Amplified Drought and Flood Risk Under Future Socioeconomic and Climatic Change

The economic stress and damage from natural hazards are escalating at an alarming rate, calling for anticipatory risk management. Yet few studies have projected flood and drought risk, owing to large uncertainties, strong non‐linearities, and complex spatial‐temporal dynamics. Here, we develop an integrative global risk analysis framework encapsulating future changes in flood and drought hazards as well as associated exposure and vulnerability dimensions. Flood characteristics are quantified by fitting a generalized extreme value distribution (GEV) to the annual flow maxima time series, while drought properties are characterized by the standardized precipitation evapotranspiration index (SPEI) and the standardized precipitation index (SPI). The drivers of drought and flood risk changes at the global and regional scales are explored, and the wide cascade of uncertainties in the risk assessment is decomposed. We find a substantial increase in both flood and drought risk towards the end of the century over most of the globe, driven by compounding changes in exposure, vulnerability, and hazard. A shift from a fossil‐fueled development to a sustainable one decreases the global area facing a risk doubling from 61% to 33% for flood and from 41% to 23% for drought. South America and Africa are identified as hotspot regions where a concomitant, large increase in both flood and drought risk are projected. The hazard quantification method is ubiquitously the dominant uncertainty source for drought risk changes, while the contribution of uncertainty sources for flood risk changes is highly variable in space.

In previous global flood risk studies, vulnerability has been considered by using a limited number of proxy indicators. This is partly due to the limited availability of future gridded socioeconomic data at the global scale, while a wide range of indicators is available at the local and regional scales to determine flood or drought vulnerability (Engström et al., 2020;Kim et al., 2015;Nasiri et al., 2016;Scheuer et al., 2011). Most global flood studies to date have solely considered future population exposure and economic vulnerability as expressed by gross domestic product (GDP) (Hinkel et al., 2014;Jongman et al., 2015;Kinoshita et al., 2018;Liao et al., 2019;Tanoue et al., 2016;Vafeidis et al., 2019;Ward et al., 2017). Yet GDP alone is a poor proxy of vulnerability because the vulnerability of communities and/or people can be very large even in rich countries, as exemplified by the summer 2003 European heatwave with a total of 68,312 fatalities (Formetta and Feyen, 2019). Other studies have projected changes in global flood risk using static exposure and/or vulnerability and dynamic hazards (Alfieri et al., 2017;Hirabayashi et al., 2013;Lange et al., 2020;Willner et al., 2018). Proper attention to the dynamic nature of exposure and vulnerability is, however, extremely important considering that the design and implementation of risk management and adaptation policies may lead to a risk probability decrease in short-term, but an increase over longer-term (Cardona et al., 2012). In this context, the future risk increase of some extreme events was found to be driven by future population expansion in hazard-prone regions rather than hazard intensification (Knorr et al., 2016). Apart from the population exposure and economic vulnerability, additional dimensions of flood vulnerability such as urbanization and agricultural value have been investigated in a handful future Jongman et al., 2012;Vousdoukas et al., 2018Vousdoukas et al., , 2020Winsemius et al., 2013Winsemius et al., , 2016 and historical  flood risk assessments.
For future drought risk, a relatively larger number of global studies have been carried out using a broad variety of approaches and indicators. Nevertheless, previous drought risk assessments mostly considered historical hazard, exposure, and vulnerability (Carrão et al., 2016;Meza et al., 2020), dynamic future hazards under static exposure and/or vulnerability (Gu et al., 2020;Lange et al., 2020;Li et al., 2009;Pokhrel et al., 2021) and dynamic hazards with only a dynamic exposure (Arnell et al., 2018;Gu et al., 2020;Liu et al., 2018;Smirnov et al., 2016). Yet none of the previous studies have considered dynamic hazard, exposure, and vulnerability for future risk assessments of global droughts.
The knowledge gaps in drought and flood risk assessments demonstrate the need to develop an integrated global vulnerability and risk framework to capture a greater range of dimensions and factors of vulnerability and disaster risk (e.g., social, cultural, economic, physical, and environmental aspects), the nonlinear relations of the factors and the dynamics of the vulnerability and risk. The limited dimensions of vulnerability used in previous studies may constitute a serious limitation in global risk estimates. The importance of integrating different vulnerability indicators was particularly emphasized for future climate change impact assessments, e.g., combining three indicators of national-scale education level, the gender gap in education, and GDP to characterize the quality and efficiency of governance for climate change adaptation (Andrijevic et al., 2020).
Risk analysis also involves several methods which introduce uncertainty to the analysis results, such as climate modeling, hydrological modeling, and hazard and vulnerability assessments (Hosseinzadehtalaei et al., 2020). The uncertainty quantification in the field of risk assessment, however, remains largely underdeveloped, despite important information it provides for a better interpretation of climate change impacts as well as for informed policies to mitigate the associated risk (Webster et al., 2003). Furthermore, the relative importance of exposure, vulnerability, and hazard in projected flood and drought risk has rarely been quantified.
In this study, we provide a comprehensive assessment of the risk changes for both flood and drought globally. To this end, an integrative global framework is developed to combine changes in different flood and drought characteristics and in different exposure and vulnerability components. We define the risk as the multiplication of the hazards and exposure-vulnerability metrics. Three scenario combinations for future risk assessments are considered: "Sustainability" (shared socioeconomic pathway (SSP) SSP1 combined with representative concentration pathway (RCP) RCP2.6), "Regional rivalry" (SSP3 combined with RCP6.0), and "Fossil-fueled development" (SSP5 combined with RCP8.5). The Sustainability and Fossil-fueled development scenarios both envision a world with a relatively high income growth, rapid urbanization, and higher investments in health and education. While in the former this relatively optimistic trend for future economic and social developments is obtained by an increasing shift towards sustainable practices, it is achievable by an energy-intensive, fossil fuel-based economy in the latter. The Regional rivalry scenario describes pessimistic human development with a lack of investments in education and health, a fast-growing population, and more inequality (O'Neill et al., 2016).
