A random-forest model to assess predictor importance and nowcast severe storms using high-resolution radar goes satellite lightning observations

Few studies have assessed combined satellite, lightning, and radar databases to diagnose severe storm potential. The research goal here is to evaluate next-generation, 60-s update frequency geostationary satellite and lightning information with ground-based radar to isolate which variables, when used in concert, provide skillful discriminatory information for identifying severe (hail $ 2.5 cm in diameter, winds$ 25ms21, and tornadoes) versus nonsevere storms. The focus of this study is predicting severe thunderstorm and tornado warnings. A total of 2004 storms in 2014 15 were objectively tracked with 49 potential predictor fields related to May, daytime Great Plains convective storms. All storms occurred when 1-min Geostationary Operational Environmental Satellite (GOES)-14 super rapid scan data were available. The study used three importance methods to assess predictor importance related to severe warnings and used random forests to provide a model and skill evaluation measuring the ability to predict severe storms. Three predictor importance methods show that GOES mesoscale atmospheric-motion-vector-derived cloud-Top divergence and above-Anvil cirrus plume presence provide the most satellite-based discriminatory power for diagnosing severe warnings. Other important fields include Earth Networks Total Lightning flash density, GOES estimated cloud-Top vorticity, and overshooting-Top presence. Severe warning predictions are significantly improved at the 95% confidence level when a few important satellite and lightning fields are combined with radar fields, versus when only radar data are used in the random-forest model. This study provides a basis for including satellite and lightning fields within machine-learning models to help forecast severe weather.