Questions related to historical and future water resources and scarcity have been addressed by several macroscale hydrological models. One of these models is the Variable Infiltration Capacity (VIC) model. However, further model developments were needed to holistically assess anthropogenic impacts on global water resources using VIC.
Our study developed VIC-WUR, which extends the VIC model using (1) integrated routing, (2) surface and groundwater use for various sectors (irrigation, domestic, industrial, energy, and livestock), (3) environmental flow requirements for both surface and groundwater systems, and (4) dam operation. Global gridded datasets on sectoral demands were developed separately and used as an input for the VIC-WUR model.
Simulated national water withdrawals were in line with reported
Food and Agriculture Organization (FAO) national annual withdrawals (adjusted
VIC-WUR has potential for studying the impacts of climate change and anthropogenic developments on current and future water resources and sector-specific water scarcity. The additions presented here make the VIC model more suited for fully integrated worldwide water resource assessments.
Questions related to historical and future water resources and scarcity have been addressed by several macroscale hydrological models over the last few decades (Liang et al., 1994; Alcamo et al., 1997; Hagemann and Gates, 2001; Takata et al., 2003; Krinner et al., 2005; Bondeau et al., 2007; Hanasaki et al., 2008b; van Beek and Bierkens, 2009; Best et al., 2011). Early efforts focused on the simulation of natural water resources and the impacts of land cover and climate change on water availability (Oki et al., 1995; Nijssen et al., 2001a, b). Recently, a larger focus has been on incorporating anthropogenic impacts, such as water withdrawals and dam operations, into water resource assessments (Alcamo et al., 2003; Haddeland et al., 2006b; Biemans et al., 2011; Wada et al., 2011b; Hanasaki et al., 2018).
Global water withdrawals have increased 8-fold over the last century and are projected to increase further (Shiklomanov, 2000; Wada et al., 2011a). Although water withdrawals are only a small fraction of the total global runoff (Oki and Kanae, 2006), water scarcity can be severe due to the variability of water in both time and space (Postel et al., 1996). Severe water scarcity is already experienced by two-thirds of the global population for at least part of the year (Mekonnen and Hoekstra, 2016). To stabilize water availability for different sectors (e.g. irrigation, hydropower, and domestic uses), dams and reservoirs have been built, which are able to strongly affect global river streamflow (Nilsson et al., 2005; Grill et al., 2019). In addition, groundwater resources are being extensively exploited to meet increasing water demands (Rodell et al., 2009; Famiglietti, 2014).
One of the widely used macroscale hydrological models is the Variable Infiltration Capacity (VIC) model. This model was originally developed as a land surface model (Liang et al., 1994), but it has mostly been used as a stand-alone hydrological model (Abdulla et al., 1996; Nijssen et al., 1997) using an offline routing module (Lohmann et al., 1996, 1998a, b). Where land surface models focus on the vertical exchange of water and energy between the land surface and the atmosphere, hydrological models focus on the lateral movement and availability of water. By combining these two approaches, VIC simulations are strongly process based; this provides a good basis for climate-impact modelling.
VIC has been used extensively in studies such as coupled regional climate model simulations (Zhu et al., 2009; Hamman et al., 2016), combined river streamflow and water temperature simulations (van Vliet et al., 2016), hydrological sensitivity to climate change research (Hamlet and Lettenmaier, 1999; Nijssen et al., 2001a; Chegwidden et al., 2019), global streamflow simulations (Nijssen et al., 2001b), flow regulation and redistribution research (Voisin et al., 2018; Zhou et al., 2018), and real-time drought forecasting (Wood and Lettenmaier, 2006; Mo, 2008). Several studies have also used VIC to simulate the anthropogenic impacts of irrigation and dam operation on water resources (Haddeland et al., 2006a, b; Zhou et al., 2015, 2016) based on the model set-up of Haddeland et al. (2006b). However, further developments were needed to holistically assess anthropogenic impacts on global water resources using this model (Nazemi and Wheater, 2015a, b; Döll et al., 2016; Pokhrel et al., 2016).
Firstly, the VIC model did not include groundwater withdrawals or water withdrawals from domestic, manufacturing, and energy (thermoelectric) sources. Although these sectors use less water than irrigation (Shiklomanov, 2000; Grobicki et al., 2005; Hejazi et al., 2014), they are important actors locally (Gleick et al., 2013), especially for the water–food–energy nexus (Bazilian et al., 2011). Sufficient water supply and availability are essential for meeting a range of local and global sustainable development goals related to water, food, energy, and ecosystems (Bijl et al., 2018). Secondly, environmental flow requirements (EFRs) have often been neglected (Pastor et al., 2014), even though they are “necessary to sustain aquatic ecosystems which, in turn, support human cultures, economies, sustainable livelihoods, and well-being” (Arthington et al., 2018). Anthropogenic alterations already strongly affect freshwater ecosystems (Carpenter et al., 2011), with more than a quarter of all global rivers experiencing very high biodiversity threats (Vorosmarty et al., 2010). By neglecting EFRs, the sustainable water availability for anthropogenic uses is overestimated (Gerten et al., 2013). Lastly, while the model set-up of Haddeland et al. (2006b) already included important anthropogenic impact modules (i.e. irrigation and dam operation), these were not fully integrated. Therefore, multiple successive model runs were required (see Sect. 2.1), which was computationally expensive – especially for global water resources assessments.
