9.4.2.2 Studies Based on Space-Time Patterns

Global-scale analyses using space-time detection techniques (Section 9.4.1.4) have robustly identified the influence of anthropogenic forcing on the 20th-century global climate. A number of studies have now extended these analyses to consider sub-global scales. Two approaches have been used; one to assess the extent to which global studies can provide information at sub-global scales, the other to assess the influence of external forcing on the climate in specific regions. Limitations and problems in using smaller spatial scales are discussed at the end of this section.

The approach taken by IDAG (2005) was to compare analyses of full space-time fields with results obtained after removing the globally averaged warming trend, or after removing the annual global mean from each year in the analysis. They find that the detection of anthropogenic climate change is driven by the pattern of the observed warming in space and time, not just by consistent global mean temperature trends between models and observations. These results suggest that greenhouse warming should also be detectable at sub-global scales (see also Barnett et al., 1999). It was also shown by IDAG (2005) that uncertainties increase, as expected, when global mean information, which has a high signal-to-noise ratio, is disregarded (see also North et al., 1995).

Another approach for assessing the regional influence of external forcing is to apply detection and attribution analyses to observations in specific continental- or sub-continental scale regions. A number of studies using a range of models and examining various continental- or sub-continental scale land areas find a detectable human influence on 20th-century temperature changes, either by considering the 100-year period from 1900 or the 50-year period from 1950. Stott (2003) detects the warming effects of increasing greenhouse gas concentrations in six continental-scale regions over the 1900 to 2000 period, using HadCM3 simulations. In most regions, he finds that cooling from sulphate aerosols counteracts some of the greenhouse warming. However, the separate detection of a sulphate aerosol signal in regional analyses remains difficult because of lower signal-to-noise ratios, loss of large-scale spatial features of response such as hemispheric asymmetry that help to distinguish different signals, and greater modelling and forcing uncertainty at smaller scales. Zwiers and Zhang (2003) also detect human influence using two models (CGCM1 and CGCM2; see Table 8.1, McAvaney et al., 2001) over the 1950 to 2000 period in a series of nested regions, beginning with the full global domain and descending to separate continental domains for North America and Eurasia. Zhang et al. (2006) update this study using additional models (HadCM2 and HadCM3). They find evidence that climates in both continental domains have been influenced by anthropogenic emissions during 1950 to 2000, and generally also in the sub-continental domains (Figure 9.11). This finding is robust to the exclusion of NAO/Arctic Oscillation (AO) related variability, which is associated with part of the warming in Central Asia and could itself be related to anthropogenic forcing (Section 9.5.3). As the spatial scales considered become smaller, the uncertainty in estimated signal amplitudes (as demonstrated by the size of the vertical bars in Figure 9.11) becomes larger, reducing the signal-to-noise ratio (see also Stott and Tett, 1998). The signal-to-noise ratio, however, also depends on the strength of the climate change and the local level of natural variability, and therefore differs between regions. Most of the results noted above hold even if the estimate of internal climate variability from the control simulation is doubled.

Figure 9.11. Scaling factors indicating the match between observed and simulated decadal near-surface air temperature change (1950–1999) when greenhouse gas and aerosol forcing responses (GS) are taken into account in ‘optimal’ detection analyses (Appendix 9.A), at a range of spatial scales from global to sub-continental. Thick bars indicate 90% confidence intervals on the scaling factors, and the thin extensions indicate the increased width of these confidence intervals when estimates of the variance due to internal variability are doubled. Scaling factors and uncertainties are provided for different spatial domains including Canada (Canadian land area south of 70°N), China, Southern Europe (European land area bounded by 10°W to 40°E, 35° to 50°N), North America (North American land area between 30°N and 70°N), Eurasia (Eurasian land area between 30°N and 70°N), mid-latitude land area between 30°N and 70°N (labelled NH-land), the NH mid-latitudes (30°N to 70°N including land and ocean), the NH, and the globe. The GS signals are obtained from CGCM1 and CGCM2 combined (labelled CGCM, see Table 8.1 of the TAR), HadCM2 (see Table 8.1 of the TAR), and HadCM3 (see Table 8.1, this report), and these four models combined (‘ALL’). After Zhang et al. (2006) and Hegerl et al. (2006b).

