CWatM (Community Water Model) is a distributed hydrological model that can be implemented at regional to global scales (Burek et al., 2020). The model is developed with the Python programming language and is open-access (https://cwatm.iiasa.ac.at/, last access: 13 June 2022). CWatM aims to reproduce the main hydrological processes, including water management. In this study, CWatM is applied in two regions and coupled with a finer-resolution groundwater flow model, MODFLOW (Harbaugh, 2005). CWatM is used at a spatial resolution of 1 km in Seewinkel and ∼ 0.9 km (30 arcsec) in Bhima. Because the Bhima basin covers a large area, CWatM is applied in a geographic projection system (WGS84), while square cells are used for Seewinkel (Lambert azimuthal equal area projection system ETRS89). MODFLOW is applied at 100 m in Seewinkel and 250 m in Bhima using square cells in both regions.

The following sections more extensively describe how groundwater processes and pumping are modeled. The coupled CWatM–MODFLOW uses a computationally efficient approach combining MODFLOW 6's Basic Model Interface (BMI) (Hughes et al., 2017) and Python packages Flopy (Bakker et al., 2016) and xmipy (Russcher et al., 2020). The coupling of CWatM and groundwater is illustrated in Fig. 1. In particular, it is necessary to use high-resolution groundwater models, as topography controls the variability in the water table. For example, Fig. 1 contains an illustration of 1 CWatM cell containing 100 MODFLOW cells (10 shown within the cross-section), and it highlights the fact that different land cover types are considered in each CWatM cell. Soil processes that occur within each land cover part are modeled independently before being averaged according to their respective coverage area within the CWatM cell. This approach allows consideration of the subgrid variability without reducing resolution.

The Seewinkel region is a subpart of the Burgenland region in Austria (Karner et al., 2019; Hatvani et al., 2014; Mitter and Schmid, 2021). The study area, of around 573 km2, is limited in the west by Lake Neusiedl (Neusiedlersee in German) and in the south and in the east by the border between Austria and Hungary. The Seewinkel region has been classified as semi-arid (Karner et al., 2019; Magyar et al., 2021), even if the aridity index, defined as the ratio between mean annual precipitation and potential evapotranspiration, reaches 1.0. Evapotranspiration shows a strong seasonal pattern, while precipitation is relatively homogeneous throughout the year (Table 1). The region is an important agricultural area, and around 16 % of it is allocated to irrigated crops (maize, potatoes, vineyards, vegetables, and wheat). Only groundwater is used to satisfy irrigation demands. The study area hosts several natural ponds of environmental interest. These ponds mainly depend on the water table in the aquifer. To keep the ponds safe, pumping from irrigation wells is limited when the water table reaches a threshold depth (Magyar et al., 2021). The Seewinkel region and associated groundwater are modestly impacted by human water management.

The CWatM–MODFLOW model (shown in black in Fig. 3a) is applied at 1 km resolution, associated with a MODFLOW resolution of 100 m from a 10 m resolution digital elevation model (DEM; Comprehensive digital field model (DGM) of the state of Burgenland, 2019). To capture the influence of regional groundwater flow, the modeled area is larger than the study area (marked in red in Fig. 3a). Therefore, small ponds are included in the study area, but Neusiedlersee is not. Outside the study area, the percentage of river within each MODFLOW cell is defined from a river network computed with the finer 10 m resolution DEM assuming that rivers are created when the drainage area exceeds 1 km2. There is no river inside the study area, but there are several canals draining groundwater. Note that the MODFLOW DRAIN boundary condition on these canals is set at 1 m below the altitude inferred from the 100 m DEM. Groundwater leaving the aquifer below canals and lakes directly reaches the surface water network (see the baseflow in Fig. 1). Using field data, the Seewinkel aquifer is modeled as one aquifer layer with a thickness of 20 m. Regional characteristics and model properties are synthesized in Table 1.

One pumping well is set up every 1 km2 at the center of each CWatM cell. Irrigation demand and associated pumping rates depend on the deficit between potential evapotranspiration demand and current soil water contents within the CWatM cell (see the Supplement about the influence of irrigation efficiency and spatial density of pumping wells). Finally, the simulated CWatM–MODFLOW pumping rate at the Seewinkel scale for 2015 is 48 % smaller than the abstraction limits imposed for the year 2015 in Seewinkel. Therefore, the model seems to appropriately represent groundwater use.

