Attribution of regional-scale temperature changes to anthropogenic and natural causes



[1] The causes of twentieth century temperature change in six separate land areas of the Earth have been determined by carrying out a series of optimal detection analyses. The warming effects of increasing greenhouse gas concentrations have been detected in all the regions examined, including North America and Europe. In most regions, cooling from sulfate aerosols counteracts some of the greenhouse warming, and there is some evidence for reduced net aerosol cooling in Asia, possibly as a result of warming from black carbon.

1. Introduction

[2] An increasing body of evidence indicates that global warming that has been observed over the course of the last century cannot be explained by natural externally forced or internal variability [Mitchell et al., 2001]. A number of recent studies [eg., Stott et al., 2001; Tett et al., 2002] have used optimal detection [Hasselmann, 1997], a form of linear regression [Allen and Tett, 1999], to estimate the contributions, with their uncertainties, of anthropogenic and natural forcings to recent temperature changes. Optimal detection studies consistently show that anthropogenic forcings were the dominant factor controlling global warming in the latter half of the 20th century, leading the IPCC to conclude in the Third Assessment Report that “most of the warming observed over the last 50 years is attributable to human activities”.

[3] Most detection studies investigating atmospheric temperature changes have considered global scale patterns of change. [Zwiers and Zhang, 2003] showed that the combined effects of greenhouse gases and sulfate aerosols may be detected on sub-global scales, and showed a detectable anthropogenic influence on warming in Eurasia and North America. Our aim here is to carry out an attribution analysis of regional climate change throughout the globe, incorporating the most important multiple contributors to climate change over the 20th century.

2. Methods

[4] We compare simulations of the coupled climate model, HadCM3 [Johns et al., 2003], with the CRUTEM2(v) dataset of near-surface air temperatures over land [Jones and Moberg, 2003]. We analyze the 100 year period from 1st December 1899 to 30 November 1999, the HadCM3 model year starting on 1st December. We consider decadal means of temperature changes in six large continental-scale regions: N America, Asia, S America, Africa, Australia, Europe (see Table 1), each of which is divided into 2 or 3 sub-regions. These sub-regions are based on regions that were defined by [Giorgi, 2002], in order to represent different climatic regimes while approximately covering all land areas except Antarctica and of a sufficient spatial scale to exclude wavelengths smaller than those resolved by the HadCM2 climate model (which has the same 2.5 in latitude by 3.75 degree in longitude atmospheric resolution as HadCM3). In some cases we combine the regions of [Giorgi, 2002] so as to have no more than three sub-regions within each continental region. This allows a more stable estimation of the optimal fingerprint, which is a 10 decade spatio-temporal pattern, and for which we have only a limited amount of data with which to estimate the characteristics of internal variability (see below).

Table 1. Definition of Continental and Sub-Continental Scale Regions With Standard Deviations of Decadal Variations (K) in Each Region From Observations, Observations With a Polynomial (3rd order) Trend Removed, Observations With the all Forcings Ensemble-Mean Removed, and From CONTROL
  1. a

    Error estimates represent uncertainty at the 90% confidence interval based on resampling from CONTROL.

North America170W–105W30N–75N0.330.170.180.16 ± 0.06
105W–60W30N–50N0. ± 0.09
105W–10W50N–85N0.290.160.370.28 ± 0.18
Asia65E–1450E10S–30N0. ± 0.04
40E–180E30N–50N0.330.090.210.15 ± 0.08
40E–180E50N–70N0.420.120.170.24 ± 0.09
Southern and Central America80W–35W20S–10N0. ± 0.07
75W–40W55S–20S0. ± 0.04
115W–85W10N–30N0. ± 0.06
Africa20W–50E10S–20N0. ± 0.04
10W–50E35S–10S0. ± 0.06
20W–65E20N–30N0. ± 0.07
Australia110E–155E45S–28S0. ± 0.05
110E–155E28S–10S0. ± 0.08
Europe10W–40E30N–50N0. ± 0.09
10W–40E50N–75N0.330.190.300.31 ± 0.16
Global (land and ocean) ± 0.02

[5] We consider ensembles of climate model simulations, each ensemble consisting of four simulations starting from a different initial condition taken from a long multi-century control run of HadCM3 (CONTROL). The three ensembles we use in our optimal detection analyses are forced with changes in well-mixed greenhouse gases (GHG), changes in well-mixed greenhouse gases plus anthropogenic sulphur emissions and their implied changes to cloud albedos, and tropospheric and stratospheric ozone changes (ANTHRO), and changes in both solar irradiance and stratospheric aerosol following volcanic eruptions (NAT). These simulations are those described in more detail by [Tett et al., 2002]. In addition we compare the observed temperature changes in our 16 sub-continental regions with an ensemble of four simulations containing all the forcings included in the three other ensembles (GHG, ANTHRO and NAT). This ensemble of simulations, denoted ALL, was shown by Stott et al. [2000] to consistently simulate the temporal evolution of global-mean temperatures during the 20th century.

