The main objective of this paper is to test the suitability of several meteorological drought indices to estimate the impacts of climate change on the water cycle of a small catchment. Seven drought indices, described in Sect.  and in the Supplement, are investigated. The information on drought severity (as computed by these indices) is compared to three simulated hydrological impacts of drought: (1) the mean annual discharge at the outlet, (2) the mean annual hydraulic heads in 12 observation wells of the local aquifer, and (3) the water deficit (WD), which is a simplified representation of how well the water demand of the crops can be met : WD[%]=100×ETc-AETETc, where ETc is the annual crop evapotranspiration under standard conditions with no soil moisture limitation and AET is the simulated actual evapotranspiration, calculated on one daily timescale and aggregated for each year.

The time series of the drought impacts listed above are obtained using the outputs from a calibrated, integrated, pde-based, hydrological model (Sect. ) forced by present and future meteorological time series (Sect. ) and daily irrigation scenarios (Sect. ). Five climate scenarios (one based on present climate and four based on the projections of regional climate models) and three irrigation scenarios are constructed and combined with each other in our simulations. The length of the simulation is 180 years for each combination of (present and future) climate and irrigation scenarios. This is equivalent to a total of 2700 simulated years. From these 2700 simulated years, we extract time series of discharge, hydraulic heads, and water deficit.

These time series are directly used to represent the drought impacts on hydrology. They are compared to the time series of meteorological drought indices (Sect. ): we first compute the Pearson correlation coefficients between the drought indices and the hydrological variables. Next, we analyze changes in the (assumed) linear relationship between hydrological variables and drought indices. These comparisons are repeated in present and future climate for the different irrigation scenarios. A suitable drought index for climate-change studies would have a large correlation coefficient with all hydrological variables and the relationships between this index and the hydrological variables would be identical in present and future climate. The results and the interpretation of these quantitative studies are presented in Sects.  and .

This study is focused on annual droughts. We choose the annual timescale because it is often used when predicting future droughts e.g., and because it is the most dominant precipitation cycle worldwide . Even though seasonal and sub-annual timescales are essential for drought management e.g.,, we aim here to test the capabilities of drought indices to predict future hydrological impacts, not to produce direct predictions of future drought impacts. For our purpose, an annual timescale is sufficient and enables a detailed analysis of the differences between the correlation coefficients and the linear relationships, which are at the center of this study.

The Lerma catchment is situated within the Ebro basin in Spain with an altitude varying between 330 and 490 ma.s.l. and an area of ∼ 7.3 km2 (Fig. ). Its climate is classified as semi-arid, with a mean precipitation of ∼ 400 mmyear-1 (2004–2011) and a mean potential evapotranspiration rate of ∼ 1300 mmyear-1 (2004–2011) . Precipitation and temperature have been measured since 1988 at the meteorological station of Ejea de los Caballeros (∼ 5 km north of the study area). Solar irradiance, wind speed, and relative humidity have been measured since 2003. Annual precipitation is highly variable, ranging from 268 to 558 mm (2004–2011). Because of the limited water resources, drought is a serious natural hazard in the region .

The catchment underwent a rapid transition from non-irrigated to irrigated agriculture between 2006 and 2008. The majority of the fields within the catchment are now irrigated, with an annual irrigation of 286 mm in 2011 . This transition was closely monitored and crop types, monthly hydraulic head data, daily discharge, and irrigation volume are available. In addition, a vertical–electrical–sounding campaign was conducted to better understand the local geology. Two main hydrologically relevant layers were identified: the top layer is composed of clastic and unconsolidated Quaternary deposits and forms a shallow aquifer. Underneath lies an aquitard composed of lutite and marlstones (Fig. ). Soils are relatively shallow, with depths below ground surface ranging between 0.3 and 0.9 m , and are classified as inceptisols.

