The predicted future changes in LULC over the Plana Oropesa-Torreblanca are of greater magnitude than the historical ones and could drastically modify the rural and urban landscape. The already-approved tourist developments (the public urbanization work – PAI – for the Marina d'Or Golf in Oropesa and Cabanes, and the General Town Plan – PGOU – for Torreblanca) anticipate an increase in population of more than 130 000 inhabitants, as well as the disappearance of most of the agricultural activity in the area. These significant changes in LULC will produce significant impacts on water demands, and, therefore in pumping and recharge and so to the hydrodynamics of the aquifer. In contrast, there are no significant changes to LULC anticipated in the area belonging to the municipality of Alcalà de Xivert, also situated on the Plana.

The General Town Plan for Torreblanca (PGOU Torreblanca, 2009) approves the conversion of 70 % of the municipality's area included in the General Town Plan, which is currently used for citrus agriculture, into land classified as buildable residential or industrial. Even along the coast, to the north of Prat de Cabanes, is the projected urbanization – included in the so-called Integrated Activity Plan (PAI) – of Doña Blanca Golf (Fig. 5).

The municipalities of Cabanes and Oropesa in the southern inland part of the Plana de Oropesa-Torreblanca have approved PAIs for the Marina d'Or Golf, which will include three golf courses, private residential developments, hotel complexes and associated garden areas. Once all these planned constructions are completed, they will cover some 16 km2 (Fig. 5). Two of the conditions that the Valencian Government imposed before it approved the PAI Marina D'Or Golf are that the whole of the area that requires irrigation (both golf course and garden areas) must use recycled residential water from the wastewater treatment works; and that all water destined for urban supply must be sourced from a desalination plant. Thus, as the various individual projects are built, the groundwater abstractions in the area are falling, until they will cease completely at the end of the PAI development period. Over the neighbouring Maestrazgo aquifer, we will assume that there are no changes in LULC.

We propose a method to generate consistent potential future climate scenarios for a short-term horizon (2011–2035) from the historical (1973–2010) data (see Sect. 2.2) and the climate model simulations performed in the framework of the EU CORDEX project (2013). It requires an analysis of the results obtained by applying different downscaling techniques. A multi-criteria analysis of some statistics of these series was performed to identify the best simulations of the historical data. Different ensemble hypotheses have been adopted to define more representative potential future climate scenarios to be employed in the groundwater impact study. Our main target is to provide an estimate of the most representative plausible future climate scenarios. For this reason we propose generating and propagating four plausible representative climate scenarios defined by an ensemble of different climate models, which provide a better approach of future climate scenarios than taking directly a scenario defined by a single model. The “ensembles” coalesce and consolidate the results of individual climate projections, thus allowing for more robust climate projections that are more representative than those based on a single model (Spanish Meteorological Agency, AEMET, 2009). In this paper we do not intend to perform a detailed analysis of hydrological uncertainty. In this case, in order to assess the uncertainty on hydrological impacts it would be more appropriate to obtain results from each individual climate model.

In this work we have focused on the information available for the most pessimistic emission scenario (RCP8.5). We analyse the information coming from the EU CORDEX project (2013), where we find nine climate-change scenarios (see Table 1) defined with the simulations (control and future series) obtained with five regional climate models (CCLM4-8-17, RCA4, HIRHAM5, RACMO22E and WRF331F) nested with some GCMs (four GCMs were available).

We have obtained representative lumped series of these simulations for our system, by weighting the values in each CORDEX cell according to its surface in the domain. Figure 6 shows a significant bias between the historical data and the control simulation that will force us to apply a correction technique to generate future scenarios.

We also observe important differences between the statistics of future (2011–2035) and control (1976–2000) series (rainfall and T) for each RCM (Fig. 7).

In accordance with the hypotheses assumed to define future climate series in a water resource system (starting from the climate model simulations) we can consider two different kinds of downscaling techniques: bias correction techniques and delta change techniques (Räisänen and Räty, 2012). In the present study we apply both conceptual approaches (bias correction and delta change techniques).

