Advances in Water Resources 64 (2014) 21–33

Contents lists available at ScienceDirect

Advances in Water Resources

journal homepage: www.elsevier .com/ locate/advwatres

Dynamic attribution of global water demand to surface waterand groundwater resources: Effects of abstractions and returnflows on river discharges

0309-1708/$ - see front matter � 2013 Elsevier Ltd. All rights reserved.http://dx.doi.org/10.1016/j.advwatres.2013.12.002

⇑ Corresponding author. Tel.: +31 0302532183.E-mail address: i.e.m.degraaf@uu.nl (I.E.M. de Graaf).

I.E.M. de Graaf a,⇑, L.P.H. van Beek a, Y. Wada a, M.F.P. Bierkens a,b

a Department of Physical Geography, Utrecht University, Utrecht, The Netherlandsb Unit Soil and Groundwater Systems, Deltares, Utrecht, The Netherlands

a r t i c l e i n f o a b s t r a c t

Article history:Received 8 August 2013Received in revised form 5 December 2013Accepted 7 December 2013Available online 19 December 2013

Keywords:Water abstractionsGroundwaterSurface waterReturn flowsGlobal hydrological modelRiver low flows

As human water demand is increasing worldwide, pressure on available water resources grows and theirsustainable exploitation is at risk. To mimic changes in exploitation intensity and the connecting feed-backs between surface water and groundwater systems, a dynamic attribution of demand to waterresources is necessary. However, current global-scale hydrological models lack the ability to do so. Thisstudy explores the dynamic attribution of water demand to simulated water availability. It accounts foressential feedbacks, such as return flows of unconsumed water and riverbed infiltration. Results showthat abstractions and feedbacks strongly affect water allocation over time, particularly in irrigated areas.Also residence time of water is affected, as shown by changes in low flow magnitude, frequency, and tim-ing. The dynamic representation of abstractions and feedbacks makes the model a suitable tool for assess-ing spatial and temporal impacts of changing global water demand on hydrology and water resources.

� 2013 Elsevier Ltd. All rights reserved.

1. Introduction

Worldwide water demand increased substantially over the pastdecades as a results of growing population numbers, expanded irri-gation areas, and economic development, raising water scarcity inmany parts of the world [1]. As a result, an increasing number ofrivers run dry for substantial periods of the year before reachingthe sea [2]. In regions with frequent water stress and large aquifersystems groundwater is often used as an additional resource tomeet water demands. In many of these areas groundwater abstrac-tions exceed groundwater recharge, depleting existing groundwa-ter stores, thereby negatively affecting stream flow of groundwaterfed rivers, ecosystems, and depths of local attainable groundwater[3].

Previous model studies that focused on global-scale water con-sumption and its effects had to deal with the fact that little to noinformation exists on the attribution of water demand to surfacewater and groundwater abstraction. Therefore, between studiesdifferent assumptions have been made about this attribution. Lim-iting ourselves to models that explicitly account for human waterabstractions, examples of attribution rules are: H08 [4], where sur-face water is preferentially abstracted, WBMplus [5], where water

from reservoirs and groundwater is preferentially abstracted,LPJmL [6], where irrigation demand is attributed to surface waterand groundwater resources using temporal invariant fractions,WaterGap [7] where sector specific abstractions are calculatedwith temporally invariant but country-specific fractions of totalwater demand, and PCR-GLOBWB [8] where also sector specificgross and net abstractions are calculated, and where groundwaterabstractions are constrained to reported values (i.e. IGRACwww.un-igrac.org). Thus, none of these attribution rules take intoaccount the abundance of both surface water and groundwater re-sources at the same time. The distribution rules of these modelsare potentially not very robust under changes in water availability,be it from climate change or water consumption. Wada et al. [9] at-tempted to include such changes in PCR-GLOBWB and improvedthe previous scheme to account for feedbacks between water sup-ply and demand. The fraction between baseflow and long-termaverage river discharge is used to allocate water demand to surfacewater and groundwater resources. However, the used fraction doesnot reflect actual changes in available surface water. Also returnflows are still static and thus do not affect actual water availabilityin groundwater and surface water resources.

The goal of this study is to explore a dynamic attributionscheme that is able to mimic changes in exploitation intensity ofsurface and groundwater. This scheme explicitly considers thefeedbacks connecting surface water and groundwater systemsand their exploitation. More specifically, compared to previous

22 I.E.M. de Graaf et al. / Advances in Water Resources 64 (2014) 21–33

work the following features are added in this study: (1) dynamicattribution of water demand to surface water and groundwaterbased on the actual availability of surface water and groundwater;(2) including the effects of groundwater and surface water abstrac-tions as well as the return flows on the surface water and ground-water system at run time (full integration of the hydrologicalcycle); (3) including a two-way interaction between surface waterand groundwater by allowing both drainage of groundwater to sur-face water bodies, as well as suppletion of groundwater by surfacewater through riverbed infiltration. By means of these additions wewill be able to simulate adjustments of preferred water use basedon changes in availability of surface water and groundwater due toclimate change or increased water consumption. Apart from thevalidation of the abstraction rates produced by this scheme wespecifically focus on the effects of return flows on river discharge.

We use the global hydrological model PCR-GLOBWB [10] tosimulate water storages and river discharges over the period1960–2010 (daily time step, 0.5� resolution, about 50 km at theequator). In this study, total water demand stems from irrigation,industry, and domestic use and defines the total abstraction ifsufficient water is available. Abstractions are variably taken fromsurface water and groundwater driven by simulated water avail-ability and are sector independent. Return flows of unconsumedabstracted water are simulated at the same time as the abstrac-tions. Return flows are sector specific and return to a single source;those of irrigation to the groundwater, those of industry anddomestic use to the surface water. In other words, return flowscause a redistribution of abstracted water over the water re-sources. Three model runs are used to test for the effects ofabstractions and return flows on water allocation and river dis-charges: (1) no abstractions (NA), (2) abstractions only (AB), (3)abstractions and return flows (ABRE).

