Conjunction of wavelet-entropy and SOM clustering for multi-GCM statistical downscaling
Abstract Important issues in statistical downscaling of general circulation models (GCMs) is to select dominant large-scale climate data (predictors). This study developed a predictor screening framework, which integrates wavelet-entropy (WE) and self-organizing map (SOM) to downscale station rainfall. WEs were computed as the representatives of predictors and fed into the SOM to cluster the predictors. SOM-based clustering of predictors according to WEs could lead to physically meaningful selection of the dominant predictors. Then, artificial neural network (ANN) as the statistical downscaling method was developed. To assess the advantages of different GCMs, multi-GCM ensemble approach was used by Can-ESM2, BNU-ESM, and INM-CM4 GCMs. Moreover, NCEP reanalysis data were used to calibrate downscaling model as well for comparison purposes. The calibration, validation, and projection of the proposed model were performed during January 1951 to December 1991, January 1992 to December 2005 and January 2017 to December 2100, respectively. The proposed data screening model could reduce the dimensionality of data and select appropriate predictors for generalizing future rainfall. Results showed better performance of ANN than multiple linear regression (MLR) model. The projection results yielded 29% and 21% decrease of rainfall at the study area for 2017–2050 under RCPs 4.5 and 8.5, respectively.