Intercomparison of statistical downscaling models: a case study of a large-scale river basin

Abstract

Climate change assessment at a local scale requires downscaling of general circulation models (GCMs) using various approaches. In this study, statistical downscaling using established machine learning techniques is compared with the proposed extreme gradient boosting decision tree (EXGBDT) technique. The Cauvery river basin in southern peninsular India, which is known for its frequent droughts and floods, was considered in this study. The ACCESS 1.0 CMIP5 historical GCM simulation was used for downscaling the local climate with the help of daily observation data from 35 stations located in the study zone. An intercomparison of model performance in predicting daily weather variables such as precipitation and average, maximum, and minimum temperatures over the upper, middle, and lower Cauvery river basin was performed. The findings show that mean-variance is around 15% and bias is negligible for the proposed EXGBDT model, which is better than other models under consideration. The NSE and R2 values range from 0.75-0.85 for both training and testing periods. The intercomparison of monthly mean values of observed and downscaled data for different sub-basins and parameters suggests higher model efficiency. The lower variance observed in the comparison of CLIMDEX indices suggests that the EXGBDT model performance is better in representing the local climatic condition.