To simulate the hydrological response of the Lerma catchment, we use HydroGeoSphere , a three-dimensional, fully coupled, integrated hydrological model, based on partial differential equations. In HydroGeoSphere , water flow in the variably saturated subsurface is modeled using the three-dimensional Richards' equation, while overland flow is simulated by the diffusive-wave approximation of the Saint-Venant equations. We use the Mualem–van Genuchten parametrization to relate relative permeability and water saturation to capillary pressure in the vadose zone. The surface and subsurface domains are coupled using a dual-node approach, where the coupling between the domains is conceptualized as a virtual thin layer of porous material. Potential evapotranspiration is computed using the FAO Penman–Monteith equation , and time-varying crop coefficients are used to account for the spatial variability of crops (see the Supplement for more information). The model choice is based on the necessity of modeling the transition to irrigation, which has a large impact on the hydrology of the catchment. Moreover, HydroGeoSphere allows us to simultaneously study the impact of droughts on the surface and subsurface components of water flow. The underlying equations have been reviewed by von Gunten et al. (, ) and are not repeated here.

The conceptual model of our study area and its calibration have also been presented by and thus are only presented here briefly. We divide the subsurface catchment into six zones, two zones representing the aquitard, one representing the aquifer, and three representing the different soil zones (Fig. ). The model parameters are homogeneous in each zone and the saturated hydraulic conductivity is 1 order of magnitude smaller in the vertical direction than in the horizontal one to account for anisotropy. The surface domain is divided into 55 zones, representing the different farm fields. Daily irrigation volume, Manning's parameters, seasonal leaf area index, and rooting depth are specified separately for each surface zone, based on crop types and irrigation data. Precipitation is given as daily input, apart from days with intense rainfall (> 25 mmday-1). In this case, precipitation data are given as a 3 h mean during summer and spring, and as a 9 h mean during autumn and winter, to mimic intense convection events , which are frequent in the region. A no-flow boundary condition is assumed at the lateral and bottom boundaries of the subsurface domain. Critical flow depth is used for the lateral boundaries of the surface flow domain.

We calibrated the parameters of the model using three computational grids of increasing resolution . The calibrated parameters are the hydraulic conductivity in all zones, apart from the “weathered aquitard” zone (Fig. ), the porosity of the aquifer, and the van Genuchten parameters of the soil zones. The calibration period is from 2006 to 2009 and the validation period is from 2010 to 2011. The model is calibrated on the measured discharge at the outlet and on the hydraulic heads in 8 observation wells (12 observation wells were used during validation). The model reproduces the measurements satisfactorily . For example, the Nash–Sutcliffe efficiency of discharge is 0.74 during the calibration period and 0.92 during the validation period. The model performs similarly well under all irrigation conditions. Because the model was able to reproduce the response in both discharge and groundwater tables to the changes in irrigation practice, we are confident that it can also predict the response to changes in meteorological forcing projected by climate models.

More than 150 drought indices have been developed in the past and it would be unrealistic to include all of them in this study. Therefore, we have selected seven well-known and commonly used drought indices, based on the reviews by , , , , and . Our choice was guided by the required data input and the popularity of the indices in recent studies related to climate change. The selected indices are

the standardized precipitation index (SPI): SPI is a widely used drought index whose computation is based on fitting long-term precipitation data to a probability distribution. This probability distribution is then transformed into a normal distribution.

We present the selected indices in more detail in the Supplement and provide a summary in Table . We generally consider meteorological drought indices that aggregate data annually (Sect. ). The exceptions are the Palmer drought indices (PDSI and PHDI), whose time length depends on an empirical estimation of the start and the end of drought periods .

Potential evapotranspiration (ET0) is needed to compute SPEI, PDSI, PHDI, and RDI. To obtain this variable, we use the FAO Penman–Monteith equation , which is presented in the Supplement along with additional explanations on the calculation of ET0.

To compare how well the drought indices can predict the chosen hydrological variables in present and future climate, we use two approaches. First, we compute Pearson's linear correlation coefficient r, which quantifies how well the variability in one time series can be explained by the variability of another time series, assuming a linear relationship between the two variables. In the context of this study, it indicates whether the drought indices have the capability of finding periods with discharge or hydraulic heads lower than usual and periods with a water deficit higher than usual. It is defined as follows: r=cov(DI,x)σDIσx, in which cov is the covariance, σi is the standard deviation of the variable i, DI is the value of the drought index, and x is the hydrological variable under consideration.

