Coastal aquifer recharge and groundwater–seawater exchanges using downscaled GRACE data: case study of the Djeffara plain (Libya–Tunisia)

The Djeffara plain, extending over Tunisia and Libya (48,000 km2), experiences unsustainable exploitation of groundwater, the severity of which needs to be determined. Here, we used data from Gravity Recovery and Climate Experiment (GRACE) and Global Land Data Assimilation System (GLDAS) to assess the groundwater storage variation (∆GWS) for the period 2002–2016. ∆GWS values, downscaled over a 0.5°×0.5° grid using pumping wells spatial distribution and rates, were introduced in a discrete set of groundwater and salt budget equations to ascertain the recharge and the coastal fluxes. An average recharge of 4.3±2.7 mm·yr−1 representing 2.2±1.4% of the domain-averaged annual rainfall was obtained. A saline intrusion of 147.3×106 and 46.9×106 m3·yr−1 was calculated for the Tripoli (Libya) and Djerba (Tunisia) areas, respectively. The overall Submarine Groundwater Discharge (SGD) of freshwater along the 450 km long Djeffara coast represents about 339.6×106 m3·yr−1 for the 2002–2016 period.


Introduction
In the Djeffara plain, like in other areas of the Maghreb, demographic and economic growths as well as the development of irrigated agriculture contribute to massive and unsustainable groundwater exploitation. In the transboundary multilayer aquifers system underlying the Djeffara plain shared 298 Julio Gonçalvès et al. to determine the degree of (un)sustainability of groundwater exploitation and to guide management policies. The few available values for the recharge in the Djeffara plain are derived from hydrogeological modeling [OSS, 2006]. With the exception of the studies by Allemmoz and Olive [1980] and Zouari et al. [1985] in the regions of Tripoli and Gabès (see Figure 1), alternative approaches using geochemical or hydrogeophysical methods, for instance, are missing so far. While the estimation of a still unclear current recharge is a common issue for all the circum-Sahara regions, coastal aquifers such as the Djeffara system raise specific problems. Heavy and increasing withdrawals in coastal domains lead to salinization problems caused by seawater intrusion [Steyl and Dennis, 2010]. This has been clearly established in the Tripoli area (Libya) with saline intrusion due to increasing urbanization and agricultural activities [Alfarrah andWalraevens, 2018, Sadeg andKarahanoglu, 2001]. The identification of groundwater-seawater fluxes, termed Submarine Groundwater Discharge (SGD), is another issue specific to coastal aquifers. The quantitative importance of SGD in comparison to surface water fluxes for oceans mass balance is a subject of controversy but their qualitative impact is largely accepted [e.g. Luijendijk et al., 2020]. SGD is often associated with a substantial supply of nutrients that affects the ecological equilibrium of coastal areas [e.g. Rodellas et al., 2015].
A commonly used approach to derive SGD fluxes on local to regional scales is to close a radionuclide mass balance, using predominantly radon and radium as SGD tracers. Local SGD fluxes (on a bay/beach scale) were estimated using these tracers at many locations along the northern (European) coast of the Mediterranean Sea, but there is a conspicuous lack of any data-based SGD estimates in the southern Mediterranean basin along north African coastlines, including those adjacent to the Djeffara plain. Equally important, such radionuclide tracer-derived estimates include all advective fluxes across the sediment water interface in the term SGD [e.g. Rodellas et al., 2015]. It has been widely demonstrated that the major process contributing to radionuclide fluxes (>80% in most cases) is the recirculation of seawater through coastal sediments, and freshwater SGD generally are one order of magnitude lower than such total SGD fluxes [e.g. Stieglitz et al., 2013]. While radionuclide-derived values are useful from a geochemical perspective, they do not yield estimates comparable to freshwater flow estimates based on hydrogeological or geophysical approaches, unless combined with other tracers such as salinity [e.g. Stieglitz et al., 2013, Weinstein et al., 2007. Similar to the quantification of recharge, no estimates other than those produced by hydrogeological modeling are available for freshwater SGD [e.g. Zektser et al., 2007] or saline intrusion.
