Online climate change projections report Annex 2.3
The above tests consider variations in specific aspects of our methodology, however it is also important to consider how different the results could have been had we chosen an entirely different approach. Here, the first point is that while a number of methods for probabilistic climate projection have been published in the research literature, we are not aware of any that have been designed to sample uncertainties as comprehensively as is done in UKCP09 (for example there are several methods which sample uncertainties in physical climate system processes, but none which combines these with uncertainties in both carbon cycle processes and downscaling). This is because it is acceptable in academic studies to explore methodologies which are conditional upon the omission of important known sources of uncertainty, however this would not be acceptable in a project like UKCP09, since our aim is to produce information suitable to support user decisions in the real world. So we cannot compare UKCP09 against some competing approach designed to produce probabilities with the same level of decision-relevance.
However, by omitting some elements of the UKCP09 approach we can compare it against alternative methodologies conditional on sampling similar subsets of the uncertainties in climate projection. For example, a number of approaches have been suggested in which probabilistic projections are derived purely from results from a multi-model ensemble of global coupled ocean-atmosphere models of typically 10–20 members (Tebaldi and Knutti (2007) review several of these), rather than our approach of using larger ensembles of model variants specifically designed to sample uncertainties, with multi-model ensemble results playing a significant but more subsidiary role. Some of the multi-model approaches are nevertheless similar to ours in their basic character, in that they seek to construct a range of alternative projections which express the effects of uncertainties arising from modelling errors, and then adjust these according to some set of observational constraints. Another class of approaches seeks to project future changes explicitly designed to be consistent with uncertainties in some set of observations of recent climate, using climate model results to provide the necessary relationships between historical observations and future changes (e.g. Piani et al. (2005), Knutti et al. (2006), Sanderson et al. (2008)). Closely related to these are approaches which seek to project future changes by assuming a linear relationship between errors in past and future changes, constraining future changes according to the range of past errors consistent with observations (Allen et al. (2000), Stott and Kettleborough (2002), Stott et al. (2006a)).
We compare our projections for annual mean temperature with those made by a method of the latter type, based on Stott et al. (2006a). Their method uses model simulations and historical observations of changes in surface temperature during the 20th century to derive a distribution of alternative scaling factors which can be applied to the simulated changes to fit the observed changes to a level consistent with uncertainties in the latter. The distribution of scaling factors is then applied to the future model response to produce a probabilistic climate projection. Stott et al. (2006a) produced two versions of this technique: The first version projected future regional changes according to past changes in the same region (thus obtaining relatively conservative estimates of uncertainty by neglecting possible constraints from aspects of past change remote to the region of interest); the second version scaled future regional changes according to errors in past spatial and temporal patterns of change over the whole globe (thus obtaining narrower estimates of uncertainty, although this does not take account of possible errors in the regional pattern of response, since it scales the model’s pattern of response over the whole globe by the same factor, with uncertainties, for each region). We use an updated version which accounts for past changes in global patterns of surface temperature, thus removing the contrasting limitations of the two earlier techniques. The Stott et al. method provides projections for large regions (no downscaling method is included), and does not account for uncertainties in future changes in radiative forcing arising from carbon cycle processes. Therefore, we consider a like-for-like comparison of projections of spatially averaged temperature for the whole of northern Europe, applying the UKCP09 methodology without downscaling, and with no sampling of the effects of future uncertainties in climate feedbacks involving the carbon cycle (by holding these feedbacks fixed at values diagnosed from the standard published variants of the relevant configurations of HadCM3). Both methods assume that there is a negligible effect from other possible sources of uncertainty in either historical forcing (e.g. black carbon) or future changes (e.g. methane cycle) — see Box 2.1, Chapter 2.
We applied the Stott et al. method to each of the 17 members of our PPE_A1B ensemble of perturbed variants of HadCM3 (Section 3.2.4 and Figure 3.2), obtaining projections with associated uncertainties from each ensemble member, and combining these to form probabilistic projections shown by the blue curves in Figure A2.4. The results show that the median projection of future changes is slightly smaller in the UKCP09 method. The UKCP09 method also produces a slightly wider spread from 2010 onwards, but a somewhat narrower spread during the historical period. Uncertainties from UKCP09 broaden by including a more complete sampling of the possible uncertainties arising from parameter choices in models and structural model errors common to model projections, and narrow by including a wider range of observational constraints, whereas the Stott et al. uncertainties rely on linear scaling of available model simulations based on a more limited range of observational constraints. Such differences could serve to broaden or narrow the UKCP09 uncertainty ranges relative to the Stott et al. uncertainty ranges, dependent on their competing influences. A detailed examination of these differences is beyond the scope of this report.
The Stott et al. method is set up to provide projections which are relatively conservative (in the sense that only one relatively well understood observational constraint is used), and which minimise their dependence on the set of climate model simulations used to produce them (Stott et al. 2006b). Projections derived from this technique will be determined by the scaling factors, and associated uncertainties, found by matching simulated and observed realisations of the past climate warming attributable to human activity. On the other hand, the UKCP09 approach is based on a different philosophy which seeks to place more weight on detailed aspects of climate system physics, both by sampling possible variations in these more widely, and then seeking to constrain them with a wider range of observations. It is therefore reassuring that two methods based on different principles and assumptions should give relatively similar projections in practice. This further supports the results of Figure A2.1 in indicating that the UKCP09 projections are likely to be reasonably robust to the key assumptions involved in their generation.