IPCC Fourth Assessment Report: Climate Change 2007
Climate Change 2007: Working Group I: The Physical Science Basis Characterising and Quantifying Regional Uncertainty Review of regional uncertainty portrayed in the TAR

In the TAR, uncertainties in regional climate projections were discussed, but methods for quantifying them were relatively primitive. For example, in the TAR chapter on regional projections (Giorgi et al., 2001a), uncertainties in regional projections of climate change (e.g., large or small increases/decreases in precipitation) from different GCMs were qualitatively portrayed based only on simple agreement heuristics (e.g., seven of the nine models showed increases). Early examples of quantitative estimates of regional uncertainty include portraying the median and inter-model range of a variable (e.g., temperature) across a series of model projections and attaching probabilities to a group of scenarios on a regional scale (Jones, 2000; New and Hulme, 2000). Using multi-model ensembles

A number of studies have taken advantage of multi-model ensembles formed by GCMs that have been driven by the same forcing scenarios to generate quantitative measures of uncertainty, particularly probabilistic information at a regional scale. Table 11.3 summarises aspects of the methods reviewed in this section and in Section The results highlighted in Section 10.5 and Box 10.2 on climate sensitivity demonstrate that multi-model ensembles explore only a limited range of the uncertainty. In addition, the distribution of GCM sensitivities is not by construction a representative sample from those probability distributions and thus the regional probabilities generated using multi-model ensembles will not represent the full spread of possible regional changes.

Table 11.3. Methods for generating probabilistic information from future climate simulations at continental and sub-continental scales, SRES-scenario specific. Results from the methods of Greene et al. (2006) and Tebaldi et al. (2004a,b) are displayed in Figure 11.26.

Reference   Experiment  Input Type Spatial Scale   Methodological Assumptions  
Time Resolution Synthesis Method and Results Model Performance Evaluation 
Furrer et al. (2007)  Multi-model Ensemble  Grid points (after interpolation to common grid)  Seasonal multi- decadal averages  Bayesian approach. AOGCMs are assumed independent. Large-scale patterns projected on basis functions, small-scale modelled as an isotropic Gaussian process. Spatial dependence fully accounted for by spatial model.  Model performance not explicitly brought to bear.  
       PDFs at grid point level, jointly derived accounting forspatial dependence   
Giorgi and Mearns (2003)  Multi-model Ensemble  Regional averages (Giorgi and Francisco, 2000)  Seasonal multi- decadal averages  Cumulative Distribution Functions (CDFs) derived by counting threshold exceedances among members, and weighing the counts by the REA method.  Model performance (bias and convergence)explicitly quantified in each AOGCMs’ weight.Observable at same spatial scale and time resolution, for period 1961 to 1990.  
       Stepwise CDFs at the regional levels   
Greene et al. (2006)  Multi-model Ensemble  Regional averages (Giorgi and Francisco, 2000)  Seasonal and annual averages  Bayesian approach. AOGCMs dependence is modelled. Linear regression of observed values on model’s values (similar to Model-Output-Statistics approach used in weather forecasting and seasonal forecasting) withcoefficients estimates applied to future simulations.  Model performance measured on 1902 to 1998historical trend reproduction at same spatialscale and time resolution. 
       PDFs at regional level    
Harris et al. (2006)  Perturbed Physics Ensemble (PPE)  Grid points  Seasonal multi- annual averages  Scaled equilibrium response patterns from a large slab-model Perturbed Physics Ensemble (PPE), using transient responsesof an Energy Balance Model driven by PPE climatefeedbacks. Quantifying scaling error, against a smaller PPEof transient simulations, to include in PDFs. All model versions assumed equally likely. 
       PDFs at arbitrary level of aggregation   
Stott et al. (2006a)  Single Model (HadCM3)  Continental averages  Annual decadal averages  Linear scaling factor estimated through optimal fingerprinting approach at continental scales or at global scale and appliedto future projections, with estimated uncertainty. Naturalvariability estimated from control run added as additionaluncertainty component.  Not applicable 
       PDFs at the continental-scale level   
Tebaldi et al. (2004a,b)  Multi-model Ensemble  Regional averages (Giorgi and Francisco, 2000)  Seasonal multi- decadal averages  Bayesian approach. AOGCMs are assumed independent. Normal likelihood for their projections, with AOGCM-specific variability.  Model performance (bias and convergence) implicitly brought to bear through likelihoodassumptions. Observable at same spatial scaleand time resolution, for period 1961 to 1990 in original papers, for period 1980 to 1999 for results displayed in this report.  
       PDFs at the regional level  

Räisänen and Palmer (2001) used 17 GCMs forced with an idealised annual increase in atmospheric CO2 of 1% to calculate the probability of exceedance of thresholds of temperature increase (e.g., >1°C) and precipitation change (e.g., <–10%). These were used to demonstrate that a probabilistic approach has advantages over conventional deterministic estimates by demonstrating the economic value of a probabilistic assessment of future climate change. Giorgi and Mearns (2002) developed measures of uncertainty for regional temperature and precipitation change by weighting model results according to biases in their simulation of present-day climate and convergence of their projections to the ensemble’s mean. Their Reliability Ensemble Average (REA) method was applied to the nine GCMs assessed in the TAR to provide uncertainty estimates separately for the SRES A2 and B2 emission scenarios for 22 large sub-continental regions.