We develop a composite exposure and vulnerability index by consolidating changes in social, economic, and physical proxy indicators. Because the future gridded socioeconomic data relevant for flood and drought vulnerability are only available for population, GDP, and land use, they are used here to derive the indicators. Flood and drought hazard indices are also established by integrating changes in intensity, duration, and frequency of the hazards. Our framework provides a comprehensive risk assessment given that disregarding any of the determinants of exposure, vulnerability, and hazard components may lead to an underestimation of the risk. For instance, a drought of moderate severity but with a long duration is hardly recognized as an extreme event, while it may lead to large socioeconomic losses due to its potential to rapidly deplete stored water and thereby diminish resilience to following droughts (Lehner et al., 2017). This issue is efficiently addressed in our hazard index which combines changes in different characteristics.
We characterize drought properties by the standardized precipitation evapotranspiration index (SPEI) and the standardized precipitation index (SPI) using simulations of 18 global climate models (GCMs) from the Coupled Model Intercomparison Project Phase 5 (CMIP5), (Taylor et al., 2012) under RCP2.6, RCP6.0, and RCP8.5 scenarios. Flood characteristics are quantified by fitting a generalized extreme value distribution (GEV) to the annual flow maxima time series from 84 realizations (seven global hydrological models, GHM, fed by four CMIP5 GCMs under RCP2.6, RCP6.0, and RCP8.5) from phase 2b of the Inter-Sectoral Impact Model Intercomparison Project (ISIMIP2b; Frieler et al., 2017). The drivers of drought and flood risk changes at the global and regional scales are explored by analyzing the changes in the hazard and exposure-vulnerability components. Finally, the wide cascade of uncertainties in the risk assessment is decomposed at the regional scale.

Data
For the drought risk assessment, we analyze monthly precipitation, maximum, minimum and mean air temperature simulations from 18 GCMs participating in CMIP5 (Table S1 in Supporting Information S1). The 18 models were chosen based on their data availability. To allow all multi-model ensemble members to receive equal weight in ensemble median (i.e., unweighted median), only the first realization (r1) of each model is considered. The selected historical period is from 1971 to 2000 and the future period from 2070 to 2099 forced by three representative concentration pathways (RCPs), (Van Vuuren et al., 2011) 2.6, 6.0, and 8.5. These pathways refer to a stringent mitigation scenario resulting in low radiative forcing levels (RCP2.6), a medium-high scenario (RCP6.0), and a high-end emission scenario (RCP8.5). Although the drought risk analysis may have limited practical implications for deserts and hyper-arid regions, it may provide useful information for a physical understanding of climate change impact in these regions. The drought risk analysis is hence performed on the entire global land area as opposed to some studies which excluded deserts and hyper-arid regions (Carrão et al., 2016;Naumann et al., 2018).
We also analyze daily discharge simulations from seven ISIMIP global hydrological models (GHMs) forced by climate projections from four CMIP5 GCMs (GFDL-ESM2M, HadGEM2-ES, IPSL-CM5A-LR, and MI-ROC5) for climate change impact assessment on flooding. The GHMs are CLM4.5 (Lawrence et al., 2011; 10.1029/2021EF002295 4 of 24 Thiery et al., 2017), H08 (Hanasaki et al., 2008), JULES-W1 (Best et al., 2011), LPJmL (Bondeau et al., 2007), MATSIRO (Takata et al., 2003), ORCHIDEE (Guimberteau et al., 2014), and WaterGAP2 (Mueller Schmied et al., 2016). A comprehensive overview of the modeling frameworks of these GHMs is provided by Telteu et al. (2021). We consider the same historical and future periods and concentration pathways as the CMIP5 simulations for the ISIMIP2b simulations. Model grid cells with annual flow maxima of <1 m 3 s −1 of the historical period are excluded for further analysis, because of the smaller importance for flood analysis and inadequate data for distribution fitting. The Southern Hemisphere high-latitude landmass (>60°S) is also screened out as it is almost uninhabited and not subject to flooding.
The annual population density, urbanized areas, and GDP data on a 0.5° × 0.5° grid available through ISIMIP for the historical  and future  periods are used. The annual country-level GDP data were downscaled to 0.5° grid resolution based on the methodology described in Murakami and Yamagata (2019). Historical population density and urbanized areas data are based on the HYDE3.2 database (Klein Goldewijk et al., 2017). Future population data are derived from the national population projections based on the shared socioeconomic pathways (SSPs), (O'Neill et al., 2014) SSP1, SSP3, and SSP5 as described in Samir and Lutz (2014). The future urbanization patterns are based on the projections from the MAgPIE land-use model (Popp et al., 2014;Stevanovic et al., 2016) according to SSP2 and RCP2.6 and RCP6.0. To make a smooth transition between the historical (HYDE3.2) and future land-use patterns, a harmonization algorithm was applied to the MAgPIE data. We use the gridded land use data on a 0.5° × 0.5° grid from AIM-SSP/RCP Gridded Emissions and Land-use data set (Fujimori et al., 2018) to obtain the proportion of total croplands. The cropland proportion data for three combinations of SSP1-RCP2.6, SSP3-RCP6.0, and SSP5-RCP6.0 are utilized (SSP5-RCP8.5 combination is not available).

Assessment of Risk Changes
The change in the risk of drought and flood is calculated for each model grid cell by multiplying the change rate of hazard and exposure-vulnerability using Equation 1.
In which,  due to its indirect relation with vulnerability. The DHI and FHI will be defined in the following section. To avoid having zero values in the multiplicative approach of the CEVI E quantification, the proxy indicators are normalized between 0.05 and 0.95 for historical and future periods separately. Due to the limited relevance of urbanization for drought risk, it is not considered for the drought CEVI E . The risk assessment is performed under three combined scenarios: "Sustainability" (SSP1 combined with RCP2.6), "Regional rivalry" (SSP3 combined with RCP6.0), and "Fossil-fueled development" (SSP5 combined with RCP8.5).

Hazard Quantification Procedure
The change in the drought hazard index for each grid cell ( is computed as the multiplicative combination of the change in three drought properties of intensity, frequency, and duration. Equation 2a can then be written as follows (Equation 4): where at each grid cell g for future (f) and historical (h) periods, I, F, and D are drought intensity, frequency, and duration, respectively. We use an equal weight for the three properties as they are considered equally important in generating the hazardous impact of drought (Maccioni et al., 2015).
To characterize drought, we compute the time series of the Standardized Precipitation Evapotranspiration Index (SPEI), (Vicente-Serrano et al., 2010) values for each grid cell and each GCM. SPEI uses the difference between precipitation and evapotranspiration as a measure of drought or sustainable water availability, better characterizes the occurrence of dryer conditions and considers the impact of global warming (Peña- Gallardo et al., 2019). Evapotranspiration is calculated by the Hargreaves-Samani method (Hargreaves and Samani, 1985) that effectually incorporates solar radiation by its indirect approximation from maximum and minimum temperatures.
