Comparing the performance of multiple statistical downscaling approaches using a perfect model framework

Empirical statistical downscaling approaches quantify relationships between large-scale predictors and high-resolution predictands. Though based on historical simulations and observations, these relationships are assumed to hold true under an altered climate. Here, we use an innovative new “perfect model” approach (Dixon et al. 2016) to test that assumption of stationarity in four different commonly-used statistical downscaling techniques: delta, empirical quantile mapping, parametric quantile mapping, and non-parametric kernel density estimation. The perfect model approach uses high-resolution 25 km simulations by GFDL HiRAM as “observations” and coarsened 250 km resolution simulations as “model” for the periods 1979-2008 and 2086-2095 under the higher RCP 8.5 scenario. We evaluate the performance of the statistical downscaling models using impact-relevant threshold and intensity metrics specifically tied to agriculture, ecosystems, human health, infrastructure, energy demand, and water availability to demonstrate how the validity of the stationarity assumption can vary by downscaling approach, geographic location, season, and the quantile(s) of the distribution reflected by a given metric.

We find that a simple seasonal delta approach is adequate for non-extreme values over the continental U.S. but has strong biases at more northern latitudes, some as large as or greater than the absolute change in temperature projected to occur. Empirical quantile mapping is stationary out to the 90th quantile of the distribution of wet days; beyond that, precipitation biases increase rapidly to the point where biases are as large as projected changes for the 99.9th quantile. Empirical quantile mapping has no trend in bias but substantial salt-and-pepper noise in the bias at the tails and a large positive temperature bias at the coastlines that indicate a non-stationarity effect likely associated with coastal winds. Finally, the non-parametric approach significantly reduces bias relative to every other method, demonstrating improved ability and validity of the stationarity across geography, quantile, and variable.