Recently, version 5 of the VIC model (VIC-5) was released (Hamman et al., 2018), which focused on improving the VIC model infrastructure. These improvements provide the opportunity to fully integrate anthropogenic impacts into the VIC model framework while also reducing computation times. Here, the newly developed VIC-WUR model is presented (named after the development team at Wageningen University and Research). The VIC-WUR model extends the existing VIC-5 model using several modules that simulate the anthropogenic impacts on water resources. These modules will implement previous major works on anthropogenic impact modelling and will also integrate environmental flow requirements into VIC-5. The modules include the following: (1) integrated routing, (2) surface and groundwater use for various sectors (irrigation, domestic, industrial, energy, and livestock), (3) environmental flow requirements for both surface and groundwater systems, and (4) dam operation.
The next section first describes the original VIC-5 hydrological model (Sect. 2.1), which calculates natural water resource availability. Subsequently, the integration of the anthropogenic impact modules, which modify the water resource availability, is described (Sect. 2.2). Global anthropogenic water uses for each sector are also estimated (Sect. 2.3). To assess the capability of the newly developed modules, the VIC-WUR results were compared with Food and Agriculture Organization (FAO) national water withdrawals by sector and by source (FAO, 2016); with Huang et al. (2018), Steinfeld et al. (2006), and Shiklomanov (2000) data on water withdrawals by sector; with Gravity Recovery and Climate Experiment (GRACE) terrestrial water storage anomalies (NASA, 2002); with Global Runoff Data Centre (GRDC) streamflow time series (GRDC, 2003); and with Yassin et al. (2019) and Hanasaki et al. (2006) data on reservoir operation (Sect. 3.2). VIC-WUR simulation results are also compared with various other state-of-the-art global hydrological models. Lastly, the impacts of adhering to surface and groundwater environmental flow requirements on water availability are assessed (Sect. 3.3). This assessment is included to indicate the effects of the newly integrated surface and groundwater environmental flow requirements on worldwide water availability.
Schematic overview of the VIC-WUR model that includes the VIC-5 model
The basis of the VIC-WUR model is the Variable Infiltration Capacity model version 5 (VIC-5; Liang et al., 1994; Hamman et al., 2018). VIC-5 is an open-source macroscale hydrological model that simulates the full water and energy balance on a (latitude–longitude) grid. Each grid cell accounts for sub-grid variability in land cover and topography as well as allowing for variable saturation across the grid cell. For each sub-grid, the water and energy balance is computed individually (i.e. sub-grids do not exchange water or energy with one another). The methods used to calculate the water and energy balance are summarized in Appendix A and are mainly based on the work of Liang et al. (1994). For a description of the global calibration and validation of the water balance, the reader is referred to Nijssen et al. (2001b).
VIC version 5 (Hamman et al., 2018) upgrades did not change the model representation of physical processes, but they improved the model infrastructure. Improvements include the use of NetCDF (network common data form) for input and output processing and the implementation of parallelization through message passing interface (MPI). These changes increase computational speed and make VIC-5 better suited for (computationally expensive) global simulations. The most significant modification that enables new model applications is that VIC-5 also changed the processing order of the model. In previous versions, all time steps were processed for a single grid cell before continuing to the next cell (time before space). In VIC-5, all grid cells are processed before continuing to the next time step (space before time). This development allows for interaction between grid cells every time step, which is important for full integration of the anthropogenic impact modules, especially water withdrawals and dam operation.
For example, surface and subsurface runoff routing to produce river streamflow was typically done as a post-processing operation (Lohmann et al., 1996; Hamman et al., 2017a), due to the time-before-space processing order of previous versions. In order for reservoirs to account for downstream water demand, an irrigation demand initialization was required. This initialization could either be an independent offline dataset (Voisin et al., 2013) or multiple successive model runs (Haddeland et al., 2006b). As VIC-5 uses the space-before-time processing order, irrigation water demands and runoff routing could be simulated each time step. The routing post-process was replaced by our newly developed routing module, which simulates routing sequentially (upstream to downstream) based on the Lohmann et al. (1996) equations.
VIC-WUR extends the existing VIC-5 through the addition of several newly implemented anthropogenic impact modules (Fig. 1). These modules include sector-specific water withdrawal and consumption, environmental flow requirements for both surface and groundwater systems, and dam operation for large and small (within-grid) dams.
In VIC-WUR, sectoral water demands need to be specified for each grid cell (Sect. 2.3). To meet water demands, water can be withdrawn from river streamflow, small (within-grid) reservoirs, and groundwater resources. Streamflow withdrawals are abstracted from the grid cell discharge (as generated by the routing module), and reservoir withdrawals are abstracted from small dam reservoirs (located in the cell). Groundwater withdrawals are abstracted from the third-layer soil moisture and an (unlimited) aquifer below the soil column. Aquifer abstractions represent renewable and non-renewable abstractions from deep groundwater resources. Subsurface runoff is used to fill the aquifer if there is a deficit.
