1 Introduction

Water scarcity is a major global issue. Existing pressures on water resources will be exacerbated by increases in population and also by climate change. Various studies have explored how both these factors might affect global water scarcity in the future by using population projections and simulated changes in climate from global climate models (GCMs) with water resources models (Alcamo et al. 2007; Arnell 2004; Arnell et al. 2011; Gosling et al. 2010; Hayashi et al. 2010; Oki and Kanae 2006).

Most of this work acknowledges that projections of water scarcity are dependent upon not only scenarios of population change and emissions, but also the number of GCMs used for simulating the future climate. There are over 20 GCMs included in the Coupled Model Intercomparison Project Phase 3 (CMIP3) multi-model dataset (Meehl et al. 2007a), which all simulate different but plausible climates under identical emissions scenarios (Meehl et al. 2007b). Previous global assessments of the impact of climate change on water scarcity have used various numbers of CMIP3 GCMs ranging from one to six (Arnell 2004; Hayashi et al. 2010). Additional uncertainty can arise from the measure used to define water scarcity (Alcamo et al. 2007).

This paper presents the most comprehensive global-scale assessment to date of the impact of climate change on water scarcity. “Comprehensiveness” is interpreted as taking account of the various factors that affect climate change impacts assessment, including; the range of possible future climates projected by different GCMs, the different magnitudes of possible climate change due to different emissions trajectories, different population projections and water withdrawals, and differing methods for estimating water scarcity. To achieve this, we use climate change patterns from 21 CMIP3 GCMs, four socio-economic and emissions scenarios, and two different measures of water scarcity. We only use a single hydrological model, however, so we do not sample hydrological modelling uncertainty.

The overarching aim of the paper is to assess how our more comprehensive estimates of global water scarcity compare to previous assessments. Additionally, in light of The Copenhagen Accord and Durban Platform, which state ambitions of limiting the increase in global mean temperature to 2 °C and 1.5 °C above the pre-industrial value respectively, we assess how different amounts of global mean warming up to 4 °C might affect global water scarcity.

2 Methodology

2.1 Introduction

The experimental design involved applying the patterns of climate change from 21 GCMs to a global hydrological model to calculate average annual runoff, which in turn was used to estimate global water scarcity with a water resources model.

2.2 Climate change scenarios

Two types of scenarios are considered for each of the 21 GCMs; 1) four SRES emissions scenarios (B1, B2, A1B and A2) for three 30-year time horizons centred on 2020, 2050 and 2080; and 2) seven prescribed changes in global mean temperature relative to present (0.5, 1.0, 1.5, 2.0, 2.5, 3.0 and 4.0 °C of 30-year duration each). The current climate is characterised by the CRU TS3.1 data set (Harris et al. 2012) for the 1961–1990 time horizon, which is approximately 0.3 °C above pre-industrial.

Spatial and temporal climate change scenarios at 0.5° × 0.5° resolution were constructed by pattern-scaling output from 21 CMIP3 GCMs (Meehl et al. 2007a) with ClimGen (Todd et al. 2011). ClimGen uses the change pattern for any given GCM to perturb a historical dataset (CRU TS3.1 in this study) to ensure minimal bias with respect to observations. This is often referred to as the delta method (Arnell and Gosling 2013). Further details on pattern-scaling and the 21 GCMs we used are in Online Resource 1 in the Electronic Supplementary Materials (ESM) file. Not all 21 GCMs are independent of each other and they are considered equally plausible in this analysis. To help explore the effects of using different GCMs, we selected HadCM3 as an illustrative ‘marker’ GCM, which presents a plausible characterisation of the spatial variability of change in climate (Arnell et al. 2013).

2.3 The global hydrological model and water resources model

Simulations of average annual runoff across the global domain at a spatial resolution of 0.5° × 0.5°, for every climate change pattern and scenario, were performed with an established global hydrological model, Mac-PDM.09 (Gosling and Arnell 2011). Mac-PDM.09 is used in several recent studies (Arnell and Gosling 2013; Arnell et al. 2013; Hagemann et al. 2013; Thompson et al. 2013). Further details about Mac-PDM.09 are included in Online Resource 2.