To explore the influence of excluding air temperature for the drought characterization, drought is also quantified using the standardized precipitation index (SPI), (McKee et al., 1993), which is the most frequently used index for drought characterization (Alborzi et al., 2018;Chiang et al., 2018;Hirschi et al., 2011;Tabari et al., 2012), also recommended by the World Meteorological Organization (WMO, 2012) and the National Drought Mitigation Center (NDMC, 2018). The empirical probabilities of precipitation for SPI and precipitation-evapotranspiration difference for SPEI are computed to avoid any assumption on the underlying distribution function of the data (Farahmand and AghaKouchak, 2015). The difference between the SPEI and SPI results is analyzed in terms of hazard quantification uncertainty in the uncertainty decomposition of future drought risk. The Wilcoxon rank-sum test is also used for testing the statistical significance of the difference.
To account for annual drought patterns, a 12-month accumulation period is used which represents anomalies of accumulated precipitation or precipitation-evapotranspiration difference for the given month and the 11 previous months. This is the time scale of the globally most extreme drought events in the last decades (Sheffield and Wood, 2011) and was shown to have the best predictive ability of the likelihood of impact occurrence in different sectors (Blauhut et al., 2016). Similar to Spinoni et al. (2020), we select the entire period (1971-2099) as a baseline period to compute the anomalies in the SPI and SPEI methods. Using a longer baseline period ensures a more robust characterization of standardized drought indicators (Wu et al., 2005), as the selection of a shorter period with a different drought frequency and severity may affect the indicator values over the entire period, resulting in underestimation or overestimation of climate change impact on drought. Furthermore, using past data as a baseline period to explore future drought events may introduce bias in calculated climate change signals, as normal conditions in the past will change in the future, leading to unrealistically extreme events at the end of the twenty-first century.
We derive the drought characteristics (i.e., drought duration, frequency, and intensity) using the run theory where a run represents a part of time series wherein drought index is either below (negative run) or over (positive run) a given drought defining threshold. In the current study, a drought event occurs every time the SPI/SPEI indicator falls below −1 for at least two consecutive months, and it ends when the indicator rises above zero (McKee et al., 1993). The drought threshold of −1 corresponds to the higher limit of a moderate drought. Drought in this study, therefore, includes moderate (−1.5 < SPI/SPEI < −1), severe (−2 < SPI/SPEI<−1.5), and extreme (SPI/SPEI<−2) droughts. The drought frequency is defined as the number of drought events for the historical  and future (2070-2099) periods. The duration of a drought event refers to the number of months between its start and end month. The drought intensity (dimensionless) is the average indicator value of a drought event. For each GCM, we calculate drought frequency and the medians of drought duration and intensity across all drought events at each grid cell for historical and future periods. The ratio of the magnitude of drought characteristics in the future period over that of the corresponding characteristics in the historical period is then used in Equation 4. To compute the ensemble median, the CMIP5 GCMs with different horizontal resolutions are resampled from their native scale to a 0.5° (latitude) ×0.5° (longitude) grid. We use the last-step procedure for the resampling where climate change signals are calculated directly on the native grids of GCMs, and then resampled. The error of the last-step procedure is less than the first-step procedure in which precipitation is first resampled, and then climate change signals are computed (Diaconescu et al., 2015).
Similar to DHI, the change in flood hazard index (FHI) for each grid cell ( is calculated as the multiplicative combination of the change rate of flood intensity and frequency using Equation 5. Note that the change in flood duration cannot be determined using the available data. Equation 2b can accordingly be written as follows (Equation 5): The notations have the same meaning as before. The change in flood frequency is computed as the change in the return period (T). Because an increase in the return period means a decrease in frequency and vise versa, the relative change in return period is reversed in Equation 5 to obtain the change in flood frequency.
To exclude the possible extrapolation bias in estimating flow quantiles from the 30-year data (1971-2000 and 2070-2099) and to enable a direct comparison with similar studies (e.g., Arnell et al., 2018;Dankers et al., 2014), the river flow with a 30-year return period (average occurrence of once every 30 years) at each model grid cell is used as a proxy for flooding. To this end, a generalized extreme value distribution (GEV) is fitted to the annual flow maxima time series for each native grid cell of the GHMs for both historical and future periods (Tabari, 2021). The river flow intensity associated with the return period T is subsequently computed by: where T E I is the river flow for return period T (30 in our case), and  E ,  E , and  E are the location, scale, and shape parameters of the GEV distribution. The ratio of the flood intensity of a 30-year return period between future and historical periods is utilized in Equation 5 as the change in flood intensity.
To calculate the frequency of flood events, the future return period of the historical extreme flow of T = 30 years is derived from Equation 7.
is either 0 or 1. The historical return period (30 in this study) is divided over the obtained future return period in Equation 5 as the change factor in flood frequency.

Decomposing the Cascade of Uncertainties
We partition the cascade of uncertainties in flood and drought risk changes at a regional scale following the variance decomposition-same sample size (VD-SSS) method (Tabari et al., 2019) as traditional VD methods (Greve et al., 2018) artificially inflates uncertainty contributions from sources with larger sample sizes (Hosseinzadehtalaei et al., 2017). The total uncertainty in the projections of future flood and drought risk is partitioned into contributions from climate model response (GCM uncertainty), global hydrological models for flood and the multi-scaler indices for drought (hazard quantification uncertainty), emission scenarios for the future hazard projections (RCP uncertainty), and SSP scenarios for the future CEVI projections (SSP uncertainty).
In the iterative sampling-theory based bootstrapping procedure of VD-SSS, let n denotes the smallest sample size among the uncertainty components (3 for flood risk and 2 for drought risk) and N be the ensemble size. After taking the median across the other uncertainty components, n is randomly drawn from N and the standard deviation across the bootstrap samples is calculated. This process is repeated many times (1,000 iterations in our case) and the median of the empirical bootstrap distribution of sample standard deviation signifies the uncertainty. The traditional VD method is applied for the uncertainty component of size n (here, hazard quantification uncertainty for drought and RCP and SSP uncertainties for flood). The fractional uncertainties are defined as the uncertainty of each source divided by the total uncertainty and the total uncertainty equals the sum of the uncertainty contributions (Tabari et al., 2019).