The ability of models to simulate many features of the observed temperature changes and variability at continental and sub-continental scales and the detection of anthropogenic effects on each of six continents provides stronger evidence of human influence on climate than was available to the TAR. A comparison between a large ensemble of 20th-century simulations of regional temperature changes made with the MMD at PCMDI (using the same simulations for which the global mean temperatures are plotted in Figure 9.5) shows that the spread of the multi-model ensembles encompasses the observed changes in regional temperature changes in almost all sub-continental regions (Figure 9.12; see also FAQ 9.2, Figure 1 and related figures in Chapter 11). In many of the regions, there is a clear separation between the ensembles of simulations that include only natural forcings and those that contain both anthropogenic and natural forcings. A more detailed analysis of one particular model, HadCM3, shows that it reproduces many features of the observed temperature changes and variability in the different regions (IDAG, 2005). The GFDL-CM2 model (see Table 8.1) is also able to reproduce many features of the evolution of temperature change in a number of regions of the globe (Knutson et al., 2006). Other studies show success at simulating regional temperatures when models include anthropogenic and natural forcings. Wang et al. (2007) showed that all MMD 20C3M simulations replicated the late 20th-century arctic warming to various degrees, while both forced and control simulations reproduce multi-year arctic warm anomalies similar in magnitude to the observed mid 20th-century warming event.

Figure 9.12. Comparison of multi-model data set 20C3M model simulations containing all forcings (red shaded regions) and containing natural forcings only (blue shaded regions) with observed decadal mean temperature changes (°C) from 1906 to 2005 from the Hadley Centre/Climatic Research Unit gridded surface temperature data set (HadCRUT3; Brohan et al., 2006). The panel labelled GLO shows comparison for global mean; LAN, global land; and OCE, global ocean data. Remaining panels display results for 22 sub-continental scale regions (see the Supplementary Material, Appendix 9.C for a description of the regions). This figure is produced identically to FAQ 9.2, Figure 1 except sub-continental regions were used; a full description of the procedures for producing FAQ 9.2, Figure 1 is given in the Supplementary Material, Appendix 9.C. Shaded bands represent the middle 90% range estimated from the multi-model ensemble. Note that the model simulations have not been scaled in any way. The same simulations are used as in Figure 9.5 (58 simulations using all forcings from 14 models, and 19 simulations using natural forcings only from 5 models). Each simulation was sampled so that coverage corresponds to that of the observations, and was centred relative to the 1901 to 1950 mean obtained by that simulation in the region of interest. Observations in each region were centred relative to the same period. The observations in each region are generally consistent with model simulations that include anthropogenic and natural forcings, whereas in many regions the observations are inconsistent with model simulations that include natural forcings only. Lines are dashed where spatial coverage is less than 50%.

There is some evidence that an anthropogenic signal can now be detected in some sub-continental scale areas using formal detection methods (Appendix 9.A.1), although this evidence is weaker than at continental scales. Zhang et al. (2006) detect anthropogenic fingerprints in China and southern Canada. Spagnoli et al. (2002) find some evidence for a human influence on 30-year trends of summer daily minimum temperatures in France, but they use a fingerprint estimated from a simulation of future climate change and do not detect an anthropogenic influence on the other indices they consider, including summer maximum daily temperatures and winter temperatures. Min et al. (2005) find an anthropogenic influence on East Asian temperature changes in a Bayesian framework, but they do not consider anthropogenic aerosols or natural forcings in their analysis. Atmosphere-only general circulation model (AGCM) simulations forced with observed SSTs can potentially detect anthropogenic influence at smaller spatial and temporal scales than coupled model analyses, but have the weakness that they do not explain the observed SST changes (Sexton et al., 2003). Two studies have applied attribution analysis to sub-continental temperatures to make inferences about changes in related variables. Stott et al. (2004) detect an anthropogenic influence on southern European summer mean temperature changes of the past 50 years and then infer the likelihood of exceeding an extreme temperature threshold (Section 9.4.3.3). Gillett et al. (2004a) detect an anthropogenic contribution to summer season warming in Canada and demonstrate a statistical link with area burned in forest fires. However, the robustness of these results to factors such as the choice of model or analysis method remains to be established given the limited number of studies at sub-continental scales.

Knutson et al. (2006) assess temperature changes in regions of the world covering between 0.3 and 7.4% of the area of the globe and including tropical and extratropical land and ocean regions. They find much better agreement between climate simulations and observations when the models include rather than exclude anthropogenic forcings, which suggests a detectable anthropogenic warming signal over many of the regions they examine. This would indicate the potential for formal detection studies to detect anthropogenic warming in many of these regions, although Knutson et al. (2006) also note that in some regions the climate simulations they examined were not very realistic and showed that some of these discrepancies are associated with modes of variability such as the AO.