The runoff simulations from Mac-PDM.09 were applied to a water resources model (Arnell et al. 2011; Gosling et al. 2010) to estimate water scarcity. For every climate change pattern we calculated two well-known measures of water scarcity (Rockström et al. 2009; Oki and Kanae 2006); 1) “Water Crowding Index” (WCI; a measure of the annual water resources per capita in a watershed) and 2) “The Water Stress Index” (WSI; a measure of the ratio of water withdrawals to resources). A WCI threshold of <1,000 m3/capita/year and a WSI of >0.4 were used to indicate exposure to water scarcity (Rockström et al. 2009). For both measures, available water resource in each of 1339 watersheds across the globe was calculated by summing simulated average annual runoff from each 0.5° × 0.5° grid cell within a given watershed. WCI is heavily dependent upon population size while the WSI accounts for variations in withdrawals across watersheds and therefore tends to highlight pressures in watersheds with large amounts of irrigation.

The water resources model calculates four metrics (Arnell et al. 2011) that isolate the sole impact of future climate change on water scarcity (i.e. they represent the additional impact of climate change on top of population and/or withdrawals pressure). This involves calculating first, future water scarcity in the absence of climate change (i.e. due to future population and/or withdrawals pressure only) and then subtracting this from the water scarcity that occurs due to the combined effects of future climate change and population and/or withdrawals pressure. The four metrics are:

  1. 1)

    The number of people in a region who live in watersheds with no water scarcity in the absence of climate change but that enter water scarcity due to climate change.

  2. 2)

    The number of people in a region who live in watersheds with water scarcity in the absence of climate change but that move out of water scarcity due to climate change.

  3. 3)

    The number of people in a region living in watersheds with water scarcity in the absence of climate change who see a “significant” decrease in runoff due to climate change.

  4. 4)

    The number of people in a region living in watersheds with water scarcity in the absence of climate change who see a “significant” increase in runoff due to climate change but who still remain in water scarcity.

A “significant” change in runoff is defined to be greater than the standard deviation of average annual runoff due to natural multi-decadal climatic variability. This was calculated from multiple estimates of the 30-year average annual runoff using climate scenarios constructed from a long unforced simulation with the HadCM3 climate change pattern (Arnell and Gosling 2013).

For each of the two water scarcity measures, the impact of climate change on exposure to water scarcity (i.e. on top of future population and/or withdrawals pressure) is summarised by summing 1) and 3) to characterise population exposed to a potential increase in water scarcity due to climate change, and summing 2) and 4) to characterise population with a potential reduction in water scarcity due to climate change. Throughout the manuscript, for conciseness, these are referred to as an “increase in scarcity” and a “decrease in scarcity”, respectively, and the estimates for each can be considered as relative to the situation in the future (e.g. 2050) where there is higher population and withdrawals than present.

2.4 Socio-economic scenarios

When exploring water scarcity due to climate change under SRES scenarios, future population was taken from the IMAGE v2.3 representation of the B1, B2, A1B and A2 storylines (Van Vuuren et al. 2007) and used with the appropriate runoff simulations in the water scarcity model. When exploring the prescribed warming scenarios, future population was assumed to be equivalent to that under SRES A1B for the 30-year time horizon centred on 2050, because these climate projections are not associated with any specific time in the future or SRES scenario. Withdrawals were estimated by rescaling Shen et al.’s (2008) projections to match the population projections used here. Note that different projections of future withdrawals would give different indications of future water resources scarcity.

Watershed population exposure to water scarcity was aggregated to the national scale and then to the regional scale, so water scarcity can be expressed regionally in absolute (millions of people) and relative (as a percentage of regional population) terms. The countries included in each region are listed in Online Resource 3 and displayed in a map in Online Resource 4.

3 Results

3.1 Water scarcity in the absence of climate change

In the year 2000, depending on the measure of water scarcity, 1.6 (25 % of global population) and 2.4 billion (39 %) people are estimated to be living in watersheds exposed to water scarcity (Table 1). The greatest proportions of populations living in water-scarce watersheds are located in East Asia (660 and 666 million) and South Asia (491 and 1004). More people fall into the water scarcity category with the WSI than with the WCI (Table 1; see also Online Resource 5 for global maps). Notable regions where the two water scarcity measures result in opposing results (i.e. one measure results in water scarcity and the other measure does not), include parts of the US and Australasia.

