The impact of climate change on groundwater resources Open Access

Changes in precipitation and evapotranspiration (ET), at different spatial and temporal scales, will affect the global water cycle (Brutsaert, 2005). Surface water bodies like rivers respond fast to climate change, as the residence time of water in rivers is relatively low. The effects of climate change on river discharge have been more intensively studied than the impact of climate change on aquifers. For instance, it is expected that the summer river discharge of the main European rivers will decrease, whereas winter river discharge is likely to increase with more severe floods (Arpe and Roeckner, 1999; Pfister et al., 2004).

Assessing the impact of climate change (Intergovernmental Panel on Climate Change (IPCC), 2007) on groundwater resources is an important issue that has not received too much attention in the scientific literature until now. Aquifers react much slower to climate change than surface water bodies, but on the long‐term the average groundwater level might be much lower or higher than under the current conditions, with serious implications for drinking water supply, for example. The drinking water supply for a large part of the population on Earth relies on groundwater resources. It is well known that certain aquifers in the past, during different climate conditions, had much higher or lower groundwater levels than currently.

Aquifers for which the impact of climate change was studied are situated nearly exclusively in Northern America (Cohen et al., 2006; Scibek and Allen, 2006a, b; Jyrkama and Sykes, 2007; Loaiciga, 2003; Allen et al., 2004) and Europe (Eckhardt and Ulbrich, 2003; Woldeamlak et al., 2007). For an aquifer in southern British Columbia in Canada only a small impact of climate change on groundwater levels is expected (Scibek and Allen, 2006a), whereas Jyrkama and Sykes (2007) predicted an increased recharge for an aquifer in Ontario (Canada). The Ellensburg basin in the state of Washington (USA) was calculated to have a reduction of more than 25 per cent of the median annual recharge for the scenario that the amount of carbon dioxide in the atmosphere doubles, as compared to the atmospheric carbon dioxide concentrations in 1990 (Vaccaro, 1992). The extended High‐Plains aquifer (450,000 kilometer² and underlying the states of South Dakota, Wyoming, Colorado, Nebraska, Kansas, Oklahoma, Texas and New Mexico in the USA) is expected to receive less recharge for all considered future climate series (Rosenberg et al., 1999). The Edwards Balcones Fault Zone aquifer in Texas (USA) is also expected to receive less recharge with severe impacts on water resources (Loaiciga et al., 2000).

Concerning Europe, for an aquifer close to Grenoble (France) ground water levels were estimated to decrease up to four meters, making impossible the current practice of irrigated agriculture in the future (Bouraoui et al., 1999). A chalk aquifer in Belgium showed for some of the future climate scenarios drops in groundwater level of up to seven metres (Brouyère et al., 2004; Woldeamlak et al., 2007). For the UK, particularly in the Southern part a reduction in groundwater recharge is expected, in spite of the increased winter rainfall (Arnell, 1998). In the exceptional warm and dry summer of 2003, in Switzerland fast decreases of groundwater levels were observed, for example in a piezometer close to Uster, Canton of Zurich, the decrease was more than five metres in only nine months (AWEL, 2004). The combination of high water demand and reduced aquifer recharge, resulted in problems of water supply (e.g. in Winterthur). It is expected that summers like 2003 will be rather common in Switzerland at the end of the twenty‐first century (Schär et al., 2004).

Green et al. (2007) estimated for two aquifers in Australia a large uncertainty regarding future recharge to the groundwater. For the Mediterranean climate zone in West Australia an increase of the recharge is likely, while for the subtropical climate zone in Queensland a 37 per cent rainfall increase is likely to at least double the amount of groundwater recharge.

