Bias Estimation and Downscaling for Regional Climate Models Using Gaussian Process Regression

General circulation models are widely used to predict changes in regional climate under various pathways of projected carbon emissions. These global models are spatially coarse and are commonly downscaled to provide useful predictions at regional scales. As part of this process, any regional biases should be removed. Our research compares the accuracy of downscaling and bias removal applied to daily maximum temperature via two methods, empirical quantile mapping and Gaussian process regression. We find that the temporal stationarity assumptions within empirical quantile mapping result in a higher RMSE and distributional mismatch between observed and predicted bias than the application of Gaussian processes when applied to a dynamically downscaled global circulation model. The area of study is a sub-Arctic region covering a portion of the Copper River watershed in Alaska.