Tebaldi et al. (2004a,b) used a Bayesian approach to define a formal statistical model for deriving probabilities from an ensemble of projections forced by a given SRES scenario. Using the Giorgi and Mearns (2002, 2003) approach, model bias and convergence criteria determine the shape and width of the posterior probability density functions (PDFs) of temperature and precipitation change signals. Expert judgement can be incorporated in the form of prior distributions that have the effect of assigning different relative weights to the two criteria (Tebaldi et al., 2004b; Lopez et al., 2006). The method developed by Furrer et al. (2007) to combine GCM output at the grid point scale into probabilistic projections is described in detail in Chapter 10. By straightforward area averaging, PDFs of climate change at the regional scale can be obtained. When this is done for the Giorgi and Francisco (2000) regions, the regional PDFs from Furrer et al. (2007) agree overall with the empirical histogram of the ensemble projections and the Tebaldi et al. (2004b) PDFs, with relatively small differences in spread and generally no clear difference in location.

Greene et al. (2006) used a Bayesian framework to model an ensemble of GCM projections under individual SRES scenarios by an extension of methods used for seasonal ensemble forecasting. The set of GCM simulations of the observed period 1902 to 1998 are individually aggregated in area-averaged annual or seasonal time series and jointly calibrated through a linear model to the corresponding observed regional trend. The calibration coefficients and their uncertainty are estimated and then applied to the future projections to provide probabilistic forecasts of future trends. Two critical assumptions are responsible for this method’s results being so different from the ensemble projections or the PDFs produced by Tebaldi et al. (2004a,b) (see Figure 11.26 and Supplementary Material Figures S11.33 to S11.35). Firstly, the method attributes large uncertainty to models that are unable to reproduce historical trends despite the uncertainty in the relatively weak forcings in the historical period and the large natural variability at regional scales. Second, a strong stationarity assumption is required to extrapolate the relationship derived over the historical record to future trends, which involve a different combination of and some significantly stronger forcings. The significantly smaller warming and the large width of the PDFs (at times including negative values) labelled by a ‘G’ in Figure 11.26 are then interpretable as a result of this stationarity constraint and the large uncertainty in the fitting of the trends. They contrast starkly with the larger warming represented in the histograms of model projections and their synthesis in the Tebaldi et al. (2004a,b) and in the Furrer et al. (2007, not shown) PDFs. This is particularly so in the lower-latitude regions of Africa, South Asia and the SH, possibly as a consequence of particularly weak trends in the observations and/or relatively worse performance of the GCMs.

Figure 11.26

Figure 11.26. Map comparing PDFs of change in temperature (2080 to 2099 compared to 1980 to 1999) from Tebaldi et al. (2004a,b) and Greene et al. (2006) as well as the raw model projections (represented by shaded histograms) for the Giorgi and Francisco (2000) regions. Areas under the curves and areas covered by the histograms have been scaled to equal unity. The scenario is SRES A1B and the season is NH winter (DJF). Asterisks adjacent to ARC and ANT regions indicate that only the Tebaldi et al. results were available.

Dessai et al. (2005) apply the idea of simple pattern scaling (Santer et al., 1990) to a multi-model ensemble of AOGCMs. They ‘modulate’ the normalised regional patterns of change by the global mean temperature changes generated under many SRES scenarios and climate sensitivities through the Model for the Assessment of Greenhouse-Gas Induced Climate Change (MAGICC), a simple probabilistic energy balance model (Wigley and Raper, 2001). Their work focuses on measuring the changes in PDFs as a function of different sources of uncertainty. In this analysis, the impact of the SRES scenarios turns out to be the most relevant for temperature changes, particularly in the upper tail of the distributions, while the GCM weighting does not produce substantial differences. This result is probably dependent on the long horizon of the projections considered (late 21st century). Arguably, the emission scenario would be less important in the short to mid-term. Climate sensitivity has an impact mainly in the lower tail of the distributions. For precipitation changes, all sources of uncertainty seem relevant but the results are very region-specific and thus difficult to generalise. More work to test the robustness of these conclusions is needed, especially when these are obviously not consistent with the results in Figure 10.29. For example, the use of pattern scaling is likely to underestimate the range of projections that would be obtained by running a larger ensemble of GCMs (Murphy et al., 2004).