The suggested dynamic allocation scheme is validated by com-paring simulated groundwater abstraction magnitudes and frac-tions for the year 2000 to reported values on global andcontinental scale. The impacts of abstractions on river discharges,especially during low flows, are analyzed globally by looking atchanges in flow magnitudes, frequency, and timing of low flows.Past trends of abstractions are given. These trends show temporalchanges in abstraction intensity under the influence of feedbackmechanisms.

2. Methods

2.1. Hydrological model

The global hydrological model PCR-GLOBWB [10] is used to cal-culate water storages and fluxes of the terrestrial part of the hydro-logical cycle for the period 1960–2010. A schematic representationof the model is given in Fig. 1. Only a summarized model descrip-tion is given here, for details we refer to [10].

PCR-GLOBWB is a grid-based hydrological model (here 0.5� gridresolution globally) that operates at a daily time step. Each grid cellcontains surface water elements and a vertically structured repre-sentation of the canopy, two soil layers, and an underlying ground-water reservoir. Sub-grid variability is used to represent fractionsof different vegetation (i.e. short and tall), saturated soil (to quan-tify surface runoff and lateral outflow from the unsaturated zone),and surface water (i.e. lakes, reservoirs, wetlands, floodplains). Pre-cipitation can be stored as canopy interception and as snow whentemperatures are below 0 �C. Throughfall and meltwater arepassed to the upper soil layer. Actual evapotranspiration is calcu-lated from potential evaporation and soil moisture conditions. Ver-tical exchange between the soil and groundwater layers occurs bypercolation and capillary rise. Drainage from the soil column to the

river network takes place as overland flow, subsurface flow fromthe two soil layers, and baseflow from the groundwater reservoir.This last reservoir is parameterized based on lithology and topog-raphy, and is represented as a linear reservoir model [11]. Thus, foreach time step and each grid cell the water balance of the soil col-umn is calculated. The combined runoff is accumulated and routedas river discharge along the drainage network based on DDM30[12] using a kinematic wave approximation on a sub-daily timescale. Open water evaporation, water storage in lakes, and attenu-ation by floodplains and wetlands are taken into account withinthe routing scheme. Reservoirs are located on the river networkbased on GLWD1 [13]. Reservoir storage and release are dynami-cally calculated by evaluation of the downstream water demand.This encompasses all blue water demand within an area 600 kmdownstream of the reservoir outlet (approximately a week withan average discharge velocity of 1 m s�1). When more than one res-ervoir is present directly upstream, demand is partitioned propor-tionally to reservoir capacity. PCR-GLOBWB was forced with dailyfields of precipitation, temperature, and reference potential evapo-transpiration over the period 1960–2010. For the period 1960–2000 precipitation and air temperature were prescribed by theCRU TS 2.1 monthly dataset [14,15], which was downscaled todaily fields by using the ERA-40 reanalysis [16]. For the period2000–2010 climate data were retrieved for the ERA-Interim [17]reanalysis. Reference potential evapotranspiration was calculatedusing the Penman–Monteith equation according to FAO guidelines[18]. For the period 1960–2000 radiation and wind speed wereprescribed by CRU CLIM 1.0 climatology data [19]. For compatibil-ity, data from ERA-Interim (precipitation, temperature, potentialevaporation) was bias-corrected by scaling the long-term monthlymeans of these fields to the CRU TS 2.1 dataset over the overlap-ping period (1979–2001).

The model concept of PCR-GLOBWB and used allocationscheme. Middle part: the soil compartment, divided into two soillayers (S1 and S2) and one linear groundwater reservoir (S3). Pre-cipitation (Prec) falls as rain or snow (temperature dependent) andcan be stored in canopy or as snow accumulation (Ss). Verticaltransport within the soil column appears from percolation or cap-illarity rise (P). The total local gains from all cells, i.e. drainage(QDr), subsurface flow (QSf), and baseflow (Qbf), are routed alongthe drainage direction to yield the channel discharge (Qchannel).In every grid cell water can be abstracted from surface water orgroundwater. Return flows go to surface water or groundwater,dependent on the water use.

2.2. Water demand

Throughout the paper total water demand is used to denote thewater requirements for three sectors: irrigation, industry, andhouseholds. It denotes potential water withdrawal, i.e. the waterthat would be abstracted if sufficient water were available (grosswater demand).

Data on sectoral water demand for the model period wereadopted from the previous study of Wada et al. [8]. To overcomethe lack of available spatially explicit data, sectoral water demandswere estimated using country statistics on the extent of irrigatedareas and population numbers downscaled to 0.5� resolution. Toapproximate economic development over the period 1960–2010data of Gross Domestic Product (GDP), electricity produced, andhousehold consumption were used.

Industrial demand is kept constant over the year. Domestic de-mand reflects seasonal variability according to air temperaturefluctuations. Water recycling ratios for industry and domestic useare adopted from [8] and were calculated per country, on the basisof by GDP and the level of economic development, i.e. high income(80%), middle income (65%), and low income (40%) economies.

Fig. 1. The model concept of PCR-GLOBWB and used allocation scheme. Middle part: the soil compartment, divided into two soil layers (S1 and S2) and one lineargroundwater reservoir (S3). Precipitation (Prec) falls as rain or snow (temperature dependent) and can be stored in canopy or as snow accumulation (Ss). Vertical transportwithin the soil column appears from percolation or capillarity rise (P). The total local gains from all cells, i.e. drainage (QDr), subsurface flow (QSf), and baseflow (Qbf), arerouted along the drainage direction to yield the channel discharge (Qchannel). In every grid cell water can be abstracted from surface water or groundwater. Return flows goto surface water or groundwater, dependent on the water use.

I.E.M. de Graaf et al. / Advances in Water Resources 64 (2014) 21–33 23

Irrigation water demand is calculated assuming optimal cropgrowth (i.e. maximum crop transpiration), and taking into accountbare soil evaporation and soil water availability. The former issimulated in PCR-GLOBWB as actual transpiration from the unsat-urated zone (S1 and S2 in Fig. 1). Losses during abstraction andtransport are calculated by multiplying the required irrigationwater demand with a country specific efficiency factor taken fromRohwer et al. [20]. Irrigation water demand is increased with10–40% upon implementing this efficiency factor. Evaporationlosses during application are not accounted for in the estimationof irrigation water demand. Domestic irrigation (e.g., gardens) isnot considered, as it is small compared to crop irrigation waterdemand.

