Performance of the XGBoost algorithm in downscaling temperature and relative humidity in a temperate climate: a case study in Kermanshah

Abstract

Climate change requires a precise analysis of local data, and statistical downscaling using machine learning algorithms such as XGBoost can enhance the spatial accuracy of General Circulation Models (GCMs). This study examined the performance of XGBoost in the daily downscaling of temperature and relative humidity at the synoptic station of Kermanshah during the period from 1990 to 2014. In order to investigate the performance of the XGBoost model in downscaling the climate variables of temperature and relative humidity, local and large-scale data were divided into two training and testing sections, so that the years 1990 to 2007 were considered as the training section and the years 2007 to 2015 were considered as the testing section. The present research showed that the XGBoost algorithm, as one of the advanced machine learning methods, performed very well in downscaling climatic parameters, especially temperature, and to some extent relative humidity. To evaluate the model's performance, the metric values were examined separately in two sections: training and testing. For temperature using ECMWF-ERA5 data, the KGE, NSE, and R² values in the training section were 0.98, 0.98, and 0.99, respectively. For temperature using MPI-ESM1-2-HR data, the values of these metrics in the training section were 0.93, 0.94, and 0.97, and in the testing section, they were 0.91, 0.88, and 0.94, respectively. Additionally, for relative humidity using ECMWF-ERA5 data, the metric values in the training section were 0.82, 0.93, and 0.97, and in the testing section, they were 0.7, 0.72, and 0.87. Finally, for relative humidity using MPI-ESM1-2-HR data, the values of these metrics in the training section were 0.72, 0.67, and 0.82, and in the testing section, they were 0.70, 0.65, and 0.81. Graphical analyses confirmed the superiority of the model in simulating intermediate and extreme temperature values, especially with ECMWF-ERA5, but limitations in reproducing extreme values were observed in relative humidity