Probabilistic UK Climate Projections Conditioned on Global Warming Levels

Probabilistic projections from the UK Climate Projections 2018 are presented for four global warming levels (GWLs) at 1.5, 2, 3, and 4°C above the 1850–1900 baseline. Our results show how uncertainties associated with climate models and four representative concentration pathways (RCP) emission scenarios translate to UK regional scale changes in maximum temperature and precipitation, with data also available for minimum and mean temperatures, humidity and surface net downward shortwave radiation flux. We compare weighting the likelihood of RCPs based on (hypothetical) policy decisions, against our baseline assumption that each RCP is equally likely. Differences between weighted and unweighted GWL distributions are small, particularly in relation to the full breadth of uncertainties that are incorporated into the probabilistic projections. Finally we quantify the relative importance of scenario, model and internal variability on regional projected GWLs and show that uncertainty associated with an uncertain climate response to forcings dominates at all GWLs.


Introduction
Probabilistic projections from the latest UK climate projections 2018 (UKCP18, G. R. Harris et al., 2022;Lowe et al., 2018;Murphy et al., 2018Murphy et al., , 2020) ) provide information on known uncertainties in future climate changes for a set of 21st century emission scenarios represented by four representative concentration pathways (RCPs, Moss et al., 2010).These describe future changes in atmospheric constituents, such as carbon dioxide, that drive radiative forcing arising from changes in the balance between incoming and outgoing radiation to the Earth's atmosphere.Whilst study of time-dependent responses to specific RCP pathways remains of considerable interest, there has been increasing focus on an impacts-based narrative to describes future climate change.Discussion of climate change in terms of the global-mean temperature relative to the pre-industrial period , often referred to as global warming levels (GWLs), are one such narrative.
GWLs are now an intrinsic way of defining climate change both globally, such as in the IPCC Special Report on the impacts of global warming of 1.5°C (IPCC, 2022), and nationally, such as in the UK's third Climate Change Risk Assessment (Betts & Brown, 2021;CCC, 2021) or as a supplement to HM Treasury guidance (DEFRA, 2020).Many studies have used GWLs to contextualize extreme events (e.g., Dosio et al., 2018;Hanlon et al., 2021;Klutse et al., 2018;Perry et al., 2022), climate risk indicators (Arnell et al., 2021), or to address the response of increased global temperatures in specific sectors such as human health (Jenkins et al., 2022), transport infrastructure (e.g., Mulholland & Feyen, 2021) or catastrophe risk modeling (e.g., Bates et al., 2023).However, in most of these studies the choice of data sources provides only limited sampling of the uncertainties in regional changes known to be associated with future global temperature changes.This caveat is acknowledged by authors with varying degrees of clarity.
In this paper, we use the UKCP18 probabilistic projections to provide a more representative assessment of uncertainty when framing future regional climate changes in terms of GWLs, than has otherwise been presented in most previous studies.For a given emissions pathway, uncertainty in the UKCP18 probabilistic projections incorporates the effects of internal climate variability and modeling uncertainty in atmosphere, ocean and carbon cycle processes.These uncertainties are estimated using perturbed parameter ensembles (PPEs) and multi-model ensembles and are constrained using a set of observational metrics of model performance (G.R. Harris et al., 2013;Murphy et al., 2018;Sexton et al., 2011).These UKCP18 probability distributions provide a broad representation of potential future changes for the UK and Europe.This is demonstrated, for example, by comparisons with independent estimates of uncertainty derived from other probabilistic methods (Brunner et al., 2020) and the latest CMIP6 multi-model data set (Murphy et al., 2023).We re-frame the probabilistic data to allow UKCP18 users to answer questions such as: "What will changes in UK summer rainfall look like in a 2°C world?".Note that our evaluation of GWLs is not time-bound, as per the IPCC definition (e.g., Allen et al., 2018).In the context of UKCP anomalies, the present results summarize regional climate change anomalies relative to the period 1981-2000 (UKCP18 baseline), at a GWL (e.g., 2°C) relative to 1850-1900 pre-industrial conditions.The results do not represent the regional climate change induced by a GWL of (e.g., 2°C relative to 1981-2000, see further discussion in James et al., 2017).
The data used in this paper are based exclusively on the probabilistic projections component of UKCP.Therefore, references to "Probabilistic-Global" and "Probabilistic-UK" data refer to global average surface temperature and UK data respectively from the probabilistic projections.UKCP18 also provides separate Global, Regional and Local projections products.However, data from these are not used here.

