Exploring the Sensitivity of Projected Climate Indices to Statistical Downscaling Choices

Despite steady advancements in global climate model (GCM) spatial resolution and simulation quality, the volumes of model output that comprise collections such as the Coupled Model Intercomparison Project Phase 5 (CMIP5) online archive are often deemed to be inappropriate for direct use in many climate impacts studies. Unacceptably large climate model biases and too coarse spatial resolution are two factors frequently cited as reasons that GCM outputs need to go through downscaling processing to produce altered climate projection data sets suitable for use as input to a range of applications. In practice, a variety of statistical downscaling (SD) techniques have been employed to generate value-added climate projections via methods that refine GCM output based upon comparisons with observation-based data, with each SD technique having its own strengths and weaknesses. Still, it is not uncommon for climate impacts studies to explore a range of potential futures by using a set of climate projections drawn from several GCMs and multiple future emissions scenarios, but downscaled by a single SD method. Focusing on the conterminous United States, here we use a set of different SD techniques to downscale daily surface variables. The downscaling cases are conducted both using a perfect model framework and in the more familiar approach of using gridded observational products and GCM output. From each set of downscaled results, a number of climate indices and other derived quantities are calculated and compared to one another, in order to investigate the sensitivity of results to the choice of SD method. Notable geographical and seasonal variations exist. For some indices, the performance of certain SD methods degrade markedly when applied to late 21st-century conditions while other SD methods exhibit far less degradation. Additionally, SD method-to-method variations tend to be greater for climate indices that that are more closely linked to the tails of the distribution. This finding is prompting further investigation not only of the SD methods' core algorithms, but also the more ad hoc adjustments embedded within some SD techniques that aim to deal with what might otherwise be considered outliers.