[6] The observed decadal means are constructed from monthly mean values, requiring data for at least 8 out of 12 months in at least half the years, otherwise the data are set to missing. The model simulated decadal data are masked by the observational decadal data mask and regional means constructed from the available data.

[7] We apply standard optimal detection methodology, expressing observed decadal-mean near-surface temperature changes y as a linear sum of simulated changes from x1 (GHG), x2 (ANTHRO), x3 (NAT) plus noise, υ0:

display math

where βi is the vector of unknown scaling factors to be estimated in the regression. We take into account the uncertainty introduced by taking the model-simulated responses from a finite ensemble, which differ from the underlying noise-free responses that would be obtained from a hypothetical infinite ensemble. We do this by including the additional noise term υi in the regression equation [Allen and Stott, 2003].

[8] We carry out the optimal detection analysis in each of the six regions using the sub-regions (defined in Table 1) to represent the spatial pattern of change within each of the 6 continental scale regions. Intra-ensemble variability from the four ensembles GHG, ANTHRO, NAT, and ALL is used to construct an estimate of the covariance structure of internal variability, image which is needed for optimization, as described by [Tett et al., 2002]. A 1830 year segment of a control run of HadCM3, in which external forcings have no year-to-year variations, is used to estimate an independent covariance estimate, image for significance testing.

[9] The regression analysis is carried out in the space spanned by the leading p empirical orthogonal functions of image We maintain the maximum truncation possible, consistent with the residual being consistent with model simulated natural internal variability [Allen and Tett, 1999] and scaling factors, βi, being stable across a range of truncations.

[10] From the runs available, (GHG and ANTHRO), we wish to extract the pure greenhouse component (denoted G) and the non-greenhouse sulfate and ozone components (SO). Assuming that the climate response to these two different forcings is linearly additive [Haywood et al., 1997], it is possible to calculate the scaling factors on G and SO [see Tett et al., 2002]. By far the largest component of the non-greenhouse gas anthropogenic forcing at the surface comes from sulfate aerosols, with the indirect effect, through increases in cloud albedos, being dominant in HadCM3 [Tett et al., 2002].

3. Internal Climate Variability

[11] An important requirement for detecting climate change signals in the observed record is that the model's representation of internal variability should be adequate at the scales considered. Part of the rationale for choosing the regions defined in Table 1 is that these regions are of the appropriate spatial scale to represent regional patterns of climate change while avoiding small scales that are not well represented by climate models. [Stott and Tett, 1998] found that models underestimate climate variability at spatial scales less than 2000 km; the regions used here vary in size from a few to several thousand km.

[12] As a further test of the model's ability to represent internal variability in the regions considered, decadal variability from the detrended decadal mean observations is compared in Table 1 with decadal variability calculated from 100-year samples taken from the HadCM3 control, where CONTROL decadal data are masked by the observational decadal data mask in the same way as the transient simulations. The observations in each region have been detrended in two ways in order to obtain an estimate of unforced variability: by fitting a 3rd order polynomial and calculating the residuals after removing this fit, and by subtracting the ensemble mean temperature changes in the ALL ensemble from the observed changes. Whichever way the observations are detrended, the observed decadal variability is estimated to be generally consistent with decadal variability in the control run (Table 1).

4. Estimated Forced Contributions to Observed Temperature Changes

[13] The scaling factors for the anthropogenic (G and SO) and natural (NAT) components of observed near-surface temperature change are shown in Figure 1 for each of the six continental regions. If a bar (which represents the 5 to 95 percentile uncertainty range) includes 1, this indicates that the model simulation of the climatic response to this forcing is consistent with the estimated observed response. Where the 5 to 95 percentile uncertainty range does not include zero, this indicates that the relevant signal has been detected at the 5% significance level (ie there is less than a 5% chance that natural variability rather than the forcing is responsible for the observed change). A climatic response to greenhouse gases is detected in all the continental regions. In all these regions the scaling factors for greenhouse gas forcing, G, are consistent with 1, indicating that the model's simulation of the climatic response to greenhouse gas forcing is consistent with that observed, with the best estimates being close to 1 in all regions except Australia. In addition, we find evidence of a detectable natural influence on the regional scale in N and S America and Europe (Figure 1), although natural forcings are weaker and more uncertain than the anthropogenic forcings, reflected in the larger uncertainty bars in Figure 1. Our confidence in a significant naturally-forced component to the temperature changes observed in these regions is therefore not as great as that in a significant anthropogenically-forced component.

Figure 1.

Ranges of scaling factors for the three signals (G, SO, NAT) for the six continental-scale regions defined in Table 1. The bars show the 5 to 95 percentile uncertainty ranges, the stars mark the best estimates, and the dashed lines mark scaling factors equal to unity.

