Impacts of Multiscale Predictors on Random Forest Based Probabilistic Forecasts of Severe Weather Hazards

Severe weather hazards such as damaging winds, large hail, and tornadoes, are dependent on meteorological features ranging from convective to synoptic scale. Although numerical weather prediction (NWP) models are capable of resolving the full range of spatial scales, severe weather typically occurs on still unresolvable scales. RF has been shown to produce skillful forecasts of severe weather whilst using NWP outputs as predictors. Previous studies utilizing RF for prediction of severe weather have typically used NWP predictors processed to the same resolution as the RF prediction. As such, this does not allow the RF to account for large scale structure of the predictors.

This research focuses on the impacts of including predictors defined over increasingly large spatial scales in the context of generating severe weather probabilistic forecasts using RF with NWP model predictors. The NWP predictors are obtained from the convection-allowing ensembles generated by the OU MAP lab during the 2018 and 2019 Hazardous Weather Testbed Spring Forecasting Experiments. The RF-based probabilities are generated for day 1 (12-12 UTC) and 4 hour (21-00 UTC) forecast lead times and several severe and significantly-severe weather hazards including wind, hail, and tornadoes. Two experiments were conducted. A control experiment (CTL) was first performed utilizing the established method of remapping both the RF model input and output variables onto an 80km grid. A set of experiments (EXP) were then introduced and evaluated, including predictors that directly account for features over a range of spatial scales. In particular, the multi-scale predictors were calculated by taking the average or the maximum of each predictor over increasingly large spatial scales.

Results show a general increase in skill in the EXP forecasts over the CTL forecasts for many predictands. The most statistically significant increases are found for the day 1 forecast lead time any severe and wind and for the 4 hour forecast lead time significant wind, significant tornado, and any significant severe weather predictands.

Further diagnostics of the multi-scale predictors, such as tree interpreter (TI) and case studies, are ongoing to gain insight into how and why the multi-scale predictors are important and the physical meaning of the key mutli-scale forecast predictors.