The partitioning of water withdrawals between surface and ground water resources is data driven (similar to e.g. Döll et al., 2012; Voisin et al.,
2017; Hanasaki et al., 2018). Partitioning was based on the study by Döll et al. (2012), who estimated groundwater withdrawal fractions for each
sector in around 15 000 national and subnational administrative units. These groundwater fractions were mainly based on information from the
International Groundwater Resources Assessment Centre (IGRAC;
Water can also be withdrawn from the river streamflow of other “remote” cells in delta areas. As rivers cannot split in the routing module, the model is unable to simulate the redistribution of water resources in dendritic deltas. Therefore, streamflow at the river mouth is available for use in delta areas (partitioned based on demand) to simulate the actual water availability. Delta areas were delineated by the global delta map from Tessler et al. (2015).
In terms of water allocation, under conditions where water demands cannot be met, water withdrawals are allocated to the domestic, energy, manufacturing, livestock, and irrigation sectors in that order. Withdrawn water is partly consumed, meaning the water evaporates and does not return to the hydrological model. Consumption rates were set at 0.15 for the domestic sector and 0.10 for the industrial sector, based on data from Shiklomanov (2000). The water consumption in the energy sector was based on Goldstein and Smith (2002) and varies per thermoelectric plant based on the fuel type and cooling system. For the livestock sector, the assumption was made that all withdrawn water is consumed. Unconsumed water withdrawals for these sectors are returned as river streamflow. For the irrigation sector, consumption was determined by the calculated evapotranspiration. Unconsumed irrigation water remains in the soil column and eventually returns as subsurface runoff.
Water withdrawals can be constrained by environmental flow requirements (EFRs). These EFRs specify the timing and quantity of water needed to support terrestrial river ecosystems (Smakhtin et al., 2004; Pastor et al., 2019). Surface and groundwater withdrawals are constrained separately in VIC-WUR, based on the EFRs for streamflow and baseflow respectively. EFRs for streamflow specify the minimum river streamflow requirements, whereas EFRs for baseflow specify the minimum subsurface runoff requirements (from groundwater to surface water). As baseflow is a function groundwater availability, baseflow requirements are used to constrain groundwater (including aquifer) withdrawals.
Various EFR methods are available (Smakhtin et al., 2004; Richter et al., 2012; Pastor et al., 2014). Our study used the variable monthly flow (VMF) method (Pastor et al., 2014) to calculate the EFRs for streamflows. VMF calculates the required streamflow as a fraction of the natural flow during high (30 %), intermediate (45 %), and low (60 %) flow periods, as described in Appendix B. The VMF method performed favourably compared with other hydrological methods, in 11 case studies where EFRs were calculated locally (Pastor et al., 2014). The advantage of the VMF method is that the method accounts for the natural flow variability, which is essential to support freshwater ecosystems (Poff et al., 2010).
EFR methods for baseflow have been rather underdeveloped compared with EFR methods for streamflow. However, a presumptive standard of 90 % of the natural subsurface runoff through time was proposed by Gleeson and Richter (2018), as described in Appendix B. This standard should provide high levels of ecological protection, especially for groundwater-dependent ecosystems. Note that part of the EFRs for baseflow are already captured in the EFRs for streamflow, especially during low-flow periods that are usually dominated by baseflows. However, the EFRs for baseflow specifically limit local groundwater withdrawals, whereas EFRs for streamflow include the accumulated runoff from upstream areas. Moreover, the chemical composition of groundwater-derived flows is inherently different, making them a non-substitutable water flow for environmental purposes (Gleeson and Richter, 2018).
Due to the lack of globally available information on local dam operations, several generic dam operation schemes have been developed for macroscale hydrological models to reproduce the effect of dams on natural streamflow (Haddeland et al., 2006a; Hanasaki et al., 2006; Zhao et al., 2016; Rougé et al., 2019; Yassin et al., 2019). In VIC-WUR, a distinction is made between “small” dam reservoirs (with an upstream area smaller than the cell area) and “large” dam reservoirs, similar to Hanasaki et al. (2018), Wisser et al. (2010), and Döll et al. (2009). Small dam reservoirs act as buckets that fill using surface runoff from the grid cell they are located in, and reservoirs' storage can be used for water withdrawals in the same cell. Large dam reservoirs are located in the main river and use the operation scheme of Hanasaki et al. (2006), as described in Appendix C.
The scheme distinguishes between two dam types: (1) dams that do not account for water demands downstream (e.g. hydropower dams or flood protection dams) and (2) dams that do account for water demand downstream (e.g. irrigation dams). For dams that do not account for demands, dam release is aimed at reducing annual fluctuations in discharge. For dams that do account for demands, dam release is additionally adjusted to provide more water during periods of high demand. The operation scheme was validated by Hanasaki et al. (2006) for 28 reservoirs and has been used in various other studies (Hanasaki et al., 2008b, 2018; Döll et al., 2009; Pokhrel et al., 2012; Voisin et al., 2013). Here, the scheme was adjusted slightly to account for monthly varying EFRs and to reduce overflow releases, which is described in Appendix C.