[14] Reconstructions of the decadally varying contributions of the three climate forcing factors (G, S, NAT) to the observed temperature changes in the six continental-scale areas are shown in Figure 2. Greenhouse gases provide an accelerating warming in all regions (Figure 2). Aerosols produce a general net cooling in all areas except Australia, with cooling starting earlier in Europe than in N or S America (temperatures decreasing by the 1930s in Europe, by the 1940s in N America, and the 1950s in S America). We estimate there is likely to be less aerosol cooling in Europe and Asia than simulated by HadCM3. The best estimates of the scaling factors for SO for Europe and Asia are 0.33 and 0.53 respectively (Figure 1) compared to 1.08 and 1.04 for N and S America, although there are large uncertainties in these amplitudes, due to degeneracy between signals and internal variability, which is larger at these scales than global scales. Despite the uncertainties, our estimates of the scaling factors indicate that it is more likely that HadCM3 over-emphasizes rather than under-emphasizes the aerosol cooling in Eurasia, according to our analysis.

Figure 2.

Reconstructions of the contributions from G, SO and NAT to observed warming in the six continental-scale regions defined in Table 1. The shading centered on the observations shows the uncertainty region due to internal variability (two-sigma decadal variability estimated from image

[15] As we would expect from the scaling factors on G, SO, NAT being largely consistent with 1, the ALL ensemble of simulations captures many features of temperature changes seen in the 16 regions considered over the 20th century. However, there are three main discrepancies between HadCM3 and the observations.

[16] Australian temperatures are relatively stable for the first half of the century [see Figure 2, Giorgi, 2002] but this is not well captured by the ALL ensemble mean which warms more in these five decades than is observed. The regression scheme attempts to fit the observed temperatures by balancing a gradually increasing but downweighted greenhouse contribution with small aerosol and natural contributions (Figure 2). The inability of the un-scaled model simulations to reproduce observed temperature anomalies in this region imply that, whether because of model or data problems, it is not possible to reliably attribute Australian temperature changes by this analysis.

[17] The ALL ensemble mean does not capture the pattern of warming that was observed in large parts of North America and the North Atlantic in the early part of the century, and has a large region of significant differences in the tropical Atlantic ocean (Figure 3 of [Stott et al., 2000]). The exaggerated warming and cooling pattern seen in the reconstructions of N American temperature changes in the 1930s (top left panel of Figure 2) could be a spurious fit to an observed warming that is not captured by the model, either because of a missing forcing in the model or because the warming was caused mainly by a warm phase of decadal variability, perhaps due to variations in the thermohaline circulation [Delworth and Knutson, 2000].

[18] In Mid Asia (40–180E, 30–50N) the ALL ensemble mean cools during the middle of the century whereas the observations show a gradually increasing warming. This is consistent with the model systematically overestimating the climate response in Asia to sulfate aerosol in the middle of the century, as indicated by the scaling factors for SO (top middle panel of Figure 1) which are much more likely to be less than rather than greater than 1. If, on the relatively large spatial scales considered here, the pattern of climate response to black carbon aerosols is similar to that due to sulfate aerosols, the low scaling factor on SO could be related to an off-setting of the real-World cooling due to sulfate aerosols by warming from black carbon.

5. Conclusions

[19] Our results show significant anthropogenic warming trends in all the continental regions analyzed. In all these regions, greenhouse gases are estimated to have caused generally increasing warming as the century progressed, balanced to a greater or lesser degree, depending on the region, by cooling from sulfate aerosols in the middle of the century. Our analysis indicates that HadCM3 appears to overestimate aerosol cooling in Asia, although there are large uncertainties in our estimates of the scaling factors required to multiply the modelled climatic response to agree with the observed response. Uncertainty due to internal variability become greater as spatial scale reduces from the global scale to continental scales. Degeneracy between signals also contributes to uncertainty, with warming trends being consistent with greater or lesser greenhouse warming if balanced by greater or lesser aerosol cooling.

[20] In common with other attribution studies, these results are subject to uncertainties introduced by neglecting forcings, such as land use changes, and depend on internal variability being well simulated by the model. In addition, although we account for errors in the magnitude of patterns of temperature change in response to the forcings included, we do not account for errors in the patterns themselves. These sources of uncertainties are likely to become more important as spatial scale is reduced. Multi-model analyses have been shown to reduce uncertainty in attribution results by averaging different models' systematic errors [Gillett et al., 2002] and it would be interesting to investigate the robustness of these results to using different climate models. Nevertheless, with a single model, we detect anthropogenic climate change in all the continental land areas examined. Cooling from sulfate aerosols counteracts some of the greenhouse warming in many regions. It is possible that the model's overestimate of cooling from aerosols in Asia could be related to missing warming from black carbon.


[21] This work was funded by the UK Department for Environment, Food and Rural Affairs under contract PECD 7/12/37.