Regional climate model performance and application of bias corrections in simulating summer monsoon maximum temperature for agro-climatic zones in India

Abstract

The present study evaluates the performance of Conformal-Cubic Atmospheric Model (CCAM) simulations downscaled from six global climate models (GCMs) (i.e., ACCESS1.0, CNRM-CM5, CCSM4, GFDL-CM3, MPI-ESM-LR, and NorESM-M) and Max Plank’s Regional Model (REMO2009(MPI)) obtained from the South-Asia Coordinated Regional Climate Downscaling Experiment (CORDEX) for analyzing the summer monsoon maximum temperature (Tmax) over the fifteen Agro-Climatic Zones (ACZs) in India. The model simulations are compared with the two sets of observed data obtained from the India Meteorology Department (IMD) and Climate Research Unit (CRU) for the period from 1981 to 2005. The results illustrate that the skill of CCAM regional climate models (RCMs) is higher than the REMO in simulating the Tmax over all the regions. The spatial patterns of Tmax in CCAM (CCSM) and CCAM (CNRM) are closer to IMD, while the Tmax distributions in CCAM (CNRM), CCAM (CCSM), and CCAM (BCCR) agree well with the CRU, and correlation coefficient (CC) is more than 0.6; however, large positive biases in all RCMs are depicted over the Himalayan regions. The inter-comparison among all the RCMs suggest that the CCAM (CNRM) and CCAM (CCSM) are rendering as the foremost models in simulating Tmax over different ACZs. Performances of these two models also infer the usefulness of the model products for impact studies over the individual ACZs. However, the existing systematic biases in the RCMs impeded the model performance and it is necessary to remove the model bias prior to some real-time application. In this study, two bias correction methods, i.e., linear scaling (LS) and distribution mapping (DM), have been used to correct RCM output bias. It is found that the model performance using DM correction is better than LS method. The performance validations are evaluated based on the probability density function (PDF), CC, and standard deviation (SD) with 95% confidence level. The model evaluation has also been justified using mean absolute error (MAE) index, Nash-Sutcliffe coefficient (NS) index, percent bias (Pbias), and the Willmott’s index of agreement (d) which confirm the research findings. The results are providing an effective guidance on the usefulness of bias corrected RCMs over a particular ACZs for impact assessment.