2.3. Water abstractions and feedbacks

Total water demand can be met from three resources: (1) sur-face water, (2) groundwater, and (3) desalinated water. Althoughglobal desalinated water use is very small (0.2% of global waterabstractions [21]) it can be an important resource of water locally(e.g. Middle East). Country scale data of desalinated water use aretaken from FAO AQUASTAT [21], and downscaled onto a coastalribbon (around 40 km) based on gridded population intensities[22]. Desalinated water use is assumed to be constant over the yearand subtracted from the total water demand, as it is assumed to beconsumed entirely.

The remaining water demand is dynamically attributed to sur-face water and groundwater resources, depending on simulatedwater availability in the resources. Return flows (i.e. unconsumedabstracted water) and riverbed infiltration affect water availabilityand were simulated at the same time as the abstractions, and in-cluded at runtime in the hydrological model. The outer arrows of

Fig. 1 show the developed allocation scheme. In every grid cell,total abstractions (solid) are variably and sector-independentlytaken from surface waters and groundwater. With ‘sector-independent’ is meant that irrigation, industry, and domesticdemands are satisfied from groundwater and/or surface waterresources, and that possible local preferences are not accountedfor. Return flows (dashed) contribute to a single source, surfacewater or groundwater, and are sector dependent. Return flows thusredistribute abstracted water over the water resources. This redis-tribution causes a shift from the fast surface water to the slowergroundwater, or the other way around, and affects (long-term)local and downstream water availability and residence times.

As a first estimate of surface water and groundwater abstrac-tions the total water demand is allocated to surface water andgroundwater resources, using the ratio of two-year running-aver-ages of local baseflow (QBf) over upstream accumulated discharge(Qchannel). This ratio changes over time and reflects the dynamicsof water availability in the surface water and groundwater re-sources. Also, these abstraction estimates are not dependent ondata. It is estimated that close to the main river branches the ratioof QBf over Qchannel is small and potentially more surface water isabstracted than groundwater. Farther from the main riverbranches, the ratio is larger and potentially more groundwater isabstracted than surface water. The two-year running-average isused to smoothen out one-year fluctuations, e.g. a dry year be-tween two wet years, and to highlights longer-term trends. It isused based on the assumption that more structural water manage-ment actions (e.g. infrastructural actions) are taken when watershortage last at least two years. Total water demand is dividedby the grid-cell, WDtot [L], and potential surface water and ground-water abstractions (GWApot and SWApot [L]) are calculated as:

GWApot ¼ ðQBf =Q channelÞ �WDtot ð1Þ

24 I.E.M. de Graaf et al. / Advances in Water Resources 64 (2014) 21–33

SWApot ¼ ð1� Q Bf =Q channelÞ �WDtot ð2Þ

Additionally, SWA is limited by an environmental flow condi-tion, in this study defined as the 90th percentile of the 7-day run-ning averaged low flow of the no-abstraction scenario (Q90NA)over the period 1960–2010. Surface water abstraction (SWA) issimulated by evaluating potential surface water abstractionagainst simulated daily surface water availability (routed alongdrainage network).

Simulated surface water availability is negatively affected byabstractions and losses through riverbed infiltration (Inf), but onthe other hand positively affected by groundwater discharge andreturn flows from surface water and groundwater abstractions

Fig. 2. Continental scale validation of simulated (gross) groundwater abstraction rates, ofractions (–). Root Mean Square Error, coefficient of determination (R2) and regression c

for industry and domestic use (SWretID and GWretID respectively).Groundwater abstraction possibly results in riverbed infiltration.To avoid simulating extremely large riverbed infiltration overhydrologically inactive groundwater reservoirs, we limit riverbedinfiltration to areas where under the no abstractions scenariobaseflow is present. For these areas riverbed infiltration occurswhenever groundwater abstractions exceed baseflow. Riverbedinfiltration is parameterized with saturated conductivity of theunderlying sediments or bedrock (based on [23]). Most likely thisleads to an overestimation because the resistance of bottom sedi-ments is not accounted for; however infiltration is small comparedto aquifer size.

f AB, ABRE, and W2012, to reported AQUASTAT data (106 m3 y�1) and groundwateroefficient (a) are given.

I.E.M. de Graaf et al. / Advances in Water Resources 64 (2014) 21–33 25

The actual change in surface water storage (Qloc) for the presentstep in the sub-loop due to local fluxes is calculated as:

Q loc ¼ Q in � SWA� Inf þ SWretID þ GWretID ð3Þ

where Qin is inflow from baseflow and specific runoff (all in L3 T�1).River discharge (Qchannel L3 T�1) is calculated from Qloc using thekinematic wave approximation.

If the available surface water is not sufficient to satisfy the po-tential surface water abstraction, the deficit is abstracted from thegroundwater. In this study groundwater can be abstracted fromtwo types of resources: (1) renewable groundwater and (2) nonre-newable or nonlocal water. Renewable groundwater is waterstored in the groundwater reservoir (S3 in Fig. 1) and is naturallyor artificially recharged. In this study nonrenewable groundwateror nonlocal water is assumed to be an additional unlimitedsource of water, abstracted when insufficient renewable water isavailable.