The Bhima basin is located in India and hosts 19 million people. The basin area is 46 000 km2. The average aridity index is 0.75 (Table 1); however, precipitation shows strong seasonality due to the monsoon. Both surface water and groundwater satisfy domestic, industrial, and agricultural water demand. Irrigation is by far the most important groundwater use. Around 30 % of the basin area is allocated to irrigated crops. The river discharge recorded at the most downstream gauging stations (Takli) indicates a flow of 120 mm yr-1 on average, but this flows for only 2 months on average.

CWatM is applied at 30 arcsec resolution (∼ 900 m), associated with a MODFLOW resolution of 250 m from a 90 m resolution DEM (Yamazaki et al., 2019). Lakes and reservoirs are marked using HydroLAKES (Messager et al., 2016) and are refined and expanded upon with local data made available by the Pune Irrigation Circle. The percentage of river within each MODFLOW cell is defined from a river network computed with a 90 m resolution DEM, assuming that rivers are created when the draining area exceeds 1 km2. In Seewinkel, the aquifer upper limit under the rivers is set at 1 m below the altitude inferred from the 250 m DEM. This correction allows us to better take into account the river incision, which is strongly smoothed when resolution is upscaled to 100 or 250 m. When the water table reaches the top of the aquifer, groundwater is sent either to the soil or to the surface water network (see the groundwater feeding soils and baseflow indicated in Fig. 1) based on the river percentage map. In Bhima, canal networks are added to bring water from reservoirs toward cultivated areas where water is required for irrigation (see Sect. 2.2.3). Leakage is considered to occur below these canals through a water conveyance efficiency of 70 %.

The Bhima basin aquifer was modeled with MODFLOW in Surinaidu et al. (2013) at 1 km resolution, assuming spatially uniform recharge and pumping, deriving each as a linear relationship with precipitation, incorporating different data sources. The ranges for hydraulic conductivity and specific yield used for calibration were also used for this study, and we refer the reader to Surinaidu et al. (2013) for a more detailed description of the region. The model presented here expands on this by including surface water hydrology and management, spatially distributed groundwater recharge and outflow (pumping, capillary rise, and baseflow), and spatially distributed non-irrigation demand, as well as by modeling crop-specific irrigation demand and increasing the spatial resolution. While the Seewinkel alluvial aquifer was identified, the Bhima basin is much larger and is not composed of one homogeneous aquifer alone. The area hosts crystalline rocks, where weathering may potentially enhance permeability and porosity. We considered one homogenous aquifer layer with a thickness of 50 m (Surinaidu et al., 2013). We set up one pumping well in every MODFLOW cell (see the Supplement regarding the influence of irrigation efficiency and the spatial density of pumping wells). As for the Seewinkel model, pumping rates depend on the deficit between potential evapotranspiration demand and soil water content in each CWatM cell. Finally, the imposed pumping rate depends on the fraction of irrigated areas above each pumping well. Pumping is prevented when the water table falls below a depth of 15 m (Surinaidu et al., 2013).

Table 1 provides mean simulated water withdrawals. Annual surface water and groundwater withdrawals within the Bhima basin are estimated for different sectors for 2013 in the upper Bhima subbasin draft report. The model simulates these withdrawals closely (5 % higher), with higher groundwater use and lower surface water use. Through discussions with local water managers and engineers, including through stakeholder workshops (Karutz et al., 2022), it was found that groundwater use is generally underreported and underestimated for both irrigation and non-irrigation purposes. Further, canal leakage is not included in the report, producing an underestimation of groundwater availability. In agreement with the uncertainty of the reported values and following the expert opinions mentioned above, this model appropriately represents the region's annual surface water and groundwater use.