Pearson's correlation coefficient indicates the degree of linear dependence between two variables. However, if this correlation coefficient is calculated under different climatic conditions, it does not indicate possible changes in the coefficients of the (assumed) linear dependencies. To investigate the changes in the linear dependency between the two climates, we perform a linear regression between a drought index and a hydrological variable in the present climate. Then, we use this linear relationship to predict the hydrological variables from the same drought index in future climate. We conduct this analysis for each combination of drought index and hydrological impact in all irrigation scenarios. By this, we aim to investigate whether drought indices in future climate represent on average a similar drought (i.e., a drought with similar hydrological impacts) than in present climate. This is important because many drought studies e.g., only report changes in drought indices, implicitly assuming identical drought impacts for identical drought-index values in present and future climate. However, a drought described by a SPI value of -1, for example, may have different consequences for discharge and water deficit in projected future climate than under current climate conditions (see Sect. ).

To quantify the changes in the linear dependencies between hydrological variables and drought indices, two performance metrics were selected: the relative model bias Brel and the normalized root mean square error (NRMSE). The relative model bias is the sum of the differences between the predicted and actual values of the hydrological variable, divided by its mean value. Brel=100%1n∑i=1nVmod,i∑i=1nVstat,i-Vmod,i, in which Vstat,i indicates the predicted value of discharge or water deficit based on the linear regression, Vmod,i represents the value of the same variable predicted by the hydrological model, and n is the length of the time series.

The NRMSE is the root mean square error divided by the standard deviation of the least-square regression in present climate σpres: NRMSE=1σpres∑i=1nVstat,i-Vmod,i2n.

In the present climate, the variability of the differences between the outputs from the hydrological model and the linear regression is smaller than 12 % of the average difference between model outputs and the linear regression. Hence, the error of the linear model in the present climate can be considered homoscedastic; i.e., σpres is considered constant in the subsequent analysis.

The climate scenarios used in this study have been presented by and are thus only summarized here.

Our future climate scenarios cover the time period of 2040–2050, using the A1B IPCC emission scenario . They are based on four regional climate models from the ENSEMBLES project driven by two global climate models (Table ). As it is not advisable to use the direct outputs from climate models as input for a small-scale hydrological model , we have downscaled the outputs from the climate models using a weather generator, i.e., a statistical model reproducing the characteristics of the observed climatic time series . We calibrated the weather generator using the observed time series of the closest meteorological station (Ejea de los Caballeros). Then, the parameters of the weather generator were modified using the differences between the control and future simulations of the regional climate models. These change factors, described in , are an indication of future changes of the mean and variability of precipitation, temperature, radiation, and relative humidity. The weather generator is run using the updated parameters to create the future climate scenarios. In this study, we use the RainSim weather generator for precipitation and the EARWIG weather generator for ET0 .

The chosen downscaling procedure has the advantage of producing longer time series, compared to the relatively short (23-year) climate record in the Lerma catchment. Moreover, it reproduces future changes in the precipitation variability, and not only in the precipitation mean, which is an important criterion when studying future droughts.

Nevertheless, the downscaling of climate model outputs is a complex task and the choice of a particular downscaling method can have a large impact on the results . Our study is not an exception and the downscaling process presented here might introduce uncertainties in the climate scenarios. We have mitigated this issue using three different approaches: (a) we prepared both present and future time series of meteorological inputs using the weather generator. Hence, the potential bias resulting from the weather generator is reproduced in the present and future time series. (b) We compared the future time series of precipitation and ET0 downscaled with the weather generator with the corresponding time series downscaled with a simpler bias correction method . The time series were found to be generally similar regardless of the downscaling method . (c) The time series of present precipitation and ET0 have been extensively tested against measurements to control the quality of the weather generator outputs .

In addition to the reproduction of the meteorological forcing mentioned in Sect. , the weather generator should also reproduce the frequency distribution of the studied drought indices. Here, we compare these frequency distributions in the observed climate record with the corresponding frequency distribution computed from the weather generator outputs in the current climate.

