Spatial-Scale Dependence of Climate Model Performance in the CMIP3 Ensemble

About 20 global climate models have been run for the Intergovernmental Panel on Climate Change (IPCC) Fourth Assessment Report (AR4) to predict climate change due to anthropogenic activities. Evaluating these models is an important step to establish confidence in climate projections. Model evaluation, however, is often performed on a gridpoint basis despite the fact that models are known to often be unreliable at such small spatial scales. In this study, the annual mean values of surface air temperature and precipitation are analyzed. Using a spatial smoothing technique with a variable-scale parameter it is shown that the intermodel spread, as well as model errors from observations, is reduced as the characteristic smoothing scale increases. At the same time, the ability to reproduce small-scale features is reduced and the simulated patterns become fuzzy. Depending on the variable of interest, the location, and the way that data are aggregated, different optimal smoothing scales from the gridpoint size to about 2000 km are found to give good agreement with present-day observation yet retain most regional features of the climate signal. Higher model resolution surprisingly does not imply much better agreement with temperature observations, in particular with stronger smoothing, and resolving smaller scales therefore does not necessarily seem to improve the simulation of large-scale climate features. Similarities in mean temperature and precipitation fields for a pair of models in the ensemble persist locally for about a century into the future, providing some justification for subtracting control errors in the models. Large-scale to global errors, however, are not well preserved over time, consistent with a poor constraint of the present-day climate on the simulated global temperature and precipitation response.