User Engagement
Engagement with the UKCP18 user network played a key role in defining the remit and outcomes of this work.Two separate user engagement sessions were conducted: a group discussion session in June 2022 and targeted interviews with 13 users in February 2023.Amongst a range of insights, these sessions highlighted that users are principally interested in 2 and 4°C GWLs, taking their lead from the UK Committee on Climate Change reports, and that there is interest in a wide range of time horizons, from near term (2030s) to long term (2080s).
Discussion with users also revealed that they are typically unsure how best to deal with multiple RCPs in their own analysis, and most frequently rely on focusing on a single RCP such as RCP8.5 (or sometimes two contrasting RCPs).Although in some use cases, focusing on RCP8.5 can be justified, a lack of guidance, methodology or summary data set was often cited as reasons for not engaging with the full (albeit limited) spectrum of RCP scenarios.These results aim to help users incorporate a wider (more realistic) range of scenarios into their downstream applications.

Method and Data
The two key conceptual aspects to the methodology are: (a) aggregating climate information across RCP scenarios, and (b) mapping GWLs to the UK, using the UKCP18 probabilistic projections of global-mean surface temperature (Probabilistic-Global) and UK variables (Probabilistic-UK).Broadly speaking, the "sub-selection" approach (as described in James et al., 2017) is the basis of Equation 2.
For a given RCP scenario, the probabilistic projections consist of 3,000 individual realizations of time-dependent climate change.These represent equally likely but different outcomes, designed to represent uncertainties derived from 360 climate model projections (G.R. Harris et al., 2022;Murphy et al., 2018).Each realization provides time series of projected monthly and seasonal anomalies (relative to 1981-2000) for a select set of UK climate variables, alongside a corresponding time series of global average surface temperature anomalies.G. R. Harris et al. (2022) use a set of emulation techniques, trained on global and regional simulations from the HadCM3 model, to expand the sampling of parametric model uncertainties (for further details, see the Supporting Information S1).Each realization replicates approximately the spatial, temporal and inter-variable relationships found in raw climate model data, within the limits of the statistical method.A linkage between global average temperature and UK variables is built into each realization through a linear pattern-scaling relationship used in one of Geophysical Research Letters 10.1029/2024GL108507 the emulation steps, ensuring that each realization provides a physically coherent relationship between global and regional changes based on climate model results.By identifying subsets of realizations whose global temperature outcomes match a given GWL, we can therefore diagnose a corresponding range of potential outcomes for UK variables.
To calculate GWLs, the 1981-2000 baseline of the Probabilistic-Global data set is first adjusted to define GWL anomalies relative to a pre-industrial baseline of 1850-1900 (see Figure S1a in Supporting Information S1).The adjustment is calculated from observations.Specifically, we use the mean difference between HadCRUT.5.0.2.0 (v202401, Morice et al., 2021) temperature anomaly periods of 1981-2000 and 1850-1900:   HadCRUT5 1981 2000 HadCRUT5 1850 1900 = 0.60°C, (1) which is consistent with the CCRA3 estimate (CCC, 2021, p. 37) and other studies (e.g., Allen et al., 2018).Since there is approximate inter-variable coherence within each probabilistic realization, the UKCP18 Probabilistic-UK regional-scale data (v20220425), joined by sample ID with the baseline-adjusted annually averaged Probabilistic-Global data set (v20190429), constitutes a basis for analysis of regional changes conditioned on GWLs of choice.For this purpose, we create a pooled data set consisting of all 3,000 sample IDs for each year and each RCP scenario (Figure S1b in Supporting Information S1).
Following the CCRA third Technical Report (Betts & Brown, 2021), a 20-year rolling mean (an interval of 9/ +10 years around the "central" date) is used to calculate GWLs (e.g., Rogelj et al., 2017) (Figure S1c in Supporting Information S1).The 20-year rolling mean is calculated for both Probabilistic-Global global mean surface temperature (GMST) and Probabilistic-UK regional data in each realization sample, and for each RCP.From the 20-year rolling mean data, GWL levels of 1.5, 2, 3, and 4°C with a tolerance of ±0.5°C are sub-selected across all RCP scenarios and probabilistic samples (Figure S1d in Supporting Information S1), with no temporal constraints on GWLs imposed, so as to maximize the available data.For example, for the 4°C GWL we do not stipulate that this level has to be reached on or after 2080 (e.g.).The 0.5°C tolerance for GMST follows the CCRA reporting preference, and will contribute to the GWL-specific ranges for UK variables through the leading-order influence of GMST on patterns of regional climate change (e.g., Tebaldi & Arblaster, 2014).In addition, a substantial component of the range in GWL-specific UK distributions will arise from the component of uncertainty associated with regional patterns of response per unit global warming (Murphy et al., 2018).For each GWL, our distributions therefore capture a range of plausible UK changes that represent our GMST tolerance plus the combined effects of natural climate variability and uncertainty in forced signals of regional climate change (Murphy et al., 2018).There is likely to be significant overlap between these regional ranges for different GWLs, particularly between the 1.5 and 2°C GWLs since the difference is the same as the designated tolerance level for sampling global temperature changes (James et al., 2017).
For the probabilistic projections data, relative probabilities in the top and bottom 5% of the distribution are considered less reliable than those in the remainder of the distribution (Met Office, 2019).Here, we consider quantiles between 0.1 and 0.9, calculated by each sub-selected GWL across all RCPs and samples, for the Probabilistic-UK variable and year (Figure S1e in Supporting Information S1).This workflow is applied separately to each region (e.g., river basins and 25 km grid cells) and temporal aggregation (months, seasons, annual mean) for which probabilistic projection data is available (Section 2.2).
In terms of high-impact low-likelihood events (e.g., Wood et al., 2023), we believe our results will have captured some of these occurrences despite only considering quantiles between 0.1 and 0.9, because (a) the distributions derived for the probabilistic projection tend to be wider than other comparable products (Brunner et al., 2020) due to our comprehensive sampling approach, and (b) we apply a ±0.5°C tolerance to our GWL definitions, so our 4°C results (e.g.) will still include some 4.5°C outcomes, as the full distribution (before clipping) will have extended to 4.5°C.

