Optimal Selection of Predictor Variables in Statistical Downscaling Models of Precipitation

Abstract:

Statistical downscaling models developed for precipitation rely heavily on predictors chosen and on accurate relationships between regional scale predictand and GCM-scale predictor for providing future precipitation projections at different spatial and temporal scales. This study provides two new screening methods for selection of predictor variables for use in downscaling methods based on predictand-predictors relationships. Methods to characterize predictand-predictors relationships via rigid and flexible functional relationships using mixed integer nonlinear programming (MINLP) model with binary variables and artificial neural network (ANN) models respectively are developed and evaluated in this study. In addition to these two methods, a stepwise regression (SWR) and two models that do not use any pre-screening of variables are also evaluated. A two-step process is used to downscale precipitation data with optimal selection of predictors and using them in a statistical downscaling model based on support vector machine (SVM) approach. Experiments with the proposed two new methods and three additional methods based on correlation between predictors and predictand and the other based on principal component analysis are evaluated in this study. Results suggest that optimal selection of variables using MINLP albeit with linear relationship and ANN method provided improved performance and error measures compared to two other models that did not use these methods for screening the variables. Of all the three screening methods tested in this study, SWR method selected the least number of variables and also ranked lowest based on several performance measures.