“Dendrology” in Numerical Weather Prediction: What Random Forests and Logistic Regression Tell Us about Forecasting Extreme Precipitation

Abstract

Three different statistical algorithms are applied to forecast locally extreme precipitation across the contiguous United States (CONUS) as quantified by 1- and 10-yr average recurrence interval (ARI) exceedances for 1200–1200 UTC forecasts spanning forecast hours 36–60 and 60–84, denoted, respectively, day 2 and day 3. Predictors come from nearly 11 years of reforecasts from NOAA’s Second-Generation Global Ensemble Forecast System Reforecast (GEFS/R) model and derive from a variety of thermodynamic and kinematic variables that characterize the meteorological regime in addition to the quantitative precipitation forecast (QPF) output from the ensemble. In addition to encompassing nine different atmospheric fields, predictors also vary in space and time relative to the forecast point. Distinct models are trained for eight different hydrometeorologically cohesive regions of the CONUS. One algorithm supplies the GEFS/R predictors directly to a random forest (RF) procedure to produce extreme precipitation forecasts; the second also employs RFs, but the predictors instead undergo principal component analysis (PCA), and extracted leading components are supplied to the RF. In the last algorithm, dimension-reduced predictors are supplied to a logistic regression (LR) algorithm instead of an RF. A companion paper investigated the quality of the forecasts produced by these models and other RF-based forecast models. This study is an extension of that work and explores the internals of these trained models and what physical and statistical insights they reveal about forecasting extreme precipitation from a global, convection-parameterized model.