Limitations and Assumptions
There are some notable limitations to this approach.The use of the "sub-selection" method, means our results are not immediately comparable with studies based on other approaches, such as pure "pattern scaling" (e.g., Mitchell, 2003) or "time slicing" (e.g., Hanlon et al., 2021), in which GWL are associated with particular future time periods.A detailed comparison of these different approaches is summarized by James et al. (2017), albeit not in the context of UKCP data.
Although conceptually aggregating across RCPs precludes the user from considering different emissions scenarios, the summary (quantile) statistics remain constrained by the available range of possible future scenarios.It is impossible to capture the full uncertainty with respect to future global emissions based on these four scenarios alone (e.g., Knutti et al., 2008;Stephenson et al., 2012 and discussion therein).However, these four RCPs do span a variety of storylines, ranging from strong mitigation (RCP2.6) to fossil-fuel intensive (RCP8.5)emissions.Furthermore, the probabilistic projections for RCP4.5 and RCP6.0 were derived by scaling results derived from climate model simulations for RCP8.5 (described in Murphy et al., 2018, p. 28).For this reason, more complex methodological approaches to summarizing on GWLs were rejected as they would not overcome these limitations and would unnecessarily obfuscate the data provenance.As per general UKCP18 advice (Met Office, 2018), RCPs are neither forecasts nor policy recommendations, and this analysis does not represent the updated IPCC AR6 use of Shared Socioeconomic Pathways.
The RCP scenarios are not assigned relative likelihoods, therefore we weight them equally when pooling the projections data across them (e.g., as per Wigley & Raper, 2001).However, this does implicitly result in the quantile calculations for higher degrees of global warming having relatively fewer contributions from low emission scenario samples (e.g., RCP2.6). Figure 1 illustrates this effect when considering the time horizons for different GWLs.Note the circular pictograms that illustrate the number of RCP probabilistic samples that reach each of the four GWLs.For 4°C there are zero RCP2.6 and only 604 RCP4.0 probabilistic samples that reach this warming threshold.
Finally, when making the baseline adjustment of Equation 1 as the first step of the workflow (Figure S1a in Supporting Information S1), the construction of the probabilistic sample range is based on climate models simulating a range of historical changes in GMST, due to sampling of alternative values for factors such as equilibrium climate sensitivity that influence time-dependent global warming.Ideally the adjustment of the baseline across the probabilistic projections should vary to reflect differences in the historical GMST changes associated with each of the sampled realizations.However, this would have required significant methodological development, while the constant, observation-based adjustment used here follows common practice used in several recent studies.