Projected Changes in Drought Risk
The global map of the future changes in drought risk based on the multi-model ensemble median shows strong increases (>threefold) in the risk by the end of the 21st century in most of South America, the Caribbean, central America, the United States, southern Europe, Middle East and North Africa (MENA), southern Africa, and Australia ( Figure 1). Projected drought risk decreases in the rest of the global land area, mainly in the Russian Federation, northern Europe, southern Alaska, and Canada, but also parts of South America, and eastern and southeastern Asia. Of the global land grid cells, 50%, 62%, and 61% show an increase in the drought risk under the Sustainability, Regional rivalry, and Fossil-fueled development scenarios, respectively. For the respective scenarios, the drought risk is projected to at least double in 23%, 39%, and 41% of the global landmass and at least triple in 13%, 28%, and 31% of the area. As expected, the area experiencing an increased drought risk increases with rising GHG emissions and socioeconomic developments.
To further quantify the spatial distribution of the risk changes, the changes are analyzed for 21 continental and subcontinental land regions (Giorgi and Francisco, 2000) ( Figure S1 in Supporting Information S1). The drought risk is expected to increase in 13, 17, and 16 regions under the Sustainability, Regional rivalry, and Fossil-fueled development scenarios, respectively (Figure 2a). The projected changes in the drought risk increase from the Sustainability to the Fossil-fueled development scenario, with drought risk tripling in 2, 5, and 7 regions under the Sustainability, Regional rivalry, and Fossil-fueled development scenarios, respectively. Apart from the Sahara region, the highest increases in the drought risk are found in southern Africa, the Mediterranean basin, central Asia, and central North America where the future risk is more than five times larger than the historical risk under the Fossil-fueled development scenario. In these regions, 100% of the area experience a drought risk increase under the Regional rivalry and Fossil-fueled development scenarios, while for the Sustainability scenario 83%, 89%, 91%, and 100% of the area of central Asia, central North America, Mediterranean basin, and southern Africa face a drought risk increase (Figure 3). Apart from the area under risk, an increasing trend magnitude of the risk changes with GHG emissions and socioeconomic developments is also evident for all the regions (Figure 2a). In the four hotspot regions, the risk change under the Fossil-fueled development scenario is 2-2.7 times larger than that under the Sustainability scenario (Figure 3). The results from the exploration of the drivers of drought risk changes demonstrate that the risk changes ( Figure 1) are mostly driven by shifts in the drought hazard index (DHI; Figures S2-S4 in Supporting Information S1) rather than by alterations in the composite exposure-vulnerability index (CEVI) ( Figure S5 in Supporting Information S1). The spatial pattern of the drought risk changes mirrors the pattern of the changes in DHI and drought properties. We then decompose the regional changes in the drought risk into exposure-vulnerability and hazard components. A comparison between the regional changes in the DHI ( Figure 2b) and CEVI (Figure 4a) with the corresponding scenarios reveals that the drivers vary with the scenario and in space. No change in CEVI is projected for 8, 6, and 9 regions under the Sustainability, Regional rivalry, and Fossil-fueled development scenarios, respectively, of which, Alaska, Australia, Greenland, north Asia, and Tibet are common among the scenarios. In the regions with no CEVI change, the drought risk change is due to the DHI change. Only in northern Europe under the Fossil-fueled development scenario,  Table S2 in Supporting Information S1 for the full names of the abbreviated regions. Violin plots depict the probability density of drought ensemble (18 GCMs), and the ensemble median is shown by black cross. no CEVI change is accompanied by no DHI change, leading to no risk change. Southeast Asia under the Sustainability and Regional rivalry scenarios and south Asia under the Fossil-fueled development scenario show an opposite pattern, where the risk change is totally attributable to the CEVI change as no change happens in DHI. The drought risk changes in 2 regions (northern Europe and western Africa) under the Sustainability scenario and in 4 regions (eastern and western Africa, eastern North America, and northern Europe) under the Regional rivalry scenario are driven by changes in both DHI and CEVI with a comparable magnitude. In 2 regions (eastern and southern Africa) under the Sustainability scenario, in 8 regions (Amazon and Mediterranean basins, central and western America, central North America, southern South America, Sahara, and southern Africa) under the Regional rivalry scenario and in 5 regions (central and western North America and eastern, southern, and western Africa) under the Fossil-fueled development scenario, the DHI increase is far larger than the CEVI increase. The drought risk change in these regions is thus more attributed to the DHI changes. In contrast, a larger increase in CEVI is observed in 2 regions (east and south Asia) under the Sustainability scenario and in one region (eastern North America) under the Fossil-fueled development scenario, attributing drought risk changes to the CEVI changes. Opposite signs of DHI and CEVI changes are seen in 6 regions (Amazon and Mediterranean basins, central North America, central America, southern South America, and central Asia) under the Sustainability scenario, in 2 regions (east and south Asia) under the Regional rivalry scenario and in 5 regions (central America, southern South America and central, east, and southeast Asia) under the Fossil-fueled development scenario, whereby the projected decrease in CEVI partially offsets the increase in DHI. The contribution of the DHI changes to the drought risk change is, therefore, larger in these regions.
We further explore the drivers of the drought risk change by analyzing the changes in the components of DHI and CEVI. The drought properties in order of decreasing magnitude of relative changes are frequency, duration, and intensity. The spatial distribution of the DHI change follows that of the drought frequency change ( Figures S2-S4 in Supporting Information S1). The drought risk pattern becomes more conspicuous by a partly similar global pattern of changes in drought duration and intensity. Regionally, the drought properties hold the same order of the change magnitude in almost all regions ( Figure 5). The changes in drought frequency indeed show the most distinct regionality, with the ensemble median change ranging from −11% in northern Europe to 162% in the Sahara for the Sustainability scenario, from −23% in Greenland to 193% in the Sahara for the Regional rivalry scenario, and from −33% in Greenland to 201% in the Sahara for the Fossil-fueled development scenario ( Figure 5). The regionality is also noticeable in drought duration changes, where the ensemble median change varies between −10% in Alaska and 45% in the Sahara for the Sustainability scenario, −13% in Alaska and 87% in the Mediterranean basin for the Regional rivalry scenario, and −17% in Alaska and 115% in the Mediterranean basin for the Fossil-fueled development scenario. In contrast, the spatial variation in drought intensity changes is small, ranging between −1% and 8%.
