ANN-based statistical downscaling of climatic parameters using decision tree predictor screening method

Abstract

In this paper, artificial neural network (ANN) was used for statistically downscale the outputs of general circulation models (GCMs) to assess future changes of precipitation and mean temperature in Tabriz synoptic station at north-west of Iran. Since one of the significant subjects in statistical downscaling of GCMs is to select the most dominant large scale climate variables (predictors) among huge number of potential predictors, the predictors screening methods including decision tree, mutual information (MI) and correlation coefficient (CC) were used to statistically downscale mean monthly precipitation and temperature. Three GCMs were used, including Can-ESM2 and BNU-ESM from IPCC AR5 models and CGCM3 from IPCC AR4 models. The results of downscaling in the base period (1951–2000) indicated that among feature extraction methods decision tree had superiority to MI and CC techniques. Therefore, the future projection of precipitation and mean temperature during 2020–2060 was implemented using ANN-based simulation according to the most efficient downscaling model (i.e., decision tree-based calibration). Different results according to different GCMs and scenarios were obtained for precipitation projection. In this way, the Can-ESM2 model under RCP8.5 showed 29.78% decrease in annual precipitation and CGCM3 model under B1 indicated 1.06% increase of annual precipitation. Temperature projection outcomes denoted that annual mean temperature will increase over the region and the most increase in mean temperature was determined by BNU-ESM model under RCP8.5.