Water table data from 81 and 373 monitoring boreholes were gathered for the Seewinkel and Bhima areas, respectively. These observed datasets come from the eHYD (https://ehyd.gv.at/, last access: 3 January 2021​​​​​​​) database in Austria (BMLRT, 2020) and the Central Groundwater Board in India. They have been preprocessed to make them compatible with the simulated water table. First, boreholes close to the model's boundary limits were removed, given that the simplified no-flow boundary condition in the groundwater model could lead to an unrealistic water table in these parts. Second, boreholes were removed where time data covered less than 50 % of the simulation period. Third, we did not consider 5 % of the remaining boreholes with the greatest discrepancy between simulation and observation because they are impacted by specific local conditions that are not well reproduced by the model. This prevents them from influencing the calibration. Indeed, because of model assumptions such as homogeneous aquifers or pumping well locations, it can be inferred that some boreholes could not be well represented by the model due to the vicinity of pumping wells, rivers, or local heterogeneities. Finally, we kept 62 and 351 boreholes for the Seewinkel and Bhima areas (Figs. 4a and 5a), respectively. In the Bhima basin, we also used daily discharge data aggregated at a weekly scale at five gauging stations from the National Hydrology Project in India.

To fully benefit from the informative content of the observed water table, we separately evaluated time-averaged water tables (static part) and water table time fluctuations (transient part). This was done to better understand how each parameter is sensitive to both static and transient parts of observed data. As described above, the calibration relies on 62 and 351 boreholes for Seewinkel and Bhima, respectively.

First, we assessed the model's ability to reproduce the mean spatial variability of the water table driving the lateral groundwater flow. Observed and simulated water tables were averaged along the calibration period at each borehole. However, a comparison between observed and simulated water tables is not relevant and not sensitive to parameters, as the water table mimics the surface elevation, which extends to several orders of magnitude, as noted by Gleeson et al. (2021) and Reinecke et al. (2020). Indeed, the observed water table varies from 430 to 1000 m in Bhima, while observed water table depth varies from 1 to 20 m. Moreover, model topography can differ from actual topography. On the other hand, water table depths contain more discriminant information and are essential for the interaction between groundwater, soil, and surface. Therefore, we compared the water table depth (WTD) instead of the water tables. As a criterion, we used the “normalized mean water table depth difference” [%], defined by the following equation (Eq. 2): 2Cmean=100×1n×∑i=1n|WTD‾obs,i-WTD‾sim,iWTD‾obs,i|, where n is the number of monitoring boreholes, obs and sim refer to observed and simulated depths, respectively, and WTD‾ refers to the time-averaged WTD. Note that WTD‾obs,i ranges from 0.25 to 12 m and from 1 to 20 m for Seewinkel and Bhima, respectively. Thus, shallow water table depths do not have too much weight on Cmean.

The second criterion focuses on time fluctuations in water tables observed in boreholes, whatever the mismatch value between observed and simulated time-averaged water table depth. Water table fluctuations correspond to the initial signal after the time average is removed. To simplify, we averaged observed and simulated water table fluctuations from all boreholes instead of comparing the fluctuations at each borehole. This approach smooths local behaviors and informs the water table fluctuations (WTFs) at the basin scale at the first order. As a criterion, we used the normalized root mean square error [%] on average time fluctuations, defined by (Eq. 3): 3nRMSE=100SD(WTFobs)1nobs×∑t=1nobs(WTFobs,t-WTFsim,t)2, where nobs refers to the number of observations in time (t). Thus, the nRMSE corresponds to the root mean square error (RMSE) expressed as a percentage of the standard deviation (SD) of the observed data.

CWatM–MODFLOW simulation starts in 1981 and ends in 2016 for Seewinkel, while it begins in 1993 and ends in 2009 for Bhima. For both regions, models are run the first time during the whole period to initialize the water table. For Bhima, hydraulic conductivity and specific yield variables were included first in a calibration (Fortin et al., 2012) using five daily discharge stations (2000–2009), and ranges for calibration were derived from Surinaidu et al. (2013). The top 20 % from the discharge calibration was further analyzed for water table fluctuation and depths. Groundwater parameters hydraulic conductivity and porosity are further calibrated by comparing the simulated and observed water table recorded monthly in boreholes from 1983 to 2016 and twice a year from 1997 to 2009, respectively, for Seewinkel and Bhima. Standard CWatM parameters were used in Seewinkel. Therefore, only groundwater parameters are calibrated in this study.