Data for Users
We provide probabilistic data and graphics for impacts analysis for each of the four GWLs identified above (Steptoe, 2024).These are provided for each of the variables covered by the probabilistic projections, namely daily mean, maximum and minimum surface air temperature, precipitation, sea-level pressure, specific humidity, total cloud cover, total and net downward surface short-wave radiation and net downward long-wave radiation.These are provided for each of four sets of UK regions, defined on national, administrative, river-basin and gridded (25 km) bases (Fung et al., 2018).
The variables available from the probabilistic projections are monthly, seasonal and annual averages.Users requiring daily data for their GWL-based impacts analysis can obtain this from the UKCP Global, Regional or Local projections (e.g., Arnell et al., 2021;Hanlon et al., 2021), at the cost of accepting less complete representations of uncertainty.In order to combine these lines of information, users could consider developing statistical relationships between daily and monthly or seasonal information using the Global, Regional or Local products, and then applying these relationships to the probabilistic samples to infer a wider range of outcomes consistent with the uncertainties represented in the probabilistic results.

Probabilistic Projections on GWLs
Examining the proportion of probabilistic samples from all RCPs (n = 12,000) whose 20-year means reach a given GWL (based on 1850-1900 baseline) for the pooled samples from the Probabilistic-Global data (derived in Figure S1a in Supporting Information S1) highlights that for the 20-year period centered on 2090 (2080-2099), almost 2-in-5 (39%) probabilistic samples reach the 4°C GWL threshold (Figure S2 in Supporting Information S1).For 2 and 3°C GWL, an equivalent percentage (of probabilistic samples from all RCPs = 40%) is reached by 2027 and 2063 respectively.
Estimating time-horizons at GWLs (Figure 1) also draws on the baseline adjusted Probabilistic-Global projections (Figure S1a in Supporting Information S1) to illustrate the plausible future period in which a given GWL may be reached.It plots both separate RCP estimates and a simple combined estimated based on the quantiles of pooled RCPs.For GWLs of 1.5, 2, 3, and 4°C the median time-horizon of their occurrence is 2032 (5th-95th percentile interval [2012,2066]), 2050 (2028, 2078), 2072 (2052, 2086), and 2078 (2064, 2088) respectively.For each GWL threshold, pictograms also plot the relative proportion of probabilistic samples per RCP that reach each GWL.This emphasizes that multiple RCPs can reach a given GWL.For example, 1-in-5 RCP4.5 probabilistic samples reach the 4°C GWL threshold.
Figure S3 in Supporting Information S1 uses the same data as Figure 1, but visualizes data by decade.For each decade between 2020 and 2080, the 20-year mean changes in GMST are plotted for each RCP (colored boxplots) with the combined estimate (black boxplot).As there is no sub-selection by GWL, each decade plots the full 3,000 samples from the probabilistic data for each RCP.In the 20-year period centered on 2089 (2080-2099) the median changes in GMST is estimated as 3.2°C (1 d.p.) in the combined distribution with a 5th-95th percentile range of [1.8, 4.9]°C.By mid-century (2040-2059) the corresponding estimate is 2.0°C [1.4,2.7]°C.Figure 2 is an example of the regional contextualization of GWLs for monthly maximum near-surface air temperature anomalies (tasmaxAnom).Similar figures and associated data are available for all the variables and regions included in the UKCP18 probabilistic projections (Section 2.2).Figures S4-S14 in Supporting Information S1 show further variables for monthly and seasonal aggregations.Regional anomalies are presented in relation to their original baseline (1981)(1982)(1983)(1984)(1985)(1986)(1987)(1988)(1989)(1990)(1991)(1992)(1993)(1994)(1995)(1996)(1997)(1998)(1999)(2000), but sub-selected by GWL according to the years that fall within the GWL tolerance (Figure S1d in Supporting Information S1).As with other plots, quantiles are calculated over the pooled RCPs for all samples that meet this GWL criteria.