Table 1 Numbers of people exposed to water scarcity in the absence of climate change (i.e. due to population change only), using two measures of water scarcity, for four time horizons, under the A1B scenario
Full size table

Total global population in the years 2000, 2020, 2050 and 2080, under the A1B scenario, is estimated to be 6.1, 7.3, 8.2 and 7.8 billion respectively. This places pressures on future water resources and in the absence of climate change it is estimated that by 2050 under A1B, 3.1 (37 %) and 4.3 billion (53 %) people will be living in watersheds exposed to water scarcity, globally. The greatest absolute exposure to water scarcity is in South Asia (1.5 and 1.7 billion) and East Asia (0.7 and 1.2 billion).

3.2 Runoff changes

Simulated changes in average annual runoff by 2050 from Mac-PDM.09, when it is forced with the pattern of climate change from the HadCM3 GCM only under A1B, show runoff increases relative to present (1961–1990) across large areas of East Asia (+30 %), South Asia (+30 %), the high northern latitudes (+40 %), and East Africa (+20 %) (see Online Resource 6(a)). Some areas see up to +90 % increases in runoff. Changes in runoff in South Asia and East Asia reflect large changes in climate on a large regional runoff volume seen in the present climate. Declines in runoff are simulated for Brasil (−60 %), South America (−30 %), Southern Africa (−40 %), Western Europe (−20 %) and Central Europe (−30 %).

In line with previous assessments (Gosling et al. 2010; Milly et al. 2005), there is high consistency that runoff decreases with climate change across central Europe and that it increases in the high northern latitudes but there is less agreement between all CMIP3 model climate change patterns across many regions of the globe, including parts of South Asia and East Asia (see Online Resource 6(b)). There is little change in the pattern of consistency when different time horizons (2020 and 2080) or scenarios (B1, B2, A2) are considered (not displayed).

3.3 Water scarcity due to climate change only under SRES scenarios

Regions of the globe experience both increases and decreases in water scarcity in the future due to the sole effects of climate change (see Fig. 1; expressed as a percentage of future regional population, and Online Resource 7; expressed in millions of people). We present the estimates associated with each climate change pattern and the range across the ensemble but we do not calculate any measures of central tendency because these can be an unreliable summary indicator of climate change impacts (Gosling et al. 2012). Tables that present the absolute and relative values displayed in Fig. 1 are presented in Online Resources 8–11). In Fig. 1, an increase in water scarcity of 100 % for a region would mean that in 2050, 100 % of the people living in watersheds in that region experience an increase in water scarcity that is attributable solely to climate change (i.e. the increase is not due to changes in population or withdrawals; it is additional to them). A number of general conclusions can be drawn from these results.

Fig. 1
figure 1

Exposure to an increase and decrease in water scarcity attributable solely to climate change (i.e. the increases and decreases are additional to the effects of changes in future population or withdrawals), expressed as a percentage of future regional population, at 2050, using the WCI (1,000 m3/capita/year) and the WSI (0.4). The individual markers denote estimates for individual GCMs, with a different marker shape assigned to each SRES scenario. Filled red circles denote the HadCM3 GCM. The vertical grey bars denote the range across 21 GCMs

Full size image

When considering all CMIP3 models together, a greater global population experience an increase in water scarcity due to climate change than a decrease, but this is not necessarily the case for individual models. In 2050 and under A1B, an increase in water scarcity is experienced globally by 0.5 to 3.1 billion (WCI) and 0.8 to 3.9 billion (WSI); a decrease is experienced by 0.2 to 2.2 and 0.1 to 2.7 billion for each water scarcity measure respectively. This is consistent across SRES scenarios as well as the two measures of water scarcity. However, this result does not hold for every individual CMIP3 model. For example, with the HadCM3 pattern more people are exposed to a decrease in scarcity (1.9 and 2.7 billion) than an increase (1.0 and 1.3) by 2050.

In absolute terms, uncertainty in the effects of climate change on water scarcity due to the application of all CMIP3 models is greatest for South Asia and East Asia. Here, by 2050 and under the A1B scenario, between 52–1460 million and 0–506 million people are exposed to increased water scarcity respectively, when using the WCI. If these two estimates are expressed as a fraction of the total global population, then South Asia and East Asia account for 1–18 % and 0–6 % respectively of the global population exposed to increased water scarcity due to climate change. Thus a substantial proportion of the uncertainty in the global-scale effect of climate change on water scarcity is due to uncertainties in estimating the effects in South Asia and East Asia.

