Revealing and Assessing Statistical Downscaling and Bias Correction Method Uncertainties

Bias correction and statistical downscaling methods are often employed to address dynamical climate model shortcomings, with an aim being to produce data products that are more suitable for direct use as input to various applications, including hydrological models. In this regard, utilizing observations and statistical techniques to refine dynamical model results serves as a "value-added" process. However, various implementations of statistical processing methods each produce somewhat different results, thereby revealing uncertainties that exist in the value-added data products. Practitioners' appreciation of the existence of such uncertainties can lead to the more quantitative question, "How correct are the corrections applied to refine climate projections?"

Here we summarize results of experiments designed to test the performance of a set of common bias correction approaches used to statistically downscale climate model projections. The extent to which some methods' performance degrades when applied to 21st century projections relative to a historical training period is examined using variants of a perfect model framework. Several methods that generally perform well are found to have notable weaknesses that manifest under certain climate conditions for some variables and locations. We also illustrate, via a set of sensitivity studies, that some evaluation metrics are more sensitive than others to the choice of the observation-based precipitation product used in training. Results support the general statement that the matching of statistically processed climate projection data products to climate impacts applications can be very much application-dependent. Accordingly, better-informed choices can be made when a reckoning is made of both the strengths and weaknesses of available bias correction and statistical downscaling methods and of the sensitivities of an intended end-use application to climate data inputs.