Renewable groundwater abstraction (GWArenw) is simulated byevaluating total GWApot (sum of GWApot and surface water deficit)against available groundwater (stored in S3) and recharge (bothnatural (R) and artificial recharge, the latter stemming from irriga-tion return flow in case of the ABRE scenario). Irrigation returnflows from surface water or groundwater abstractions (SWretIRR

and GWretIRR respectively) are simplified for the modeling purpose,and directly recharge to the groundwater. Evaporation losses dur-ing application are therefore not accounted for in the simulation ofirrigation return flows. This results in an overestimation of irriga-tion return flow, however it is expected to be small compared tothe abstracted irrigation water. Groundwater storage [L] is calcu-lated as:

S3t ¼ S3t�Dt þ ðR3� GWAtot þ Inf þ GWretIRR þ SWretIRRÞDt ð4Þ

When S3t is positive, all groundwater abstractions are renewable.When S3 becomes negative, part of the groundwater demand is ab-stracted from nonrenewable or nonlocal water resources. This is as-sumed to be a virtual unlimited groundwater source calculated as

Fig. 3. Groundwater abstractions rates in the conterminous USA (106 m3 y�1) for thecompared with simulated abstractions, without return flows (AB), simulated abstractionsW2012. Statistics of the validation are given: coefficient of determination (R2), and regr

negative groundwater storage, as little to no data is available onthe actual availability of this resource. When nonrenewablegroundwater or nonlocal water is abstracted, natural and artificialrecharge is still available for GWArenw. After a period of nonrenew-able groundwater or nonlocal water abstractions, negative ground-water storage recovers when total recharge becomes larger thanGWAtot. Renewable groundwater abstractions [L] are then calcu-lated as:

If S3t > 0 : GWArenw ¼ minðGWAtot; S3t=DtÞ ð5Þ

If S3t 6 0 : GWArenw ¼ minðGWAtot ;R3þ GWretIRR þ SWretIRRÞ ð6Þ

Nonrenewable groundwater or nonlocal water (NRGNLW) abstrac-tion is then calculated as [L]:

NRGNLW ¼ GWAtot � GWArenw ð7Þ

2.4. Runs and analysis

Three model runs were formulated: no abstractions (NA),abstractions (AB), and abstractions and return flows (ABRE). Inthe two latter runs, total water demand did not change. Contraryto AB, in the ABRE case additional water becomes available for re-use by return flows, where in the AB case return flows are treatedas losses to the system. However, for industrial and domestic waternet surface water abstraction rates are applied for both cases, assurface water to surface water return flows are implicitly included.Net irrigation water abstractions are for ABRE smaller than for AB.

To gain confidence in the proposed water allocation scheme,first simulated groundwater abstraction magnitudes and ground-water fractions of the total abstracted water are validated againstreported data per country from the FAO AQUASTAT database[21]. The FAO AQUASTAT database reports water statistics at coun-try level with emphasis on irrigation and agricultural water use.Not for all countries the data are coherent and reliable, due tothe fact that national statistics are not available and water

year 2000. Reported data by U.S. Geological Survey (www.usgs.gov) per county,, including return flows (ABRE), and simulated abstractions of the previous study ofession (a) (state/county scale).

26 I.E.M. de Graaf et al. / Advances in Water Resources 64 (2014) 21–33

demands had to be estimated [21]. Simulated groundwaterabstraction magnitudes are also validated for the USA against re-ported data from the USGS database (www.usgs.gov). Results arecompared in the same way to the previous study of Wada et al.[24] where reported groundwater abstraction data from IGRAC(www.un-igrac.org) were used to condition groundwater abstrac-tions. Second, similar to Van Beek et al. [10], minimal monthlyriver discharges over the model period reported for 2219 stationsfrom the GRDC long-term dataset were compared to simulatedminimal monthly discharges for the three model runs. This showshow accurate minimal monthly flows are simulated under the con-dition of abstractions and return flows.

The effects of abstractions and return flows on water availabil-ity are analyzed by looking at river discharges globally. We focusespecially on the effects on low flows, as the contribution of base-flow is largest then. In addition, we studied in more detail four se-lected major rivers, that differ in term of climatic region, dominanttype of abstraction (groundwater vs surface water), and dominantwater use sector; (1) Rhine and (2) Danube, where industrial anddomestic water use are dominant, (3) Mississippi, with irrigation,industrial and domestic water use, and (4) Indus, with predomi-nantly irrigation. For these rivers flow duration curves (FDCs) of7-day averaged discharges and hydrographs over the period1994–2001 are compared for the three model runs. For an exten-sive validation of PCR-GLOBWB for these and many other majorrivers we refer to Van Beek et al. [10].

3. Results

3.1. Model validation

Fig. 2 shows scatter plots of the comparison of reported FAOAQUASTAT data against simulated groundwater abstraction mag-nitudes and fractions for the AB and ABRE runs per country forthe year 2000 (upper two scatters). Simulated abstractions on0.5� resolution were aggregated to country scale. In the lower scat-ters simulated results are compared with the previous study ofWada et al. [24] (from now on W2012) also per country for theyear 2000.

Fig. 4. Observed versus simulated minimal river discharges of ABRE run, for 2219 statiorepresents 1:1 line, the dashed lines indicate bound values differing by less than order

Results show larger abstractions are better estimated thansmaller abstractions (e.g. USA, Mexico, Spain, France), as seen fromthe smaller scatter for larger abstractions (closer to the 1:1 line).Estimates of ABRE are slightly, but not significantly, better thanfor AB. The coefficient-of-determination (R2) for AB and ABRE isrespectively 0.96 and 0.98, however regression coefficient (a) is0.6 and 0.62. The simulated values of W2012 are similar to theFAO AQUASTAT data (close to 1:1 line), R2 is 0.97 and a is 1.

The difference in model accuracy between this study andW2012 is explained by difference in model approach. In this studyadditional model uncertainties arise from the fact that demand isallocated purely based simulated baseflow and channel flow. InW2012 groundwater abstractions are constrained to reported datafrom IGRAC, which are comparable to the FAO AQUASTAT data.

The Root Mean Square Errors calculated for AB, ABRE, andW2012 are respectively 8.12, 7.57, and 19.28 (all � 106 m3 y�1).This shows estimates of ABRE are better than AB and both are bet-ter than W2012. From Fig. 2 it follows that in this study, comparedto W2012, better estimates are mainly caused by countries withlow abstraction rates that are simulated in this study but not in-cluded in the IGRAC database. Most of these low abstraction mag-nitudes are found for in countries with low economic development(e.g. Ghana, Congo). For these countries R2s of AB, ABRE, andW2012, are 0.8, 0.9, and 0.5 respectively with a being 0.6, 0.8,and 0.3. Here, model uncertainties of W2012 are large due to dif-ferences in reported data between the two datasets, while ABand ABRE results are not dependent on reported abstraction data.Still groundwater abstractions are underestimated with AB andABRE, possibly from lack of including user or sector specific prefer-ences, e.g. the use of clean groundwater above polluted surfacewater.