Close investigation of Fig. 2 explains why this uncertainty range is so large for South Asia and East Asia and also why it is greater for the WSI than the WCI. While the application of a single GCM (HadCM3) indicates that some watersheds in East Asia move out of water scarcity and/or see a decrease due to climate change, when using all CMIP3 models around 4–10/21 simulations indicate that water scarcity increases in these watersheds and around 11–14/21 show a decease in scarcity. Climate change affects more watersheds in East Asia when using the WSI than the WCI, which means the uncertainty across simulations showing increases and decreases in water scarcity translate into a larger uncertainty range in absolute terms.

Fig. 2
figure 2

The effect of climate change only on exposure to water scarcity in 2050 under the SRES A1B scenario, using the WCI (1,000 m3/capita/year; left panels) and the WSI (0.4; right panels). Top panels display the change in scarcity classes with the HadCM3 GCM only. The other panels show consistency across 21 simulations (with all CMIP3 models), in terms of the number of simulations out of 21 that show an increase in exposure (indicative of where there is an increase in scarcity or a watershed moves into water scarcity; middle panels), or decreases (a decrease in scarcity or move out of scarcity; bottom panels)

Full size image

The differences in projections of water scarcity across the four scenarios are relatively small when compared with the differences across the 21 simulations. Globally, the absolute (relative) increases in scarcity by 2050 with HadCM3, when using the WCI are 1.0 (13 %), 1.0 (12 %), 1.1 (12 %) and 1.4 billion (14 %), for A1B, B1, B2 and A2 respectively. By comparison, the minimum and maximum values across the 21 simulations for absolute (relative) increases in scarcity by 2050 and using the same scarcity measure are 0.5 and 3.1 (6 and 38 %), 0.5 and 2.8 (6 and 34 %), 0.6 and 3.5 (6 and 39 %), and 0.7 and 4.7 billion (7 and 46 %) for A1B, B1, B2 and A2 respectively. This means that the range across the 21 simulations with all CMIP3 models for a single scenario (e.g. A1B) is greater than the range across all four emissions scenarios for a single GCM (HadCM3). This is observed at both the global-scale and at the regional-scale.

3.4 Sensitivity to the WCI and WSI thresholds for water scarcity

So far we have described exposure to water scarcity due to climate change based upon the two measures of water scarcity (WCI and WSI) with scarcity indicator thresholds of <1,000 m3/capita/year and >0.4 respectively. A more rigorous analysis of the sensitivity of water scarcity exposure to the scarcity measure compares scarcity indicator thresholds (500, 1,000, 1,700 m3/capita/year and 0.1, 0.2, 0.4 respectively) (see Fig. 3 for relative exposure and Online Resource 12 for absolute exposure). These thresholds, which are arbitrary, have been used previously with the WCI for defining “extreme water shortage” (<500), “chronic water shortage” (<1000) and “moderate water shortage” (<1700) (Kummu et al. 2010; Arnell 2004) and with the WSI for defining “low stress” (>0.1), “medium stress” (>0.2) and “high stress” (>0.4) (Arnell 1999). This analysis demonstrates how sensitive estimates of water scarcity are to the selection of arbitrary thresholds.

Fig. 3
figure 3

Comparison of WCI and WSI water scarcity measure thresholds. Exposure to an increase or decrease in water scarcity due to climate change is expressed as a percentage of regional future population in 2050, assuming the A1B scenario

Full size image

The sensitivity of exposure to the three thresholds for each measure is appreciable. Generally, it is greater than sensitivity of exposure to SRES scenario but slightly less than sensitivity to climate change pattern. For example, the estimated increase in global exposure to water scarcity by 2050 under A1B and using the WCI (1,000 m3/capita/year) is 0.5 to 3.1 billion (6–38 %) across all CMIP3 models, while it is 0.5 to 1.5 billion (6–18 %) across the three WCI thresholds with HadCM3. This compares with 1.0 to 1.4 billion (12–14 %) across the four SRES scenarios with HadCM3. For some regions the sensitivity to indicator is comparable to sensitivity to climate change pattern (Central Asia and Central Europe).