Since 2002, a satellite monitoring of Earth's gravity field (GRACE experiment) enables groundwater mass storage (GWS) variation assessments over large regional-scale aquifer systems. Satellite approaches provide an integrated view of processes over large areas, which would be much more time-and costconsuming to obtain from ground observations, if not impossible. GRACE data have been widely used for quantifying aquifer storage changes [e.g., Ahmed et al., 2014, Bonsor et al., 2018, Carvalho Resende et al., 2019, Leblanc et al., 2009, Ramillien et al., 2014, Richey et al., 2015, Rodell et al., 2009. Some studies used a global mass balance to estimate the domain-averaged recharge for large regional aquifers of the Saharan belt [Gonçalvès et al., 2013, Mohamed et al., 2017. Similar to these studies, GRACE-derived groundwater mass variations (∆GWS) could be combined with ground data to estimate hydrological variables such as the recharge or the coastal fluxes independently from any hydrogeological model. However, all these previous applications using GRACE data have been carried out on large surface basins (>100,000 km 2 ). In fact, due to the relatively low resolution of the GRACE data (>1°) and signal-tonoise ratio issues, poor quality applications are expected for basins with a surface area of less than about 100,000 km 2 . One way to address this problem for more local-scale basins such as the Djeffara aquifers system is to downscale the GRACE data using ground-based information or data assimilation systems variables [Miro andFamiglietti, 2018, Seyoum et al., 2019].
In this context, GRACE-derived ∆GWS for the Djeffara domain are first downscaled at the resolution of the gridded annual average rainfall (AAR) data (0.5°× 0.5°) assuming a major control of pumping wells on ∆GWS. The weights in the downscaling approach proposed here are therefore based on the distance to pumping well areas and pumping rates. Downscaled GRACE-derived ∆GWS values are compared to their independent calculations of groundwater mass variations based on ground data, i.e., aquifer storativity and water level drawdown values available for the Libyan Djeffara.
The classical groundwater mass balance equation is discretized over the 0.5°× 0.5°grid introducing the fluxes to be ascertained, i.e., the recharge and the coastal fluxes as well as the downscaled ∆GWS values. Then, the linear system of discrete mass balance equations obtained was solved for recharge and saline intrusion or SGD assuming a linear relation between average annual rainfall (AAR) and recharge. The results, which are the first independent estimates from any numerical models to our knowledge, are discussed in comparison with hydrogeological modeling outputs as well as with SGD flux estimates based on radium tracer approaches.

Hydrogeological context
Located between the Mediterranean coast and the mountains of the Nafusa mountains to the south and the Dahar to the west, the Djeffara plain covers a surface area of about 48,000 km 2 over eastsouthern Tunisia and northern Libya ( Figure 1a). This region receives an annual average precipitation of about 200 mm for a potential evaporation of 2760 mm characterizing an arid climate [Trabelsi et al., 2009]. The mountains (jebels) correspond to anticlines that structurally separate the Djeffara Plain from the north-western Sahara Aquifers System (NWSAS, "Continental Intercalaire" CI and "Complexe Terminal" CT aquifers Figure 1b). The Djeffara basin is characterized by an almost continuous stratigraphic succession from the Permo-Triassic to the Quaternary [OSS, 2006]. However, in the southern Libyan part, the Triassic layers are unconformably overlain by Miocene formations (erosion surface). Therefore, from a (hydro)geological standpoint, the Djeffara plain can be subdivided into a northern domain (Tunisia) where the aquifer formations are mainly the Senonian and Mio-Plio-Quaternary units, and a southern domain where the Triassic and Miocene formations are of primary importance in addition to the Plio-Quaternary deposits. Overall, the stratigraphic succession can be closely examined to define a multilayered aquifer system with lateral variations from the Tunisian to the Libyan Djeffara. In Tunisia, the Trias is composed of sandstones, dolomites, and, in the upper part, an intercalation of clay and evaporites which extends stratigraphically until the Jurassic. The Trias and Jurassic formations are overlain by the dolomitic Turonian and the calcareous Senonian formations which constitute the main exploited and productive aquifer together with the Mio-pliocene-Quaternary sands. The Mio-pliocene-Quaternary aquifer is separated from the Senonian by more or less continuous marly gypsum and clayey formations. In Libya, the Lower Triassic consists of sandy clay then dolomite, marls, and clay layers overlain by sandstone. The Upper Triassic and Lower Jurassic are made of gypsum and anhydrite. The Jurassic and Cretaceous formations (only along Nafusa Mountains) are of small hydrogeological importance either because they are absent (erosion) or because they are not really exploited [Ben Lasmar et al., 2017]. From bottom to top, the Miocene begins with a clay layer followed by a limestone formation (sometimes referred to as the Lower Miocene Aquifer), then a sandy limestone (Middle Miocene Aquifer). Separated from the latter by a sandy clayey layer, the Mio-Plio-Quaternary is extremely heterogeneous (sand, alluvial deposits, limestones, clay) and intensively exploited in terms of groundwater.
