Statistical Downscaling of Temperature Using Global Climate Model Outputs - Effect of Bias correction

General Circulation Models (GCMs) are the primary source of knowledge for constructing climate scenarios and provide the basis for quantifying the climate change impacts at multi-scales and from local to global. However, the climate model simulations have a lower resolution than the desired watershed or hydrologic scale. Different downscaling methodologies are adopted to transform the global scale (coarser resolution) climate information to the local scale (finer resolution). One of the drawbacks of the GCM simulations is the systematic bias relative to historical observations. Bias correction is thus required to adjust the simulated values to reflect the observed distribution and statistics. In this study, the effect of bias correction is evaluated on the statistical downscaling models' performance to predict the temperature. Three statistical downscaling models are used: (i) Multi-linear Regression (MLR); (ii) Generalized Regression Neural Network (GRNN); and (iii) Cascade Neural Network (CasNN). The average daily temperature simulations generated by 25 GCMs of Coupled Model Intercomparison Project Phase-5 (CMIP5) are used in the study. The analysis is carried out at 22 stations of the Upper Thames River Basin (UTRB) in Canada during the baseline period of 1950 to 2005. The downscaling models' performance is evaluated using the Pearson Correlation Coefficient (CC) and Nash Sutcliffe Efficiency (NSE). The results indicated that bias correction had improved all the downscaling models' performance at all stations of UTRB. The respective increase in CC and NSE values for (i) MLR is 8% and 10%; (ii) GRNN is 4% and 7%; and (iii) CasNN is 4% and 8%. Among the three downscaling models, multi-linear regression and cascade neural network models have shown similar performance.