Online climate change projections report 3.2.11 Downscaling
In order to provide climate projections at the fine spatial scales required for UKCP09 (see Figure 1.2.a), a downscaling method is required to derive such information from our global climate model simulations, run using a horizontal resolution of ~300km. This was achieved by running simulations of a high resolution limited area regional climate model (RCM), configured from HadCM3 and run at 25 km horizontal resolution. A perturbed physics ensemble of 17 RCM variants was produced, eleven of which were eventually used in UKCP09 (as explained below). These simulations sampled uncertainties in the effects of varying regional physical processes on the simulation of fine scale detail. The simulations capture detailed regional effects of mountains, coastlines and variations in land surface properties, although they do not allow for variations of land surface types within a model grid box, in contrast to a more recent version (Essery et al. 2003) being used in additional work to provide a more sophisticated assessment of Urban Heat Island effects (see Annex 7).
Each ensemble member was driven from 1950–2100 by time series of lateral boundary conditions (atmospheric surface pressure, wind, temperature and moisture plus chemical species required for the calculation of sulphate aerosol concentrations) and surface boundary conditions (sea surface temperatures and sea ice extents) saved from a member of the PPE_A1B ensemble of HadCM3 simulations (Section 3.2.4)*. Parameter settings in each RCM ensemble member were chosen to be consistent with the settings used in the relevant HadCM3 simulation. For most parameters this was achieved simply by using the same values in both simulations, however in a few cases the parameters were adjusted to allow for known dependencies on horizontal resolution.
The RCM simulations used the domain shown in Figure 3.8, chosen so as to be large enough to avoid the risk that relaxation to GCM data at the lateral boundaries will damp the simulation of fine scale detail over interior regions of interest (e.g. Jones et al. 1995), yet small enough to minimize the risk that inconsistencies could develop between the simulations of large scale climate features in the driving GCM and nested RCM integrations (e.g. Jacob et al. 2007).
* The RCM simulations in UKCP09 are a significant development from those done for UKCIP02 in terms of resolution (25 km cf 50km), ensemble design (eleven simulations sampling modelling uncertainties cf three simulations sampling only initial state uncertainties), and length of simulation (covering 1951-2100 continuously, cf two “time slices” of 1961-1990 and 2071-2100). These developments allow us to sample a spread of possible realisations of fine scale detail throughout the 21st century in UKCP09, thus avoiding the assumption in UKCIP02 that a single “master pattern” for the 2080s can be scaled back in time to earlier periods.
In eleven ensemble members this experimental design succeeded in producing simulations of detailed climate variability and change over the UK which were physically plausible, and consistent with the driving GCM simulations of synoptic scale features (see Annex 3). In six ensemble members, however, the RCM simulations were found to be deficient in their simulations of storms and precipitation, exhibiting too little variability and too many dry days, especially in summer. This was traced to the impact of one of the parameter perturbations, involving a reduction in the order of the diffusive damping applied when calculating dynamical transport of heat, momentum and moisture. The GCM uses 6th order diffusion in its standard variant, whereas the RCM uses 4th order damping as standard (due to its finer grid). Some of our perturbed GCM simulations used 4th order diffusion (thus sampling the effect of increasing the spatial scale of the applied damping), leading to modest reductions in storminess and precipitation variability. An attempt was made to implement an equivalent perturbation in the relevant RCM simulations, moving from 4th to 2nd order diffusion with accompanying changes to the diffusion coefficient to achieve a corresponding change in damping characteristics based on theoretical calculations. However, in practice the changes had a much larger impact than anticipated in the RCM simulations, rendering their time series of winds and precipitation inconsistent with those of the driving GCM runs. These six ensemble members were therefore not used in the calibration of our downscaling procedure, summarized in the following paragraph.
Downscaling to UKCP09 target regions
The downscaling was implemented by developing regression relationships between changes simulated by the RCM over regions for which projections are required by UKCP09 (individual 25 km grid boxes and a set of administrative and river-based regions over land (Figure 1.2), plus a set of marine regions (Figure 1.4)), and changes simulated at nearby grid points in the GCM. This task bears some similarities to a traditional statistical downscaling approach, in which a set of large-scale predictor variables is used to obtain values of localized predictand variables, using relationships trained on historical observations (e.g. Wilby et al. 2004). Such methods assume that historical relationships persist into the future, however such an assumption is avoided in our case, as the relationships are trained using future changes in the predictor and predictand variables simulated by the GCM and RCM, since their purpose is to allow us to infer fine-scale changes for parts of the model parameter space for which no RCM simulation is available.