Comparing Weighted RCPs
When constructing pooled distributions, an alternative approach is to assign weights to RCPs that reflect some prior belief about varying future likelihoods of occurrence, instead of assuming each RCP to be equally likely.We consider a hypothetical weighting of the RCPs that assumes national policies will generally favor climate mitigation strategies such that weighting coefficients are RCP2.6 = 2, RCP4.5 = 4, RCP6.0 = 3, and RCP8.5 = 1, which would imply that RCP4.5 (a scenario including moderate mitigation measures) is the most likely pathway, RCP2.6 is half as likely as RCP4.5 to occur, and RCP8.5 a quarter as likely to occur, etc. Weighting essentially acts to replicate the observation associated with each RCP by its weight.
To assess differences in the weighted and unweighted distributions of UK changes for each GWL, we look at the difference in their means standardized by the corresponding range.This is known as the standardized mean difference (SMD) and also referred to as Cohen's d (Cohen, 1988), where: and x, x w are the unweighted and weighted means (by GWL) respectively and s, s w are the unweighted and weighted sample standard deviations.The SMD can be interpreted as the number of standard deviations that separate the two groups.When d = 0.2, one group is located 0.2 standard deviations from the other group and there is about 92% overlap between the distributions (e.g., Goulet-Pelletier & Cousineau, 2018;Magnusson, 2023).
We compare weighted and unweighted aggregations for maximum temperature and precipitation.Across UK administrative regions, maximum near-surface air temperature estimates (tasmaxAnom) from policy-weighted aggregation are universally cooler, with the greatest difference observed between and September and for GWLs at 3 and 4°C (Figure S15 in Supporting Information S1).For example, at 4°C GWL, un-weighted regional maximum surface air temperature anomalies are between 0.12°C (Isle of Man, April) and 0.62°C (London, July) warmer than the equivalent policy weighted estimates.However, SMDs are small (see Figure S16 in Supporting Information S1).For GWLs ≤ 3°C, SMD ≤ 0.07; for 4°C, the SMD ranges from 0.09 to 0.22.Broadly, this suggests that the differences in means between the policy weighted and unweighted distributions are negligible compared to the uncertainties that drive their ranges.
For precipitation, the greatest impacts of policy weighting are found for the 4°C GWL, with notably drier (May-September) and wetter (November-February) periods (Figure S17 in Supporting Information S1).North-west England sees the greatest wetting (3.3%, July) and North Scotland the greatest drying ( 3.0%, December).The SMDs (Figure S18 in Supporting Information S1) are generally smaller than those for tasmaxAnom.Even at 4°C GWL, most months and regions have SMD < 0.1, with the maximum SMDs generally found during June-August and November-February (max = 0.14, North Scotland, December).For GWLs ≤ 3°C, SMD ≤ 0.06.The smaller SMDs for precipitation are likely to arise (at least partly) from a larger relative contribution from natural climate variability to the UK ranges than for temperature (e.g., Lehner et al., 2020).
In contextual terms, small SMD values suggests that the uncertainty in maximum temperature and precipitation arising from the choice of scenario (in the context of our scenario aggregation) is small compared to other sources of uncertainty captured in the probabilistic projections, which sample alternative realizations of climate feedbacks arising from atmosphere, ocean and carbon cycle processes in 360 climate model simulations (Murphy et al., 2018).Given the challenge associated with justifying the weights of each RCP, we suggest that assuming equal weighting of RCPs is sufficient in the absence of better prior information on RCP likelihoods.

Scenario and Model Uncertainty on GWLs
For each (unweighted) GWL we also partition the model and scenario-based uncertainty, and climate variability after Hawkins and Sutton (2009), whereby the total uncertainty at each GWL (T GWL ) is assumed to be: where S GWL is the RCP scenario-based uncertainty (i.e., aleatoric uncertainty), calculated as the variance of the probabilistic sample means for the four scenarios; M GWL is the uncertainty derived from model responses (i.e., epistemic uncertainty), calculated as the variance across probabilistic samples, meaned across scenarios, and V GWL is the modeled climate variability, calculated as the mean of probabilistic sample variances.Unlike Hawkins and Sutton, all probabilistic samples are treated equally.This is because the 3,000 realizations are generated using an importance sampling approach that accounts for performance-based weighting of climate model responses applied as part of the Bayesian methodology underpinning the calculations (see Murphy et al., 2018).Identifying GWLs uses the same sub-selection approach described previously (and in Supporting Information S1) and is based on the 20-year rolling mean of the United Kingdom mean data for precipitation (prAnom) and temperature (tasAnom).Partitioning fractional uncertainty components follow Lehner et al. (2020), where fractions of a given component are S GWL /T GWL , M GWL /T GWL , and V GWL /T GWL respectively.
As noted by Lehner et al. (2020) this approach should only be regarded as a qualitative (rather than a quantitative) depiction of the multimodel multi-scenario uncertainty, as uncertainty components are not strictly independent of each other.In particular, V GWL is estimated by fitting a smooth polynomial function to time series of anomalies from each realization.The residuals include internally generated variability, but also (in general) a component arising from forcing agents capable of generating variability on decadal or longer time scales, including solar, volcanic and aerosol influences.
The relative importance of scenario uncertainty versus model response and climate variability, with respect to GWLs, is visualized in Figure 3, for UK average 20-year mean responses.We compare their time-varying contributions to total uncertainty when all data are aggregated against fractional contributions stratified by GWLs.For both surface air-temperature anomaly (tasAnom) and precipitation (prAnom) uncertainty derived from an uncertain climate response (i.e., "model" uncertainty for consistency with historical terminology) dominates the total uncertainty across all time periods and GWLs.In the UKCP18 context, this is primarily uncertainty in forced signals of regional climate change, derived by combining results from multi-model and PPEs with a set of constraints derived from atmosphere, ocean and carbon cycle observations (as noted in Section 1).For temperature, the uncertainty associated with radiative forcing uncertainty (i.e., "scenario uncertainty") becomes much more dominant toward the end of century, accounting for over half of the total fraction uncertainty by 2070-2089.But on GWLs, scenario uncertainty never accounts for more than a quarter of the total, reflecting the wide time period over which each GWL may be reached.For precipitation, uncertainty from climate variability, likely to derive mainly from the chaotic evolution of the climate system (i.e., "internal" variability) contributes a greater fraction of the total variance across all GWLs, but its contribution at 4°C GWL is approximately half of that at 2°C GWL.Overall, the contributions from climate variability are small as we are examining 20-year means rather than individual annual averages.

