NEWS-e is designed to provide ensemble predictions of individual thunderstorms and their hazards, which is well suited to feature-based verification techniques that allow forecast accuracy to be examined on a storm-by-storm basis. Application of feature-based verification of 2016/17 NEWS-e forecasts has resulted in a dataset of approximately 2.5 million predicted thunderstorm objects and 700,000 rotation track objects. Each object is classified as matched or unmatched based on its spatiotemporal distance to the nearest observed object in Doppler radar data.
The current study will extract environmental information from the inflow of each thunderstorm and rotation track object, then use a random forest to assess the sensitivity of simulated thunderstorm objects to their upstream environment. Specifically, the ability of a random forest to rank individual variables according to their relative value in determining the outcome of each decision tree will be leveraged to identify environmental conditions that produce the strongest simulated mesocyclones and those that most often result in an accurate prediction. Additionally, the regional dependence of the environment on simulated thunderstorm objects will be examined by comparing results from cases over the Great Plains to those produced during the cool season in the southeastern United States as part of VORTEX-SE.