Statistical downscaling of climatic parameters using the XGBoost model: A study on temperature and relative humidity in arid regions

Abstract

Climate change and its effects on water resources and agriculture have made the accurate prediction of climatic parameters at the local scale even more necessary. General Circulation Models, due to their low spatial resolution, require downscaling for regional analyses. This study investigates the performance of the XGBoost machine learning model in the statistical downscaling of monthly mean temperature and relative humidity at the synoptic station of Qaen during the period from 1991 to 2015. In this research, the output of the GCM model, after correcting structural errors using the Chunk Mapping method,was used as input for the XGBoost model. The model's performance was evaluated using statistical criteria KGE, NSE, NRMSE, and R² in two phases: training and testing. The results indicated that the XGBoost model exhibited very good performance in downscaling mean temperature (with R² and NSE values close to one and low NRMSE in both phases) and acceptable performance for relative humidity. The model's stability in temperature simulation was evident, although there is a need for improvement in the model training process for relative humidity. The analysis of the distribution of simulated data showed that the model faces limitations in reproducing extreme temperature values (less than -10 degrees Celsius) and very high relative humidity (more than 80 percent), showing a greater tendency to simulate median temperature values. These findings are consistent with similar research in other parts of the world and confirm the high potential of XGBoost as an efficient tool in climate change studies, especially in arid regions.