Comparative validation of statistical and dynamical downscaling models on a dense grid in central Europe: temperature

Abstract

Minimum and maximum temperature in two regional climate models and five statistical downscaling models are validated according to a unified set of criteria that have a potential relevance for impact assessments: persistence (temporal autocorrelations), spatial autocorrelations, extreme quantiles, skewness, kurtosis, and the degree of fit to observed data on both short and long times scales. The validation is conducted on two dense grids in central Europe as follows: (1) a station network and (2) a grid with a resolution of 10 km. The gridded dataset is not contaminated by artifacts of the interpolation procedure; therefore, we claim that using a gridded dataset as a validation base is a valid approach. The fit to observations in short time scales is equally good for the statistical downscaling (SDS) models and regional climate models (RCMs) in winter, while it is much better for the SDS models in summer. The reproduction of variability on long time scales, expressed as linear trends, is similarly successful by both SDS models and RCMs. Results for other criteria suggest that there is no justification for preferring dynamical models at the expense of statistical models—and vice versa. The non-linear SDS models do not outperform the linear ones.