The plots of simulated and observed groundwater abstractionfractions show a large scatter, partly explained by the scatter ingroundwater abstractions magnitudes. Due to the nature of ratiocalculations a limited mismatch in magnitude results automati-cally in larger scatter for the groundwater fraction. Again the scat-ter is smaller for ABRE than for AB and best for W2012. Forcountries with relative large groundwater abstractions(�P100 km3 y�1), and dominantly industrial and domestic use,

ns of the long-term GRDC dataset and 21 selected large river basins. The solid lineof magnitude.

I.E.M. de Graaf et al. / Advances in Water Resources 64 (2014) 21–33 27

simulated fractions correspond well with reported fractions (e.g.USA, France). For some irrigation dominant countries, groundwaterfractions are highly overestimated (e.g. Spain). This may be ex-plained by the overestimation of irrigation return flow magnitudesthat directly recharge the groundwater without accounting for dis-tribution and application. This also explains the fact that this islargest for ABRE.

For the USA groundwater abstraction magnitudes were vali-dated at two spatial scales; at state level and at county level(Fig. 3). The simulated spatial pattern of estimated groundwaterabstraction correspond well with the reported map. Large ground-water abstraction rates are e.g. simulated for the Ogallala Aquifer,Central valley aquifer, Florida, and the Columbia Plateau basalticrock aquifer. However, R2 for both scales is low (60.5) and de-creases with higher resolution. For W2012 at state level R2 is 0.9,

Fig. 5. Relative change in magnitude of low flow (7-day-Q90) compared to the no-abreturnflows (ABRE).

however at county scale R2 is 60.5. At state level, a for this studyequals 0.7 and decreases with smaller scale, for W2012 a is 1.02and does not change. At the finer resolution model uncertaintiesincrease for both model approaches. First, because a significantpart of the counties is smaller than the used 0.5� grid cells. Forcounties larger or equal to the grid cell resolution all three modelapproaches showed better estimates compared with the reporteddata. For ABRE and AB R2 is still low, 0.2 and 0.1 respectively, how-ever a increased to 0.6 and 0.62 respectively meaning that theunderestimation of groundwater abstraction is smaller for largercounties. For W2012 for the larger counties R2 remains 0.4 and ais 0.94. Second, for this study, specific local preferences for surfacewater or groundwater abstractions are not averaged out at smallerscales and deviations from global-scale datasets of water demandare likely to be larger. From Fig. 3 it follows that in general

straction (NA) situation caused by (A) abstractions (AB) and (B) abstractions and

28 I.E.M. de Graaf et al. / Advances in Water Resources 64 (2014) 21–33

groundwater abstractions for large irrigation areas in the USA (e.g.High Plains, Mississippi embayment) are underestimated by ourmethod. This is explained by the fact that due to the low modelresolution (0.5�), simulated river discharges will be higher thansimulated river discharges on higher resolution [25]. Consequently,surface water is ubiquitously available for irrigation in most of thecells, and is abstracted instead of groundwater. Groundwaterabstractions are overestimated for the Great Lake regions (Wiscon-sin, Illinois, Michigan), where industries rely on surface waterrather than on groundwater, but this is not captured by the modeldue to the inability of the model scheme to abstract standing waterfrom lakes. Again these estimates would be better when sector-specific preferences were accounted for.

Fig. 4 shows the comparison between observed and simulatedminimum monthly discharge over the model period of 2219 GRDC

Fig. 6. Simulated additional return flows (106 m3 y�1), for the year 2000: (A) surface wgroundwater return flows from surface water irrigation.

stations for the ABRE run. Differences between the three runs aresmall, all coefficients-of-determination (R2) were high (for ABand ABRE 0.9, for NA 0.8), although the scatter is large. It is alreadyknown from van [10] that the model has the tendency to overesti-mate extreme discharges, especially minimum flows. Regressioncoefficients (a) are 1.07, 1.03, and 0.8 for AB, ABRE, and NA respec-tively. Also the abstraction scheme used in this study does not re-gard reducing abstraction rates when approaching environmentalflow conditions, or poor infrastructure when river discharges de-crease. From the figure it can also be concluded that the spreadaround the 1:1 line decreases when the basin area increases andmore runoff is accumulated. The 21 indicated stations still showthis overestimation (alpha for ABRE 1.25), but the data points cen-ter more around the 1:1 line, especially for the larger basins. The R2

for these 21 stations is 0.91 for ABRE (0.88 and 0.86 for AB and NA

ater return flows from industrial and domestic groundwater abstractions and (B)

I.E.M. de Graaf et al. / Advances in Water Resources 64 (2014) 21–33 29

respectively with alphas 1.26 and 1.23). For rivers in intensivelyirrigated basins (e.g. Nile, Huang He) minimum river dischargeswere significantly better estimated compared with the NA run.This is caused by surface water and groundwater abstractionsand consequently lower baseflows to the river channel. The differ-ence for basins in dominated industrial rivers are also significant,e.g for Rhine and Danube, where large magnitudes of surface waterand groundwater are abstracted.

3.2. Effects of water abstractions on river discharges

Fig. 5 shows the relative decrease in magnitude of averaged7-day low flows after abstractions and after including return flowscompared to the scenario without water use. Water use upstreamaffects water availability locally and downstream, as local runoff isrouted along the drainage network. Surface water abstractions and

Fig. 7. Simulated average yearly non-renewable, or non-local water abstractions (106 m2010.

return flows to the surface water affect river discharges directly.Groundwater abstractions and return flows to the groundwaterinfluence groundwater storage and river discharges throughchanges in baseflows. The impact is largest during low flows whenthe baseflow contribution is dominant.

