We simulated global gridded daily discharge and runoff over the period 1960–2010 at a resolution of 0.5∘ × 0.5∘ using three global hydrological models: PCR-GLOBWB (van Beek et al., 2011; Wada et al., 2014b), STREAM (Aerts et al., 1999; Ward et al., 2007) and WaterGAP (Müller Schmied et al., 2014), forced with WATCH Forcing Data – ERA Interim (WFD-EI) daily precipitation and temperature data (0.5∘ × 0.5∘) (Weedon et al., 2014) for the period 1979–2010 and WATCH forcing data ERA40 (WFD) for the period 1960–1978 (Weedon et al., 2011). In order to compensate for offsets in long-term radiation fluxes between the two data sets, as found by Müller Schmied et al. (2014), WFD down-welling shortwave and long-wave radiation were adjusted for use in WaterGAP to WFD-EI long-term means following the approach of Haddeland et al. (2012). Daily values were aggregated to time series of monthly discharge and runoff. Using global hydrological models gives us the advantage of a global coverage, whereas the portfolio of observed data sets (water availability and consumptive water use) is bounded by its biased regional distribution (Hannah et al., 2011; Ward et al., 2010, 2014a). However, we are aware of the caveats using these types of models to estimate water availability as all large-scale hydrological models have their own strengths and shortcomings (Gudmundsson et al., 2012; Nazemi and Wheater, 2015a, b). Therefore, we constructed ensemble-mean time series of both monthly discharge and runoff capturing the three global hydrological models. The results of the individual modelling efforts were used to evaluate the modelling agreement (Sects. 2.4 and 3.5).

Water availability is expressed in this paper as the sum of monthly runoff per food producing unit (FPU). FPUs represent a hybrid between river basins and economic regions for which it is generally assumed that water scarcity issues can be solved internally (Cai and Rosegrant, 2002; de Fraiture, 2007; Kummu et al., 2010; Rosegrant et al., 2002). We used here an updated version of the FPU used by Kummu et al. (2010), which consists of 436 FPUs, excluding small island FPUs. For FPUs located within one of the world's larger river basins, we redistributed runoff in order to avoid local over- or underestimations in water availability. Runoff was redistributed across the FPUs within these larger river basins, proportionally to the discharge distribution of that large river basin (Gerten et al., 2011; Schewe et al., 2014): WAi=Rb∗Qi∑Qi, whereby WAi is the monthly water availability within FPU i,Rb is the total monthly runoff within large river basin b,Qi is the monthly discharge in FPU i, and ∑Qi is the sum of the monthly discharge over all cells within a large river basin b.

Subsequently, we calculated the annual water availability by aggregating the simulated ensemble-mean monthly water availability time series using hydrological years. The use of hydrological years is necessary in this assessment, as ENSO tends to develop to its fullest strength during the period December–February, which intersects with the standard calendar year boundaries (Ward et al., 2014a, b). Hydrological years are referred to by the year in which they end, e.g. hydrological year 1961 refers here to the period October 1960–September 1961. Within this study we follow Ward et al. (2014a) and distinguish two hydrological years on the basis of long-term monthly maximum water availability per river basin: October–September (standard) and July–June (for river basins that have their long-term monthly maximum water availability in September, October or November). The river basin delineation used here was derived from the WATCH project (Döll and Lehner, 2002) and is equal to the river basin delineation that is used as the input for the FPU classification used within this study. We used the hydrological years setting determined at grid level, using the WATCH river basins, as input for the distinction between hydrological years at FPU scale. If an FPU consisted of more than one river basin we based the choice of hydrological year on the month (with long-term maximum water availability) with the highest prevalence within this FPU (see Supplement Fig. S1).