The work described above has involved either large-area averages of temperature and precipitation change or statistical modelling at the grid box scale. Good and Lowe (2006) show that trends in large-area and grid-box average projections of precipitation are often very different from the local trends within the area. This demonstrates the inadequacy of inferring the behaviour at fine scales from that of large-area averages. Using perturbed physics ensembles

Another method for exploring uncertainties in regional climate projections is the use of large perturbed physics ensembles (described in detail in Chapter 10). These allow a characterisation of the uncertainty due to poorly constrained parameters within the formulation of a model. Harris et al. (2006) combined the results from a 17-member ensemble (Collins et al., 2006) with a larger perturbed physics ensemble, investigating the equilibrium climate response to a doubling of atmospheric CO2 (Webb et al., 2006). They developed a bridge between spatial patterns of the transient and equilibrium climate response by way of simple pattern scaling (Santer et al., 1990), allowing results from the large ensemble to be translated into PDFs of time-dependent regional changes. Uncertainties in surface temperature and precipitation changes are derived (Supplementary Material Figures S11.36 and S11.37), which arise from the poorly constrained atmospheric model parameters, internal variability and pattern scaling errors. The latter are quantified by comparing the scaled equilibrium response with the transient response for 17 model versions with identical parameter settings. Errors introduced by the pattern-scaling technique are largest when the transient response varies nonlinearly with global temperature, as is the case for precipitation in certain regions. Other approaches to quantifying regional uncertainty

As described in Chapter 10, Stott and Kettleborough (2002) provide PDFs of future change in climate by making use of the robust observational constraints on a climate model’s response to greenhouse gas and sulphate aerosol forcings that underpin the attribution of recent climate change to anthropogenic sources. The study by Stott et al. (2006a) is the first to adapt this method for continental scales. It considers two methods of constraining future continental temperature projections, one based on using observed historical changes only over the region of interest and one based on using observed changes in global temperature patterns. The first approach produces wider PDFs, since the uncertainty of detection at the regional scale is larger. The second approach incorporates more information, hence reducing the uncertainty, but assumes that the GCM represents correctly the relationship between global mean and regional temperature change. In contrast to the studies of Section, this work uses projections from a single GCM (HadCM3), although Stott et al. (2006b) have confirmed the results of this methodology for other models.

In general, the regional sections of this chapter assess the uncertainty in regional changes based on expert understanding of the relevant processes, rather than by formal probabilistic methods, which are still in their infancy and currently do not provide definitive results. An approach to a process-based assessment of the reliability of modelled climate change responses and thus uncertainties in its future projections has been proposed by Rowell and Jones (2006). They perform an assessment of the physical and dynamical mechanisms responsible for a specific future outcome, in their case European summer drying. Their analysis isolates the contribution of the four major mechanisms analysed: the spatial pattern of warming, other large-scale changes, reduced spring soil moisture and summer soil moisture feedbacks. In certain regions, the second process makes a minor contribution with the first and third dominating. This leads to the conclusion that the sign of the change is robust as confidence in the processes underlying these mechanisms is high. Combined uncertainties: General Circulation Models, emissions and downscaling techniques

It is important to quantify the relative importance of the uncertainty arising from the downscaling step (from the RCM formulation or the assumptions underlying an empirical SD method) against the other sources of uncertainty. For example, in the application of SD methods to probabilistic scenarios, Benestad (2002b, 2004a) used a multi-model ensemble coupled to SD to derive tentative probabilistic scenarios at a regional scale for northern Europe.

The PRUDENCE project (Box 11.2) provided the first opportunity to weigh these various sources of uncertainty for simulations over Europe. Rowell (2006) Errata evaluated a four-dimensional matrix of climate modelling experiments that included two different emissions scenarios, four different GCM experiments, multiple ensemble members within the latter to assess internal variability, and nine different RCMs, for the area of the British Isles. He found that the dynamical downscaling added a small amount of uncertainty compared to the other sources for temperature evaluated as monthly/seasonal averages. For precipitation, the relative contributions of the four sources of uncertainty are more balanced. Déqué et al. (2005, 2007) show similar results for the whole of Europe, as do Ruosteenoja et al. (2007) for subsections of Europe. Kjellström et al. (2007) find that the differences among different RCMs driven by the same GCM become comparable to those among the same RCM driven by different GCMs when evaluating daily maximum and minimum temperatures. However, mean responses in the PRUDENCE RCMs were often quite different from that of the driving GCM. This suggests that some of the spread in RCM responses may be unrealistic due to model inconsistency (Jones et al., 1997). However, it should be noted that only a few of the RCMs in PRUDENCE were driven by more than one GCM, which adds further uncertainty regarding these conclusions. Other programs similar to PRUDENCE have begun for other regions of the world, such as NARCCAP over North America (Mearns et al., 2005), Regional Climate Change Scenarios for South America (CREAS; Marengo and Ambrizzi, 2006), and the Europe-South America Network for Climate Change Assessment and Impact Studies (CLARIS; http://www.claris-eu.org) over South America.