9.4.1.4 The Influence of Greenhouse Gas and Total Anthropogenic Forcing on Global Surface Temperature
Since the TAR, a large number of studies based on the longer observational record, improved models and stronger signal-to-noise ratio have increased confidence in the detection of an anthropogenic signal in the instrumental record (see, e.g., the recent review by IDAG, 2005). Many more detection and attribution studies are now available than were available for the TAR, and these have used more recent climate data than previous studies and a much greater variety of climate simulations with more sophisticated treatments of a greater number of both anthropogenic and natural forcings.
Fingerprint studies that use climate change signals estimated from an array of climate models indicate that detection of an anthropogenic contribution to the observed warming is a result that is robust to a wide range of model uncertainty, forcing uncertainties and analysis techniques (Hegerl et al., 2001; Gillett et al., 2002c; Tett et al., 2002; Zwiers and Zhang, 2003; IDAG, 2005; Stone and Allen, 2005b; Stone et al., 2007a,b; Stott et al., 2006b,c; Zhang et al., 2006). These studies account for the possibility that the agreement between simulated and observed global mean temperature changes could be fortuitous as a result of, for example, balancing too great (or too small) a model sensitivity with a too large (or too small) negative aerosol forcing (Schwartz, 2004; Hansen et al., 2005) or a too small (or too large) warming due to solar changes. Multi-signal detection and attribution analyses do not rely on such agreement because they seek to explain the observed temperature changes in terms of the responses to individual forcings, using model-derived patterns of response and a noise-reducing metric (Appendix 9.A) but determining their amplitudes from observations. As discussed in Section 9.2.2.1, these approaches make use of differences in the temporal and spatial responses to forcings to separate their effect in observations.
Since the TAR, there has also been an increased emphasis on quantifying the greenhouse gas contribution to observed warming, and distinguishing this contribution from other factors, both anthropogenic, such as the cooling effects of aerosols, and natural, such as from volcanic eruptions and changes in solar radiation.
A comparison of results using four different models (Figure 9.9) shows that there is a robust identification of a significant greenhouse warming contribution to observed warming that is likely greater than the observed warming over the last 50 years with a significant net cooling from other anthropogenic forcings over that period, dominated by aerosols. Stott et al. (2006c) compare results over the 20th century obtained using the UKMO-HadCM3, PCM (see Table 8.1 for model descriptions) and Geophysical Fluid Dynamics Laboratory (GFDL) R30 models. They find consistent estimates for the greenhouse gas attributable warming over the century, expressed as the difference between temperatures in the last and first decades of the century, of 0.6°C to 1.3°C (5 to 95%) offset by cooling from other anthropogenic factors associated mainly with cooling from aerosols of 0.1°C to 0.7°C and a small net contribution from natural factors over the century of –0.1°C to 0.1°C (Figure 9.9b). Scaling factors for the model response to three forcings are shown in Figure 9.9a. A similar analysis for the MIROC3.2 model (see Table 8.1 for a description) finds a somewhat larger warming contribution from greenhouse gases of 1.2°C to 1.5°C offset by a cooling of 0.6°C to 0.8°C from other anthropogenic factors and a very small net natural contribution (Figure 9.9b). In all cases, the fifth percentile of the warming attributable to greenhouse gases is greater than the observed warming over the last 50 years of the 20th century (Figure 9.9c).
The detection and estimation of a greenhouse gas signal is also robust to accounting more fully for model uncertainty. An analysis that combines results from three climate models and thereby incorporates uncertainty in the response of these three models (by including an estimate of the inter-model covariance structure in the regression method; Huntingford et al., 2006), supports the results from each of the models individually that it is likely that greenhouse gases would have caused more warming than was observed over the 1950 to 1999 period (Figure 9.9, results labelled ‘EIV’). These results are consistent with the results of an earlier analysis, which calculated the mean response patterns from five models and included a simpler estimate of model uncertainty (obtained by a simple rescaling of the variability estimated from a long control run, thereby assuming that inter-model uncertainty has the same covariance structure as internal variability; Gillett et al., 2002c). Both the results of Gillett et al. (2002c) and Huntingford et al. (2006) indicate that inter-model differences do not greatly increase detection and attribution uncertainties and that averaging fingerprints improves detection results.