Monthly gridded water consumption (0.5∘ × 0.5∘) was estimated for the sectors livestock, irrigation, industry, and domestic within PCR-GLOBWB using daily WFD-EI precipitation and temperature data in combination with yearly information on livestock densities; the extent of irrigated areas; desalinated water use; non-renewable groundwater abstractions; and past socioeconomic developments, namely GDP, energy and electricity production, household consumption, and population growth (Wada et al., 2011b, 2014b). For a complete description and extensive discussion of the methodological steps taken to compose these monthly consumptive water use time series, we refer to Wada et al. (2011b, 2014b). Time series of desalinated water use and non-renewable groundwater abstractions were subtracted from the total consumptive water use estimates as they lower the need for blue water. Subsequently, we aggregated gridded monthly consumptive water use into yearly totals per FPU (WCi,yr), following the hydrological years. Since the resulting transient consumptive water use estimates are partially driven by changing socioeconomic conditions (population, GDP, and growth in irrigated areas), and therefore disguise any possible correlations with ENSO-driven climate variability; we repeated the steps above whilst we fixed the socioeconomic parameters at 1961 levels (following the hydrological year naming convention). These fixed consumptive water use estimates were used to evaluate the sensitivity to ENSO-driven climate variability (Sects. 3.1 and 3.2), whereas the transient water consumption time series were used to evaluate the development of water scarcity conditions under changing socioeconomic conditions (Sect. 3.3).

Blue water scarcity refers to the imbalance between blue water availability (i.e. water in rivers, lakes, and aquifers) and the needs for water over a specific time period and for a certain region (Falkenmark, 2013). Although water scarcity could also relate to the green (water in the unsaturated soil), white (part of rainfall that feeds directly back into the atmosphere), and deep blue (fossil groundwater) water sources (Savenije, 2000), we focus here on blue water scarcity (hereafter: water scarcity) only. Within this study we applied two complementary indicators to express water scarcity conditions per FPU: the water crowding index (WCI) for population-driven water shortage and the consumption-to-availability ratio (CTA ratio) for demand-driven water stress (Brown and Matlock, 2011; Rijsberman, 2006). The WCI quantifies the yearly water availability per capita (Falkenmark et al., 1989, 2007; Falkemark, 2013), whereby water demands are based on household, agricultural, industrial, energy, and environmental water consumption (Rijsberman, 2006). Like previous studies (e.g. Alcamo et al., 2007; Arnell, 2003; Kummu et al., 2010), we used 1700 m3 capita-1 per year as the threshold level to evaluate water shortage events. The CTA ratio evaluates the ratio between consumptive water used and water availability in a specific region and is a derivative from the withdrawal-to-availability (WTA; Raskin et al., 1997) ratio. Usually, a region is said to experience water stress events when water withdrawals comprises ≥ 40 % of the available water resources, whilst moderate water stress conditions occur if 20 % ≥ WTA ≤ 40 % (Raskin et al., 1997). The use of the WTA ratio is widely quoted and applied in previous research contributions, e.g. by Alcamo et al. (2003, 2007), Arnell et al. (1999), Cosgrove and Rijsberman (2000), Hanasaki et al. (2013), Kiguchi et al. (2015), Kundzewicz et al. (2007), Oki et al. (2001), Oki and Kanae (2006), and Vörösmarty et al. (2000). Hoekstra et al. (2012) and Wada et al. (2011a) applied this WTA ratio in an adapted form, using blue water footprints and potential consumptive water use estimates respectively to assess water stress conditions: the CTA ratio. This approach accounts for the share of water that has been recycled (industry) or not used (irrigation) and which flows back into the natural system. The threshold level for water stress using these consumptive water demands is therefore conceived to be lower than the threshold level for water stress as estimated using withdrawals. Following Hoekstra et al. (2011, 2012), Richter et al. (2011), and Wada et al. (2011a), we applied a threshold level of 0.2 to indicate water stress events. Equations (2) and (3) show the use of the WCI (WCIi,yr) and the CTA ratio (CTAi,yr), respectively, WCIi,yr=WAi,yrPi,yr(watershortageeventifWCIi,yr≤1700),

CTAi,yr=WCi,yrWAi,yr(waterstresseventifCTAi,yr≥ 0.2), whereby WAi,yr is the water available per spatial unit i and hydrological year yr, Pi,yr is the population, and WCi,yr is consumptive water use. Water scarcity conditions were assessed here at the FPU scale. The FPU scale is seen as an appropriate spatial scale to study water scarcity conditions as it is generally assumed that lower-scale water scarcity issues can be overcome by the reallocation of water demand and supply within this spatial unit (Kummu et al., 2010). However, one should keep in mind that, due to the assumption of full exchange possibilities – both from an infrastructural and water management perspective and its relative large spatial scale, analysis executed at the FPU scale may disguise lower-scale water scarcity issues (Kummu et al., 2010; Wada et al., 2011a).