The Global Reservoir and Dam (GRanD) database (Lehner et al., 2011) was used to specify location, capacity, function (purpose), and construction year of each dam. The capacity of multiple (small and large) dams located in the same cell were combined.
VIC-WUR water withdrawals are based on the irrigation, domestic, industry, energy, and livestock water demand in each grid cell. Water demands represent the potential water withdrawal, which is reduced when insufficient water is available. Irrigation demands were estimated based on the hydrological model, whereas water demands for other sectors were provided to the model as an input. Domestic and industrial water demands were estimated based on several socioeconomic predictors, whereas energy and livestock water demands were derived from power plant and livestock distribution data. Due to data limitations, the energy sector was incomplete, and energy water demands were partly included in the industrial water demands (which combined the remaining energy and manufacturing water demands). For more details concerning sectoral water demand calculations, the reader is referred to Appendix D.
Irrigation demands were set to increase soil moisture in the root zone so that water availability was not limiting crop evapotranspiration and growth. The exception is rice paddy irrigation (Brouwer et al., 1989), where irrigation was also supplied to keep the upper soil layer saturated. Water demands for rice paddy irrigation practices are relatively high compared with conventional irrigation practices due to increased evaporation and percolation. Therefore, the crop irrigation demands for these two irrigation practices were calculated and applied separately (i.e. in different sub-grids). Note that multiple cropping seasons are included based on the MIRCA2000 land use dataset (Portmann et al., 2010; see Sect. 3.1 in this paper for more details).
Total irrigation demands also included transportation and application losses. Note that transportation and application losses are not “lost” but rather returned to the soil column without being used by the crop. The water loss fraction was based on Frenken and Gillet (2012), who estimated the aggregated irrigation efficiency for 22 United Nations subregions. Irrigation efficiencies were estimated based on the difference between AQUASTAT-reported irrigation water withdrawals and calculated irrigation water requirements (Allen et al., 1998), using data on crop information (e.g. growing season and harvest area) from AQUASTAT.
Domestic and industrial water withdrawals were estimated based on the gross domestic product (GDP) per capita and the gross value added (GVA) by industries respectively (from Bolt et al., 2018; Feenstra et al., 2015; and World Bank, 2010; see Appendix D in this paper for more details). These drivers do not fully capture the multitude of socioeconomic factors that influence water demands (Babel et al., 2007). However, the wide availability of data allows for the extrapolation of water demands to data-scarce regions and future scenarios (using studies such as Chateau et al., 2014).
Domestic water demands per capita (used for drinking, sanitation, hygiene, and amenity uses) were estimated in a similar fashion to Alcamo et al. (2003). Demands increased non-linearly with the GDP per capita due to the acquisition of water-using appliances as households become richer. A minimum water supply is needed for survival, and the saturation of water-using appliances sets a maximum on domestic water demands. Industrial water demands (used for cooling, transportation, and manufacturing) were estimated in a similar fashion to Flörke et al. (2013) and Voß and Flörke (2010). Industrial demands increased linearly with the GVA (as an indicator of industrial production). As industrial water intensities (i.e. the water use per production unit) vary widely between different industries (Flörke and Alcamo, 2004; Vassolo and Döll, 2005; Voß and Flörke, 2010), the average water intensity was estimated for each country. Both domestic and industrial water demands were also influenced by technological developments that increase water use efficiency over time, as in Flörke et al. (2013).
Domestic water demands varied monthly based on air temperature variability as in Huang et al. (2018), based on Wada et al. (2011b). Using this approach, water demands were higher in summer than in winter, especially for counties with strong seasonal temperature differences. Domestic water demand per capita was downscaled using the HYDE3.2 gridded population maps (Goldewijk et al., 2017). Industrial water demands were kept constant throughout the year. Industrial demands were downscaled from national to grid cell values using the NASA Back Marble night-time light intensity map (Roman et al., 2018). National industrial water demands were allocated based on the relative light intensity per grid cell for each country.
Energy water demands (used for the cooling of thermoelectric plants) were estimated using data from van Vliet et al. (2016). Water use intensity for generation (i.e. the water use per generation unit) was estimated based on the fuel and cooling system type (Goldstein and Smith, 2002), which was combined with the generation capacity. Note that the data only covered a selection of the total number of thermoelectric power plants worldwide. Around 27 % of the total (non-renewable) global installed capacity between 1980 and 2011 was included in the dataset due to lack of information on cooling system types for the majority of thermoelectric plants. To avoid double counting, energy water demands were subtracted from the industrial water demands.