We also determine which climate variables have contributed to the drought changes. A large increase in the drought hazard in the Amazon and Mediterranean basins, Australia, central America, southern Africa, and the Sahara is due to compounding total precipitation decrease and total evapotranspiration increase (Figures S6 in Supporting Information S1). In other regions, a concomitant increase in both evapotranspiration and precipitation leads to either an increase or decrease in drought hazard. In 11 regions, the increase rate of evapotranspiration is larger than the increase rate of precipitation with the most pronounced cases in central and western North America and central Asia, again leading to a drought hazard increase. A drought hazard decrease is seen in 4 regions where the increase rate of evapotranspiration is smaller than that of precipitation.
A comparison between the spatial distributions of the drought CEVI (Figures S5 in Supporting Information S1) and its components ( Figures S7-S8 in Supporting Information S1) indicates that the change in cropland proportion is the leading spatial mode of variability in the CEVI change. In fact, the high vulnerability to drought in Africa under all scenarios, in the Americas under the Regional rivalry and Fossil-fueled development scenarios, in Europe under the Fossil-fueled development scenario mainly comes from this component. High CEVI values in Africa and the Middle East are in part due to the projected population growth in the region. Thanks to a decreasing population and croplands, partly to an improving economy, the vulnerability to drought in most of Europe is low under the Sustainability scenario. In central, southern,  Table S2 in Supporting Information S1 for the full names of the abbreviated regions. and eastern Asia, GDP increase and somewhat controlled population growth and cropland expansion make the CEVI value low.

Projected Changes in Flood Risk
The world map of the future flood risk changes exhibits a dramatic rise in the risk across most of the global land area (Figure 6). In contrast, the flood risk will halve mainly in central and eastern Europe and partly in Nordic countries, southern South America, and central North America. Globally, the flood risk is projected to increase by the end of the 21st century in 65%, 78%, and 78% of the analyzed landmass for the Sustainability, Regional rivalry, and Fossil-fueled development scenarios, respectively. The twofold flood risk increase is found in 33%, 56%, and 61% of the unmasked global area for the respective scenarios. For the same order  Table S2 in Supporting Information S1 for the full names of the abbreviated regions. Violin plots depict the probability density of drought ensemble (18 GCMs), and the ensemble median is shown by black cross. of scenarios, 17%, 41%, and 79% of the area exhibit a threefold increase in the flood risk. The flood risk shows a similar order of sensitivity as the drought risk to rising GHG emissions and socioeconomic developments, highlighting the benefits of reducing GHG emissions and controlling socioeconomic changes.
The regional analysis of the change in the flood risk shows an increase in all regions except northern Europe under all scenarios, central Asia under the Sustainability and Fossil-fueled development scenarios, central America, east Asia, the Mediterranean basin, and Tibet under the Sustainability scenario (Figure 7a). Based on the ensemble median, the flood risk is likely to double at the end of the century in 4, 13, and 16 regions under the Sustainability, Regional rivalry, and Fossil-fueled development scenarios, respectively. For the respective scenarios, the flood risk is expected to triple in 0, 9, and 10 regions and to quadruple in 0, 5, and 5 regions. Eastern and southern Africa and south and southeast Asia exhibit the largest increases in the future flood risk, respectively, with at least four-fold increases for the Regional rivalry and Fossil-fueled development scenarios. Comparing the area of these regions with a quadruple flood risk increase indicates that the quadruple increase under the Fossil-fueled development scenario is observed in 79%, 69%, 65%, and 67% of the area in eastern and southern Africa and south and southeast Asia, respectively ( Figure 8). For the Sustainability and Regional rivalry scenarios, the percentage area of the quadruple flood risk increase is generally smaller, ranging between 53% and 74% for the Sustainability scenario and between 17% and 28% for the Regional rivalry scenario.
The exploration of the drivers of flood risk changes shows that the flood risk changes (Figure 6) in most of the land area are generated mainly by the changes in FHI (Figures S9-S11 in Supporting Information S1) rather than the CEVI changes (Figures S5 in Supporting Information S1). In contrast, the spatial distribution of the flood risk changes follows the spatial distributions of both the FHI and CEVI changes which resemble each other. At the regional scale, however, the drivers of the flood risk change vary with region and scenario (comparing Figures 4 and 7b) as in drought risk. In 10 regions under the Fossil-fueled development scenario, the large FHI increases are muted by either a weaker increase (central and eastern North America and eastern and southern, and western Africa) or a decrease (central America, southern South America, and east, south, and southeast Asia) in CEVI. A reduction of the magnitude of the flood risk increase because of a weaker increase or decrease in CEVI compared to FHI is seen respectively in 6 and 1 regions under the Regional rivalry scenario and in 1 and 5 regions under the Sustainability scenario. In 7, 7, and 8 regions respectively under the Sustainability, Regional rivalry, and Fossil-fueled development scenarios, the flood risk change stems from the FHI change because of no CEVI change. The CEVI decrease controls the flood risk decrease in central America, central and east Asia, and the Mediterranean basin under the Sustainability scenario, and in central Asia under the Fossil-fueled development scenario. A decrease in both FHI and CEVI contributes to the flood risk decrease in northern Europe under the Sustainability and Regional rivalry scenarios. The flood risk increase in western North America under the Regional rivalry and Fossil-fueled development scenarios is driven by a CEVI increase. The CEVI contribution to the flood risk increase in 4 regions (eastern North America, Mediterranean basin, and southern and western Africa) under the Regional rivalry scenario is larger than that of FHI. The FHI and CEVI have a comparable contribution to the flood risk increase in 2 regions (southern and western Africa) under the Fossil-fueled development scenario. The global pattern of the flood characteristics resembles each other and that of FHI (Figures S9-S11 in Supporting Information S1). The change magnitude of flood frequency is far larger than that of flood intensity. Regionally, a larger change in flood frequency than in flood intensity is seen for all regions (Figure 9), accounting for, on average, 32%, 58%, and 94% under the Sustainability, Regional rivalry, and Fossil-fueled development scenarios, respectively. The magnitude of the regional changes in flood frequency vary from −28% in northern Europe to 77% in south Asia for the Sustainability scenario, from −34% in northern Europe to 230% in south Asia for the Regional rivalry scenario, and from −55% in northern Europe to 416% in eastern Africa for the Fossil-fueled development scenario, while those for flood intensity vary from −5%, −6%, and −13 in northern Europe to 17%, 42%, and 58% in south Asia for the respective scenarios.
