ProbSevere Version 3: Improved Exploitation of Data Fusion and Machine Learning for Nowcasting Severe Weather

Abstract National Oceanic and Atmospheric Administration (NOAA)/Cooperative Institute for Meteorological Satellite Studies (CIMSS) ProbSevere, or the “probability of severe” models, provide next-hour probabilistic guidance on severe weather in the United States (large hail, strong wind gusts, and tornadoes) using machine learning and meteorological data from remotely sensed platforms and short-term numerical weather models. ProbSevere has been widely used by NOAA’s National Weather Service (NWS) since 2016. ProbSevere version 3 employs multiplatform, multiscale storm object identification and tracking to collect storm features and then uses gradient boosting decision trees to make predictions of convective hazards from the extracted predictors. ProbSevere v3 was trained, validated, and tested on data from 2018 to 2023. Results demonstrate improved performance over ProbSevere v2 for the majority of the United States across all seasons, spanning a wide range of meteorological regimes. The greatest improvements are for the hail, wind, and any severe models, whereas the tornado model demonstrated modest improvement. ProbSevere v3 models were evaluated by NWS forecasters in simulated real-time experiments from 2021 to 2023. The majority of forecasters preferred ProbSevere v3 over v2 due to improved confidence in predictions and greater lead times for the issuance of severe weather warnings. An analysis of predictor importance revealed several radar, lightning, satellite, and numerical model fields that strongly contribute to the model predictions, which is consistent with past research and knowledge of severe thunderstorm forecasting. This reinforces the importance of environmental data fusion for nowcasting severe weather associated with midlatitude convection. Significance Statement This study details the updates made to a U.S.-wide, real-time, probabilistic severe-weather guidance system aimed to improve forecaster confidence and the accuracy of the U.S. National Weather Service severe weather warnings. An evaluation of the system demonstrated improved performance over the operational predecessor, which was driven by radar, ground-based lightning, satellite, and numerical model data sources.