Many rivers across the world show a decrease in low flow mag-nitude due to abstractions (Fig. 5A), while the effect of return flowsdiffers, depending on the dominant sector (Fig. 5B). For areaswhere industrial use is dominant (e.g. USA, Europe) the impact ofreturn flows is largest, as the largest part of the abstracted waterflows back to surface waters raising low flows. This results in high-er low flows and thus higher surface water availability. For exam-ple, the decrease of low flow after abstractions is between 25–50%for the Danube, upstream Mississippi, and downstream Rhine(Fig. 5A). Including return flows largely minimizes the effects ofabstractions, e.g. for the Danube the decrease is 5–12.5%,

3 y�1), accounting for return flows, over the periods (A) 1960–1970 and (B) 2000–

30 I.E.M. de Graaf et al. / Advances in Water Resources 64 (2014) 21–33

downstream Mississippi 0–5% (Fig. 5B. For e.g. the Rhine and Elbelow flows even increase (Fig. 5B) as a result of additional returnflows from groundwater abstractions by industry (Fig. 6A). The ef-fect is that groundwater discharge, that would otherwise occur la-ter, is bypassed by abstraction and return flow to the surface water.As such, it increases low flows at the expense of discharge later onin time.

For irrigated basins, e.g. the Indus, Ganges, Yellow River, the ef-fects of abstractions are large. Low flows for the downstream Indusdecrease with 25–50%, for the Ganges more than 50%, and for theYellow River between 12.5–25% and 25–50% (Fig. 5A). The effectof return flows is not significant. For the Indus and Ganges the ef-fect is minimal, while for the Yellow River low flow decreases12.5–25%, both up- and downstream (Fig. 5B). However, for theseintensively irrigated basins, additional irrigation return flow tothe groundwater is large, as can be seen from Fig. 6B. For these ba-sins, the additional groundwater recharge from surface waterabstractions, does not lead to an increase of baseflow, but the avail-ability of renewable groundwater is increased and the allocation ofwater demand changes. As a result, less nonrenewable groundwa-

Fig. 8. seven day averaged flow duration curves of the four selected rivers for the differeflows (ABRE). (River name, and GRDC station number): Rhine 6335020, Mississippi 412

ter or nonlocal water will be abstracted for ABRE, compared to AB.Fig. 7 shows the spatial and temporal development of nonrenew-able groundwater or nonlocal water for ABRE. Hotspots are presentover well-known intensively irrigated areas (see e.g. [26]); India,Pakistan, China, Saudi-Arabia, USA, Mexico.

For the four selected rivers Flow Duration Curves (FDCs) andhydrographs are plotted for GRDC-stations closest to the river out-let (Figs. 8 and 9). All FDCs of the AB-run show a decrease in lowflows. However, the effects of including return flows are differentbetween rivers. For the Rhine and the Mississippi, additional returnflows from surface water irrigation (Fig. 6B) increase the contribu-tion of baseflow during low flows (flatter FDC for ABRE than forAB). For the Danube the negative effects of abstractions are clearlyreduced by return flows, although baseflow contribution duringlow flows slightly decreases (steeper FDC). Also for the Indus, re-turn flows have a positive effect, although small.

The hydrographs indicate a similar response and additionallyshow changes in timing and regime. Under ABRE the Rhine showsa decrease in peak flows due to surface water abstractions and therelative small contribution of baseflow to these events. For periods

nt model runs, no abstractions (NA), abstractions (AB), and abstractions and return7800, Danube 6742900, Indus 2335950.

Fig. 9. Hydrographs of monthly discharge over the period 1994–2000 for the four selected rivers. For the Danube and Indus no GRDC observations over the given period wereavailable. (River name, GRDC station number).: Rhine 6335020, Mississippi 4127800, Danube 6742900, Indus 2335950.

I.E.M. de Graaf et al. / Advances in Water Resources 64 (2014) 21–33 31

directly following low flows (e.g. 1996, 1997), river discharge ofABRE is higher than NA. Here, the direct recharge of abstractedgroundwater for industry or domestic to the river leads to a directincrease of low flows. The same holds for the Mississippi. For theIndus, the effects of irrigation return flows are clear. Before a lowflow (e.g., 1996, 1997, 1998), the baseflow component is clearlyhigher for ABRE than for AB. Surface water used for irrigation addi-tionally recharges the groundwater after which it contributes viabaseflow to the river discharge. As a result, the timing of peak flowsand low flows changes, both for AB and ABRE (e.g. 1995, 1996 and1996, 1997). To a smaller extent this is also found for the Danube(e.g. 1994, 1999).

3.3. Global trends in abstraction rates

Fig. 10 presents global-scale trend of demand, surface water,and groundwater abstractions over the period 1960–2010 for theAB and ABRE runs (numbers are given in Table 1).

For both runs, the ratio of surface water abstraction to totalgroundwater abstractions is close to 50%. However, for ABRE moresurface water is abstracted, and the relative increase in surfacewater abstraction in slightly larger compared to AB, 48% and 42%respectively. The figure shows that around 40% of the total ab-stracted water can be reused via surface water return flows.

A clear difference between the two model scenarios is found forthe water allocation to renewable groundwater and nonrenewableor nonlocal water resources. For ABRE roughly 50% of the total ab-stracted groundwater comes from renewable groundwater re-sources, for AB this is 40%. As already shown in Fig. 6, hotspotsof NRGNLW are intensively irrigated areas. Especially in these re-gions, return flows limit the expansion of NRGNLW areas and lessgroundwater depletion takes place. Our numbers on abstractionscorrespond well with the previous study of W2012 (Table 1). Forthe year 2000 we estimate for ABRE a total groundwater abstrac-tion of 910 km3 y�1, of which 456 km3 y�1 is NRGNLW. Previousstudies found for Example 1070 km3 y�1 total groundwater

Fig. 10. Global trends, over the period 1960–2010 of simulated demand, abstractions, and return flows, for (a) abstractions, excluding return flows (AB) and (b) abstractions,including return flows (ABRE).