With help of a water balance model a rough estimate of changes in soil moisture availability over the continent of Africa was estimated for the period 2010‐2039. Besides, climate change, soil degradation is expected to have an important impact on reduced soil moisture availability and reduced recharge to the groundwater over large parts of the continent (Feddema and Freire, 2001). First rough assessments were also made of the changes in groundwater recharge over Russia. Over large parts of Russia an increase of the groundwater levels is expected during the period until 2025 (Semenov et al., 2004). For the Pingtung Plain in Southwestern Taiwan available groundwater resources are expected to decline during the next two decades (Hsu et al., 2007). The impact of climate change on groundwater resources is of special concern in semi‐arid regions. In these regions, already a water deficit exists and climate models expect a reduction of future precipitation amounts (for example in the Mediterranean area). In addition, in these regions an important part of the drinking water is pumped from aquifers. The combination of a reduced aquifer recharge and a strongly increasing water demand (Vörösmarty et al., 2000) makes the situation here especially delicate. For the coastal Chaouia aquifer in Morocco, it is found that piezometric head levels decreased considerably during the last decennia, due to increased water extraction and in spite of a small increase in the amount of precipitation (Moustadraf et al., 2008).

Climate change can impact the groundwater levels through various processes that affect the water balance. The hydrological water balance for a catchment can be written as: Equation 1 where P is precipitation [L³T⁻¹], ET is evapotranspiration [L³T⁻¹], R is discharge (surface runoff and river discharge) [L³T⁻¹], Q is human water extraction [L³T⁻¹] and ΔS is change in storage (positive if storage increases) [L³T⁻¹]. Under stationary climate conditions, and over a longer time period, ΔS=0. Changing climate conditions yield changes in precipitation, ET, discharge, water extraction and storage. These storage changes affect the unsaturated and saturated zones and storage in the form of snow and ice. If we consider the storage changes for a compartment like an aquifer, the interactions with the other hydrological compartments also have to be considered. Changing interactions with surface water bodies have a significant impact on changes in aquifer storage and should be included in the aquifer water balance: Equation 2 where L_in and L_out are fluxes from the surface water bodies (i.e. rivers and lakes) to the aquifer and viceversa [L³T⁻¹]. Climate change assessment studies for groundwater solve normally a distributed form of equation (2). However, ΔP and ΔET are normally taken from climate simulations, and only if integrative hydrological models are used, the other terms on the right‐hand side of equation (2) are explicitly solved.

In the next sections, we focus on using ΔP and ΔET from GCM's for assessing the impact of future climate change on groundwater resources. The focus will be on the state of the art of integrated models that predict this impact. Other aspects of climate change and groundwater resources, like observed recent changes in groundwater levels or estimated historical groundwater levels from paleoclimate data are not discussed here.

Coupled global atmosphere‐ocean‐sea ice‐land models can help to predict the future climate. The set of governing equations for atmospheric fluid motion is discretized by finite difference or spectral methods and propagated forward in time, for example for the next century. The maximum time step is limited so that the fastest disturbances (gravity waves or jet streams; sound waves are filtered away from the equations) cannot pass one grid cell in one time step. This in turn limits the maximum grid size because of CPU limitations. Currently, most climate models have a typical resolution of 100‐200 kilometer. This coarse grid resolution makes that various sub‐grid processes are not solved physically, but parameterised (Kalnay, 2003). These parameterisations are verified experimentally (i.e. for the current weather and climate conditions). Processes that are parameterised are convection, cloud formation, land‐surface processes including hydrology, radiation and turbulence, among others. The coarse resolution also limits the proper representation of the topography and soil and vegetation properties. These simplifications especially impact the modelling of the hydrological cycle in climate models.

Different climate models use different parameterisation schemes, have different spatial discretizations and also different soil and vegetation properties (among others), and therefore also give different climate predictions. The European study PRUDENCE compared the simulations of the future European climate with a large number of different climate models, and found large differences in the projected amounts of precipitation (Déqué et al., 2007). For the IPCC A2 scenario, the precipitation predictions for Northern Switzerland showed a considerable uncertainty. The 95 per cent probability interval of summer precipitation around 2050 is between a decrease of 31 and 7 per cent. For winter precipitation, the variation is between −1 and +21 per cent (Frei, 2007). These confidence intervals capture some, but most probably not all of the conceptual uncertainty. Coupled atmosphere‐ocean‐sea ice‐land models are increasingly used for seasonal predictions, and there it was found that a multi‐model ensemble approach gave better predictions than a single model ensemble approach, and covered better the span of uncertainty (Hagedorn et al., 2005). For future climate predictions, the uncertainty with respect to the evolution of greenhouse gas emissions is of course, an additional relevant uncertain factor. Studies that assess the impact of climate change on groundwater resources did in general not work with future climate scenarios from different climate models.

