Regional Assessment of Sampling Techniques for More Efficient Dynamical Climate Downscaling

Dynamical downscaling is a computationally expensive method whereby finescale details of the atmosphere may be portrayed by running a limited area numerical weather prediction model (often called a regional climate model) nested within a coarse-resolution global reanalysis or global climate model output. The goal of this study is to assess using sampling techniques to dynamically downscale a small subset of days to approximate the statistical properties of the entire period of interest. Two sampling techniques are explored: one where days are randomly selected and another where representative days are chosen (or targeted) based on a set of selection criteria. The relative merit of using random sampling versus targeted random sampling is demonstrated using daily mean 2-m air temperature (T2M). The first two moments of dynamically downscaled T2M can be approximated within 0.3 K using just 5% of the population of available days during a 20-yr period. Targeted random sampling can reduce the mean absolute error of these estimates by as much as 30% locally. Estimation of the more extreme values of T2M is more uncertain and requires a larger sample size. The potential reduction in computational cost afforded by these sampling techniques could greatly benefit applications requiring high-resolution dynamically downscaled depictions of regional climate, including situations in which an ensemble of regional climate simulations is required to properly characterize uncertainty in the model physics assumptions, scenarios, and so on.