Probst, P. , R. Rizzi, E. Tosi, V. Lucarini, T. Maestri, 2012: Total cloud cover from satellite observations and climate models. Elsevier, Atmospheric Reasearch, in press, doi:10.1016/j.atmosres.2012.01.005 [not yet available on line at http://www.sciencedirect.com/science/journal/aip/01698095]
with the abstract [highlight added]
Global and zonal monthly means of cloud cover fraction for total cloudiness (CF) from the ISCCP D2 dataset are compared to same quantities produced by the 20th century simulations of 21 climate models from the World Climate Research Programme’s (WCRP’s) Coupled Model Intercomparison Project phase 3 (CMIP3). The comparison spans the time frame from January 1984 to December 1999 and the global and zonal averages of CF are studied. It is shown that the global mean of CF for the PCMDI-CMIP3models, averaged over the whole period, exhibits a considerable variance and generally underestimates the ISCCP value. Large differences among models, and between models and observations, are found in the polar areas, where both models and satellite observations are less reliable, and especially near Antarctica. For this reason the zonal analysis is focused over the 60° S60° N latitudinal belt, which includes the tropical area and midlatitudes. The two hemispheres are analysed separately to show the variation of the amplitude of the seasonal cycle. Most models underestimate the yearly averaged values of CF over all the analysed areas, whilst they capture, in a qualitatively correct way, the magnitude and the sign of the seasonal cycle over the whole geographical domain, but overestimate the amplitude of the seasonal cycle in the tropical areas and at mid-latitudes,when taken separately. The inter annual variability of the yearly averages is underestimated by all models in each area analysed, and also the interannual variability of the amplitude of the seasonal cycle is underestimated, but to a lesser extent. This work shows that the climate models have a heterogeneous behaviour in simulating the CF over different areas of the Globe, with a very wide span both with observed CF and among themselves. Some models agree quite well with the observations in one or more of the metrics employed in this analysis, but not a single model has a statistically significant agreement with the observational datasets on yearly averaged values of CF and on the amplitude of the seasonal cycle over all analysed areas
The conclusion has the text
In this paper the monthly mean of total cloud cover fraction (CF) is chosen as benchmark for intercomparing and validating climate models included in the PCMDI-CMIP3 project, which have contributed decisively to the preparation of IPCC AR4 (Solomon et al., 2007). As observational counterpart, the satellite observations of clouds constituting the ISCCP D2 dataset for the 1984–1999 time frame, are considered. These data are compared to the corresponding period of the standard 20th century simulations of 21 climate models.
Whilst some models agree quite well with the observations in one or more of the metrics employed in this analysis, not a single model shows a statistically significant agreement with the observational dataset of yearly averaged values of CF and on the amplitude of the seasonal cycle on both tropical and extratropical regions. Our results highlight that the representation of the basic statistical properties of clouds in state-of-the-art climate models is still incomplete, as relevant systematic errors are present for most models in both tropical and extratropical regions. Typically, the climate models underestimate both the global CF and the zonal averaged CF over almost all zonal bands.
The range of model results is very wide since the annual and global averaged CF ranges from about 47% to 73%, with a mean difference with D2 observations of about 7%. The largest differences among models in the zonal averages are found in the tropical region and in the two polar regions, where the relative spread of models’ outputs reaches 0.4 (Tropics), 0.6 (Arctic region) and 0.9 (Antarctica). One must however also consider that it is likely that the error in the CF properties in the observational dataset is largest in the polar regions.
Looking at higher order statistics, it is shown that the interrannual variability of global averaged CF are quite strongly underestimated in all models with respect to observations,whilst the interannual variability of the seasonal signal is only slightly underestimated.
The documented differences between the observational dataset and the models constitute a problemsince the statistical properties of clouds play a decisive role in the earth climate, by providing a first order contribution to the energy budget at the top of the atmosphere (Solomon et al., 2007) and at the surface. It is therefore a feature that influences many physical processes inside the real atmosphere and inside models. Since most models are tuned to provide a TOA energy balance as close as possible to the measured record, the systematic deviations between a model and the CF observational dataset imply compensating deviations in a range of physical processes occurring almost everywhere in the system. The documented systematic inter-model discrepancies provide an indication of the effect of diverse mix of physical processes on CF. The authors believe that this is not an healthy situation.
The results presented in this paper provide a natural complement to the analyses shown in Pincus et al. (2008), who discussed the second moments of the statistics of the CF but did not show the results of the mean climatology. Often, the results obtained from different climate models are averaged under the assumption that the model biases will partially compensate, so that a more realistic estimate of the climate properties are achieved by the so-constructed “mean model”. As discussed in, e.g., Lucarini (2008) such a procedure, even if commonly used, is not really well defined in a probabilistic sense, and should be interpreted only in a qualitative sense. Since in our case most of the models have biases of the same sign with respect to observations, the ensemble mean (constructed in our case with a simple un-weighted averaging) does not provide good agreement with observations for the considered statistical estimators, with discrepancies in most case larger than one standard deviation of the single model outputs.