The conjunction of feature extraction method with AI-based ensemble statistical downscaling models

In this study, two general circulation models (GCMs) (Can-ESM2, BNU-ESM) were used to simulate the future precipitation of Tabriz city. The weakness of GCMs is the coarse resolution of climate variables in which the different methods of downscaling is about to solve this deficiency. In this study, the Artificial Intelligence (AI) models i.e., Artificial Neural Network (ANN) and Adaptive neuro fuzzy inference system (ANFIS) were used to statistically downscale the climate variables of GCMs. Without any doubt, the most important step during the use of these models, is selecting of the dominant inputs among huge of large-scale GCM data. So in this study for the selection of dominant inputs, decision tree and mutual information (MI) feature extraction methods were used. Also, the ensemble techniques were used to evaluate the efficiency of downscaling models and to decrease the uncertainties. Comparison the result of downscaling models indicated that the ensemble technique (i.e., hybrid of ANN and ANFIS) with dominant inputs based on decision tree feature extraction method presents better performance. In both GCMs, the application of the ensemble downscaling couple with dominant predictors based on decision tree model in precipitation downscaling showed 10%-38% increase in DC in versus the individual ANN and ANFIS downscaling models. The projection precipitation of Tabriz synoptic station for future (2020-2060) by proposed ensemble AI- based model indicated 30%-40% precipitation decreases under RCP4.5 and RCP8.5 scenarios.