Table 1Simulated global abstractions (km3 y�1) of nonrenewable groundwater or nonlocalwater (NRGNLW), renewable groundwater (renew.GW), and surface water (SW) fortwo model runs (AB and ABRE) compared with results of Wada et al. [24].

AB ABRE W2012

1960 2000 2010 1960 2000 2010 1960 2000

NRGNLW 285 546 561 193 446 432 171 494renw.GW 178 344 422 271 463 517 271 582SW 446 803 1000 450 896 1116 456 843

32 I.E.M. de Graaf et al. / Advances in Water Resources 64 (2014) 21–33

abstractions of which 494 km3 y�1 is NRGNLW [24], and 400–800 km3 y�1 [1] and 450 km3 y�1 [27] for NRGNLW.

The basic assumption made in this study is that the NRGNLWreservoir is unlimited, and all water demand is satisfied. As a firstattempt to limit this resource, an maximum exploitation depth of100 m was set based on the assumption that below 100 m ground-water becomes unattainable [28]. For the year 2000, 141 km3y�1

becomes unattainable and has to be satisfied by nonlocal resourcesor deeper aquifers. The regions where this occurs are the inten-sively irrigated areas. In order to simulate deep groundwaterabstractions and groundwater dynamics properly, the existingmodel should be extended with a groundwater flow model.

4. Conclusions and discussion

As human water demand is increasing worldwide, it is impor-tant to simulate how water demand is potentially allocated to sur-face water and groundwater resources and to account for thedynamic effects of abstractions and return flows on water avail-ability. This study is the first to develop a dynamic allocationscheme to simulate surface water and groundwater abstractionsand corresponding feedbacks, including return flows and a two-way groundwater-surface water interaction. It allows us to analyzethe temporal and spatial trends of water availability affected byabstractions and return flows on the global scale.

In the dynamic scheme local water availability is affected by lo-cal and upstream surface water and groundwater abstractions andreturn flows. Return flows cause a redistribution of unconsumedwater over the water resources and change residence times ofthe water (channel discharge, baseflow). The advantage, comparedto existing schemes, is that this scheme does not rely on

preconceived assumptions about the preferences of surface waterover groundwater abstractions or vice versa, or on fixed ratios be-tween the two. As a result, the distribution rules of our scheme aremore robust under changes in water availability caused by waterconsumption or climate. However, the spread in simulated waterabstractions increases compared to previous schemes because ofthe dynamic attribution instead of the use of data or fixed fractionfor surface water or groundwater abstractions. Especially for inten-sively irrigated regions groundwater abstractions are overesti-mated, caused by e.g. the overestimation of groundwateravailability due to return flows that directly recharge to thegroundwater; i.e. irrigation return flows may also end up in surfacewater or evaporate.

Changes in river discharges and river regimes show the effectsof abstractions and return flows. Largest impacts are found duringlow flows, when the contribution of groundwater through base-flow is largest.

For industrial regions, the negative effects of abstractions interms of magnitude largely diminish when return flows are in-cluded. Up to 80% of the abstracted water flows directly back tothe channel as return flow. For irrigated regions, return flows aresmaller (up to 60%) and the impact of abstractions on river dis-charges is larger. However, return flows do change timing andduration of low flows for these regions. By the redistribution ofwater, baseflow is maintained longer and the timing and durationof low flows changes. Also, for these regions, due to the increase inlocal attainable groundwater by return flows, less nonrenewablegroundwater, or nonlocal water is abstracted. The largest impactof return flows on the temporal and spatial expansion of nonre-newable groundwater or nonlocal water abstraction are thereforefound in irrigation hotspots like India, Pakistan, and China.

Obviously large model uncertainties remain. Not only in theproposed abstraction and return flow scheme, but also the forcingas well as abstraction data. In previous studies, using the samemodel, uncertainties in recharge and abstractions are analyzed[26], as well as uncertainties in meteorological forcing [9]. And in-deed the uncertainties are considerable, although not so large thattrends and variability are obscured. Uncertainties arising from theabstraction scheme can only be reduced by testing and improvingthe underlying assumptions with empirical evidence from localcase studies. Improvements can be expected from making the ratioof groundwater to surface water abstractions not only dependenton the availability of these resources, but also sector specific anddependent on local preferences. Likewise, increasing model

I.E.M. de Graaf et al. / Advances in Water Resources 64 (2014) 21–33 33

resolution would lead the better estimates of abstracted water byremoving scale effects such that more local preferences of ground-water to surface water abstraction and availability can be takeninto account. However, this requires information about accessibil-ity of water, e.g. available and required infrastructure, which is notan issue on 0.5� resolution. Also distance to surface water and res-ervoirs, between-basin water transfer as well as specific yield andgroundwater depth are factors to consider.

Acknowledgments

We thank three reviewers, Huan Wu Ph.D. of the University ofMaryland and two anonymous reviewers, for their constructiveand thoughtful suggestions, which substantially helped to improvethe quality of the manuscript. This study was funded by the Neth-erlands Organization for Scientific Research (NWO) in the projectPlanetary Boundaries Fresh Water Cycle.

References

[1] Vörösmarty CJ, Green P, Salisbury J, Lammers RB. Global water resources:vulnerability from climate change and population growth. Science2000;289(5477):284–8. http://dx.doi.org/10.1126/science.289.5477.284.

[2] Postel SL. Entering an era of water scarcity: the challenges ahead. Ecol. Appl.2000;10(4):941–8. http://dx.doi.org/10.1890/1051-0761(2000)010%5B0941:EAEOWS%5D2.0.CO;2.

[3] Gleeson T, VanderSteen J, Sophocleous MA, Taniguchi M, Alley WM, Allen DM,et al. Groundwater sustainability strategies. Nat. Geosci. 2010;3(6):378–9.http://dx.doi.org/10.1038/ngeo881.

[4] Hanasaki N, Inuzuka T, Kanae S, Oki T. An estimation of global virtual waterflow and sources of water withdrawal for major crops and livestock productsusing a global hydrological model. J. Hydrol. 2010;384(3–4):232–44. http://dx.doi.org/10.1016/j.jhydrol.2009.09.028.