All seven drought indices used in our study are normalized (Sect. ) so that they can be used in different regions. If the normalization would have been carried out separately in the observed and simulated data, the frequency distributions of the drought indices would be similar, regardless of the similarity of the time series. To provide a meaningful comparison, we compute the normalization on the simulated data (weather generator) and we use the same normalization for the observed data (current climate record).

To compute each drought index, we use the measured time series, which has a length of 23 years (1988–2011). In addition, we compute the drought indices using the simulated data. To get a comparable length between measured and modeled data, the time series of drought indices based on the weather generator are separated into 15 periods with a duration of 23 years each (totaling 354 years). The final length of this time series is chosen such that it is about twice the length of the hydrological simulations (180 years). We then prepare 15 empirical cumulative distribution functions (ecdf) based on the outputs of the weather generator and compare them with the ecdf based on the current observed climate record (Fig. ).

The ecdf of all drought indices based on measurements fall into the region defined by the 15 modeled ecdf. Hence, differences between the observed and simulated data were small compared to the difference between the 15 modeled ecdf. In addition, we used a two-sided Kolmogorov–Smirnov test to compare the time series based on modeled and measured data. This test e.g., is a non-parametric statistical test that quantifies the maximum distance in cumulative probability between two distributions and tests how likely it is that the two samples are drawn from the same distribution. All drought indices pass this test; i.e., the null hypothesis of identical ecdf between measured and simulated data is not rejected at a 5 % significance level. Therefore, the drought indices based on the time series of the weather generator outputs show a reasonable agreement with the observed time series to be used in present climate. Weather generators are commonly operated to produce time series of future hydro-meteorological variables e.g.,, and we are also confident of using the weather generator to produce future time series of drought indices.

Consistent with our earlier study , we use three irrigation (or land-use) scenarios that can be summarized as follows:

scenario NOIRR: without irrigation and without agriculture;

Future precipitation (Fig. ) is predicted to decrease in summer and spring (between 3 and 39 % of the current precipitation, depending on the regional climate model). By contrast, in winter and autumn, an increase in precipitation is predicted (between 1 and 55 %). Change in total annual precipitation depends on the regional climate model. MPI and UCLM predict a wetter future, while ETHZ and METO predict a dryer one (see Table  for the references of the regional climate models). The coefficients of variation increase in spring (between +3 and +6 %), decrease in winter and autumn (between -0.1 and -10 %), and do not show a clear trend in summer (between +5 and -5 %).

Because of the higher temperature, potential evapotranspiration (ET0) increases (between 9 and 22 % in the annual average) in all regional climate models for all months. This increase might impact droughts, regardless of the precipitation changes.

The hydrological responses of the Lerma catchment to climate change under different irrigation conditions have been modeled previously by . As this study extends these results, we will shortly recall them here. Overall, the catchment responses to climatic change strongly depend on the irrigation scenarios and on the considered regional climate model. For all considered climate scenarios, the increase in temperature and the decrease in summer precipitation result in a lower groundwater table and in a decrease in low-flow discharge (defined as the total discharge during dry periods). This decrease is more intense in scenarios with irrigation than in the scenario without irrigation. Peak discharge decreases if irrigation is present. However, it often increases in scenarios without irrigation, notably because the lack of vegetation results in lower infiltration and higher surface runoff during thunderstorms. Spring and summer actual evapotranspiration increases if the catchment is irrigated because of the increase in ET0 and the relatively large soil moisture. Without irrigation, changes in annual actual evapotranspiration depend on the annual precipitation. In climate scenarios where precipitation decreases, actual evapotranspiration decreases because of the lower water availability. By contrast, if annual precipitation increases, actual evapotranspiration also increases. More details on the modeling of hydrological impacts of climate change are available in .

In this section, we analyze the correlation between the different drought indices for the 180 years of each scenario and the corresponding simulated mean annual discharge, water deficit, and hydraulic heads. For this purpose, we use the Pearson linear correlation coefficient r between the drought indices and the hydrological variables (Sect. ). We conduct the same analysis for present and future climate, and for the different irrigation scenarios. Here, we present only the main results of this comparison (details are available in the Supplement).

