Feedback on Machine-Learning-Based Spatial Severe Weather Probabilities from the 2023 Hazardous Weather Testbed Spring Forecasting Experiment

Over the past year, machine-learning-based 30-minute severe hail, wind, and tornado spatial probabilities were developed using predictors from the Warn-on-Forecast System (WoFS) and spatiotemporally extrapolated ProbSevere Version 2 (PS2) objects. Early work has indicated that this product, known as WoFS-PHI, produces more skillful forecasts than comparable machine learning forecasts using either WoFS or PS2 predictors alone, at lead times from 30 minutes to 3 hours. Spatial probability products such as WoFS-PHI have the advantage of highlighting specific areas threatened by severe weather at specific times. However, spatial probabilities, by definition, require a spatial radius over which they are valid, and it is not always clear how best to set this radius to balance sharpness, probability of detection, and false alarm rate, especially since the “optimal” radius may change based on lead time, the specific forecast scenario, and user preference.

To that end, feedback on WoFS-PHI spatial hazard probabilities was sought during a next-day evaluation activity in the 2023 Hazardous Weather Testbed Spring Forecasting Experiment (HWT SFE). The primary goals of this activity were to: 1) obtain general feedback on the usefulness of WoFS-PHI for severe weather forecasting, and 2) determine the spatial radii (from 7.5 to 39 km) that participants most preferred as a function of lead time and WoFS initialization time. Preliminary results suggest that, overall, participants preferred the 15 and 30 km radii and were more likely to prefer smaller neighborhoods at earlier lead times and later WoFS initializations, when the uncertainty in the spatial placement of storms was lower. Overall, participants liked the spatial precision of WoFS-PHI and found the product most useful when storms were just beginning to initiate. However, many participants stated a desire to see higher-magnitude probabilities. This feedback will help inform future iterations of WoFS-PHI.