Comparing the spatial distributions of the flood CEVI and its components reveal that the change in cropland proportion is the largest (Figures S7-S8 in Supporting Information S1). The CEVI changes for flood  Table S2 in Supporting Information S1 for the full names of the abbreviated regions. Model grid cells with annual flow maxima of <1 m 3 s −1 of the historical model period are screened out, excluding the entire Sahara region. Violin plots depict the probability density of flood ensemble (28 experiments resulting from combination of 7 GHMs and 4 GCMs), and the ensemble median is shown by black cross. and drought have an almost identical geographical pattern (Figures S6 in Supporting Information S1), making the overall impact of urbanization, which was used as an additional proxy indicator for the flood risk assessment, small. Regionally, under the Sustainability scenario, the change magnitude of the flood CEVI in 8 regions (Amazon and Mediterranean basins, central North America, central America, southern South America, and east and southeast Asia) is higher compared to the drought CEVI, while it is lower for 2 regions (northern Europe and western Africa) (Figure 4). Regionally, the change magnitude of the flood CEVI is higher compared to the drought CEVI wherever they are increasing (1, 8, and 3 regions for the Sustainability, Regional rivalry, and Fossil-fueled development scenarios, respectively) ( Figure 4). The opposite situation is seen for decreasing CEVI (7, 3, and 7 regions for the Sustainability, Regional rivalry, and Fossil-fueled development scenarios, respectively): a higher decrease magnitude for the drought CEVI in comparison with the flood CEVI. In the rest of the cases, there is no change in both flood and drought CEVI or the difference between them is negligible (<5%).

Uncertainty in Drought and Flood Risk Projections
The future projections of drought and flood risk are subject to a large uncertainty (Figures S12 in Supporting Information S1). The highest total uncertainties in drought risk projections are seen in southern Africa, the Sahara, central Asia, and the Mediterranean basin, and the lowest ones in southeast Asia, south Asia, and northern Europe, and east Asia. The spatial distribution of the total uncertainty for flood risk projections is different, with the largest uncertainties in eastern, western, and southern Africa and southeast Asia, and the smallest ones in central Asia, northern Europe, and the Mediterranean basin.
The total uncertainty in the projections of future flood and drought risk is decomposed into GCM, hazard quantification method (choice between GHMs for flood and between multi-scaler indices for drought), RCP, and SSP uncertainties. For drought risk projections, hazard quantification uncertainty dominates everywhere except over southeast Asia where SSP is the main uncertainty source (Figure 10a). The hazard quantification uncertainty is larger than 50% in 12 out of the 21 regions and ranges from 24% in southeast Asia to 67% in north Asia. In fact, the changes in the drought hazard derived from SPEI are 2, 2.8, and 3.6 times (and similar ratios for the risk) larger than those from SPI for RCP2.6, RCP6.0, and RCP8.5, respectively (Figures S13 in Supporting Information S1). It shows that the difference between SPEI-and SPI-based drought risk changes becomes more obvious under higher warming levels. A larger discrepancy between the results of different drought indices under the stronger warming of RCP8.5 is consistent with previous studies (Lehner et al., 2017). In addition to the change magnitude, an opposite sign of change (hazard decrease by SPI vs. hazard increase by SPEI) is found in 11 regions, implying that an increase in precipitation in these regions is outbalanced by an increase in atmospheric evaporative demand forced by higher temperatures. The difference between the SPI and SPEI results is statistically significant in all the regions for the three scenarios except southeast Asia under RCP2.6 ( Table S2 in Supporting Information S1). The  Table S2 in Supporting Information S1 for the full names of the abbreviated regions. Model grid cells with annual flow maxima of <1 m 3 s −1 of the historical model period are screened out, excluding the entire Sahara region. Violin plots depict the probability density of flood ensemble (28 experiments resulting from combination of 7 GHMs and 4 GCMs), and the ensemble median is shown by black cross. uncertainty decomposition analysis also reveals that scenarios (RCP and SSP) contribute least to the total uncertainty in all regions except in central America where GCM uncertainty has the smallest contribution (Figure 10a). The contribution of GCM uncertainty ranges between 14% in north Asia and 38% in Australia.
In contrast to drought risk projections, the contribution of uncertainty sources for flood risk projections has a high spatial variability (Figure 10b). While hazard quantification dominates the total uncertainty in 9 out of the 20 considered regions for the flood analysis, it is the least important contributor in seven regions. A closer look at uncertainty decomposition of flood risk projections results reveals that hazard quantification uncertainty contributes most in the regions located in mid and high latitudes in the Northern Hemisphere (8 out of 11 regions). Overall, hazard quantification explains 4%-59% of the total variability in flood risk projections over the 20 considered regions. GCM and RCP are the most important source of uncertainty in 5 regions, with contributions of 6%-72% and 2%-51%, respectively. Only a small subset of available CMIP5 GCMs was used to force the GHMs and therefore might underestimate this source of uncertainty. SSP uncertainty is the main source in only one region, explaining 35% of the total uncertainty in southern Africa.

On Drought and Flood Risk
Our results show that the drought risk increases in most of the global land area, with a decrease mainly found at high latitudes (e.g., Russia, Canada). The drought risk changes are driven by compounding changes in exposure, vulnerability, and hazard and primarily by hazard. An amplification of drought conditions in a warmer climate is anticipated owing to the theorized expectation of an increase in evaporative demand at the rate of 1.5%-4% (Scheff and Frierson, 2014) per degree warming according to the Clausius-Clapeyron relationship. This scaling rate is not, however, uniform across the globe because of the intervention of other factors such as moisture availability (Berg & McColl, 2021;Tabari, 2020). Our results also point to an increase in evaporative demand in all 21 regions of the world, confirming recent findings (Greve et al., 2019;Konapala et al., 2020). We show a more severe drought-prone condition in the future in several regions such as southern Africa and the Mediterranean basin, posing large challenges for existing socio-hydrological systems there and increasing the risk of environmental refugee displacement (Myers, 2002) and water conflicts (Hsiang et al., 2013). Based on our analysis, a larger drought increase in these regions is due to concurrent precipitation decrease and evapotranspiration increase.