Livestock water demands (used for drinking and animal servicing) were estimated by combining the Gridded Livestock of the World (GLW3) map (Gilbert et al., 2018) with the livestock water requirement reported by Steinfeld et al. (2006). Eight varieties of livestock were considered: cattle, buffaloes, horses, sheep, goats, pigs, chicken, and ducks. Drinking water demands varied monthly based on temperature as described by Steinfeld et al. (2006) – drinking water requirements were higher during higher temperatures.
VIC-WUR results were generated between 1979 and 2016, excluding a spin-up period of 1 year (analysis period from 1980 to 2016). The model used a
daily time step (with a 6-hourly time step for snow processes), and simulations were executed on a 0.5
Nijssen et al. (2001c) used the Advanced Very High Resolution Radiometer vegetation type database (Hansen et al., 2000) to spatially distinguish 13 land cover types. The land cover type “cropland” in the original land cover dataset was replaced by cropland extents from the MIRCA2000 cropland dataset (Portmann et al., 2010). MIRCA2000 distinguishes the monthly growing area(s) and season(s) of 26 irrigated and rain-fed crop types around the year 2000. Crop types were aggregated into three land cover types: rain-fed, irrigated, and rice paddy cropland. The natural vegetation was proportionally rescaled to make up for discrepancies between the natural vegetation and cropland extents.
Cropland coverage (the cropland area actually growing crops) varied monthly based on the crop growing areas of MIRCA2000. The remainder was treated as bare soil. Cropland vegetation parameters (e.g. leaf area index – LAI, displacement, vegetation roughness, and albedo) vary monthly based on the crop growing seasons and the development-stage crop coefficients of the Food and Agricultural Organization (Allen et al., 1998).
The latest WATCH forcing data ERA-Interim (aggregated to a 6-hourly temporal resolution), developed by the EU Water and Global Change (WATCH; Harding et al., 2011) project, was used as climate forcing (WFDEI; Weedon et al., 2014). The dataset provides gridded historical climatic variables of minimum and maximum air temperature, precipitation (as the sum of snowfall and rainfall, GPCC bias corrected), relative humidity, pressure, and incoming shortwave and longwave radiation.
For naturalized simulations, only the routing module was used. For the anthropogenic impact simulations, the sectoral water withdrawals and dam operation modules were turned on in the model simulations. For the EFR-limited simulations, water withdrawals and dam operations were constrained as described.
In order to validate the VIC-WUR anthropogenic impact modules, water withdrawal, terrestrial total water storage anomalies, and streamflow and reservoir operation simulations were compared with observations. The validation specifically focused on the effects of the newly included anthropogenic impact modules, meaning that streamflow and total water storage anomaly results are shown for river basins that are strongly influenced by human activities. A general validation for streamflow and terrestrial total water storage anomalies (including basins with limited human activities) is shown in Appendix E.
Simulated global domestic, industrial, livestock, and irrigation mean water withdrawals were 310, 771, 36, and 2202
Comparison between simulated national annual water withdrawals and the FAO-reported values for the
Simulated domestic, industrial, and irrigation water withdrawals correlated well with reported national water withdrawals, with an adjusted
When domestic, industrial, and livestock water withdrawals were compared to other studies, results were mixed (Fig. 3a, b, c). Simulated domestic
withdrawals followed a similar trend in time. However, simulated domestic water withdrawal trends were somewhat underestimated overall, with a mean
bias of 54
Comparison between the simulated monthly regional water withdrawal and values from Huang et al. (2018), Shiklomanov (2000), and Steinfeld et al. (2006) for the
Withdrawal differences for the domestic and industrial sector are probably due to the limited data availability. Our approach to compute water demands was data-driven and sensitive to data gaps (as opposed to Huang et al., 2018, who also combined model results). For example, domestic withdrawal data for China were not available before 2007, and industrial withdrawal data were limited before 1990. Moreover, data on the disaggregation of industrial sectors (e.g. energy and mining), which can be important sectors in the water–food–energy nexus, were limited.
For livestock water withdrawals, there is a large discrepancy between Huang et al. (2018) and Steinfeld et al. (2006). Both studies used similar livestock maps, but there was large differences in livestock water intensity (in units of litres per animal per year). As our study used values from Steinfeld et al. (2006) to estimate livestock water intensity, our results were closer to their values (slightly higher due to the inclusion of buffaloes, horses, and ducks). Note that Huang et al. (2018) shows trends in livestock water withdrawals, whereas our study uses static livestock maps.
Simulated irrigation water withdrawals were within the range of other macroscale hydrological model estimates (Table 1). Simulated monthly variability in irrigation water withdrawals is reduced compared with the compiled results of Huang et al. (2018) (Fig. 3d), especially in Asia. Moreover, trends in time are less pronounced, as can be seen in Africa. These differences may indicate a relative low weather or climate sensitivity of evapotranspiration in VIC-WUR, as annual and interannual weather changes affect irrigation water demands to a lesser degree.
Average annual global irrigation water withdrawals as calculated by several global hydrological models.