[5] Wisser D, Frolking S, Douglas EM, Fekete BM, Vrsmarty CJ, Schumann AH.Global irrigation water demand: variability and uncertainties arising fromagricultural and climate data sets. Geophys. Res. Lett. 2008;35(24). http://dx.doi.org/10.1029/2008GL035296.

[6] Rost S, Gerten D, Bondeau A, Lucht W, Rohwer J, Schaphoff S. Agricultural greenand blue water consumption and its influence on the global water system.Water Resour. Res. 2008;44(9). http://dx.doi.org/10.1029/2007/WR006331.

[7] Döll P, Hoffmann-Dobrev H, Portmann FT, Siebert S, Eicker A, Rodell M, et al.Impact of water withdrawals from groundwater and surface water oncontinental water storage variations. J. Geodyn. 2011;59–60:143–56. http://dx.doi.org/10.1016/j.jog.2011.05.001.

[8] Wada Y, Van Beek LPH, Bierkens MFP. Modelling global water stress of therecent past: on the relative importance of trends in water demand and climatevariability. Hydrol. Earth Syst. Sci. 2011;15(12):3785–808. http://dx.doi.org/10.5194/hess-15-3785-2011.

[9] Wada Y, Wisser D, Bierkens MFP. Global modeling of withdrawal, allocationand consumptive use of surface water and groundwater resources. Earth Syst.Dyn. Discuss 2013;4:355–92. http://dx.doi.org/10.5194/ESDD-4-355-2013.

[10] Van Beek LPH, Wada Y, Bierkens MFP. Global monthly water stress. 1: Waterbalance and water availability. Water Resour. Res. 2011;47(7). http://dx.doi.org/10.1029/2010WR009791.

[11] Kraijenhoff van de Leur D. A study of non-steady groundwater flow withspecial reference to the reservoir-coefficient. De Ingenieur 1958;19:87–94.http://dx.doi.org/10.1029/JZ067i011p04347.

[12] Döll P, Lehner B. Validation of a new global 30-min drainage direction map. J.Hydrol. 2002;258(1–4):214–31. http://dx.doi.org/10.1016/S0022-1694(01)00565-0.

[13] Lehner B, Döll P. Development and validation of a global database of lakes,reservoirs and wetlands. J. Hydrol. 2004;296(1–4):1–22. http://dx.doi.org/10.1016/j.jhydrol.2004.03.028.

[14] Mitchell TD, Jones PD. An improved method of constructing a database ofmonthly climate observations and associated high-resolution grids. Int. J.Climatol. 2005;25(6):693–712. http://dx.doi.org/10.1002/joc.1181.

[15] New M, Hulme M, Jones P. Representing twentieth-century space-time climatevariability. Part II: Development of 1901–96 monthly grids of terrestrialsurface climate. J. Clim. 2000;13(13):es2217–38. http://dx.doi.org/10.1175/1520-0442(2000)013%3C2217:RTCSTC%3E2.0.CO;2.

[16] Uppala SM, Kallberg PW, Simmons AJ, Andrae U, da Costa Bechtold V, FiorinoM, et al. The era-40 re-analysis. Q. J. R. Meteorol. Soc. 2005;131(612):2961–3012. http://dx.doi.org/10.1256/qj.04.176.

[17] Dee DP, Uppala SM, Simmons AJ, Berrisford P, Poli P, Kobayashi S, et al. Theera-interim reanalysis: configuration and performance of the data assimilationsystem. Q. J. R. Meteorol. Soc. 2011;137(656):553–97. http://dx.doi.org/10.1002/qj.828.

[18] Allen RG, Pereira LS, Raes D, Smith M. Crop evapotranspiration-guildelines forcomputing crop water requirements. FAO Irrigation and Drainage Paper 56,FAO, Rome; 1998.

[19] New M, Lister D, Hulme M, Makin I. A high-resolution data set of surfaceclimate over global land areas. Clim. Res. 2002;21(1):1–25. http://dx.doi.org/10.3354/cr021001.

[20] Rohwer J, Gerten D, Lucht W. Development of functional irrigation types forimproved global crop modelling. No. 104 in, PIK Report; 2007.

[21] FAO Aquastat on-line database. Food and Agricultural Organization, Rome,Italy, Available: <http://faostat.fao.org>; 2010.

[22] Wada Y, Van Beek LPH, Viviroli D, Dürr HH, Weingartner R, Bierkens MFP.Global monthly water stress. 2: Water demand and severity of water stress.Water Resour. Res. 2011;47(7). http://dx.doi.org/10.1029/2010WR009792.

[23] Dürr HH, Meybeck M, Drr SH. Lithologic composition of the earth’s continentalsurfaces derived from a new digital map emphasizing riverine materialtransfer. Global Biogeochem. Cycles 2005;19(4). http://dx.doi.org/10.1029/2005GB002515.

[24] Wada Y, Van Beek LPH, Bierkens MFP. Nonsustainable groundwater sustainingirrigation: a global assessment. Water Resour. Res. 2012;48(1). http://dx.doi.org/10.1029/2011WR010562.

[25] Haddeland I, Matheussen BV, Lettenmaier DP. Influence of spatial resolutionon simulated streamflow in a macroscale hydrologic model. Water Resour.Res. 2002;38(7):291–2910. http://dx.doi.org/10.1029/2001WR000854.

[26] Wada Y, Van Beek LPH, Van Kempen CM, Reckman JWTM, Vasak S, BierkensMFP. Global depletion of groundwater resources. Geophys. Res. Lett.2010;37(20). http://dx.doi.org/10.1029/2010GL044571.

[27] Pokhrel Y, Hanasaki N, Koirala S, Cho J, Yeh PJF, Kim H, et al. Incorporatinganthropogenic water regulation modules into a land surface model. J.Hydrometeorol. 2012;13(1):255–69. http://dx.doi.org/10.1175/JHM-D-11-013.1.

[28] Gleeson T, Wada Y, Bierkens M, Van Beek L. Water balance of global aquifersrevealed by groundwater footprint. Nature 2012;488(7410):197–200. http://dx.doi.org/10.1038/nature11295.