Optimal aquifer permeability and porosity (Table 1) are compared to different datasets as described below. The permeability and porosity in Seewinkel are 5 × 10-5 m s-1 and 0.07, respectively. These values are respectively 3.10-4 m s-1 and 0.19 in the global GLHYMPS database (Gleeson et al., 2014; Huscroft et al., 2018). The imposed aquifer thickness (20 m) agrees with the global depth-to-bedrock map (Shangguan et al., 2016), where the average value for the study area is around 22 m.

In Bhima, the permeability and porosity are 1.2 × 10-5 m s-1 and 0.018, respectively. These values are respectively 3.16 × 10-6 m s-1 and 0.09 in the global GLHYMPS database. The global depth-to-bedrock values consider thinner aquifers in this region (from zero to a few meters) compared to the imposed aquifer thickness (50 m). Note that for similar permeability (1 × 10-5 m s-1) and thickness (50 m), Surinaidu et al. (2013) calibrated a porosity of 0.01–0.03 in a regional model. This suggests that permeability is underestimated and porosity overestimated in GLHYMPS over this basin.

The model reproduces the time-averaged water table depths recorded in boreholes in Seewinkel (Fig. 4b), with a Cmean value of 38 %. Therefore, we can infer that the hydraulic gradient and water table depth are reasonably well reproduced. The results for the Bhima basin show more contrast (Fig. 5b), as the Cmean is 69 % and simulated water table depth is too shallow on average. Indeed, the simulated mean water table averaged over all monitoring wells is 1.7 m higher than the observed mean in Bhima (Fig. 6b), while the model shows a slight bias of 0.5 m in Seewinkel (Fig. 6a). The reason for this bias is examined in the Discussion section.

In Seewinkel, the observed mean water table lies above a 1 m (2 m) depth for 17 % (47 %) of monitoring wells, while this fraction is 29 % (56 %) in the model (Fig. 4b). Thus, the model overestimates very shallow conditions (<1 m depth). The total land fraction where the simulated mean water table is shallower than a 1 m (2 m) depth is 25 % (46 %). Thus, it can be inferred that a non-negligible land fraction hosts shallow groundwater in Seewinkel.

In Bhima, the observed mean water table lies above a 1 m (2 m) depth for 0 % (2 %) of monitoring wells, while this fraction is 17 % (32 %) in the model (Fig. 5b). Thus, we conclude that the model overestimates shallow conditions in Bhima. Moreover, the total land fraction where the simulated mean water table is shallower than a 1 m (2 m) depth is 6 % (10 %), indicating that monitoring wells are preferentially located in shallow areas.

Both models successfully reproduce water table fluctuations averaged across all monitoring boreholes (Figs. 4c and 5c). The nRMSE reaches 52 % and 41 % for the Seewinkel and Bhima models, respectively. While the primary seasonal and interannual behavior is well reproduced in both areas, some events or periods are not well captured. For example, the Seewinkel model underestimates simulated fluctuations from 1986 to 1988. In addition, fluctuations are overestimated from 2000 to 2005, but the seasonal amplitude is slightly underestimated. In the Bhima basin, it also appears difficult to reproduce several interannual fluctuations and seasonal behavior at the same time. For example, changing the hydraulic parameters to increase seasonal amplitudes from 2003 to 2005 would decrease the base level even more during this period (Fig. 5c). Likewise, increasing or decreasing pumping rates would improve the mean water table bias but would increase the nRMSE focused on water table fluctuations (Supplement).

A comparison between observed and simulated weekly discharge in the Bhima basin is provided in the Supplement. The Kling–Gupta efficiency (KGE) (Gupta et al., 2009) values are 0.75, 0.68, 0.54, 0.76, and 0.66 from upstream to downstream stations. These results and the associated good criterion on water table fluctuations (nRMSE) make us confident that the model simulates rainfall partitioning between discharges and evapotranspiration wells.

The impact of pumping on the water table is a function of mean pumping rate and aquifer properties (permeability, saturated thickness, and porosity). Groundwater pumping theoretically decreases the water table within and around irrigated areas. The results show that groundwater pumping enhances the seasonal water table fluctuations by intensifying water table recession during the dry season in both regions (Fig. 6). It is also shown in Fig. 6 that irrigation accentuates water table depletion during the driest years. Interestingly, during the recharge seasons, the water table within the borehole network is not strongly influenced by pumping in the two study regions, as the water tables with and without pumping are very close (Fig. 6). Because the water table is a proxy for groundwater storage, this result indicates that groundwater storage recovers rapidly and/or irrigation enhances net groundwater inputs. However, this is not the case for the driest periods (i.e., 2004–2005 in Seewinkel and 2003–2005 in Bhima).

