Changing population dynamics and uneven temperature emergence combine to exacerbate regional exposure to heat extremes under 1.5 °C and 2 °C of warming

This analysis employs atmosphere-only model simulations using an experimental framework developed under the 'Half a degree Additional warming, Prognosis and Projected Impacts' (HAPPI) project (Mitchell et al 2016b). The HAPPI model framework prescribes sea surface temperatures and other boundary conditions (sea ice, greenhouse gases, aerosols) that would be consistent with the current climate, as well as for future worlds under 1.5 °C and 2 °C of global mean warming (see Mitchell et al (2017) for further details of the experimental setup). We apply these boundary conditions, via the Weather@Home (W@H) distributed computing framework, to produce thousands of atmosphere-only regional climate model simulations over a spatial domain encompassing a large region of Asia, the Middle East and Eastern Africa (hereafter SAEA region, figure 1). The W@H project utilises the spare computing power of volunteers to run thousands of simulations of the global atmosphere-only model HadAM3P (horizontal resolution of 1.875° × 1.25°), with a nested regional model (HadRM3P) operating over SAEA at a horizontal resolution of approximately 50 kilometres (see Guillod et al (2017) for further details of the W@H modelling framework). This combination of very large model ensembles (between 1500 and 2500 model years per experiment, see table S1) with high-resolution climate data provides a unique opportunity to accurately characterise population exposure to exceptionally severe heat events under future warming.

Zoom In Zoom Out Reset image size

To quantify changes in extreme heat across the SAEA region under future warming scenarios, we first calculate five-day running mean values of daily maximum temperatures at each grid point in the region, and then extract the maximum value found across the period spanning April to September for each model year (hereafter Tx5DX). This yields 2692, 1785 and 1863 estimates of Tx5DX at each grid point for model simulations of the present-decade climate, 1.5 °C and 2 °C worlds, respectively (see supplementary table S1 available at stacks.iop.org/ERL/13/034011/mmedia). The choice of a five-day window is subjective, but is intended to reflect the approximate length of heat events for which severe impacts manifest themselves across the regions of interest.

The threshold of Tx5DX which occurs for one-in-ten model years under each of the future scenarios is then identified for each grid cell, and we then quantify the probability of this same threshold being exceeded in the present-decade runs. We then calculate the ratio of the probability of occurrence in the future (one in ten, by definition) relative to the probability of occurrence today, and denote this term the risk ratio (or RR₁₀). Hence, this metric represents the increase in frequency of future extreme heat, relative to the present-day climate.

While previous studies provide confidence in the capability of HadAM3P to faithfully simulate high-temperature extremes (Uhe et al 2016, Guillod et al 2017), the decision to present changes between warming levels in a relative framework (rather than in terms of absolute temperatures) ensures raw model output can be used directly, and avoids the need to make subjective decisions about what bias correction techniques might be most appropriate (Sippel et al 2016).

Finally, while the risk ratio metric applied here remains consistent with the approach employed by many previous studies (Stott et al 2016, NAS 2016), it is noted that our method differs by defining an event threshold with respect to the future world, rather than the present-day world. This choice was made to ensure that estimates of very small probabilities will be calculated with the experiment which has the largest ensemble size. Further, it is noted that the choice of a one-in-ten year event in the future is an arbitrary threshold, but enables consideration of the most severe heat extremes which could occur in the future.

Zoom In Zoom Out Reset image size

Figures 1(a) and (b) present ensemble-mean increases in Tx5DX across the SAEA region under the 1.5 °C and 2 °C experiments, relative to the present-decade experiment. Both results show regions of faster-than-average warming in higher latitudes and in continental interiors, while less rapid signals are found in coastal regions and at lower latitudes—such patterns are consistent with expected changes to absolute temperature under future warming (Joshi et al 2008, Collins et al 2013).

To interpret the relative significance of these future changes in extreme heat, figures 1 (c) and (d) show the same results as for panels a and b, but this time normalised by the variability in Tx5DX experienced in the current climate (taken as the standard deviation across all 'present-decade' model runs, calculated separately for each individual grid box). When warming is kept to 1.5 °C, changes remain relatively modest across the full region, with most places experiencing changes in Tx5DX of less than 1σ (with the exception of parts of the Middle East). But after 2.0 °C of warming, changes exceeding 2 standard deviations can be found across large regions of equatorial Africa, as well as large sections of Indonesia, the Middle East and Iran. Such changes are all the more striking when ensemble-mean estimates of RR₁₀ are considered (figures 1(e) and (f)): while under 1.5 °C of warming, most regions experience risk ratios less than ten, increases in the frequency of severe local Tx5DX escalate to well beyond 50 f old for many regions in the 2.0 °C-warmer world. In fact, the regions in black denote those locations for which the threshold of temperature seen every ten years in the future was never seen to occur across more than 2600 model years in the present-day climate. Such results corroborate previous studies, which found significant changes to the frequency of heat extremes are possible after only 2.0 °C of warming, though such dramatic increases could be minimised if global mean warming is kept to 1.5 °C (Perkins-Kirkpatrick and Gibson 2017, Russo et al 2017, Lewis et al 2017).

