machine learning techniques are utilized to better understand and extend the predictability of severe weather events across various regions of the United States. First, a convolutional neural network (CNN) is constructed and trained on practically perfect severe weather hindcasts, Climate Data Record tropical outgoing longwave radiation anomalies, and ERA-5 synoptic and mesoscale data. The CNN is used to predict the likelihood of severe weather events such as tornadoes, severe wind, and hail occurring over the U.S. at timescales varying from one to three weeks. Subsequently, through the application of explainable AI (XAI), we identify crucial features and locations within the input data that have significant impact on the model's predictions. The application of various XAI techniques lend insight into the processes and teleconnections strongly tied to improved predictability of severe weather.