Precipitation is calculated at the coarse grid of the climate model, and is affected by the mentioned parameterizations and the smoothing of topography, soil and vegetation properties. Precipitation estimated by a climate model has to be downscaled to the local hydrological model.

Downscaling can be done with dynamical models or statistical methods. Dynamical downscaling is carried out with regional climate models (RCMs) and the RCM is along its boundaries forced by the global circulation model (GCM). The RCM calculates the future climate on a higher resolution grid and is expected to give better results as the representation of the topography and soil and vegetation properties is better than in a (GCM). However, the RCM is embedded in the GCM and one‐way coupled, which means that the RCM cannot feedback on the global circulation. More over, even if a RCM is run on a high‐resolution grid there is still a need for statistical downscaling. There are two main reasons why there is still a need for statistical downscaling:

• 1.

  the spatial resolution of the RCM and the (groundwater) hydrological model do not coincide; and

• 2.

  bias in the RCM due to, for instance, the mentioned parameterizations.

The bias tends to show a complex temporal pattern. For example, often it is observed (by comparing model simulations with observations (Frei et al., 2006) that the frequency of precipitation events is overestimated, whereas the intensity is underestimated. For Switzerland, it was found that convective precipitation is underestimated and orographically enforced precipitation overestimated. If climate model predictions of precipitation would only be corrected on the basis of the mean deviation between simulated and measured precipitation, the bias of future precipitation would not be assessed correctly. To put a very simple example: if for instance for the current climate conditions both the contribution of orographically induced precipitation and convective precipitation to the overall precipitation amount is 50 per cent, and the bias is, respectively, +30 and −10 per cent, the overall bias is +10 per cent. If for future climate conditions the contribution of the orographically induced precipitation reduces to 25 per cent and the convective precipitation increases to 75 per cent, the overall bias of the simulated future time series would not be +10 but 0 per cent.

Statistical downscaling techniques link observations at a meteorological station with the atmospheric flow pattern, for instance through methods like canonical correlation analysis (Busuioc et al., 1999; Hertig and Jacobeit, 2007). Through the established links, which are assumed to remain valid under a changing climate (but have a physical basis), values for the meteorological variables of interest for the future can be obtained. For instance, for estimating daily precipitation at a certain location, simulated fields of specific humidity, pressure, divergence and vorticity at different levels in the atmosphere at grid boxes surrounding the location, could be used as predictors. With help of longer historical time series, it is also possible to investigate the stationarity of these multivariate statistical relationships (Hertig and Jacobeit, 2007).

The more sophisticated downscaling techniques for precipitation like canonical correlation analysis might yield considerably different future precipitation amounts compared with simple bias corrections. Some hydrological studies which investigated the impact of climate change dealt with this problem more seriously, for example by comparing the simulation results of a GCM in a reference period (typically the present climate) with a re‐analysis data set (Scibek and Allen, 2006a) or a measured time series (Loaiciga, 2003). Often, however, a simpler approach is followed and some very general trends, extracted from a GCM, are used to generate future climate data. Many studies that address the impact of climate change on groundwater resources, focus on the influence of spatial variable processes in the soil, and their impact on percolation and recharge, with a limited attention on obtaining a realistic representation of future time series of meteorological forcing (Rosenberg et al., 1999; Woldeamlak et al., 2007; Jyrkama and Sykes, 2007). Besides uncertainty and downscaling, there are other relevant issues concerning precipitation. However, they are even more difficult to take into account and were not included in studies of the assessment of climate change on groundwater resources. One aspect is the temporal resolution of future precipitation. Even if the total amount of precipitation is unchanged, a different distribution over the year might result in a larger or smaller ET. It is also possible that an increase of extreme precipitation intensities would result in a larger surface runoff and less‐aquifer recharge.