Comparing models with and without irrigation allows us to infer the impact of pumping on the water table. However, groundwater recharge is impacted by irrigation, as we elaborate in the next section. Pumping reduces the average water table at boreholes from 0 during the wet season to around 30 cm and 1–2 m during the growing season in Seewinkel and Bhima, respectively (Fig. 6). The impact of pumping also varies spatially, depending on whether the monitoring borehole is located near the irrigation areas where pumping occurs. Drawdown due to pumping obtained on the time-averaged water table in monitoring boreholes varies between 0 and 50 cm and between zero and a few meters for Seewinkel and Bhima, respectively.

The impact of pumping on the water table is more pronounced in Bhima than in Seewinkel (Fig. 6). This can be explained in two ways. First, aquifer permeability, thickness, and porosity are higher in Seewinkel (5 × 10-5 m s-1, 20 m, and 0.07, respectively) than in Bhima (1.2 × 10-5 m s-1, 50 m, and 0.018). Therefore, water table drawdowns around pumping wells are less pronounced in Seewinkel. Second, the mean pumping rate is smaller in Seewinkel (31 mm yr-1) than in Bhima (107 mm yr-1).

As expected, irrigation increases the total actual evapotranspiration in the two study regions. In Seewinkel, evapotranspiration in irrigated areas increases from 77 to 87 mm yr-1. In Bhima, evapotranspiration in irrigated areas increases significantly from 102 to 209 mm yr-1. However, these changes in evapotranspiration are substantially smaller than the allocated irrigation withdrawals due to conveyance and application losses and the fact that some of the irrigated water reaching the soil percolates toward the groundwater without being used by crops. Indeed, groundwater pumping amounts to an annual average of 31 mm yr-1 in Seewinkel, while annual groundwater and surface water withdrawal amounts to 107 and 115 mm yr-1 in Bhima, respectively. The increase in evapotranspiration is associated with a decrease in river discharge. In Seewinkel, irrigation reduces streamflow from 158 to 152 mm yr-1. However, pumping increases the net lateral groundwater inflow to the study area by around 2 mm yr-1. In Bhima, irrigation reduces streamflow from 257 to 167 mm yr-1. In all, 3 and 14 mm yr-1 from groundwater pumping reach the river network due to irrigation losses in Seewinkel and Bhima, respectively.

Irrigation also significantly increases groundwater recharge in both areas. Recharge increases by 24 % and 50 % in Seewinkel and Bhima, respectively. This is counterbalanced by groundwater pumping such that irrigation causes a reduction of net groundwater input of 12 and 5 mm yr-1 in Seewinkel and Bhima, respectively. However, the net recharge reduction is more pronounced without irrigation from the surface water bodies in Bhima. Indeed, around 40 mm yr-1 of the recharge rise is attributed to irrigation from surface water bodies in Bhima. Note that in both models, irrigation increases recharge in irrigated lands, not only during the irrigation season but also during the wet season (Appendix A1).

Recharge increases not only in irrigated areas (92 % and 94 % for Seewinkel and Bhima, respectively) but also in grasslands (6 % and 5 %) and in groundwater-supported grasslands (2 % and 1 %). These results can be explained by the influence of groundwater pumping, which reduces the water table outside the irrigated areas. When irrigation is not applied, recharge is limited in some periods due to the high water table in groundwater-supported grasslands and episodically in other land cover types. The water table is deeper when irrigation pumping is used, thus enhancing water percolation from the soils to the aquifer. Consequently, soil humidity and evapotranspiration rates decrease in groundwater-supported areas due to irrigation, while recharge increases. This mechanism is illustrated by Fig. 1, which shows a 1D section of a water table with and without irrigation.

