Statistical Downscaling of Precipitation Using Machine Learning with Optimal Predictor Selection

Abstract

Various methods have been proposed to downscale the coarse resolution general circulation model (GCM) climatological variables to the fine-scale regional variables; however, fewer studies have been focused on the selection of GCM predictors. Additionally, the results obtained from one downscaling technique may not be robust and the uncertainties related to the downscaling scheme are not realized. To address these issues, the writers employed independent component analysis (ICA) for predictor selection that determines spatially independent GCM variables. Cross-validation of the independent components is employed to find the predictor combination that describes the regional precipitation over the upper Willamette basin with minimum error. These climate variables, along with the observed precipitation, are used to calibrate three downscaling models: multilinear regression (MLR), support vector machine (SVM), and adaptive-network-based fuzzy inference system (ANFIS). The presented method incorporates several GCM grids in the downscaling process that allows considering more predictors in the model calibration and removes the predictors correlation and dependence by ICA. Also, the study uses several downscaling techniques to develop an ensemble of precipitation time series that can be used in hydrologic climate impact assessment. The performance assessment of the results indicates that the procedure is successful in choosing the predictors for downscaling the GCM data both in monthly and seasonal timescales. The study shows that by choosing proper predictors the MLR model is an efficient method for precipitation downscaling.