Conclusion
This study summarizes new regional climate change information over the UK, based on the UKCP18 probabilistic climate projections expressed in terms of GWLs at 1.5, 2, 3, and 4°C.Results are based on 12,000 probabilistic

Geophysical Research Letters
10.1029/2024GL108507 STEPTOE AND MURPHY samples collated across four RCP scenarios, and show that nearly two-in-five probabilistic samples reach the 4°C GWL threshold by end of the century.No samples from RCP2.6 reach the 4°C GWL.With this exception, however, significant fractions of the samples from each RCP reach each GWLs, illustrating that arbitrary exclusion of RCPs could lead to results that do not fully reflect the true uncertainty associated with translating GWLs to regional climate changes.
We illustrate the translation of GWLs to regional-level changes over the UK for maximum temperature and precipitation anomalies.The results demonstrate a more complete handling of modeling uncertainty for the available seasonal and monthly average climate variables, compared with alternative GWL products for the UK.Data for the full set of regions and variables provided by the probabilistic projections are available in Steptoe (2024).
Comparing GWL information derived from RCPs weighted by a hypothetical policy likelihood against information assuming equally weighted RCPs, show that differences between the two sets of aggregated estimates are small compared to the uncertainty in the probabilistic ensemble.Partitioning uncertainty components shows that uncertainty associated with an uncertain climate response, and in the UK-Probabilistic data context especially the uncertainty in forced signals of regional climate change dominates across all GWLs and to a greater extent than when estimating the climate change signal in time only.

Figure 1 .
Figure 1.Time horizons for global warming levels (GWLs) (relative to the pre-industrial period, 1850-1900).Combined representative concentration pathways (RCP) estimates (black boxplots), with contributing RCP components (colored boxplots) are based on UKCP18 probabilistic projection data.The contribution of each RCP is assumed to be equally likely in the combined estimate but note that the number of RCP projection members reaching each GWL varies, as shown by circular pictograms.Their annotations quantify what percentage of the 3,000 samples per RCP reach each GWL.

Figure 2 .
Figure 2.An example of combined estimates of monthly percentiles of maximum surface air temperature anomalies (tasmaxAnom, °C) relative to the 1981-2000 baseline for South West England regional average, under different global warming levels (GWLs).The anomaly value plotted at each month, for each warming level, is the combined estimate of changes pooled from all four representative concentration pathways, for those samples that reach each GWL threshold.

Figure 3 .
Figure 3. Fractional contributions to uncertainty in the Probabilistic-UK data partitioned into model response ("model," blue), representative concentration pathways scenario-based uncertainty ("scenario," green) and climate variability ("variability," orange) afterHawkins and Sutton (2009, 2011).Data is based on the 20-year rolling means of the United Kingdom (spatial-)mean annual data for precipitation (prAnom) and temperature (tasAnom).Left panels show uncertainty of the 20-year rolling mean data, and right panels the aggregated uncertainty partitions by global warming level.