In Seewinkel, the mean soil water content in groundwater-supported areas decreases from 85 to 79 cm due to pumping (Fig. 7a). The associated evapotranspiration rate diminishes from 595 to 573 mm yr-1 per unit area of groundwater-supported areas. In Bhima, the mean soil water content in groundwater-supported areas decreases from 48 to 46 cm when pumping is applied, leading to a drop in the associated evapotranspiration rate from 920 to 752 mm yr-1 per unit area of groundwater-supported areas (Fig. 7b).

The water cycle is slightly impacted by irrigation at an annual scale in Seewinkel, and we can expect a more significant impact through the dry season when irrigation withdrawals occur. Focusing on the dry summer of 2003, the soil water content in groundwater-supported areas decreased from 70 to 65 cm due to pumping (Supplement). Consequently, the evapotranspiration rate in groundwater-supported areas decreased from 292 to 274 mm m-2 and increased by 40 % in irrigated areas.

Here, we tested the implication of irrigation using the coupled CWatM–MODFLOW model at very high resolutions and including a fine representation of water management practices. In Seewinkel, groundwater is the only source of irrigation, while surface water in Bhima satisfies half of the irrigation demand. Our experiment to infer the impact of the irrigation highlights the fact that irrigation does not significantly impact the average annual water balance in Seewinkel. However, the impact of irrigation is more pronounced during dry periods (e.g., the dry summer of 2003). The low average pumping rate can explain the small effect of pumping in Seewinkel (31 mm yr-1) compared to the recharge rate (80 mm yr-1). Moreover, the net water balance of the aquifer only drops by 12 mm yr-1 because recharge increases by 19 mm yr-1 due to irrigation. In addition, permeability and porosity in Seewinkel are high enough to limit the water table drawdowns induced by pumping. Finally, the soil storage is larger in Seewinkel than in Bhima (average soil water content in groundwater-supported areas is 85 and 48 cm, respectively), involving less dependence on groundwater (less irrigation and less groundwater recharge).

In Bhima, the experiment reveals that irrigation strongly increases evapotranspiration from irrigated areas and slightly reduces available water in groundwater-supported areas. Consequently, river discharge is strongly diminished. Like Seewinkel, groundwater recharge increases in both irrigated and non-irrigated lands. The inclusion of water transfer within the unsaturated zone, thickened due to pumping, could reduce effective recharge, as pointed out by Cao et al. (2016). However, due to the pumping rate and the thickness of the aquifer, the water table drawdown was much more significant in their study.

Assuming negligible groundwater interaction within irrigated lands, 70 % and 47 % of applied irrigation (without accounting for evaporation and runoff losses) percolates below soils and increases groundwater recharge in Seewinkel and Bhima, respectively. Similar values are identified in several other studies. Estimates of recharge from excess irrigation are variable, range from 0 % to 80 % (Groundwater Resource Committe, 2009; Le Coz et al., 2013; Meixner et al., 2016; Ochoa et al., 2013; Roark and Healy, 1998; Scanlon et al., 2005; Willis and Black, 1996), and depend on the efficiency of irrigation techniques (Grafton et al., 2018; Meredith and Blais, 2019).

In Seewinkel and Bhima, we find that groundwater storage is replenished each year during the wet season in spite of pumping during dry seasons because water tables with and without pumping are very similar during wet seasons (Fig. 6). We infer that this result depends on the studied regions, as explained below. Indeed, the Seewinkel and Bhima aquifers are shallow, and pumping is limited by water table depth thresholds. Consequently, groundwater resources can be replenished during wet seasons. In Seewinkel, pumping is limited to protect the natural ponds sustained by groundwater. In Bhima, pumping is limited due to economic constraints and decreases in aquifer permeability with depth.

Recharge is enhanced where the water table is deep, while soil water content and evapotranspiration increase where it is shallow (Fig. 1) (Martínez-de la Torre and Miguez-Macho, 2019; Swenson et al., 2019). Therefore, the spatial distribution of the water table depth is critical when focusing on the interaction between groundwater and soils. Our results suggest that simulated water tables are overestimated. This point is discussed in the next section. In addition, the model could be improved by allowing crops to draw water directly from groundwater systems where their roots exceed the soil depth and where the water table is shallow. Finally, water table depth representation is also relevant from an irrigation perspective. First, pumping redistributes groundwater, potentially impacting baseflow or neighboring areas. Second, water table depth is a proxy for available groundwater resources. Finally, it affects pumping costs (Turner et al., 2019).