The values of the correlation coefficients between the hydrological variables and the drought indices depend on the drought indices. For example, the correlation coefficient between water deficit and EDI is 0.47, while the correlation coefficient between this variable and RAI is 0.78 in the present climate. However, the correlation coefficients for a particular drought index and a particular hydrological variable are similar for all irrigation scenarios in present and future climate. For example, let us consider the correlation coefficients between drought indices and discharge (Fig. ). In present climate, SPEI, RDI, and RAI have the highest correlation with discharge in the PIRR scenario (0.77<r<0.80) as well as in the NOIRR scenario (0.81<r<0.83). These indices also have similar correlation coefficients in future climate (0.79<r<0.84). If we consider the correlation of a particular drought index with discharge over all climate/irrigation scenarios, the difference in r is < 0.1.

Water deficit exhibits a similar behavior to discharge when correlation coefficients are examined. When the absolute values of correlation coefficients are large in present climate, they will be similarly large in future climate or in another irrigation scenario. SPEI, RDI, and RAI have the largest correlation coefficients with water deficit in all scenarios (0.78<|r|<0.81).

Correlation coefficients between drought indices and groundwater heads in a particular observation well are similar for all drought indices considered. However, the correlation coefficients are very different from one observation well to another (see the Supplement for more information).

Seasonal differences in the correlation coefficients are not considered here, even though these correlations might be influenced by the annual cycle. Our analysis is focused on annual droughts.

The previous section has shown that the linear correlations between drought indices and hydrological variables are relatively similar under all climatic and irrigation conditions. Hence, a particular drought index is able to identify the dry periods in present and future climate. However, this does not indicate whether the droughts in future climate have similar hydrological impacts to those in present climate. Correlation coefficients quantify how well a relationship between two variables can be expressed by an (assumed) linear equation, without considering the actual coefficients of the linear equation. The latter are commonly evaluated by linear regression.

Identifying changes in the regression coefficients of the relationships between drought indices and hydrological variables is important when making hydrological predictions based on meteorological drought indices in a changing climate. Only when the regression coefficients do not change does the same value of a drought index have the same hydrological impact. To this end, we compare changes in the (assumed) linear regressions between drought indices and discharge or water deficit (Sect. ). In the subsequent analysis, we do not consider hydraulic heads because the results almost entirely depend on the position of the observation well.

The stability of the relationship between drought indices and hydrological variables strongly depends on the chosen drought index and the irrigation scenario. In Fig. , we exemplify the relationship between SPEI and discharge for two irrigation scenarios in present and future climate. In the lower panel of Fig.  (scenario FUTIRR), the relationship between SPEI and discharge is relatively stable in different climates. A drought with a similar intensity (as defined by SPEI) has similar impacts on discharge in present and future climate. In the top panel, the bias is larger. In this case, a drought with a particular SPEI value results in a different annual mean discharge in present and future climate.

As outlined above, we use two different performance metrics to quantify this bias, the relative model bias Brel and the NRMSE (Sect. ). Figure  shows these two metrics for all indices and the two hydrological variables. Overall, our results suggest that the relationships between the chosen meteorological drought indices and hydrological variables are not stable under a changing climate. The computed model biases between drought indices in present and future climate appear important. In the scenario without irrigation, the largest relative model bias is 86.7 % for discharge and 3.8 % for the water deficit (mean discharge in present climate: 0.015 m3s-1; mean annual water deficit: 80 %). With irrigation, the largest relative bias for discharge is -25.2 % for the RAI drought index and 14.2 % for water deficit (mean discharge: 0.03 m3s-1; mean annual water deficit for irrigated and non-irrigated zones: 52 %). In the worst case described above (discharge without irrigation), the relative model bias is on the same order of magnitude as the value of the hydrological variable, which is a significant difference. For certain conditions, however, the bias is low. For example, water deficit in the scenario without irrigation is predicted well by the linear model (the largest bias is equivalent to only 3.8 % of the present water deficit).

