The Effect of Statistical Downscaling on the Weighting of Multi-Model Ensembles of Precipitation

Abstract

Recently, assessments of global climate model (GCM) ensembles have transitioned from using unweighted means to weighted means designed to account for skill and interdependence among models. Although ensemble-weighting schemes are typically derived using a GCM ensemble, statistically downscaled projections are used in climate change assessments. This study applies four ensemble-weighting schemes for model averaging to precipitation projections in the south-central United States. The weighting schemes are applied to (1) a 26-member GCM ensemble and (2) those 26 members downscaled using Localized Canonical Analogs (LOCA). This study is distinct from prior research because it compares the interactions of ensemble-weighting schemes with GCMs and statistical downscaling to produce summarized climate projection products. The analysis indicates that statistical downscaling improves the ensemble accuracy (LOCA average root mean square error is 100 mm less than the CMIP5 average root mean square error) and reduces the uncertainty of the projected ensemble-mean change. Furthermore, averaging the LOCA ensemble using Bayesian Model Averaging reduces the uncertainty beyond any other combination of weighting schemes and ensemble (standard deviation of the mean projected change in the domain is reduced by 40–50 mm). The results also indicate that it is inappropriate to assume that a weighting scheme derived from a GCM ensemble matches the same weights derived using a downscaled ensemble.