Climate change will also have an important impact on the ET from the Earth's surface. As a consequence, the terrestrial water cycle changes which on its turn has a strong feedback on meteorological conditions, as documented in several recent studies (Seneviratne and Stöckli, 2008). These changes are strongly related with changes in the soil moisture contents. The water cycle and the energy cycle are coupled at the land surface by the soil moisture content. For wet soils, a large part of the incoming radiation is used for ET (latent heat flux), whereas for dry soils most of the energy is used for heating the air (sensible heat flux). Therefore, soil moisture contents influence for example the formation of convective storms (Betts et al., 1996; Eltahir, 1998; Schär et al., 1999; Betts, 2004), air temperature (Seneviratne et al., 2006a; Koster et al., 2006), vegetation functioning (Ciais et al., 2005), and may have implications for seasonal forecasting (Koster and Suarez, 2001, Seneviratne et al., 2006b). It is important to understand this feedback better and develop better land‐atmosphere models in order to improve the regional climate predictions.

Hydrological models calculate almost without exception ET conceptually, for example by empirically based relations between:

• the soil depth and ET; and

• soil moisture content and the ratio actual ET/potential ET.

The necessary inputs to calculate potential ET are the measured radiation, wind speed, vapour pressure and air temperature. If less input data are available simplified formulations are used. Calculating potential ET from meteorological data might give good results for observed meteorological data, but there are serious drawbacks using future climate predictions as data input to calculate potential ET.

The net radiation, vapour pressure, air temperature and even the wind speed that are calculated by the GCM or RCM will depend on the land‐atmosphere scheme of the model. The land‐atmosphere scheme of a GCM divides the incoming net radiation over heating of the air (sensible heat flux), heating of the ground (ground heat flux) or ET of water (latent heat flux), and monitors the changes of the water storage in the soil compartment (soil moisture content). In fact, the soil moisture content governs the division of the incoming net radiation over the three different energy fluxes. The calculated sensible and latent heat fluxes determine also whether convective clouds are formed, which affects for example the incoming net radiation for the next time step. The land‐atmosphere models of GCM's and RCM's represent the hydrology strongly simplified, have uniform parameter values for large areas and do not allow, for instance, for lateral redistribution of water between grid cells. As the simulated meteorological conditions close to the Earth's surface are strongly affected by the simplified representation of hydrology, the potential ET that would be estimated from it might deviate significantly from the true value. Efforts are being made to improve the representation of hydrology in GCMs. For example, recently the community land model (CLM) was improved by explicitly accounting for groundwater in the land‐atmosphere scheme (Oleson et al., 2008; Stöckli et al., 2008).

Instead of getting the meteorological variables for calculating potential ET from the GCM or RCM, an alternative could be to downscale directly the estimated actual ET by the GCM or RCM. However, there are very few historical measurement data on actual ET that would allow inferring statistical downscaling relationships. Since more than a decade, a global network of eddy covariance data is collecting data on ET, but eddy covariance data underestimate ET (Wilson et al., 2002; Barr et al., 2006), which seems to be related to the missing of low‐frequency turbulence and large eddies (Finigan et al., 2003; Foken et al., 2006; Inagaki et al., 2006). The correction for the underestimation of ET is not trivial and shows a complex relation with the atmospheric stability and wind velocity (Hendricks Franssen et al., 2009). Also other processes, like foot print heterogeneity (i.e. the heterogeneity around the flux tower; Göckede et al., 2008) and advection introduce errors, which are typically dependent on the wind vector. More over, the eddy covariance data also show considerable random measurement errors (Richardson and Hollinger, 2005). Therefore, it is very difficult to relate GCM‐simulated ET for a large grid cell and measured ET (with systematic and random errors) for a small region in a meaningful manner, and the downscaling relationships are expected to be very complicated.

Altogether, currently it is believed that downscaling of the meteorological variables from which potential ET is calculated, despite the limitations mentioned before, is a more viable option than downscaling of actual ET. The joint statistical downscaling of the relevant variables for calculating potential ET from a GCM or RCM to a (groundwater) hydrological model is far from trivial. Not any of the studies that have been published on the impact of climate change on groundwater resources used sophisticated downscaling techniques to link GCM modelled potential ET and “measured” potential ET.