Simulated global mean surface and groundwater withdrawals were 2327 and 992
Comparison between simulated national annual water withdrawals and FAO-reported values from
Simulated surface and groundwater withdrawals correlated well with the reported national water withdrawals, with an adjusted
Simulated monthly terrestrial water storage anomalies correlated well with the GRACE observations, with a mean annual and interannual root-mean-squared
error (RMSE) of 1.9 and 3.5
Comparison between simulated monthly terrestrial total water storage anomalies and GRACE-observed values. Panels indicate the time series and multiyear mean averages for naturalized simulations (dashed), anthropogenically impacted simulations (red), and observed (black) terrestrial total water storage anomalies.
Average annual global groundwater withdrawals and depletion as calculated by several global hydrological models.
As with the FAO comparison, these results seems to indicate that withdrawal partitioning towards groundwater is overestimated. However, conclusions regarding groundwater depletion are limited by the relatively simplistic approach to groundwater used in our study (as discussed by Konikow, 2011, and de Graaf et al., 2017). For example, processes such as wetland recharge and groundwater flows between cells are not simulated, even though these could decrease groundwater depletion.
Simulated discharge was compared to GRDC station data (GRDC, 2003) for various anthropogenically impacted rivers. Stations were selected if the upstream area was
larger than 20 000
Discharge improvement from naturalized to anthropogenically impacted simulations (as a fraction of the naturalized RMSE). Circled larger stations are shown in Fig. 7.
Comparison between the simulated discharge and GRDC-observed values. Panels indicate the time series and multiyear averages for naturalized simulations (dashed), anthropogenically impacted simulations (red), and observed (black) discharge.
The inclusion of the anthropogenic impact modules improved discharge performance, measured in RMSE, for 370 of the 462 stations (80 %; Figs. 6, 7). Improvements were mainly due to the effects of reservoir operation on discharges (e.g. Cachoeira Morena and Suntar stations) but also due to withdrawal reductions (e.g. Tore station). However, reservoir effects on discharge were sometimes underestimated (e.g. Timbues station). Decreased performance was mostly related to under- or overestimations of (calibrated) natural streamflow which was subsequently exacerbated by reservoir operation and water withdrawals. For example, the Clark Fork River naturalized streamflow was underestimated, which was subsequently further underestimated by the anthropogenic impact modules (Whitehorse Rapids station). Moreover, increases in discharge due to groundwater withdrawals could increase naturalized streamflow (e.g. Turkeghat station). Further improvements to discharge performance would most likely require a recalibration of the VIC model parameters.
For individual reservoirs, operational characteristics were generally well simulated (Fig. 8), with reductions in annual discharge variations (e.g. Flaming Gorge and Garrison dams) and increased water release for irrigation (e.g. Seminoe Dam). However, due to changes in locally simulated and actual inflow, dam operation can take on different characteristics (e.g. Toktogul Dam). Furthermore, peak discharge events caused by reservoir overflow (as also described by Masaki et al., 2018) were not always sufficiently represented in the observations (e.g. Garrison Dam). These differences indicate locally varying reservoir operation strategies. Several studies have developed reservoir operation schemes that can be calibrated to the local situation (Rougé et al., 2019; Yassin et al., 2019). However, worldwide implementation of these operation schemes remains limited by data availability.
Comparison between the simulated reservoir operation and values observed by Hanasaki et al. (2006) and Yassin et al. (2019). Panels indicate the time series
In order to assess the impact and capabilities of the newly integrated environmental flow requirements (EFRs) module, simulated water withdrawals with and without adhering to EFRs were compared.
If water use was limited to EFRs, irrigation withdrawals would need to be reduced by about 39 % (851
Average annual irrigation water withdrawal reductions when adhering to EFRs as a global gross total
Reductions due to EFRs were similar to Jägermeyr et al. (2017), who calculated irrigation withdrawal reductions of 41 %
(997
The VIC-WUR model introduced in this paper aims to provide new opportunities for global water resource assessments using the VIC model. Accordingly, several anthropogenic impact modules, based on previous major works, were integrated into the VIC-5 macroscale hydrological model: domestic, industrial, energy, livestock, and irrigation water withdrawals from both surface water and groundwater as well as an integrated environmental flow requirement module and a dam operation module. Global gridded datasets on domestic, industrial, energy, and livestock demand were developed separately and used to force the VIC-WUR model.
Simulated national water withdrawals were in line with reported national annual withdrawals (adjusted
An assessment of the effect of EFRs shows that (when one adheres to these requirements) global water withdrawals would be severely limited (39 %). This limitation is especially notable for groundwater withdrawals, which, under the strict requirements used in our study, need to be reduced by 81 %. VIC-WUR has potential for studying impacts of climate change and anthropogenic developments on current and future water resources and sector-specific water scarcity. The additions presented here make the VIC model more suited for fully integrated worldwide water resource assessments and substantially decrease computation times compared with Haddeland et al. (2006a).