For discharge, model bias depends strongly on the irrigation scenario (Fig. , top panels). With irrigation, the drought indices often underestimate the changes in discharge, especially if the indices are based on precipitation only. For example, in the case of SPI, the model bias for discharge is -24.8 % with irrigation (and 6.8 % without irrigation). By contrast, drought indices that are based on ET0 and precipitation have a lower bias in the scenario with irrigation than in the scenario without irrigation. For example, SPEI has a model bias of 86.7 % with irrigation and of 11 % without irrigation. In the Lerma catchment, discharge is more sensitive to climate change when irrigation is present . Hence, drought indices that are more sensitive to climate change, notably to changes in ET0, predict changes in discharge better in irrigated cases. The discharge in the scenario without irrigation does not change significantly, and drought indices with a smaller reaction to climate change are better predictors of hydrological impacts than those with a stronger reaction (Fig. , top panels).

For the water deficit (Fig. , bottom panels), drought indices that include ET0 have a lower model bias than indices that only include precipitation. In the case of SPI with irrigation, the relative model bias is 13.9 %. In the case of RDI, which includes ET0, the model bias is 5.4 %. The lower bias for drought indices containing ET0 can be explained because ET0 is directly influencing the water-deficit calculation. The relative model bias is lower in the scenario without irrigation than in the scenario with irrigation. Indeed, irrigation is not accounted for in the calculation of the drought indices, but it influences the modeled water deficit.

The drought indices with the lowest model bias and a correlation coefficient r>0.6 are RAI for discharge in the NOIRR scenario, RDI for the water deficit in the FUTIRR/PIRR scenario, and SPEI for the water deficit in the NOIRR scenario and discharge in the FUTIRR/PIRR scenario.

In Sects.  and , we explored the relationships between the different drought indices and the selected hydrological variables in present and future climate. In the present section, we compare the drought indices in present climate to those in future climate. This is a step forward compared to previous studies because we use the information of Sects.  and  to improve the predictions of future droughts, notably to interpret differences between the predictions based on different drought indices.

Our definition of a drought is identical for present and future climate. Practically, we standardize the drought indices in the present climate and keep the same standardization (explained in Sect.  and in the Supplement) in the future climate. From a conceptual point of view, this is unexpected, as meteorological droughts can be defined as a period of exceptionally dry conditions. If the average precipitation changes, the definition of a meteorological drought should also be changed. However, from a practical point of view, drought severity depends on the water needs and on the vulnerabilities of society and agriculture. Hence, the definition of future droughts is linked to current conditions. From this perspective, using the same standardization in present and future climate is logical. Moreover, this procedure has been applied in the majority of studies on future droughts e.g.,.

Figure  shows the changes between present and future climates in the seven drought indices based on the outputs of the four regional climate models. Note that a decrease in the values of the drought indices indicates an increase in drought intensity.

When we compare the changes in drought indices between present and future climate, significant differences can be observed between the different climate scenarios (based on the four regional climate models). Indices that only contain precipitation (RAI, SPI, and EDI) predict a small increase in droughts or a small decrease depending on the climate scenario (Fig. , top panels). For example, the average SPI decreases by 0.4 when using the ETHZ climate scenario and increases by 0.2 when using the MPI scenario (for comparison, an SPI of -3 would be an extreme drought). In these scenarios, the MPI and UCLM regional climate models predict an increase in annual precipitation for the Lerma catchment . Hence, the climate scenarios based on these regional climate models result in a decrease in drought events (i.e., an increase in the drought index value) when indices are only based on precipitation (RAI, SPI, and EDI). Indices that also consider ET0 (Fig. , bottom panels) indicate an increase in droughts in all analyzed future climates. However, this increase is smaller when MPI and UCLM are used to construct the climate scenario. In the UCLM case, a decrease of 1.59 in the mean value of SPEI is computed. In contrast, when the ETHZ climate model is used, a decrease of 2.95 is computed (Fig. , bottom panel). Differences in the values of drought indices that include evapotranspiration between present and future climate follow predicted changes in ET0. Models that predict a strong increase in ET0, such as ETHZ, result in a stronger increase in drought risks. A change in the coefficient of variation of ET0 or annual precipitation is not directly related to changes in drought indices.