We also quantified the changes in drought characteristics to understand their role in the amplified drought hazard and provide insight for adoption and mitigation strategies. As the changes in drought duration and frequency are far larger than those in drought intensity, adoption and mitigation strategies should focus more on long, frequent droughts in the future. More frequent and prolonged droughts due to climate change have also been reported using regional climate model simulations (Spinoni et al., 2020;Tabari and Willems, 2018). The increases in drought duration and frequencies would bring some complications for the disaster recovery and hydraulic system designs to withstand more prolonged and frequent periods of reduced water availabilities Naumann et al., 2018). The lengthening of drought events may also lead to more frequent megadroughts (multidecadal drought) in the future, with unprecedented stress on water resources (Ault et al., 2016).
Like the drought risk, the flood risk is also projected to increase in most of the analyzed global landmass, which for most regions stem from a flood hazard increase. The spatial pattern of flood hazard changes matches well with those shown in previous studies (Arnell & Gosling, 2016;Dankers et al., 2014;Donnelly et al., 2017;Hirabayashi et al., 2013;Lange et al., 2020). The magnitude and direction of the flood hazard change spatially vary with the role of flood generating mechanisms. The largest increases are found for the Southern Hemisphere, specifically in eastern and southern Africa and south and southeast Asia regions. Inversely, the decreasing signals are obtained in regions where flood is generated by factors other than extreme precipitation, which is generally expected to increase in the future. As an example of the regions experiencing a flood decrease, soil moisture is the dominant deriving force for flood changes over Europe, followed by snowmelt or rain-on-snow (Berghuijs et al., 2019). The flood magnitude is expected to decrease in areas with a snowmelt-dominated hydrograph in spring (Dankers et al., 2014) as precipitation shifts from snowfall to rain (Burn & Whitfield, 2016;Cunderlik and Ouarda, 2009) and the magnitude of snowmelt-related floods decreases due to a snowmelt timing change under global warming (Arnell & Gosling, 2016). An unclear direction of flood intensity changes in the regions where the most important factor is antecedence soil moisture was also noted previously and attributed to the interwoven effects of evaporation, and precipitation dynamics (Bennett et al., 2018;Ivancic and Shaw, 2015;Wasko and Nathan, 2019).
Scrutinizing flood risk results shows that in all regions except 4 (eastern and southern Africa, south and southeast Asia) under both the Regional rivalry and Fossil-fueled development scenarios and in 2 regions (Amazon basin and western Africa) under the Fossil-fueled development scenario, both increases and decreases in flood risk are projected by different GCM-GHMs combinations. It implies that the use of a single combination may lead to misleading results, calling for the use of multiple GHMs and GCMs for future flood analyses. The importance of using multiple GHMs and GCMs for analyzing and projecting extreme events is evident from our uncertainty analysis and has been emphasized in earlier studies (Giuntoli et al., 2015;Gudmundsson et al., 2021;Hattermann et al., 2018;Hosseinzadehtalaei et al., 2020;Reinecke et al., 2021;Tabari et al., 2019). High dependence of projected precipitation changes on the used GCMs has also been found (Osuch et al., 2016;Xu et al., 2019), casting doubt on the credibility of the studies based on few and/ or unrepresentative models (Wu et al., 2020).
Looking at the results of both flood and drought risk indicates that a concurrent increase in both flood and drought risk is prevalent and found in 8, 16, and 14 regions under the Sustainability, Regional rivalry, and Fossil-fueled development scenarios, respectively. Adaptation and mitigation plans will therefore be challenging in these regions, as actions implemented to diminish risk from one hazard may inadvertently result in risk increase from the other (Di Baldassarre et al., 2017). For example, drought protection favors high water storage in the reservoir, which makes dams more vulnerable to overtopping or failure when an extreme precipitation event occurs . The concurrent increase in both flood and drought risk is mainly the case for South America and Africa. This emphasizes the necessity for a joint analysis of drought and flood risk in these regions to develop integrated policies and practices for a deliberate and effective DRR.

On Risk Drivers
Although the risk changes in more than half of the regions are predominantly driven by the drought and flood hazard index changes especially by the hazard frequency, the exposure-vulnerability index is the controlling factor in some regions, exemplified by the high vulnerability to drought in Africa, mainly resulting from projected cropland expansion. Considering that people in these regions are highly dependent on agriculture for their livelihoods and access to food, the vulnerability to drought can be even larger than what is obtained here. For example, west Africa's severe drought period in the 1980s resulted in over half a million fatalities because of severe food shortages (Hulme, 1996;Traore et al., 2014) or north Africa's severe drought during the 2000-2011 period pushed 2-3 million people in extreme poverty (UN-DESA, 2013). Apart from the regional consequences of droughts, the global food security and cattle farming productivity will also be influenced since e.g., Brazil and Argentina in the Amazon basin and southern South America regions grow 53% of the world's soybeans and 42% of the world production of corn. While the Southern Hemisphere regions are vulnerable to natural hazards due to their insufficient infrastructural, social, and economic capacity to manage disasters, wealthy industrialized nations in the Northern Hemisphere are less vulnerable, thanks to their controlled population growth and economic prosperity.
As for the CEVI changes, the changes in flood and drought CEVI have an almost identical geographical pattern. There is also a clear resemblance between the global patterns of flood hazard and CEVI changes. This is because changes in socioeconomic elements were considered for flow simulations by most GHMs. That is to say, the historical (varying over the historical period) and present-day (fixed at 2005) socioeconomic scenarios were used respectively for historical and future discharge simulations in all the GHMs except for CLM4.5 which used the present-day scenario for both historical and future simulations. The geographical distributions of the changes in the CEVI components are comparable, as these socioeconomic factors are connected (O'Neill et al., 2014). For example, population density is linked to economic prosperity (Gu et al., 2020). The magnitude of the changes, however, notably varies among the CEVI components. This highlights the importance of using a composite exposure-vulnerability index for a comprehensive risk assessment for hydroclimatic extremes.