Karoly and Wu (2005) compare observed temperature trends in 5° × 5° grid boxes globally over 30-, 50- and 100-year periods ending in 2002 with 1) internal variability as simulated by three models (GFDL R30, HadCM2, PCM) and 2) the simulated response to greenhouse gas and sulphate aerosol forcing in those models (see also Knutson et al., 1999). They find that a much higher percentage of grid boxes show trends that are inconsistent with model-estimated internal variability than would be expected by chance and that a large fraction of grid boxes show changes that are consistent with the forced simulations, particularly over the two shorter periods. This assessment is essentially a global-scale detection result because its interpretation relies upon a global composite of grid-box scale statistics. As discussed in the paper, this result does not rule out the possibility that individual grid box trends may be explained by different external forcing combinations, particularly since natural forcings and forcings that could be important at small spatial scales, such as land use change or black carbon aerosols, are missing from these models. The demonstration of local consistency between models and observations in this study does not necessarily imply that observed changes can be attributed to anthropogenic forcing in a specific grid box, and it does not allow confident estimates of the anthropogenic contribution to change at those scales.

Models do not reproduce the observed temperature changes equally well in all regions. Areas where temperature changes are not particularly well simulated by some models include parts of North America (Knutson et al., 2006) and mid-Asia (IDAG, 2005). This could be due to a regional trend or variation that was caused by internal variability (a result that models would not be expected to reproduce), uncertain forcings that are locally important, or model errors. Examples of uncertain forcings that play a small role globally, but could be more important regionally, are the effects of land use changes (Sections 9.2 and 9.3) or atmospheric brown clouds. The latter could be important in explaining observed temperature trends in South Asia and the northern Indian Ocean (Ramanathan et al., 2005; see Chapter 2).

An analysis of the MMD 20C3M experiments indicates that multi-decadal internal variability could be responsible for some of the rapid warming seen in the central USA between 1901 and 1940 and rapid cooling between 1940 and 1979 (Kunkel et al., 2006). Also, regional temperature is more strongly influenced by variability and changes in climate dynamics, such as temperature changes associated with the NAO, which may itself show an anthropogenic influence (Section 9.5.3.2), or the Atlantic Multi-decadal Oscillation (AMO), which could in some regions and seasons be poorly simulated by models and could be confounded with the expected temperature response to external forcings. Thus the anthropogenic signal is likely to be more easy to identify in some regions than in others, with temperature changes in those regions most affected by multi-decadal scale variability being the most difficult to attribute, even if those changes are inconsistent with model estimated internal variability and therefore detectable.

The extent to which temperature changes at sub-continental scales can be attributed to anthropogenic forcings, and the extent to which it is possible to estimate the contribution of greenhouse gas forcing to regional temperature trends, remains a topic for further research. Idealised studies (e.g., Stott and Tett, 1998) suggest that surface temperature changes are detectable mainly at large spatial scales of the order of several thousand kilometres (although they also show that as the signal of climate change strengthens in the 21st century, surface temperature changes are expected to become detectable at increasingly smaller scales). Robust detection and attribution are inhibited at the grid box scales because it becomes difficult to separate the effects of the relatively well understood large-scale external influences on climate, such as greenhouse gas, aerosols, solar and volcanic forcing, from each other and from local influences that may not be related to these large-scale forcings. This occurs because the contribution from internal climate variability increases at smaller scales, because the spatial details that can help to distinguish between different forcings at large scales are not available or unreliable at smaller scales, and because forcings that could be important at small spatial scales, such as land use change or black carbon aerosols, are uncertain and may not have been included in the models used for detection. Although models do not typically underestimate natural internal variability of temperature at continental scales over land (Figure 9.8), even at a grid box scale (Karoly and Wu, 2005), the credibility of small-scale details of climate simulated by models is lower than for large-scale features. While the large-scale coherence of temperatures means that temperatures at a particular grid box should adequately represent a substantial part of the variability of temperatures averaged over the area of that grid box, the remaining variability from local-scale processes and the upward cascades from smaller to larger scales via nonlinear interactions may not be well represented in models at the grid box scale. Similarly, the analysis of shorter temporal scales also decreases the signal-to-noise ratio and the ability to use temporal information to distinguish between different forcings. This is why most detection and attribution studies use temporal scales of 50 or more years.