Statistical downscaling in central Europe: Evaluation of methods and potential predictors

Abstract

Several statistical downscaling methods and large-scale predictors are evaluated to ascertain their potential to determine daily mean temperatures at 39 stations in central Europe. The methods include canonical correlation analysis, singular value decomposition, and 3 multiple regression models. The potential large-scale predictors are 500 hPa heights, sea level pressure, 850 hPa temperature and 1000-500 hPa thickness. The performance of the methods is evaluated using cross-validation and rootmean-squared error as a measure of accuracy. The stepwise screening of gridpoint data is found to be the statistical model that performed the best. Among the predictors, temperature variables yield more accurate results than circulation variables. The best predictor is the combination of 500 hPa heights and 850 hPa temperature. Geographical variations of the specification skill, mainly the differences between the elevated and lowland stations, are also discussed.