Combined statistical-dynamical downscaling of climate simulations

The need for climate change information at smaller scales than resolved by global General Circulation Models (GCMs) has motivated two approaches for deducing regional climate information from GCMs: semi-empirical (statistical) downscaling (SDS), and dynamical downscaling which most often uses regional climate models (RCMs) one-way nested from GCM output. A few recent studies have compared statistical and dynamical downscaling results for basic meteorological variables but so far there has been no attempt at a synthesis of the two methods.

Here we test the combined use of statistical and dynamical downscaling by using the results of a RCM simulation driven from large-scale data (NCEP/NCAR reanalyses) as predictor variables in SDS. The SDS method uses step-wise multiple linear regression to identify atmospheric variables in the RCM results to predict local precipitation. The regional model output is from a 10 year (1979-1988) simulation for the continental U.S. using the RegCM2 model at about 50-km resolution. This combined statistical-dynamical downscaling (SDDS) implementation is calibrated using observations for 1979-1983 and validated against observations for 1984-1988. Results for the SDDS are compared both to direct output of precipitation from RegCM2 (i.e., the usual dynamical downscaling approach) and to SDS using reanalysis data as input (i.e., the usual statistical downscaling approach). The SDDS method is shown to provide useful improvement in skill compared to both conventional techniques. A further advantage of SDDS is that it transforms the dynamical model output into a probabilistic framework.