Evaluating Machine Learning Updraft Retrieval Skill: Convective Storm Case Studies

Severe weather hazards (e.g., hail, flooding, tornadoes) from convective storms can threaten lives and damage property. These hazards have also been linked to storm updraft characteristics. Despite the strong relationship between updrafts and extreme weather conditions, forecasters have few tools to assess the strength and size of convective updrafts in real time. To fill the dearth of operational tools that characterize updrafts in real time, a recent study trained a machine learning (ML) model (i.e., U-Network) to translate radar reflectivity to maps of updrafts with promising results. The ML model successfully predicted the updraft area for a Colorado supercell as compared to observed updrafts using dual-Doppler retrievals.

This study extends the previous work by focusing on new cases with multi-Doppler data as references, aiming to test how well the ML model generalizes and to understand updraft characteristics and their relationships with severe weather hazards. One of the new cases is a tornadic supercell that occurred in Oklahoma on 20 May 2019. Updraft intensity and area metrics from the ML model are compared with those from the multi-Doppler data. Our analysis of the Oklahoma supercell yields encouraging results for the ML model’s predictive skill. For example, the ML updraft area correlates well with the multi-Doppler retrievals of updraft area (r ≥ 0.8). Importantly, both multi-Doppler updraft area and ML updraft area correlate well (r > 0.5) with severe hail area, demonstrating the potential utility of the ML model for forecasting severe weather hazards. Updraft intensity metrics have less predictive skill than updraft area metrics. Additional convective modes (i.e., linear convective systems, tropical cyclones) are also explored to better assess the ML model’s skill for different types of convective storms. Overall, our results demonstrate that ML outputs could be useful to operational meteorologists for predicting certain updraft characteristics and updraft-related severe weather hazards from radar reflectivity in real time.