The primary result of this study is that the coupled CWatM–MODFLOW model managed to reproduce the average water table time fluctuations (Figs. 4c and 5c) very well. This result is in line with other large-scale models (de Graaf et al., 2017; Martínez-de la Torre and Miguez-Macho, 2019; Sutanudjaja et al., 2014, 2011; Vergnes et al., 2020). Groundwater pumping appears to be necessary to reproduce time fluctuation anomalies in the water table (fluctuations are amplified by pumping in Fig. 6). We infer that, by calibrating water table fluctuations separately, we simulated a good combination between groundwater recharge, pumping, and aquifer lateral flow (driven by hydraulic diffusivity). Therefore, water table time fluctuations contain important information (Houben et al., 2022), even when reproducing the absolute water table depth remains challenging, particularly in Bhima, as also noted by Vergnes et al. (2020). Note that here, we initially averaged water table time fluctuations from all boreholes located within the basins, similarly to Long et al. (2020).

By contrast, the spatial distribution of water table depth can be challenging to reproduce and may be less informative because it is subject to spatial resolution and geological heterogeneity. The comparison between observed and simulated water table depth is satisfactory in Seewinkel (Fig. 4b) but only provides first-order satisfaction in Bhima (Fig. 5b). A first explanation for this may be that the Seewinkel aquifer is more homogenous and well known. Consequently, the model can better describe its properties, such as thickness, permeability, and porosity. In contrast, the Bhima basin is much larger and hosts mainly crystalline rocks with potentially variable weathering and fracture density, making its aquifer properties heterogeneous. Beyond the aquifer parameters, the ability to reproduce time-averaged water table depth is constrained by two critical factors in both areas: the pumping rates and the spatial resolution of the groundwater model. These two factors are more acute in Bhima, where pumping is more pronounced and the resolution is coarser (250 m). In both study areas, simulated mean annual pumping rates at the basin scale were validated. However, we acknowledge that pumping estimates could remain uncertain regarding annual averages, seasonal behavior, and pumping well locations.

Groundwater model resolution is critical because degrading resolution smoothens topographical variability. The altitude of local low points in valleys due to river incision, often corresponding to aquifer resurgences, is increased more than it is in neighboring cells when degrading resolution. Consequently, the water table is shallower than expected because the low points locally controlling the water table are higher than expected (see topography and water table comparisons at 100 and 250 m resolution in Appendix A2). This explains why very shallow conditions (<1 m depth) close to rivers are overestimated in Seewinkel despite the 100 m resolution employed. A similar bias was obtained by Maxwell et al. (2015), who used a spatial resolution of 1 km over the continental US.

The weakness of large-scale models for reproducing the spatial distribution of water table depths has been mentioned but not adequately explained in several previous studies (Fan et al., 2013; Maxwell et al., 2015; Reinecke et al., 2020). These studies indicate that mismatch between models and observed data could be due to (i) well representativity within a mesh that is too coarse (see also Martínez-de la Torre and Miguez-Macho, 2019), (ii) well sampling bias, or (iii) groundwater pumping. However, the ability of large-scale models to reproduce water tables remains primordial, and the means of evaluating them should be improved (Gleeson et al., 2021). Here, for the first time, we used a large-scale hydrological model, including groundwater flow and pumping, at a very fine resolution. We note that, as expected, regional models show a better aptitude to reproduce water table depth. For instance, Vergnes et al. (2020) reproduced a mean water table depth on 639 piezometers across several French regions with a median bias of around 3 m. Similarly, Soltani et al. (2021) simulated water tables with an RMSE of 3.8 m based on around 28 000 groundwater head stations in Denmark. Jing et al. (2018) reproduced water table depths with an RMSE of 6.3 m. In this study, we simulated mean water table depths with RMSEs of 1.0 and 5.0 m in Seewinkel and Bhima, respectively. Surinaidu et al. (2013) obtained a similar RMSE in Bhima with a simpler model at a coarser resolution as well as uniform recharge and discharge. However, a comparison with water table fluctuations, depth bias, and surface water flows, including discharge and reservoir levels, was not performed.