The sources of the differences between the climate scenarios, which result in the aforementioned differences in the values of drought indices, are uncertain. Nevertheless, two factors are often cited when discussing differences in future climate scenarios with identical emission scenarios: modeling of cloud cover and parameterization of the interactions between the land cover and the atmosphere . Both processes have a large influence on precipitation and evapotranspiration, and therefore on drought predictions.

In addition to the differences related to the chosen climate scenario, the choice of the drought index has a large influence on the prediction of future droughts. These differences in drought prediction are largely the reflection of the differences in the linear relationships between drought indices and hydrological variables discussed in Sect. . If a drought index has a negative bias for discharge (as is the case for indices that are based on precipitation only), small changes in future droughts are predicted. For example, when we average the four different climate scenarios, mean RAI in future climate shows a decrease of 0.02 when compared to RAI in present climate (Fig. , top panel, left column). Based on the linear model under present irrigation conditions, this can be translated into an increase in water deficit of 0.21 mmyear-1 and a decrease in discharge of 8.7×10-5 m3s-1. These changes are unlikely to have consequential impacts on irrigation or on the hydraulic regime of the catchment. For the indices that depend on ET0, the predicted increase in droughts becomes larger. For example, mean SPEI shows a decrease of 2.43 (average of four regional climate models). If we would use the linear model developed in present climate, the decrease in discharge in the scenario with irrigation would be 0.01 m3s-1, which is one-third of the annual mean discharge. Based on the hydrological model, the change in discharge in the FUTIRR scenario is 0.006 m3s-1 (average of the four climate models). Large uncertainties linked with climate prediction and hydrological modeling still prevail in this estimation. However, the hydrological model generally reproduces discharge and hydraulic head measurements. Moreover, it simulates many relevant processes leading to discharge generation. Hence, we assess this model to be more reliable in predicting hydrological effects of climate change than a mere comparison of meteorological drought-index values.

If we analyze the hydrological impacts of meteorological droughts (defined here as periods with an SPI and SPEI value of lower than 1), the general behavior is similar in present and future climate (Fig. ). As expected, during droughts, precipitation and discharge decrease, and actual evapotranspiration increases. In present climate, in the scenario without irrigation, discharge decreases by more than 60 % during dry periods when compared to the average conditions. In the scenario with irrigation, the decrease in discharge is less marked (24 % difference between dry and average conditions) as the irrigation water partly compensates for the lack of precipitation. By contrast, impacts of droughts on actual evapotranspiration are stronger in the scenario with irrigation than in the scenario without irrigation. In the latter case, soil moisture is simply too low to support actual evapotranspiration, regardless of the evaporative demand . In future climate, the decrease in precipitation and the increase in ET0 during droughts are more intense than in present climate (Fig. ). Hence, we could expect more intense droughts with larger hydrological impacts. If the catchment is irrigated, modeled hydrological impacts are indeed more intense, with a stronger decrease in discharge, a higher increase in actual evapotranspiration, and an additional decrease in the level of the water table, at least in the case of the observation wells under the irrigated zone. Observation wells that are away from the intensely irrigated fields, such as Po8, exhibit a more complicated behavior. However, if the catchment is not irrigated, certain hydrological impacts are less intense. For example, discharge and the distance to the water table decrease less during droughts in future climate than in the present one. A possible explanation for this behavior is linked to evaporation. In the non-irrigated case, the increase in ET0 during droughts is not transferred to an increase in actual evapotranspiration because of the dry average conditions. Consequently, the higher ET0 during drought in future climate has a low impact on the hydrology. Hence, impacts of climate change are lower under very dry conditions. This is probably also why drought indices that include ET0 are better at predicting discharge when irrigation is present, while the quality of their prediction is lower when the catchment is not irrigated: the presence of irrigation increases water availability, which increases the importance of ET0 in the hydrological impacts of droughts, notably a decrease in discharge.