Statistical Downscaling for the Northern Great Plains

Abstract:

The need for detailed, local scale information about the warming climate has led to the use of ever more complex and geographically realistic computer models as well as the use of regional models capable of capturing much finer details. Another class of methods for ascertaining localized data is known as statistical downscaling, which offers some advantages over regional models, especially in the realm of computational efficiency. Statistical downscaling can be described as the process of linking coarse resolution climate model output to that of fine resolution or even station-level data via statistical relationships with the purpose of correcting model biases at the local scale. The development and application of downscaling has given rise to a plethora of techniques which have been applied to many spatial scales and multiple climate variables. In this study two downscaling processes, bias-corrected statistical downscaling (BCSD) and canonical correlation analysis (CCA), are applied to minimum and maximum temperatures and precipitation for the Northern Great Plains (NGP, 40 - 53°N and 95 - 120°W) region at both daily and monthly time steps. The abilities of the methods were tested by assessing their ability to recreate local variations in a set of both spatial and temporal climate metrics obtained through the analysis of 1/16 degree station data for the period 1950 to 2000. Model data for temperature, precipitation and a set of predictor variables were obtained from CMIP5 for 15 models. BCSD was applied using direct comparison and correction of the variable distributions via quadrant mapping. CCA was calibrated on the data for the period 1950 to 1980 using a series of model-based predictor variables screened for increasing skill, with the derived model being applied to the period 1980 to 2000 so as to verify that it could recreate the overall climate patterns and trends. As in previous studies done on other regions, it was found that the CCA method recreated local variations in temperature and precipitation with the highest overall fidelity to the observed data, though both methods captured temperature better than precipitation. This result helps to further refine downscaling methods and processes by narrowing the available methods to those with suitable skill.