Optimal Selection of GCM Predictors Combined with a New Procedure for Statistical Downscaling of Precipitation

Abstract

Downscaling of General Circulation Models (GCM) outputs to a finer grid cell size is an important step in climate change impact and adaptation studies. Many investigations have been focused on presenting techniques to downscale GCM data utilizing statistical approaches. Nevertheless there is currently the need to present techniques on predictor selection and also to compare different downscaling models' capabilities. Hence in this study an algorithm has been developed to select GCM predictors in a subseasonal to seasonal time scale. Independent Component Analysis (ICA) was used to find statistically independent signals of CGCM3 variables in the 4 x 7 grid cells covering the Willamette river basin in Oregon, USA. Using the multi-linear regression cross validation (MLR-CV) the GCM predictors were selected for each period. The selected predictors were then applied to train the ANFIS (Adaptive Network-based Fuzzy Inference System) and the SVM (Support Vector Machine) models, and their performances were assessed on the test data. To design more robust networks that are less dependent on training data set, the cross validation was performed. Predictors with the best performance for each season in the test set (using both ANFIS and SVM models) were selected for that specific season. Employing the ICA allowed reducing the size of many dependent GCM variables in 28 grid cells considerably resulting in higher accuracy in downscaling and more effectiveness in the procedure. The uncertainties are then analyzed from the combination of 5 downscaling techniques including SVM, MLR and 3 different ANFIS models.