Climate Change 2001:
Working Group I: The Scientific Basis
Other reports in this collection

13.5.2.2 Defining climate change signals

The question of signal-to-noise ratios in climate model simulations was alluded to above, and has also been discussed in Chapters 9 and 12. The treatment of "signal" and "noise" in constructing climate scenarios is of great importance in interpreting the results of impact assessments that make use of these scenarios. If climate scenarios contain an unspecified combination of signal plus noise, then it is important to recognise that the impact response to such scenarios will only partly be a response to anthropogenic climate change; an unspecified part of the impact response will be related to natural internal climate variability. However, if the objective is to specify the impacts of the anthropogenic climate signal alone, then there are two possible strategies for climate scenario construction:


Figure 13.8: A summer (JJA) temperature change scenario for Europe for the 2020s and 2080s. Left panel is the median scaled response of five GCM experiments available on the IPCC Data Distribution Centre (http://ipcc-ddc.cru.uea.ac.uk/) and the right panel is the inter-model range (largest scaled response minus the smallest scaled response). (Source: Hulme and Carter, 2000.)

The relative strength of the signal-to-noise ratio can be demonstrated in a number of ways. Where response patterns are reasonably stable over time, this ratio can be maximised in a climate change scenario by using long (30-year or more) averaging periods. Alternatively, regression or principal component techniques may be used to extract the signal from the model response (Hennessy et al., 1998). A third technique is to use results from multi-member ensemble simulations, as first performed by Cubasch et al. (1994). Sampling theory shows that in such simulations the noise is reduced by a factor of (n), where n is the ensemble size. Using results from the HadCM2 four-member ensemble experiments, Giorgi and Francisco (2000), for example, suggest that uncertainty in future regional climate change associated with internal climate variability at sub-continental scales (107 km2), is generally smaller than the uncertainty associated with inter-model or forcing differences. This conclusion is scale- and variable-dependent, however (see Chapter 9, Figure 9.4; see also Räisänen, 1999), and the inverse may apply at the smaller scales (104 to 105 km2) at which many impact assessments are conducted. Further work is needed on resolving this issue for climate scenario construction purposes.

A different way of maximising the climate change signal is to compare the responses of single realisations from experiments completed using different models. If the error for different models is random with zero mean, then sampling theory shows that this model average will yield a better estimate of the signal than any single model realisation. This approach was first suggested in the context of climate scenarios by Santer et al. (1990) and is illustrated further in Chapter 9, Section 9.2.2. Treating different GCM simulations in this way, i.e., as members of a multi-model ensemble, is one way of defining a more robust climate change signal, either for use in pattern-scaling techniques or directly in constructing a climate scenario. The approach has been discussed by Räisänen (1997) and used recently by Wigley (1999), Hulme and Carter (2000; see Figure 13.8) and Carter et al. (2000b) in providing regional characterisations of the SRES emissions scenarios.

The second strategy requires that the noise component be defined explicitly. This can be done by relying either on observed climate data or on model-simulated natural climate variability (Hulme et al., 1999a; Carter et al., 2000b). Neither approach is ideal. Observed climate data may often be of short duration and therefore yield a biased estimate of the noise. Multi-decadal internal climate variability can be extracted from multi-century unforced climate simulations such as those performed by a number of modelling groups (e.g., Stouffer et al., 1994; Tett et al., 1997; von Storch et al., 1997). In using AOGCM output in this way, it is important not only to demonstrate that these unforced simulations do not drift significantly (Osborn, 1996), but also to evaluate the extent to which model estimates of low-frequency variability are comparable to those estimated from measured climates (Osborn et al., 2000) or reconstructed palaeoclimates (Jones et al., 1998). Furthermore, anthropogenic forcing may alter the character of multi-decadal climate variability and therefore the noise defined from model control simulations may not apply in the future.



Other reports in this collection