Reassessing model uncertainty for regional projections of precipitation with an ensemble of statistical downscaling methods

Abstract

This is the second in a pair of papers in which the performance of Statistical Downscaling Methods (SDMs) is critically re-assessed with respect to their robust applicability in climate change studies. Whereas Part I focused on temperatures (Gutierrez et al., 2013), the present manuscript deals with precipitation and considers an ensemble of twelve SDMs from the analog, weather typing, and regression (GLM) families. In the first part, we assess the performance of the methods with perfect (reanalysis) predictors, screening different geographical domains and predictor sets. To this aim, standard accuracy and distributional similarity scores, and a test for extrapolation capability based on dry observed historical periods are considered. As in Part I, the results are highly dependent on the predictor sets, with optimum configurations including information of middle tropospheric humidity (in particular Q850). As a result of this analysis, deficient SDMs are discarded in order to properly assess the spread (uncertainty) of future climate projections, avoiding the noise introduced by unsuitable models. In the second part, the resulting ensemble of SDMs is applied to four Global Circulation Models (GCMs) from the ENSEMBLES (CMIP3) project to obtain historical (1961-2000, 20C3M scenario) and future (2001-2100, A1B) regional projections. The obtained results are compared with those produced by an ensemble of Regional Climate Models (RCMs) driven by almost the same GCMs in the ENSEMBLES project. In general, the mean signal is similar with both methodologies (with the exception of Summer, where the RCMs project drier conditions) but the spread is larger for the SDM results. Finally, the contribution of the GCM and SDM-derived components to the total spread is assessed using a simple analysis of variance previously applied to the ENSEMBLES RCM ensemble. Results show that the main contributor to the spread is the choice of the GCM, except for the autumn results in the Atlantic sub-region of Spain and the Autumn and Summer results in the Mediterranean sub-region, where the choice of the SDM dominates the uncertainty during the second half of the 21st century due mainly to the different projections obtained from different families of SDM techniques. The most noticeable difference with the RCMs is the magnitude of the interaction terms, which is larger in all cases in the present study.