The bias correction techniques are based on the analysis of the statistical difference between the climatic variables in the historical data and the control simulations produced by the climate models for the same period. They aim to define a transformation function to correct the control series to obtain a better approximation of the historical statistic. They assume that in the future the bias between model and data will be the same as observed in the historical period (e.g. Watanabe et al., 2012; Haerter et al., 2011). The delta change approaches assume that the model can obtain a good approximation of the relative changes in climate variables' statistics, but do not provide a good prediction of the absolute values. Accordingly, they try to characterize the “delta change” produced in the main statistics of the climatic variables by analysing the relative difference between the future and control scenario simulations. The future series will be obtained by perturbation of the historical series in accordance with the estimated “delta change” (e.g. Pulido-Velazquez et al., 2011a, 2014; Räisänen and Räty, 2012).

When applying both correction techniques (bias and delta change) the spatial resolution of the historical data available for our systems is usually more detailed than those adopted by the climate model and, therefore, these transformations indirectly produce solutions with higher spatial resolution than the RCM one. Therefore, they are commonly known as downscaling transformations. We have applied two downscaling techniques (correction of first and second order moments) for both conceptual approaches (bias correction and delta change techniques). The correction of the first and second moments for the delta change approach was defined at monthly scale as described in Pulido-Velazquez et al. (2014). The bias correction case was also defined in an analogous way. The difference is that the perturbation is defined by modifying the statistics (first and second moments) of the control series to approximate the historical ones. It will be applied assuming to correct the future assuming that will be invariant. As example, the first, second and third moments of the series obtained with both approaches for one of the RCM (RCA4 linked to CNRM-CM5) are represented in Fig. 8.

Two multi-objective analyses are proposed: one related to the bias approaches and another to the delta ones.

In the delta change approaches we applied the multi-criteria analysis proposed by Pulido-Velazquez et al. (2014). It intends to identify the models that provided the best approximations to the main historical statistics (mean, standard deviation and asymmetry coefficients) based on the analysis of their control simulation of the historical period. The dominated solution or “inferior” models approaching the historical statistic were identified and eliminated. A model is eliminated (see Table 2) if any other model's prediction provides better approximation to the cited statistics.

In this study we also propose a multiple-criteria analysis for the bias correction approaches. This allows us to identify the best combination of model and bias correction techniques (see Table 3) to approximate the main statistics of the historical series. Since most of the combinations of model and bias correction techniques provide very good approximations to the first moment, a relative error threshold was defined to consider a corrected control to approach better a statistic when significant differences (higher than the threshold) are obtained.

We considered four options to define representative future scenarios by applying different ensembles of corrected projections. Two ensemble scenarios (Ei) were generated by a linear combination of all the future series generated by delta change (E1) or bias correction (E2). Two other options were defined by combining only the non-eliminated models (E3, for the delta change approach) or combinations of models and correction techniques (E4, for the bias correction techniques). They do not consider the eliminated options because they provide inferior approximations to the historical series that make us mistrust their predictions.

All ensemble predictions show a very similar increase in mean T (see Fig. 9). The standard deviation estimated with the delta change approaches is quite similar to the historical one, but both ensembles defined by applying bias correction show smaller standard deviations.

A reduction in future mean rainfall is predicted by all the ensembles for every month except September and October, when relative increases in rainfall are predicted. Since these are the months with the highest historical rainfall, the overall effect is a higher total annual rainfall compared to the historical. During September and October in this Mediterranean area we have important storms related to the phenomenon known as the “cold drop” (Roth, 2003). It is related to the higher total rainfall observed in these months. In accordance with the obtained potential scenarios, in the future we would have an increment in these extreme rainfall events.

As we also observed for the T, the standard deviation of the future rainfall predicted with the delta change approaches is quite similar to the historical one, but both ensembles defined by applying bias correction show significant reductions.

Based on the European Environment Agency analysis (EEA, 2014) of historical and potential future SLRs, we propose a SLR scenario to study the sensitivity of our GC simulations to a potential SLR. The EAA report states that, over the last 2 decades, satellite measurements have indicated a mean rate of SLR of more than 3.2 mm yr-1. If we assume that this rate remains constant, then, by the end of a future horizon of 25 years, we would have a rise in sea level of 0.08 m. On the other hand, model simulations for the RCP8.5 emission scenario show a rise in sea level for 2081–2100 in the range 0.45–0.81 m. If we assume a constant rate of SLR from now until 2100, the sea level would rise a maximum of 0.19 m (more than double the observed rate over the last 2 decades). This value (a rise of 0.19 m by the end of the future horizon) was used to define a very pessimistic scenario of maximum SLR. A linear SLR was considered during the future horizon (2011–2035).