The population data used for the calculation of the WCI (Eq. 2) were adopted from Wada et al. (2011a, b), who derived yearly gridded population maps (0.5∘ × 0.5∘) from yearly country-scale FAOSTAT data in combination with decadal gridded global population maps (Klein Goldewijk and van Drecht, 2006). We aggregated these gridded population maps to FPU scale for use in this study. In line with the hydrological year naming convention, population estimates were used for the year in which the hydrological year ends; e.g. for hydrological year 1961 we used population estimates of 1961 as input for the WCI and to calculate water scarcity impacts.

We examined the relationship respectively between water availability, consumptive water use, and water scarcity conditions, and ENSO-driven climate variability by means of their correlation with the Japan Meteorological Agency's (JMA) Sea Surface Temperature (SST) anomaly index (http://coaps.fsu.edu/jma). We used here 3-monthly mean values of the JMA SST over the periods October–December, November–January, December–February, and January–March, as El Niño and La Niña expressions are strongest in these months (Dettinger and Diaz, 2000). Following Ward et al. (2014b), we examined the correlation between WAann, WCann, CTAann, and WCIann, and the 3-monthly mean JMA SST values (OND, NDJ, DJF, JFM), using Spearman's rank correlation coefficient. Statistical significance was assessed by means of regular bootstrapping (n= 1000, p ≤ 0.05) while field significance, i.e. the joint statistical significance of multiple individual significance tests (Livezey and Chen, 1982; Wilks, 2006), for each of the 3-monthly JMA SST correlation values was tested using the binomial distribution (Livezey and Chen, 1982). With field significance testing, we counted the number of individual tests with a significant result and assessed the probability of yielding this result by chance given its statistical distribution (Livezey and Chen, 1982; Wilks, 2006). Subsequently, we examined the percentage anomalies in the median values of water scarcity conditions between El Niño and La Niña years, compared to the median values under all years. To distinguish between El Niño, La Niña, and neutral years we used the classification of ENSO years from the Center for Ocean–Atmospheric Prediction Studies based on the JMA SST values. Years are assigned as El Niño or La Niña years when their 5-month moving average JMA SST index values are (±)0.5 ∘C or greater (El Niño)/smaller (La Niña) for at least 6 consecutive months (including October–December). Reference to the different ENSO years was adjusted to be consistent with the naming convention used for the hydrological years (Table 1). We used a bootstrapped version of the non-parametric Mann–Whitney U test (n= 1000, p ≤ 0.05) to test the statistical differences in median values.

The critical threshold values put in place for the WCI and the CTA ratio (here 1700 and 0.2 respectively) determine whether water scarcity conditions adversely affect population or society. Per FPU we therefore evaluated which proportion of land area, for which we found a significant correlation between ENSO and water scarcity conditions, is also exposed to water scarcity events and how population is clustered in these areas compared to the general pattern of population density. Moreover, we assessed how these numbers changed through time given the changing socioeconomic conditions, relative to developments in (1) the population and land area sensitive to ENSO-driven climate variability but not exposed to water scarcity events; (2) the population and land area exposed to water scarcity events in areas that lack a significant correlation with ENSO-driven climate variability; and to (3) the total population growth.

A cross-model validation was executed in order to evaluate the modelling uncertainty whereby we compared the results from the ensemble mean with the outcomes of the individual global hydrological models (GHM). We examined the agreement among the different modelling results and the ensemble mean when looking at (1) the sensitivity of water availability and water scarcity conditions to ENSO-driven climate variability, and (2) the impacts of water scarcity events and relation to ENSO-driven climate variability under changing socioeconomic conditions.

Significant correlations of water availability to variations in JMA SST were found across 37.1 % of the global land surface (excluding Greenland and Antarctica), whilst for consumptive water use (simulated under fixed socioeconomic conditions at 1961 levels) we found significant correlations covering 8.3 % of the total land area (Fig. 1 and Table 2). Using the 3-monthly JMA SST period with the highest correlation, Fig. 1 shows for both water availability and consumptive water use its correlation coefficient with the inter-annual variation in the 3-monthly average JMA SST values. Only those correlations which reach statistical significance at a 95 % confidence interval are shown here. Field significance, the collective global significance of the total of individual local hypothesis tests (Livezey and Chen, 1982; Wilks, 2006), was tested for the individual 3-month correlation results and found to be highly significant when looking at water availability (p < 0.01) but insignificant when considering consumptive water use (p > 0.5). Positive correlations, i.e. more water available with the JMA SST index moving towards El Niño values, were found for 13.2 % of the global land surface, while negative correlations were found in FPUs covering 23.9 % of the global land surface. When looking at consumptive water use we found positive significant correlations for only 1.0 %, and negative correlations for 7.3 % of the global land surface.