A robust anthropogenic signal is also found in a wide range of climate models that do not have the full range of simulations required to directly estimate the responses to individual forcings required for the full multi-signal detection and attribution analyses (Stone et al., 2007a,b). In these cases, an estimate of the model’s pattern of response to each individual forcing can be diagnosed by fitting a series of EBMs, one for each forcing, to the mean coupled model response to all the forcings to diagnose the time-dependent response in the global mean for each individual forcing. The magnitude of these time-only signals can then be inferred from observations using detection methods (Stone et al., 2007a,b). When applied to 13 different climate models that had transient simulations of 1901 to 2005 temperature change, Stone et al. (2007a) find a robust detection across the models of greenhouse gas warming over this period, although uncertainties in attributable temperature changes due to the different forcings are larger than when considering spatio-temporal patterns. By tuning an EBM to the observations, and using an AOGCM solely to estimate internal variability, Stone and Allen (2005b) detect the effects of greenhouse gases and tropospheric sulphate aerosols in the observed 1900 to 2004 record, but not the effects of volcanic and solar forcing.
The detection of an anthropogenic signal is also robust to using different methods. For example, Bayesian detection analyses (Appendix 9.A.2) robustly detect anthropogenic influence on near-surface temperature changes (Smith et al., 2003; Schnur and Hasselmann, 2005; Min and Hense, 2006a,b). In these studies, Bayes Factors (ratios of posterior to prior odds) are used to assess evidence supporting competing hypotheses (Kass and Raftery, 1995; see Appendix 9.A.2). A Bayesian analysis of seven climate models (Schnur and Hasselman, 2005) and Bayesian analyses of MMD 20C3M simulations (Min and Hense, 2006a,b) find decisive evidence for the influence of anthropogenic forcings. Lee et al. (2005), using an approach suggested by Berliner et al. (2000), evaluate the evidence for the presence of the combined greenhouse gas and sulphate aerosol (GS) signal, estimated from CGCM1 and CGCM2 (Table 8.1; McAvaney et al., 2001), in observations for several five-decade windows, beginning with 1900 to 1949 and ending with 1950 to 1999. Very strong evidence was found in support of detection of the forced response during both halves of the 20th century regardless of the choice of prior distribution. However, evidence for attribution in that approach is based on the extent to which observed data narrow the prior uncertainty on the size of the anthropogenic signal. That evidence was not found to be very strong, although Lee et al. (2005) estimate that strong evidence for attribution as defined in their approach may emerge within the next two decades as the anthropogenic signal strengthens.
In a further study, Lee et al. (2006) assess whether anthropogenic forcing has enhanced the predictability of decadal global-scale temperature changes; a forcing-related enhancement in predictability would give a further indication of its role in the evolution of the 20th-century climate. Using an ensemble of simulations of the 20th century with GS forcing, they use Bayesian tools similar to those of Lee et al. (2005) to produce, for each decade beginning with 1930 to 1939, a forecast of the probability of above-normal temperatures where ‘normal’ is defined as the mean temperature of the preceding three decades. These hindcasts become skilful during the last two decades of the 20th century as indicated both by their Brier skill scores, a standard measure of the skill of probabilistic forecasts, and by the confidence bounds on hindcasts of global mean temperature anomalies (Figure 9.10). This indicates that greenhouse gas forcing contributes to predictability of decadal temperature changes during the latter part of the 20th century.
Another type of analysis is a Granger causality analysis of the lagged covariance structure of observed hemispheric temperatures (Kaufmann and Stern, 2002), which also provides evidence for an anthropogenic signal, although such evidence may not be conclusive on its own without additional information from climate models (Triacca, 2001). Consistently, a neural network model is unable to reconstruct the observed global temperature record from 1860 to 2000 if anthropogenic forcings are not taken into account (Pasini et al., 2006). Further, an assessment of recent climate change relative to the long-term persistence of NH mean temperature as diagnosed from a range of reconstructed temperature records (Rybski et al., 2006) suggests that the recent warming cannot be explained solely in terms of natural factors, regardless of the reconstruction used. Similarly, Fomby and Vogelsang (2002), using a test of trend that accounts for the effects of serial correlation, find that the increase in global mean temperature over the 20th century is statistically significant even if it is assumed that natural climate variability has strong serial correlation.