Based on the historical climate (rainfall and T) and recharge series for the period 1973–2010 described in Sect. 2.2, we applied the simple empirical rainfall-recharge proposed by Pulido-Velazquez et al. (2017) to generate yearly aquifer recharge series. The model assumes that rainfall and T are the most important climatic variables determining potential aquifer recharge, and the variability of both of them will determine the impacts of future potential climatic scenarios on the recharge.

It intends to define a correction function for the perturbation of the historical series defined as the difference in rainfall and evapotranspiration (ETR) (difference hereafter referred to as PE series), modifying its mean and standard deviation to make them equal to the statistic of the historical aquifer recharge previously deduced from the lysimeter measurements (see Sect. 2.2). The calibrated function is employed to obtain the precipitation recharge series (PR) also for the future period. The steps required are the following (Pulido-Velazquez et al., 2017).

Estimation of the change in the mean and standard deviation of PE and PR series for the historical period:1Δμ=μ(PR)-μ(PE)μ(PR)andΔσ=σ(PR)-σ(PE)σ(PR)

In order to approach the impact of seasonal variability, the annual rainfall-recharge values obtained using the simplified model were distributed between the 12 months maintaining the pattern of the historical rainfall-recharge series. We assume that recharge from a rainfall event will reach the aquifer in less than 1 month, so working with stress periods of 1 month means there is no delay between the rainfall and the aquifer recharge.

Two simplified rainfall-recharge models were developed, one for the Plana de Oropesa-Torreblanca aquifer and the other for the Maestrazgo aquifer.

The Maestrazgo aquifer recharge model is used to assess inflows to the carbonate basement that produces lateral inflows (LI) to the Plana Oropesa-Torreblanca aquifer under various potential GC scenarios. The Maestrazgo aquifer has an important storage capacity which is almost in a natural regime and does not present significant changes in hydraulic head. The Maestrazgo rainfall-recharge model is employed to assess future recharge, the LI to the carbonate basement being obtained by assuming a constant ratio between Maestrazgo rainfall recharge and these LI.

The LULC information was used to estimate agricultural water requirements following a procedure to compute crop water requirements based on the FAO Irrigation and Drainage Paper (Allen et al., 1998). This approximation was applied in previous CC impact research studies (e.g. Escriva-Bou et al., 2016). The irrigation values added to the rainfall constitute the total inflows coming from the surface system. A Turc (1954) model was employed to estimate the total ETR in the area considering not only rainfall, but also irrigation water. The difference between the total ETR and the ETR for the rainfall allows us to determine the ETR related to irrigation, taking into account the climate conditions (rainfall and T). The estimated irrigation demands have also been employed to assess pumping taking into account information about the origin of the water that supplies each demand. The irrigation demands are multiplied by the irrigation return coefficients obtained for the crops in this area in previous studies (Tuñon, 2000) to assess recharge from irrigation.

Based on the hydrological description in Sect. 2.1, a conceptual model was defined to approach the aquifer as an unconfined heterogeneous detritic aquifer. The inflows to the aquifer include rainfall recharge, lateral groundwater inflows from the bordering aquifers (LGI) and irrigation returns. The outflows include natural springs that feed the wetland, pumping wells and outflows to the sea. The historical evolution of the inflows and outflows is represented in Fig. 10.

The SEAWAT 3-D finite-difference numerical code (Guo and Langevin, 2002) was used to solve the coupled partial differential equations for variable-density flow and transport. It combines MODFLOW (McDonald and Harbough, 1988) and MT3DMS (Zheng and Wang, 1999) into a single code that conserves fluid mass, rather than fluid volume, and uses the equivalent freshwater head as the primary dependent variable.

The domain has been divided into 32 columns and 90 rows in the horizontal plane, with a cell size of 250 m × 250 m. Therefore, the groundwater flow and transport domain extends over a size of 8000 m × 22 500 m. We adopted a vertical discretization consist of 11 layers in order to avoid a high computational cost (e.g. Sreekanth and Datta, 2010) while maintaining reasonable levels of numerical dispersion and representing complex flow patterns near areas of high concentration gradients (Guo and Langevin, 2002). These 11 layers, were defined as confined/unconfined layers where the transmissivity varies. On the one hand, for the flow boundary conditions a prescribed head of 0 m has been assigned to all cells belonging to the sea frontier of the model. For the transport boundary conditions a prescribed concentration of 35 g L-1 has been assigned to all cells belonging to the sea frontier of the model.