n VIC, each sub-grid computes the water and energy balance individually (i.e. sub-grids do not exchange water or energy with one another). For the water balance, incoming precipitation is partitioned between evapotranspiration, surface and subsurface runoff, and soil water storage. Potential evapotranspiration is based on the Penman–Monteith equation without the canopy resistance (Shuttleworth, 1993). The actual evapotranspiration is calculated using two methods, based on whether the land cover is vegetated or not (bare soil). Evapotranspiration of vegetation is constrained by stomatal, architectural, and aerodynamic resistances and is partitioned between canopy evaporation and transpiration based on the intercepted water content of the canopy (Deardorff, 1978; Ducoudre et al., 1993). Bare soil evaporation is constrained by the saturated area of the upper soil layer. The saturated area is variable within the grid, as (as the model name implies) the infiltration capacity of the soil is assumed to be heterogeneous (Franchini and Pacciani, 1991). Saturated areas evaporate at the potential evaporation rate, whereas evaporation is limited in unsaturated areas. Surface runoff is produced by precipitation over saturated areas. Precipitation over unsaturated areas infiltrates into the upper soil layer and drains through the soil layers based on the gravitational hydraulic conductivity equations of Brooks and Corey (1964). In the first and second layer, water is available for transpiration, while the third layer is assumed to be below the root zone. From the third layer, baseflow is generated based on the non-linear Arno conceptualization (Franchini and Pacciani, 1991). Baseflow increases linearly with soil moisture content when the moisture content is low. At higher soil moisture contents, the relation is non-linear, representing subsurface storm flows.
For the energy balance, incoming net radiation is partitioned between sensible, latent, and ground heat fluxes as well as energy storage in the air below the canopy. The energy storage below the canopy is omitted if it is considered negligible (e.g. the canopy surface is open or close to the ground). The latent heat flux is determined by the evapotranspiration as calculated in the water balance. The sensible heat flux is calculated based on the difference between the air and surface temperature, and the ground heat flux is calculated based on the difference between the soil and surface temperature. As the incoming net radiation is also a function of the surface temperature (specifically the outgoing longwave radiation), the surface temperature is solved iteratively. Subsurface ground heat fluxes are calculated assuming an exponential temperature profile between the surface and the bottom of the soil column, where the bottom temperature is assumed to be constant. Later model developments included options for finite difference solutions of the ground temperature profile (Cherkauer and Lettenmaier, 1999), the spatial distribution of soil temperatures (Cherkauer and Lettenmaier, 2003), a quasi-two-layer snowpack snow model (Andreadis et al., 2009), and blowing snow sublimation (Bowling et al., 2004).
VIC-WUR used the variable monthly flow (VMF) method (Pastor et al., 2014) to limit surface water withdrawals. The VMF method (Pastor et al., 2014)
calculates the EFRs for streamflow as a fraction of the natural flow during high (Eq. B3), intermediate (Eq. B2), and low (Eq. B1) flow periods. The presumptive standard from Gleeson and Richter (2018) is used to limit groundwater
withdrawals (including aquifer groundwater withdrawals). This standard calculates the EFRs for baseflow as 90 % of the natural subsurface runoff
through time (Eq. B4). Here, daily instead of monthly EFRs were used to better capture the monthly flow variability.
VIC-WUR used a dam operation scheme based on Hanasaki et al. (2006). Target release (i.e. the estimated optimal release) was calculated at the start of the operational year. The operational year starts at the month where the inflow drops below the average annual inflow and, thus, the storage should be at its desired maximum. The scheme distinguished between two dam types: (1) dams that did not account for water demands downstream (e.g. hydropower dams or flood control) and (2) dams that did account for water demands downstream (e.g. irrigation dams). The original scheme of Hanasaki et al. (2006) also accounts for EFRs, which were fixed at half the annual mean inflow. Other studies lowered the requirements to a 10th of the mean annual inflow, increasing irrigation availability and preventing excessive releases (Biemans et al., 2011; Voisin et al., 2013). In our study, the original dam operation scheme was adapted slightly to account for monthly varying EFRs.
For dams that did not account for demands, the initial release was set at the mean annual inflow corrected by the variable EFRs (Eq. C1). For dams
that did account for demands, the initial release was increased during periods of higher water demand. If demands were relatively high compared with the
annual inflow, the release was corrected by the demand relative to the mean demand (Eq. C2). If demands were relatively low compared with the annual
inflow, release was corrected based on the actual water demand (Eq. C3).
As in Hanasaki et al. (2006), the initial target release was adjusted based on storage and capacity. Target release was adjusted to compensate for
differences between the current storage and the desired maximum storage (Eq. C4). Target release was additionally adjusted if the storage capacity was
relatively low compared with the annual inflow and was unable to store large portions of the inflow for later release (Eq. C5).
Water inflow, demand and EFRs were estimated based on the average of the past 5 years. Water demands were based on the water demands of downstream cells. Only a fraction of water demands were taken into account, based on the fraction of discharge the dam controlled. For example, if a dam controlled 70 % of the discharge of a downstream cell, 70 % of its demands were taken into account. Fractions smaller than 25 % were ignored.