3.5 Water scarcity due to climate change only under prescribed warming scenarios

Exposure to water scarcity, as measured with the WCI, increases non-linearly with global mean temperature and there are clear overlaps between the ranges for increases and decreases in exposure across CMIP3 models (as observed under the SRES scenarios) in relative (Fig. 4) and absolute (Online Resource 13) terms. Exposure to water scarcity increases steeply up to 2 °C in many regions (North Africa, East Africa, Mashriq, Arabian Peninsula and South Asia) and then stabilises by 4 °C. This is because by this point all the watersheds that see a decrease in precipitation experience a significant decrease; beyond this point there are no further areas where precipitation decreases significantly. With a higher threshold for the definition of ‘significant’ change, the limit would obviously occur at a higher global mean temperature. Global exposure to an increase (decrease) in water scarcity with HadCM3 by 2050 and assuming A1B for 1, 2, 3 and 4 °C prescribed warming, is 760 (1807), 973 (1872), 1055 (1897) and 1161 (1905) million people respectively, but the uncertainty range across all CMIP3 models is considerable.

Fig. 4
figure 4

Relationship between global temperature increase above 1961–1990 and exposure to water scarcity due to climate change, using the WCI. Expressed as a percentage of regional population, under an A1B socio-economic scenario in 2050

Full size image

4 Discussion and conclusions

Our estimate of present-day populations living within watersheds exposed to water scarcity (1.6 and 2.4 billion) is consistent with other studies that have published estimates around 2.4 billion (Oki and Kanae 2006), 1.4 and 2.2 billion (Arnell 2004), 1.7 and 2.3 billion (Revenga et al. 2000), 1.6 and 2.4 billion (Arnell et al. 2011), 1.2 billion (Hayashi et al. 2010), and 1.6 and 2.3 billion (Alcamo et al. 2007); the ranges are due to the application of different measures of water scarcity. By 2050, the effects of population increases alone mean that 3.1 and 4.3 billion people (37 and 53 %) will be living in watersheds exposed to water scarcity, which is similar to previous estimates around 3.4 and 5.6 billion (39 and 48 %) (Arnell 2004), 3.7 and 4.2 billion (42 and 47 %) (Arnell et al. 2011), and 3.8 billion (40 %) (Hayashi et al. 2010).

Direct comparisons of projected changes in exposure to water scarcity with other studies is not straightforward because of the application of different climate models, emissions and population scenarios, hydrological models, and measures of water scarcity. An additional issue that complicates inter-study comparisons is the spatial scale at which water scarcity is calculated. We estimated watershed population exposure to water scarcity and aggregated this to the national scale and then to the regional scale. Alternative approaches have conducted analyses at the scale of individual grid cells (Oki and Kanae 2006; Vörösmarty et al. 2000), countries (Oki et al. 2001) and food production units (Kummu et al. 2010). While the above comparisons do not account for these differences, a tentative comparison can be made, however, with two studies that both used a previous version of the hydrological model we applied and a WCI threshold of 1,000 m3/capita/year to define water scarcity at the watershed scale (then aggregating to country and regional scales). Arnell (2004) calculated a global increase in exposure (in billions of people) by 2055 under A2 of 1.1 to 2.8, across 6 GCMS. Our estimates by 2050 under A2 are 0.7 to 4.7, across 21 GCMs. Arnell et al. (2011) used a “Reference” scenario that is comparable to A1B and calculated a global increase in exposure by 2050 of 0.5 to 1.5 billion across 4 GCMs. Our estimate under A1B is 0.5 to 3.1 billion, across 21 GCMs.

Compared with these two studies, our estimates present a wider range that encapsulates both, but with very little difference at the lower-end and a considerable increase at the upper-end (over 1 billion people). This is because our assessment considered many more patterns of climate change from GCMs than either of the two studies. Some of the GCMs we used will simulate lower precipitation with climate change than those used by Arnell (2004) and Arnell et al. (2011). In absolute terms, uncertainty in the effects of climate change on water scarcity due to the application of all CMIP3 models is greatest for South Asia and East Asia, since these are two regions where GCMs show large differences in the magnitude, and sometimes sign, of precipitation change (Meehl et al. 2007b), and hence runoff change. Thus a substantial proportion of the uncertainty in the global-scale effect of climate change on water scarcity is due to uncertainties in estimating the effects in South Asia and East Asia.