Subsequently, we assessed how sensitive water scarcity conditions (simulated under fixed socioeconomic conditions at 1961 levels) are to ENSO-driven climate variability. Significant correlations to variations in JMA SST were found for 28.1 and 37.9 % of the global land surface when using the CTA ratio (water stress) and WCI (water shortage) respectively, while being tested under a 95 % confidence interval (Table 3). Due to the clustering of population and consumptive water use we found even higher percentages when looking at the population living in these areas, 31.4 and 38.7 % of the global population in 2010 for the CTA ratio and WCI, respectively.

Figure 2 shows the areas with a significant positive (red) or negative (blue) correlation of water stress conditions (CTA ratio) with the variation in JMA SST values, using the 3-monthly JMA SST period with the highest correlation (JMA SSTbestoff). Correlation results found for water shortage conditions, as defined by the WCI, show a similar pattern as for water stress and are given in Fig. S2 (Supplement). For both metrics, we found that, for a majority of the land area with a significant correlation to ENSO-driven climate variability, water scarcity conditions become more severe when the JMA SST index moves towards El Niño values (Table 3).

The regional variation in sensitivity of water scarcity conditions to ENSO-driven variability (Figs. 2 and S2) is clearly driven by the spatial distribution of water availability correlations as the general patterns are similar to those found in Fig. 1. The unequal clustering of water availability and consumptive water use leads, however, in some regions to a strengthening or weakening of the correlation signal, for example when comparing the regional variation in sensitivity results for water stress within the Amazon basin or in Southern Africa (Fig. 2) with the regional variation in correlation results for water availability in those areas (Fig. 1). For a selection of FPUs, we found significant correlations for both water availability and consumptive water use, while they lack significant correlations when considering water stress conditions, and vice versa. In Southeast Asia, for example, we observed significant correlations between ENSO and water availability and consumptive water use (Fig. 1), but no significant correlations between ENSO and water stress (Fig. 2). One explanation for this observation could be that if both water availability and consumptive water use increase or decrease with more or less the same strength under changing JMA SST values, the net effect on the CTA ratio could be insignificant since the ratio between both variables remains equal. All FPUs that show a significant correlation between water resources availability and ENSO-driven climate variability show as well a significant correlation with ENSO-driven variability when looking at the water shortage conditions (Fig. S2). This can be explained by the fact that the WCI is only driven by changes in water availability and population growth, of which the latter factor was fixed in this analysis.

Subsequently, we assessed the percentage anomalies in the median values of water scarcity conditions between El Niño and La Niña years, compared to the median values under all years. Significant anomalies (p ≤ 0.05, tested by regular bootstrapping n = 1000) in water scarcity conditions under El Niño and La Niña years, compared to all years, were found for 12.8 and 14.8 % of the global land area using the CTA ratio and the WCI, respectively (Table 4). The strongest anomaly signals were found during the La Niña phase for both water stress and shortage conditions.

Not all regions with a significant anomaly under El Niño years show (significant) anomalies in the opposite direction during La Niña years. For example, Fig. 3 visualizes the asymmetry in the anomalies found during the El Niño and La Niña phase for Latin America. Moreover, areas with significant correlations with the JMA SST index do not always show significant anomalies when looking at the different ENSO phases. This can be explained by the fact that only those years for which the 5-month moving average JMA SST index values are (±)0.5 ∘C or greater (El Niño)/smaller (La Niña) for at least 6 consecutive months (including October–December) are assigned as El Niño or La Niña years (see Sect. 2.5). Using this ENSO year definition thus disguises all variability in JMA SST values that falls just below the threshold set; i.e. variation that can have, however, a significant effect on water scarcity conditions.