We argue that simulated water tables tend to be shallower than the observed water table due to the coarse spatial resolution, as explained above. However, increasing the groundwater model resolution would significantly increase computing time, which makes it challenging to perform sensitivity analysis. Indeed, the same simulation performed in the Bhima basin at a 100 m resolution lasts around 2 d versus 9 h at a 250 m resolution. Figure 8 illustrates how increasing the resolution from 250 to 100 m slightly reduces water table depth bias. While the nRMSE criterion is not significantly better, highlighting the role of heterogeneity rather than spatial resolution, the mean water table depth is now overestimated by 0.5 m instead of 1.7 m (Fig. 8). Reinecke et al. (2019b) focused on the reason for water table depth mismatch and its sensitivity. They highlighted the importance of hydraulic conductivity, recharge, and surface water body elevation. The surface water body elevation can be assimilated to the top of the aquifer in the two CWatM–MODFLOW models described here. Therefore, this critical factor is linked to spatial resolution. By comparison, Reinecke et al. (2019a) improved their RMSE from 54 to 27 m using a resolution of ∼ 9 km and 900 m, respectively. Note that their experiment was located in a steep domain (New Zealand).

CWatM uses a mosaic approach to represent the subgrid variability of the land cover type. In addition, we developed a subgrid approach to improve model efficiency, specifically for groundwater exchanges, to benefit from the finer MODFLOW resolution. Indeed, each CWatM grid cell (1 × 1 km) contains a land cover fraction where soil is supported by groundwater. Therefore, we found that the impact of pumping on soil–groundwater exchanges was more significant when groundwater capillary rise focused on a fraction of each grid cell instead of distributing the flow homogenously over the whole cell. This approach enables better simulation of the link between the water table, evapotranspiration, and recharge without refining the CWatM grid.

As pointed out previously, surface water body elevation, linked to the spatial resolution here, is important for simulating absolute water table depth (Reinecke et al., 2019b). This is important, as the river network elevation strongly controls water table depth. Moreover, rivers represent groundwater outlets, as CWatM–MODFLOW generates baseflow based on the simulated water table below rivers. However, a whole cell cannot be entirely assimilated to a river due to the coarse resolution of large-scale models; thus, two critical issues can be noted. First, large-scale modellers never have a map of the actual river network, including the finest low-order and intermittent rivers. They usually generate a river network based on a digital elevation model (Yamazaki et al., 2019), so partitioning between baseflow and groundwater capillary rise is predefined before the simulation. Second, groundwater reaching the top of the aquifer can feed either rivers (as baseflow) or soils (as capillary rise). Our modeling approach relies on a finer digital elevation model used to compute a river network at a resolution finer than the groundwater model (Sects. 2.2.1 and 3). However, we acknowledge that small natural or artificial drains are not included in the model. Then, partitioning between baseflow and capillary rise is defined at the resolution of the groundwater model. This partitioning impacts the depth of the water table. Indeed, increasing the baseflow fraction decreases the water table depth and reduces the soil water availability. Our approach partly deals with this issue as the resolution is very fine (100 and 250 m) and because cells where groundwater oversaturates soils generate overland runoff, feeding rivers.

Other subgrid approaches exist. The TOPMODEL approach (Beven and Kirkby, 1979), for instance, is used in many regional-scale models. It infers subsurface flow convergence based on a finer digital elevation model. Vergnes et al. (2014) also proposed an alternative approach, in which a groundwater flow equation simulates the water table dynamic but capillary rise only occurs in a fraction of the cells based on the distribution of the topography at the subgrid level. Similar approaches are used within CWatM and PCR-GLOBWB when no lateral groundwater flow is considered. Another option is to represent coarse cells or basins with equivalent (representative) hillslopes (Fan et al., 2019; Loritz et al., 2017; Troch et al., 2003). Recently, new hillslope models have been developed for use at a large scale (Hazenberg et al., 2015; Swenson et al., 2019). However, this conceptual modeling approach cannot include regional lateral flow convergences or groundwater pumping. Finally, another approach would be to aim for hyper-resolution in each compartment (surface, soil, and groundwater). However, this would require huge computing capacity to be achievable at a continental scale.