Previous research has shown that the signal of changes to extreme heat emerges first when assessed as an average over large spatial scales (Angélil et al 2014), while higher variability at smaller scales may suppress any detectable signal of change (King et al 2015). However, multiple studies have also demonstrated how robust changes in extremes can in fact be found at small spatial scales when the results of many individual locations are aggregated together (Fischer et al 2013, Westra et al 2013, Fischer et al 2014, Schleussner et al 2017). An adapted version of this framework is applied here to quantify levels of population exposure to future changes in extreme heat, as found in figure 1, using gridded population data for the year 2015 (figure 2, CIESIN 2016).

Following equation (1), the average heat exposure for a given region is calculated across all ensemble members, N, by summing the risk ratio (RR₁₀) estimates across all grid cells (spanning latitudes ϕ₁ and ϕ₂, and longitudes λ₁ and λ₂) with a weighting proportional to the fraction of people living in that grid cell, P/P_TOT. More generally, this approach produces a probability distribution which characterises population exposure to different risk ratios after 1.5 °C and 2 °C of warming. This framework therefore emphasises the importance of changing extremes where people live (Frame et al 2017), and will take into account not only future changes to distributions of temperature, but also changes to the number of people living in various locations.

$$\begin{equation} {\rm{Exposure}} = \frac{1}{N} \sum \limits^{{\phi _2}}_{i = {\phi _1}} \sum \limits^{{\lambda _2}}_{j = {\lambda _1}} \sum \limits^N_{n = 1} \left( {{\rm{R}}{{\rm{R}}_{{{10}_{\rm i{\rm{jn}}}}}} \times \frac{{{P_{{\rm{ijn}}}}}}{{{P_{{\rm{TOT}}}}}}} \right) \end{equation} \tag{ 1 }$$

Of course, when considering changes in population exposure to heat extremes under warming targets consistent with the Paris Agreement, a key determinant concerns the range of population scenarios which would be consistent with such a future. The Shared Socioeconomic Pathways (SSPs) provide a range of plausible future pathways of population and socioeconomic changes, and encompass the full spread of corresponding climate change scenarios from the Representative Concentration Pathway (RCP) database (KC and Lutz 2017, O'Neill et al 2016, Jones and O'Neill 2016). In addition to present-day population observations, this study also examines population projections for the year 2090 under an SSP1 and SSP2 scenario: these two were selected as 'baseline' scenarios most closely aligned, respectively, with the RCP2.6 and RCP4.5 climate scenario (O'Neill et al 2016). Since linear combinations of model simulations from both these RCPs were also used to produce the 1.5 °C and 2.0 °C HAPPI experiments (Mitchell et al 2017), we hereafter consider both SSP scenarios as feasible to occur in a future world where temperatures stabilise at 1.5 °C.

Figure 3 shows changes in absolute population between 2015 and 2090 under the two SSP scenarios considered. Both population data sets were first interpolated to a common 0.25° × 0.25° resolution before calculations were made. Under both scenarios, there are two large-scale patterns of population change which emerge: (1) widespread decreases in population over many mid-latitude regions, including Central Asia and China, as well as large regions of South-east Asia; and (2) robust increases in population over the Middle East, parts of Pakistan and Afghanistan, and most notably over widespread areas of Eastern Africa, including already densely-populated regions in Ethiopia, Kenya, Tanzania and Uganda. Whilst more obvious under SSP2, smaller but more concentrated patterns of growth also exist under SSP1 over parts of East Africa, Turkey and even India—a feature likely indicative of continued urban intensification (Jiang and O'Neill 2017). Interestingly, population projections across India are more divergent than other regions: negligible changes in total population between 2015 and 2090 are found under an SSP1 scenario, while at least another 400 million people will inhabit the region under an SSP2 scenario (KC and Lutz 2017). Such large variations in trajectories between the scenarios for different regions can be primarily linked to (a) differences in present-day population age structures, and (b) the differing prospects of changing education rates for younger adult females (aged 20–39, KC and Lutz 2017).

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

To better understand how the SAEA-wide changes in exposure to extreme heat will translate down to smaller spatial scales with differing projections of future population, two sub-regions have been selected for further analysis: one encompassing the East African (11°S–15°N, 25°E–40°E; hereafter 'EAF') domain exhibiting strong growth irrespective of SSP; the other encompassing an area over India, Bangladesh and Sri Lanka (5°N–28°N, 70°E–93°E; hereafter 'IND') which displays scenario-dependent population growth characteristics.

