Identifying Dominant Southeastern United States Warm-Season Precipitation Features from Numerical Weather Prediction Forecasts

The Southeastern United States agriculture sector is heavily reliant on water resources generated from warm-season precipitation. However, the highly chaotic convective nature of warm-season precipitation creates immense challenges in forecasting its occurrence and magnitude at a local level as numerical weather prediction guidance remains insufficient in resolving convective processes. This study will employ machine learning techniques to identify warm-season rainfall predictors that exist within a database of simulated days for the Southeastern United States. Numerical weather prediction forecasts of daily precipitation were obtained for all warm season (June 1 – August 31) days from 2014 – 2020. Numerous predictors, including 850-700 mb lapse rate, moisture convergence, lifted index, 925 mb dewpoint depression, CAPE, CIN, etc., were derived from output from each simulation and used to formulate a kernel principal component analysis (KPCA) that identified predictors with high degrees of covariability and subsequent KPC loadings that effectively categorize precipitation and no-precipitation locales. Predictors that are most strongly related to the best separating KPC loadings were retained at each gridpoint in the study region. The predictor sets were then run through a genetic algorithm (GA) to identify an optimal combination of forecast parameters at each gridpoint within the simulation domain. Predictors whose frequency of return from the GA that were most prevalent were evaluated for their physical relevance to warm-season precipitation and retained as a “best-set” of warm-season predictors. These predictors will be used in future work to develop a gridded machine-learning classification model that forecasts daily warm-season rainfall for the Southeastern United States.