We expressed the simulated change in a given RCM variable at a given grid point as a univariate linear regression (with slope but no intercept) against the change in the same variable simulated in the GCM at a single nearby grid point. Values for five non-overlapping 30 year periods (1950–1979,1980–2009, 2010–2039, 2040–2069, 2070–2099) were expressed as changes relative to the UKCP09 baseline period of 1961–1990, and changes for all five periods and all eleven ensemble members were pooled into a single dataset for the calculation of the regression coefficient (and its associated uncertainty), and the residual unexplained variance. The residual is assumed to be normally distributed with zero mean. Figure 3.9 shows an example, in which the red lines represent the regression relationship, with residual obtained from the scatter of the black crosses about the red lines, which arises from a combination of uncertainty in the relationship between changes in the global and regional models, and also from locally generated internal variability in the RCM runs. This simple approach was used in order to minimise the risk of obtaining unrealistic relationships through overfitting. For non-coastal RCM locations over the mainland UK, the GCM point used in the regression was selected from UK land boxes in HadCM3, selecting the nearest point to the target RCM location unless an adjacent HadCM3 box could be found which explained a significantly greater portion of the variance found in the RCM response. For marine regions, a similar approach was taken, using predictors chosen from marine HadCM3 boxes nearest or adjacent to the target region. When considering coastal RCM mainland points, or points representing small islands (Channel Islands, Hebrides, Orkney, Shetland, etc), the predictor variables were selected from surrounding GCM land and sea points, to account for the possibility of a dominant maritime influence on climate at these locations.
* Some of the UKCP09 statistical calculations were performed using a transformed variable (here the natural logarithm of precipitation), which is subsequently converted back into the variable provided to users (here percentage changes in precipitation). This is done for reasons explained in Section 3.2.3.
Figure 3.9 shows close relationships between the global and regional model changes in winter. Figure 3.10 gives further examples, showing that strong relationships can also be found for summer changes, even for extreme variables subject to considerable internal variability, such as the 99th percentile of daily maximum temperature, Nevertheless, the strengths of the downscaling relationships do depend on which variable, season and region is being considered. Figure 3.11 plots the regression coefficients for changes in winter precipitation at 25 km grid squares around the UK. Significant regional variations are apparent: For example the coefficients exceed unity at many coastal locations, indicating enhanced responses in the RCMs compared with the corresponding GCM simulations, while smaller coefficients are found over parts of Wales, northern England and northern Scotland. Note that the occurrence of small regression coefficients does not necessarily indicate a failure of the downscaling method. For example, this can occur simply because: (i) the RCMs give systematically smaller changes than are found in the GCM simulations, perhaps due to the influence of regional surface topography in modifying changes found at larger scales; or (ii) because the responses in the RCM are dominated by locally generated internal variability. The region of small coefficients over central parts of northern Scotland, for example, occurs because the ratio of internal variability to forced changes is larger than in the driving GCM simulations. However, in some cases our reliance on a simple regression technique using only a single GCM predictor may limit the extent to which the relationship between forced changes in the RCM and GCM simulations is captured in the downscaling procedure.
Probabilistic projections for UKCP09 target regions were obtained by applying the calibrated downscaling relationships to probabilistic projections of 21st century climate change for the above-mentioned GCM grid boxes covering the UK and surrounding sea points, and hence obtaining estimates for the regions of Figure. 1.2 (see Section 3.2.12 for more details). In doing so, a number of limitations of our approach should be recognised. Firstly, we assume that the downscaling relationship (for a given target region and climate variable) is independent of the climate model parameter settings, and of the future period of interest. Secondly, we do not account for variations across parameter space in the skill in simulations of historical fine scale climate features found in our RCM simulations, hence the observational constraints applied to weight different parameter combinations in our Bayesian calculation (see Sections 3.2.7 and 3.2.9) are based purely on aspects of global model performance. Thirdly, we do not account for potential structural errors in our downscaling procedure, arising, for example, from our exclusive reliance on RCM variants configured from HadCM3, or (as noted above) from our neglect of more complex regression techniques based on multivariate GCM predictor variables. All of these limitations arise from the small size of our ensemble of RCM simulations: In particular, we do not possess enough simulations to emulate potential variations in fine scale characteristics of historical or future climate across parameter space. Further research in multivariate downscaling techniques and improvements in computing capacity may allow refined estimates of downscaling uncertainty to be produced in future.