Even if future potential ET is estimated from climate predictions and downscaling techniques, still a sophisticated, integrative hydrological model is needed to calculate actual ET. Such a hydrological model should be able to model the unsaturated and saturated zones and interaction between surface water bodies and groundwater in a fully coupled manner. It should also include the role of vegetation and processes like bare soil evaporation, transpiration, interception, evaporation from intercepted water and surface runoff.

The unsaturated zone plays a central role in the estimation of the impact of climate change on groundwater resources, mainly because the soil moisture content is the main factor that determines the Bowen ratio (the ratio between the sensible and latent heat fluxes) and therefore the actual ET. This was discussed in the anterior section. The modelling of processes in the unsaturated zone is also important for the timing of the recharge, i.e. the delay that excess precipitation water takes to reach the aquifer. Some studies suggest that modelling processes in the unsaturated zone is critical for an accurate estimate of the recharge (Hunt et al., 2008).

The interaction between aquifers and surface water bodies play locally a very significant role. For instance, the water balance of the aquifer below the city of Zurich in Switzerland shows that by far the largest contribution of water comes from the river Limmat (Doppler et al., 2007). The groundwater levels of such aquifers are strongly influenced by the river stages. In order to assess the impact of climate change on those aquifers, it is essential to estimate future time series of river stages as well (Scibek et al., 2007).

Surface runoff is also important to be included in the assessment of climate change on groundwater resources because the fraction of precipitation that becomes surface runoff for a certain area might be modified as a consequence of climate change. This fraction might be different as consequence of changes in the precipitation intensity, the temporal variation of the amount of precipitation (with for example more frequent precipitation over nearly saturated soils in winter), vegetation or land cover, frozen soil conditions and amount of precipitation that falls in the form of snow.

Other important aspects of the assessment of climate change on groundwater resources, include changes in water extraction (e.g. because of changes in the needed amount of irrigation water), land use changes (for example due to the vegetation response to climate change) and changes to soil properties (Holman, 2006).

Integrative hydrological models can consider many of the mentioned processes. In particular, physically based models like MIKE‐SHE (Refsgaard et al., 1992) or HYDROGEOSPHERE (Therrien and Sudicky, 1996) are interesting for assessing the impact of climate change on groundwater resources. An interesting development is that MIKE‐SHE also calculates the ET physically, i.e. by coupling the energy and water balance at the Earth's surface (Overgaard et al., 2006). In particular, a recent extension of MIKE‐SHE includes a two‐layer, energy‐based, land surface model (Overgaard, 2005). MIKE‐SHE was also tested for fully two‐way coupling with atmospheric models and used together with the RCM HIRHAM for simulations of regions in Denmark, the Okavango Delta in Botswana and Panama (personal communication). Also, the hydrological model HYDROGEOSPHERE is tested for fully two‐way coupled simulations with GCM's. Nevertheless, although processes are modelled in a coupled and physical fashion, these models often have to be applied at scales where the governing equation for flow in the unsaturated zone (i.e. the Richards equation) does not hold. The small scale heterogeneity of the unsaturated zone that is of importance for fast infiltration of rain water through macropores, is not represented well in such cases. Another deficit of these integrative hydrological models is that vegetation processes and boundary layer flows are not represented very well. Models for land‐atmosphere interaction that were developed in the atmospheric sciences community tend to represent vegetation processes and boundary layer flows in more detail and more mechanistically. An example is CLM that models vegetation processes in high detail (Oleson et al., 2008).