For the full SAEA region, median exposure to severe Tx5DX increases by a factor of 4.1 (2.4–9.6) under 1.5 °C of warming and when population is kept fixed to 2015 levels, while this number rises to 15.8 (5.0–135) after 2.0 °C of warming. When population changes are also accounted for, SAEA-wide exposure increases by a further 50% under the SSP2 population scenario, while increases are also seen, though less significant, in the lower-population SSP1 scenario. Interestingly, median RR₁₀ increases by a factor of four (2.1–25) if temperatures reach 2.0 °C of warming, instead of stabilising at 1.5 °C, irrespective of the population data set chosen. Such results are even more dramatic than previous-reported estimates of a doubling in extreme heat between 1.5 °C and 2.0 °C—however, those results were found using coarser-resolution models and with respect to the entire globe (Fischer and Knutti 2015).

Results for the sub-region encompassing India, Bangladesh and Sri Lanka (IND) qualitatively follow the same patterns of absolute exposure as for the wider SAEA region—such an outcome is unsurprising, given about one-third of the population across the entire SAEA domain is concentrated within the smaller IND region (table s2).

Across the East African (EAF) sub-region, increases in extreme heat are found to become 5.7 (2.1–12.2) times more likely after 1.5 °C warming if population remains fixed at 2015 levels, while risk ratios reach 9.4 (4.1–22.2) and 13.0 (5.7–30.7) if future population changes follow an SSP1 or SSP2 scenario respectively. Interestingly, there is a robust change in the frequency of extremes between the lower- and higher-warming scenarios: an estimated 5.2 fold (2.5–22) increase in extremes is expected between the 1.5- and 2 degree worlds irrespective of population scenario. These larger increases in RR overall for EAF (relative to the entire SAEA region) are indicative of the faster emergence in extreme temperatures found in figure 1, and also agree with many other studies focusing on the relative rapidity of changes in heat extremes across the African continent (Mahlstein et al 2011, Russo et al 2016, Harrington et al 2016).

While absolute changes in the frequency of heat extremes are highly informative for regional stakeholders, the methods chosen in this study enable us to further interrogate the changes in population exposure to RR₁₀, by quantifying the relative importance of three factors: (1) changes attributable to increases in global mean temperatures (RR_CLIM); (2) changes attributable to differences in total population size (RR_∆POPN); and (3) changes attributable to the spatial reorganisation of the population exposed (RR_{REDISTRIBUTION}). Combined, we find

$$\begin{equation} \rm RR_{10} = RR_{CLIM} \times RR_{\Delta POPN} \times RR_{REDISTRIBUTION} \end{equation} \tag{ 2 }$$

where RR_CLIM is found by keeping population levels fixed at 2015 observations (first column of table 1) and RR_∆POPN simply equals the percentage increase in total population for the region of interest under the different SSP scenarios considered (relative to 2015 observations). To calculate RR_{REDISTRIBUTION}, we repeat the process of spatially aggregating risk ratios weighted according to where people live, but first normalise the maps of population for all three scenarios, so individual grid cells contain information about what percentage of the region's people live there, rather than what is the total number of people living there. By removing the effect of changes in total population size, the fractional difference in risk ratio between these results using 2015 populations, and those using SSP scenarios, will equal the percentage increase in risk of extreme heat attributable to changes in where people live (and thus RR_{REDISTRIBUTION}).

Table 2 presents the range of estimates of RR_{REDISTRIBUTION} and RR_∆POPN for the various regions of interest. Results reveal that statistically significant increases in exposure to heat extremes occurs for the SAEA region as a whole, with best-guess (ensemble median) estimates of RR_{REDISTRIBUTION} of 1.09 (1.03–1.12) after 1.5 °C of warming, and best-guess estimates exceeding 1.2 for a 2.0 °C-warmer world. This result can be explained by the fact that those locations where future population growth (decline) is most significant (figure 3) also preferentially align with the regions of substantially higher (lower) risk ratios in the future (figures 1(e)–(f)). Estimates of RR_{REDISTRIBUTION} are also robust irrespective of whether an SSP1 or SSP2 scenario is considered—a result which is surprising given some of the large regional differences in population change. The approximate doubling in RR_{REDISTRIBUTION} after 2.0 °C of warming relative to the 1.5 °C experiment reflects the non-linear changes in heat extremes which occur at the higher warming level—this acceleration of risk ratios in fast-emerging locations will further exacerbate the relative impact of a higher population fraction living there in the future.

When considering the smaller sub-regions, we find negligible changes in risk ratios over India associated with the redistribution of where people live—this is unsurprising however, since population density is extremely high across the entirety of the region. Instead, the added increases in risk ratio in table 1 under SSP2 are the direct result of increases in total population size (RR_∆POPN). By contrast, there is actually a robust decrease in population exposure to heat extremes associated with changes to where people live for the EAF region, with best-guess estimates of an 8% and 17% reduction in heat exposure after 1.5 °C and 2.0 °C of warming, respectively. This effect therefore helps to mitigate some of the impact of 80% and 150% population increases expected for the region under SSP1 and SSP2 scenarios (respectively). However, as table 1 reveals, the net effect is still that of exposure to higher-than-average risk ratios being observed for the EAF region.