We assumed a density of freshwater of 1000 and 1025 kg m-3 for the seawater with 0.7143 as the slope of the linear equation of state that relates fluid density to solute concentration.

The groundwater model covers the Plio-Quaternary and Pre-Quaternary formations. The spatial heterogeneity is tackled using the concept of multiple statistical populations (Llopis-Albert and Capilla, 2010), in which the rock matrix and each fracture are represented as independent statistical populations. The random function for each structure (i.e. the aquifer matrix and fractures) is modelled based on a geostatistical analysis conditional on its own statistical distribution (i.e. hydraulic conductivity data as well as geological information and expert judgment). The random function is supposed to be multi-Gaussian for the rock matrix, while the fractures are considered non-multi-Gaussian. In this way, the rock matrix is generated by sequential Gaussian simulation using the GCOSIM3D code (Gómez-Hernández and Srivastava, 1990), while the fractures are generated by sequential indicator simulation using the ISIM3D code (Gómez-Hernández and Journel, 1993). The latter code makes use of local conditional cumulative density functions (ccdfs) defined by conductivity measurements and the corresponding indicator variograms. Therefore, the spatial heterogeneity is modelled as an equivalent porous medium (e.g. Llopis-Albert and Capilla, 2010). On the one hand, the hydraulic conductivity data for fractures present high values, greater than 1000 m day-1. This allows the reproduction of strings of extreme values of hydraulic conductivity that often take place in nature and can be crucial in order to obtain realistic and safe estimations of mass transport predictions. That is, it allows reproduction of preferential flow channels in strongly heterogeneous aquifers or fractured formations. On the other hand, the hydraulic conductivity data for the aquifer matrix cover a wide range of values, i.e. from 5 to 200 m day-1. In addition, for each cell we have defined a vertical hydraulic conductivity equal to a tenth of the horizontal hydraulic conductivity. The position of fractures is deterministically incorporated into the model based on geological information and expert judgment, thus allowing us to classify the cell models. Those cells of the model intersected by a fracture are assigned conductivities according to the intersecting fracture, and those that are not are considered to be cells belonging to the rock matrix.

The model was calibrated with the available historical data in the period (1981–2010), while the data in the period 1973–1981 were used to validate it. This was carried out by applying a trial and error procedure simultaneously considering quantity and quality. We have not used an inverse model (e.g. Llopis-Albert et al., 2016) because its computational cost would be important in order to deal with such a quantity of parameters (hydraulic conductivity, storativity, porosity, dispersion coefficients, etc.) and variables (hydraulic head and salinity) within a SEAWAT model over a long period of time (from 1973 to 2010), so that we have decided to apply a trial and error procedure.

The calibrated model parameters encompass different zones of horizontal hydraulic conductivities ranging from 5 to 200 m day-1, while the vertical hydraulic conductivities are between 0.5 and 20 m day-1. There are also different zones of specific storage values, which range from 10-5 to 5 × 10-4 1/m; specific yield, with values from 0.01 to 0.15; effective porosity values from 0.01 to 0.13; and dispersion coefficients from 50 to 100 m. We have replaced a highly heterogeneous porous medium with an upscaled “equivalent” homogeneous porous medium to represent the hydrogeological parameters since the cell size of the discretization is 250 m × 250 m (e.g. Llopis-Albert and Capilla, 2010). Then, we used a value of the effective porosity based on available data, which were subsequently calibrated and upscaled using expert judgment. The results of the calibration process prove the worth of this approach.

Reasonably good results in terms of goodness of fit to the historical hydraulic head and salinity were obtained, as shown in Fig. 11 for various head and salinity observation wells.

The root mean square value (rms) of the departures between observed and simulated values for both piezometric heads and salt concentrations is presented for the whole domain and temporal discretization due to the large number of boreholes (21 boreholes for piezometric heads and 31 for salinity). These values are ηh = 0.7 m and ηc = 391.8 mg L-1.

Note that this ηh = 0.7 m could seems to be a little high but we should take into account that we have observation wells where the historical hydraulic head measurement fluctuates sometimes more than 6 m during the same month (see for example observation well 6). Note that the high value for ηc = 391.8 mg L-1 could be also explained by the scale of the salt concentration, which range from 0 to 35 000 mg L-1, and the measurement fluctuations, which in some wells is even higher than 4000 mg L-1 during some months.

Finally, the calibrated model is used to propagate impacts of future LULC and CC scenarios.