The original dam operation scheme of Hanasaki et al. (2006) was shown to produce excessively high discharge events due to overflow releases (Masaki
et al., 2018). These overflow releases occurred due to a mismatch between the expected and actual inflow. In our study, dam release was increased
during high-storage events to prevent overflow and the accompanying high discharge events. If dam storage was above the desired maximum storage, target
dam release was increased to negate the difference (Eq. C6). If dam storage was below the desired minimum storage, release was decreased
(Eq. C7). Dam release was adjusted exponentially based on the relative storage difference: small storage differences were only corrected slightly,
but if the dam was close to overflowing or emptying, the difference was corrected strongly.
Data on irrigation, domestic, and industrial water withdrawals were based on the AQUASTAT database (FAO, 2016), the EUROSTAT database (EC, 2019), and the United Nations World Water Development Report (Connor, 2015). Data on GDP per capita and GVA were abstracted from the Maddison Project Database 2018 (Bolt et al., 2018), the Penn World Table 9.0 (Feenstra et al., 2015), and the World Development Indicators (World Bank, 2010).
Available data for domestic and industrial withdrawals were divided into a dataset used for parameter fitting (80 %) and a dataset used for validation (20 %). Domestic water demands were estimated for each United Nations subregion; thus, the data were divided per subregion to ensure a good global coverage of data. In the same manner, industrial water demand was divided per country. In cases where there was only a single data entry, the entry was added to both the fitting and validation data.
Conventional irrigation demands were calculated when soil moisture contents dropped below the critical threshold where evapotranspiration would be
limited. Demands were set to relieve water stress (Eq. D1). Rice paddy irrigation demands were set to always keep the soil moisture content of the upper
soil layer saturated (Eq. D2), similar to Hanasaki et al. (2008b) and Wada et al. (2014). For rice paddy irrigation, the saturated hydraulic conductivity
of the upper soil layer was reduced by its cubed root to simulate puddling practices, as recommended by the CROPWAT model (Smith, 1996). Total
irrigation demands were adjusted by the irrigation efficiency (Eq. D3). Rice irrigation used an irrigation efficiency of one, as the water losses
were already incorporated in the water demand calculation.
Domestic water demands were represented using a sigmoid curve for the calculation of structural domestic water demands (Eq. D4) and an efficiency
rate for the calculation of water use efficiency increases (Eq. D5). These equations differ slightly from Alcamo et al. (2003), as our study used
the base 10 logarithms of GDP and water withdrawals per capita because they provided a better fit.
Industrial water demands were represented using a linear formula for the calculation of structural industrial water demands (Eq. D6) and an
efficiency rate for the calculation of water use efficiency increases (Eq. D7).
TE was set at 0.976 and 1 for Organisation for Economic Co-operation and Development (OECD) and non-OECD countries respectively before the year 1980, 0.976 between the years 1980 and 2000, and 0.99 after the year 2000, based on Flörke et al. (2013). Industrial water intensities were estimated for the 246 United Nations countries based on the GVA and industrial water withdrawal data. In cases where insufficient data were available to calculate the industrial water intensities, either subregional (56 countries), regional (17 countries), or global (9 countries) intensity estimates were used.
For each thermoelectric power plant, the water intensity was combined with their generation to calculate the water demands (Eq. D8). Actual generation
is estimated by adjusting the installed generation capacity by 46 % for fossil fuel, 72 % for nuclear power, and 56 % for biomass power plants (based
on the U.S. Energy Information Administration (EIA) national annual generation data; EIA, 2013).
The energy water demands were subtracted from the industrial water demands at the location of each power plant. In cases where the grid cell industrial water demand was less than the energy water demand, national industrial water demands were lowered. In cases where even the national industrial water demands were less than the national energy water demand (three countries), the energy water demands were lowered instead. Energy demands were lowered until 10 % of the national industrial water demand remained in order to ensure some spatial coverage of industrial and energy water demands.
Livestock water demands were estimated by combining the livestock population with the water requirements for each livestock variety (Eq. D9).
VIC-WUR monthly discharge and monthly terrestrial total water storage anomalies were compared with observations from the GRDC dataset (GRDC, 2003) and the
GRACE satellite dataset (NASA, 2002) for eight major river basins (Amazon, Congo,
Lena, Volga, Yangtze, Danube, Columbia, and Mississippi river basins). Discharge stations
were selected if the upstream area was larger than 20 000
Comparison between simulated discharge and terrestrial total water storage anomalies and GRDC- and GRACE-observed values. Panels indicate the multiyear averages of anthropogenically impacted simulations (red) and observations (black).
All code for the VIC-WUR model is freely available at
BD and WHPF developed and tested the model additions introduced in VIC-WUR. MHPF was responsible for model maintenance, testing, and the routing module; BD implemented and tested the water use modules. MTHvV provided data for the energy sector. BD generated and analysed the results. MTHvV, BN, and FL provided general oversight and guidance. BD prepared the paper with contributions from all co-authors.
The authors declare that they have no conflict of interest.
The authors would like to thank Rik Leemans for his guidance and detailed comments.
This research has been supported by the Wageningen Institute for Environmental Climate Research (grant no. 5160957551).
This paper was edited by Wolfgang Kurtz and reviewed by three anonymous referees.