Due to the socioeconomic developments over the period 1961–2010 water scarcity conditions and their associated impacts intensified, both in the absolute and relative sense (Fig. 4 and Table 5). From 1961 to 2010, using 5-year averaged values, the total global population increased from 2.97 to 6.25 billion. At the same time, we found that the global population exposed to water scarcity events increased from 0.45 billion to 2.47 billion. The global population sensitive to ENSO-driven climate variability increased with a factor of 2.4 over the same time period whilst its proportion to the global total population remained relatively unchanged (Table 5). The population sensitive to ENSO variability and living in areas exposed to water scarcity events currently represent only a minority of the global population (11.4 %). These results are, however, contrasted with relative high growth factors (Table 5). The impact the spatial clustering of population and consumptive water use, and their unequal growth rates, on water scarcity events is shown by the fact that the share of land area exposed to water scarcity events only doubled over this same period for the CTA ratio (Fig. 4), from 7.4 up to 16.5 % of the global land surface . The results found for water shortage (WCI ≤ 1700) are roughly similar at the global scale (Supplement Fig. S3, Table S1) and therefore not discussed individually in this section.

Regional variations in the population exposed to water stress and/or being sensitive to ENSO-driven climate variability under changing socioeconomic conditions, are visualized in Fig. 5. Although these regional figures do not lend themselves to a similar growth factor analysis, such as executed on the global numbers in Fig. 4, we can distinguish by means of visual inspection different characteristic region types. The first group of regions (Latin America Australia and the Pacific, the Caribbean, and Middle and Southern Africa) experiences significant correlations with ENSO variability for a relative large share of its land area and population (≥ 25 % of the total population in 2010) whilst exposure to water scarcity events is low (< 25 % of the total population in 2010). The second group of regions shows both a relatively low sensitivity to ENSO-driven climate variability (< 25 % of the total population in 2010) and low exposure to water scarcity events (< 25 % of the total population in 2010), e.g. northern America and western Europe. For the third group of regions (the Middle East, India, Southeast Asia, and western and central Asia) we find significant water scarcity exposure (≥ 25 % of the total population in 2010) but no or relative low sensitivity to ENSO variability (< 25 % of the total population in 2010). Finally, the fourth group of regions shows relatively high exposure to water scarcity events (≥ 25 % of the total population in 2010) and abundant sensitivity to ENSO-driven climate variability (≥ 25 % of the total population in 2010), e.g. China and northern Africa. Comparing these observations with the regional figures found for water shortage events (Supplement Fig. S4), assessed by means of the WCI, we found different results for the regions western and central Asia (relative high sensitivity to ENSO variability and relative low water scarcity exposure), and middle and southern Africa, the Middle East and Southeast Asia (both experiencing relative high sensitivity to ENSO variability and high exposure to water scarcity events). Using both water scarcity metrics (i.e. CTA ratio and WCI) in combination with the observed growth rates in population and population exposed to water scarcity events enables us to identify those regions where adaptation measures, such as ENSO-based forecasting, have the largest (future) potential in coping with and possibly reducing the adverse impacts of water scarcity events: the Caribbean, Latin America, western and central Asia, middle and southern Africa, northern Africa, the Middle East, China, Southeast Asia and Australia, and the Pacific.

The cross-model validation exercise, in which we compared the outcomes of the individual global hydrological models with their ensemble-mean results, reveals that our findings considering the sensitivity of water availability, consumptive water use, and water scarcity conditions to ENSO-driven climate variability are robust in comparison to the use of different hydrological models. We found that for 22.8 % of the global land area (61.4 % of the total land area with a significant correlation under the ensemble mean) all individual GHMs show a significant correlation to variations in JMA SST in the same direction as the correlation results found under the ensemble means. Correlations found under the ensemble mean are supported by at least one of the global hydrological models for one-third (36.8 %) of the global land surface (Fig. 6), equal to 99.2 % of the land area that shows a significant correlation to the ensemble mean.

A comparison of the individual modelling results with the ensemble mean in terms of the estimated population exposed to water scarcity events and/or living in areas sensitivity to ENSO-driven climate variability shows the modelling spread at the global scale with respect to estimated impacts and their developments over time (Fig. 7). Looking at the 2010 values, we find the smallest percentage difference between models in the estimates of the population exposed to water scarcity events (+17.2 % CTA ratio, +21.8 % WCI), and the largest variations when looking at the population both being exposed to water scarcity events and living in areas sensitive to ENSO-driven climate variability (+68.9 % CTA ratio, +54.2 % WCI). Percentage deviations were found to be smaller when looking at the land area exposed (Supplement Fig. S5). As shown in Fig. 7 and Fig. S5, the inter-model comparison reveals that the impact estimates of the ensemble mean are conservative when comparing them with the individual modelling results, especially when looking at the population or land area sensitive to ENSO variability and/or being exposed to water scarcity events.