On Risk Uncertainty
We quantified uncertainties in future projections of flood and drought risk. The uncertainty analysis of drought risk projections specifies hazard quantification uncertainty, that is, the difference between SPEI and SPI results, as the main source. This comes as no surprise since air temperature plays a major role for a better modeling of moisture demand and availability (Ficklin and Novick, 2017;Su et al., 2018). In fact, the drought risk is underestimated using indices based on precipitation only. It was previously shown that the choice of the drought definition index greatly influences the direction and amplitude of changes (Burke & Brown, 2008). A higher sensitivity of the drought indices measuring the atmospheric demand for moisture to temperature changes has been reported to result in large discrepancies between the outputs of these indices and SPI (Burke, 2011;Dai, 2013;Dai and Zhao, 2017;Rhee and Cho, 2016;Touma et al., 2015;Zhao and Dai, 2017) and an underestimation of megadrought probabilities (Ault et al., 2016). The big differences in projected drought risk from SPEI and SPI obtained in this study underscore the need of incorporating air temperature for drought characterization in the context of progressive warming, as otherwise the risk of future droughts may be underestimated. In other words, drought indicators based on precipitation alone cannot adequately consider the influence of warming climate . This finding has important mitigation implications. That is, the dependence of the drought risk on air temperature highlights the benefit of aggressive GHG emission mitigation.
The uncertainty analysis of flood risk projections, in contrast, shows a spatially varying dominance of uncertainty sources. Hazard quantification dominates the total uncertainty over Northern Hemisphere midand high-latitude landmass, highlighting the importance of differences in model parameterization (evapotranspiration, snow, and runoff schemes) in this region (Do et al., 2020;Haddeland et al., 2011).

Limitations
The simple multi-dimensional risk framework provided by our study comprehensively analyzes the risk and its contributing components of hazard, exposure, and vulnerability. It highlights the need to include all hazard characteristics rather than a single characteristic which may underestimate the overall risk. There are, however, several caveats with our analysis. The drought and flood hazard indices were developed here as a multiplication of hazard intensity, duration, and frequency. Hazard characteristics can also be integrated by means of multivariate distribution functions.
We used river flow with a 30-year return period as a proxy for flooding like Dankers et al. (2014). To calculate flood depth and inundation areas, a global river routing model with an inundation scheme such as CaMa-Flood should be employed as done by Hirabayashi et al. (2013) and Lange et al. (2020). We also quantified the flood intensity and frequency using the ISIMIP GHMs which simulate flow regimes assuming spatially consistent routing parameters within model grid cells. In that way, the real climatological and hydrological systems for a specific location may not be accurately described (Arnell & Gosling, 2016). Furthermore, the GHMs were fed by climate simulations from four GCMs out of more than 30 available CMIP5 GCMs, which were selected based on data availability and representation of equilibrium climate sensitivity (ECS) of the CMIP5 ensemble (Frieler et al., 2017). The uncertainty in the flood risk projections might therefore be biased. The coarse-scale CMIP5 GCMs have also known deficiencies in the reproduction of extreme precipitation (Freitas et al., 2020;Tripathi and Dominguez, 2013) as one of the main drivers of flood changes (Tabari, 2021), enlarging the projection uncertainty in the flood hazard and thereby the flood risk. However, developing multi-model fine-resolution long-term climate simulations on a global scale is currently not computationally feasible.
For the calculation of the composite exposure-vulnerability index (CEVI), exposure and vulnerability were assumed equally important. The relative importance of the components may, however, vary spatially. The spatial distribution of vulnerability is expected to be mainly shaped by the spatial pattern of poverty and wealth. For poor communities, a hotspot of climate vulnerability is an area with high exposure to hazards along with high population density with uneven distribution, inadequately built infrastructure, and low levels of capacity to respond to disasters. For wealthy and developed countries, it is however an exposed area as well as concentrated valuable coastal real estate, economic assets, and the high density of settlements and heavy industrialization in risk-prone areas. Furthermore, for exposure and vulnerability, we integrated four proxy indicators concerning the social, economic, and physical aspects. Other proxy indicators can also be employed to assess the social vulnerabilities regarding the awareness and capacities of communities or societies to cope with and to adapt to hazards through resilience building in socio-ecological systems. In this context, a well-informed, more educated, and motivated population toward a culture of hazard prevention, safety, and resilience can result in disaster risk reduction.
Despite these limitations, this study highlights the importance of considering the dynamic and interacting nature of hazard, exposure, and vulnerability for a holistic future risk assessment of extreme events. The framework developed in our study serves as a starting point for a more systematic risk analysis of extreme events in future research. Such a framework provides valuable information to prioritize timely, integrated, and coordinated proactive adaptation and mitigation plans to hazards at the regional scales to prepare for changing climatic and socioeconomic conditions.

Conclusions
This study assessed future changes  in drought and flood risks at the global and regional scales, by integrating the changes in hazard characteristics and in social, economic, and physical vulnerabilities. The drought hazard was quantified by the SPEI and SPI indices using simulations of 18 CMIP5 GCMs under RCP2.6, RCP6.0, and RCP8.5 scenarios, and the flood hazard by the annual maxima using 84 ISIMIP2b realizations for the same scenarios. The drivers of drought and flood risk changes and the associated uncertainties were also analyzed. The results show a drought risk increase in most of the global landmass, with 50%, 62%, and 61% of the area experiencing an increase under the Sustainability (SSP1 & RCP2.6), Regional rivalry (SSP3 & RCP6.0), and Fossil-fueled development (SSP5 & RCP8.5) scenarios, respectively. The amplitude of changes in the drought risk is not uniform and varies among the regions. The highest drought risk increase of about five-fold is found for southern Africa, the Mediterranean basin, central Asia, and central North America. A similar spatial and scenario dependence is found for flood risk changes. The flood risk is projected to increase in the vast majority of the global land area (33%-61% depending on scenario), with the largest increases in eastern and southern Africa and south and southeast Asia regions. The global hotspot regions of concurrent, large drought and flood risk increase are identified in South America and Africa, where local governments should be prepared to deal with the challenges to develop informed, integrated climate policies and practices for deliberate and effective disaster risk reduction.
The drivers of drought and flood risk changes demonstrate a distinct regionality, but in more than half of the regions, the risk changes are predominantly driven by the drought and flood hazard index changes especially by the hazard frequency. Nevertheless, the exposure-vulnerability index changes control the risk changes in some regions, exemplified by the high vulnerability to drought in Africa, primarily originating from a projected cropland expansion. The decomposition of uncertainty sources for drought risk changes reveals hazard quantification methods as the main source everywhere. For flood risk changes, the role of uncertainty contributors varies spatially, but with the dominance of hazard quantification source over Northern Hemisphere mid-and high-latitude landmass.