Optimal Selection of Predictor Variables in Statistical Downscaling Models of Precipitation

Abstract

Two screening methods aimed at selection of predictor variables for use in a statistical downscaling (SD) model developed for precipitation are proposed and evaluated in this study. The SD model developed in this study relies heavily on appropriate predictors chosen and accurate relationships between site-specific predictand (i.e. precipitation) and general circulation model (GCM)-scale predictors for providing future projections at different spatial and temporal scales. Methods to characterize these relationships via rigid and flexible functional forms of relationships using mixed integer nonlinear programming (MINLP) formulation with binary variables, and artificial neural network (ANN) methods respectively are developed and evaluated in this study. The proposed methods and three additional methods based on the correlations between predictors and predictand, stepwise regression (SWR) and principal component analysis (PCA) are evaluated in this study. The screening methods are evaluated by employing them in conjunction with an SD model at 22 rain gauge locations in south Florida, USA. The predictor variables that are selected by different predictor selection methods are used in a statistical downscaling model developed in this study to downscale precipitation at a monthly temporal scale. Results suggest that optimal selection of variables using MINLP and ANN provided improved performance and error measures compared to two other models that did not use these methods for screening the variables. Results from application and evaluations of screening methods indicate improved downscaling of precipitation is possible by SD models when an optimal set of predictors are used and the selection of the predictors is site-specific.