The shape of the distribution of exposure across all CMIP3 models varies by region and type of exposure. This means that climate change impacts studies should try to use climate change projections from an ensemble of climate models that best represent the range across all GCMs available. It is not always computationally feasible to use all members of an ensemble (e.g. the CMIP3 models), however, and in such cases a careful and thoughtful selection of GCMs should be made, e.g. by considering where each GCM sits on the range of all GCMs in the ensemble. Otherwise there is a risk of selecting only GCMs that represent either the tail, or part of a bimodal distribution, for instance, and thus underestimating or overestimating the range of possible outcomes. While GCM performance metrics present a method for the selection of GCMs in impact studies (Wilby 2010), there remains a strong argument that all available GCMs should be used regardless because there is no difference in the projections from “better” GCMs when compared with “weaker” GCMs, at least where precipitation projections are of relevance (Chiew et al. 2009). Moreover, no single GCM may consistently out-perform all others when variables beyond precipitation only are considered (Gleckler et al. 2008).

We found that some climate change patterns result in more people exposed to an increase in water scarcity than people exposed to a decrease, at the global-scale (note that a “net water scarcity change” cannot be calculated by summing the increase and decrease in exposure for the reasons outlined by Arnell et al. (2011)). This contrasts with previous studies that used many less climate change patterns than we applied here and found consistently that a greater proportion of the global population is exposed to a decrease in water scarcity than an increase (Arnell 2004; Hayashi et al. 2010). The sensitivity of relative increases and decreases in exposure to climate change pattern adds weight to the argument that all available GCMs should be used in water resource climate change impact assessments, where possible.

Similar to earlier work (Gosling et al. 2010; Arnell 2004), we found that projections of water scarcity are substantially more sensitive to climate change pattern than emissions scenario. However, this assessment has highlighted an appreciable sensitivity to the water scarcity indicator threshold that is used. Moreover, this is an important caveat of our analysis, since all estimates of water scarcity presented here are based upon global-scale generalisations about what it means to be in a situation of water scarcity. In particular, the application of the WSI is traditionally based upon water withdrawals instead of actual consumption. To some extent, this could mean that water scarcity measured by the WSI is overestimated in watersheds where withdrawn water is predominantly used either several times (e.g. hydropower) or returned downstream for other users instead of being consumed (e.g. some parts of the US). However, this limitation is less relevant to locations where water withdrawals are used predominantly for irrigation, where the water is also consumed and not returned to the system.

Average annual runoff was used as input to the water resources model. Thus in our modelling approach, an increase in average annual runoff can benefit society through a decrease in water scarcity. On the other hand, however, this could be tempered by an increase in flood risk, which we did not consider in this study, but we have elsewhere (Arnell and Gosling 2013). Other caveats include assumptions involving the parameters used both in the pattern-scaling (Todd et al. 2011) and Mac-PDM.09 (Gosling and Arnell 2011), and the assumed rates of population change (for each SRES scenario only one population projection was used). Additionally, by using a single global hydrological model we considerably underestimated hydrological model uncertainty, which only recently has been shown to be appreciable (Hagemann et al. 2013) but we note that Haddeland et al. (2011) showed that under present-day climate forcing, Mac-PDM.09 was located towards the middle of the range in simulated runoff across 5 global hydrological models. The net effect of these caveats is that the estimates of exposure to water scarcity should not to be taken too literally as actual impacts or “hardship” but rather as an indication of the relative effects of different emissions, climate and population scenarios.

We conducted this assessment at a time when a new set of global change scenarios was being released; the GCMs of CMIP5 (Taylor et al. 2011) with the Representative Concentration Pathways (RCPs; van Vuuren et al. (2011)) and the Shared Socio-economic Pathways (SSPs; Kriegler et al. (2012)). A comparison of our CMIP3 SRES A2 simulations to CMIP5 RCP8.5 SSP3 (see Online Resource 14 for methods and results) suggests that our conclusions are robust across the CMIP3 and CMIP5 models because the estimates of water scarcity are broadly consistent.