Integrative hydrological models were used in a few cases to assess the impact of climate change on groundwater resources, but in most cases surface and subsurface hydrology were not modelled in a coupled manner. Green et al. (2007) used an integrated, physically based modelling approach, including the calculation of ET. Brouyère et al. (2004) used a physically based method, but ET was estimated with a conceptual approach. Cohen et al. (2006) used a simple methodology to calculate ET. Other authors calculated ET on the basis of a conceptual model (Allen et al., 2004; Scibek and Allen, 2006a, b; Scibek et al., 2007; Jyrkama and Sykes, 2007; Woldeamlak et al., 2007). Loaiciga et al. (2000) used a simple scaling methodology to calculate recharge rate. Hsu et al. (2007) estimated recharge rate as a function of precipitation and potential ET, by applying simple linear regression relations on the basis of historical data and historically calibrated values. In general, it can be said that studies on the impact of climate change on groundwater resources focus either on estimating future time series of recharge, or subsurface processes, but never both meteorological and subsurface processes are adequately modelled.

If an integrative hydrological model is used to assess the impact of climate change on groundwater resources, the calibration of such a model with historical observations is very important. In groundwater hydrology, the state‐of‐the‐art is to condition models inversely with Monte‐Carlo type methods like sequential self‐calibration (Gómez‐Hernández et al., 1997; Hendricks Franssen et al., 2003), the (regularized) pilot points method (LaVenue et al., 1995; Alcolea et al., 2006) or the representer method (Bennett, 1992; Valstar, 2001). An alternative is to search a single‐best solution with iterative techniques (Carrera and Neuman, 1986; Kitanidis, 1995), techniques that include singular value decomposition (Tonkin et al., 2007) or the moment equations method (Hernandez et al., 2003, 2006). In surface hydrology, new developments for the calibration of many model parameters include Markov Chain Monte‐Carlo methods (Vrugt et al., 2003; Vrugt and Robinson, 2007) or Ensemble Kalman filtering (Vrugt et al., 2005; Liu and Gupta, 2007). Nevertheless, little experience exists with the calibration of integrative hydrological models where both the unsaturated and saturated zones are highly discretized and all the above‐mentioned hydrological processes are considered. Therefore, there is still the need to improve the methods for the calibration of integrative hydrological models.

The studies published until now on the impact of climate change on groundwater resources are in general sensitivity studies, as no serious predictions were made that considered the problem in its full complexity. This paper highlighted some serious limitations of the studies published until now. It is important to downscale the future precipitation from GCM's with more rigorous methods. As meteorological data to calculate potential ET are calculated in a GCM in a fully coupled manner, but on a very coarse scale, also downscaling techniques are needed to downscale those variables. Moreover, it is desirable that integrative hydrological models are used that include a physically based ET calculation. Although integrative hydrological models exist that are able to model the flow in the unsaturated and saturated zones, surface runoff and interaction between surface water bodies and subsurface water in a fully coupled and physical manner, often these models have to be applied (in practice) on too coarse scales. Another limitation of integrative hydrological models is that some of the relevant processes, like vegetation functioning and boundary layer flows, are described in a very simplified manner. Integrative hydrological models are commonly available, but it is far from trivial to calibrate such an integrative hydrological model with historical data. Improved methods are needed to calibrate integrative hydrological models.

At the same time, it is desirable to develop a simplified methodology for assessing the impact of climate change on groundwater resources. This would also allow estimating for many small aquifers the response to climate change. However, first more experience should be gained with complex, fully coupled models before knowing which simplifications can be made. Probably, it is more justifiable to simplify the representation of subsurface groundwater flow than the procedures to obtain the future climate forcing.

Improved estimates of the impact of climate change on groundwater resources allow adapting the infrastructure to optimise water resources management for the future, so that problems with scarcity of drinking water can be reduced.

The author thanks two anonymous reviewers and Fritz Stauffer for helpful comments on an earlier version of this paper.

Harrie‐Jan Hendricks Franssen received an MSc from the Agricultural University of Wageningen, The Netherlands, with a specialization in soil science. Afterwards, he specialized in water resources management at the Technical University of Valencia (Spain) and received a PhD from the Technical University of Valencia in Groundwater Hydrology. Since 2001, he works at the Department of Environmental Engineering of the ETH Zurich, since 2005 as Lecturer and Senior Researcher in stochastic hydrology. In 2008, he obtained an additional MSc degree from the ETH Zurich in atmospheric and climatic sciences. Harrie‐Jan Hendricks Franssen can be contacted at: hendricks@ifu.baug.ethz.ch