The hydrostratigraphic analysis conducted by OSS [2006] leads to the definition of three aquifers which are in fact multilayer aquifers. The Upper Miocene, Pliocene, and Quaternary deposits form a continuous aquifer over the entire Djeffara (known as the phreatic or Upper aquifer). The Lower and Middle Miocene, also associated with the Jurassic and Upper Triassic when confining units are absent, form the Middle aquifer (deep coastal aquifer). The deep aquifer comprises the Lower and Middle Triassic formations.
Groundwater flow directions are generally from the reliefs of the jebels Nafusa and Dahar to the coast and therefore mostly from west to east in the Tunisian part and almost south to north in the Libyan Djeffara. All aquifers are characterized by high transmissivities between 1 × 10 −2 and 5 × 10 −2 m 2 ·s −1 and the regional hydraulic head gradient is in the order of 0.1% [OSS, 2006]. Interestingly, introducing these values in a Darcy flow calculation for a coastal length of 450 km leads to a horizontal drainage of 4.5 to 22.5 m 3 ·s −1 using the above transmissivity range which can be translated into a steady-state recharge of between 3.0 and 14.8 mm·yr −1 . Recharge occurs locally, either focused or diffuse as well as from inflows at the boundaries. Recharge estimates (natural and irrigation return flow) were obtained from modeling approaches at 6.8 mm·yr −1 [1.8 mm·yr −1 for focused recharge; OSS, 2006]. Using a water table 301 fluctuation method and soil moisture profiles, Allemmoz and Olive [1980] found a recharge of 10 mm·yr −1 in the area of Tripoli (Libya) exclusively from focused recharge (wadi flood infiltration). Zouari et al. [1985] used chloride and stable isotope of water profiles in the vadose zone near Gabès (Tunisia) to estimate infiltration (5 mm·yr −1 ) and evaporation (1 mm·yr −1 , presumably from the water table) leading to a net diffuse recharge of about 4 mm·yr −1 .
Groundwater inputs at the boundaries of the Djeffara aquifer system correspond to (i) a horizontal flow from the south in the Trias at the Nafusa [OSS, 2006] and (ii) an upward flow along faults which is one of the main outflows of the deep confined CI aquifer of the NWSAS [Gonçalvès et al., 2015]. Heavy groundwater exploitation (see Section 2.3) causes continuous seawater intrusion and salinization of groundwater. This was particularly reported in the most populated Tripoli area [Alfarrah andWalraevens, 2018, Sadeg andKarahanoglu, 2001]. Groundwater salinization by dissolution of evaporites is also documented [Trabelsi et al., 2012].

GRACE and GLDAS data
The GRACE twin-satellite system launched in 2002 by NASA and the German Aerospace Center (DLR) monitors the earth gravity on a monthly basis [Tapley et al., 2004]. The mission is based on the assumption that gravity variations are mostly due to mass variations in water bodies (groundwater, soil water, and surface water). Such data expressed in water height can thus be used for hydrological mass balance calculations at the regional scale. After more than one decade of gravity field processing using spherical harmonics (SH), an alternative interpretation using Mass Concentration blocks or Mascons (M) of discrete spherical caps at Earth's surface was proposed in recent years, GRACE-M [Watkins et al., 2015]. The main disadvantage of the GRACE-SH solution is a noisy signal which requires a filtering post-processing in which an undetermined fraction of the geophysical signal is lost [Landerer and Swenson, 2012]. A fraction of these signal losses can be recovered by using gain or scale factors obtained, for instance, by minimizing filtering signal degradation (undesired filter effect) using land surface model (LSM) water mass outputs [Landerer and Swenson, 2012]. Conversely, GRACE-M requires less post-processing treatment due to lower geophysical signal losses. However, mascon gravity data interpretation with a native resolution of 3°× 3°for the Jet Propulsion Laboratory (JPL) data center solution used here (see below) produces a smooth signal, which is a serious limitation for hydrological applications that require a higher resolution and thus downscaling approaches . Two alternative solutions can therefore be used: (1) GRACE-SH, which is noisier but whose resolution (1°× 1°) is closer to the target one (0.5°× 0.5°), and (2) GRACE-M, which is less noisy but requires further downscaling efforts. In this study, we used both solutions for comparison. Monthly GRACE-derived Terrestrial Water Storage (TWS) anomalies (difference between a TWS value and its average value calculated for the period January 2004-December 2009) were retrieved from the JPL data center (http://grace.jpl.nasa.gov). GRACE-SH data correspond to the RL05 solution from the Centre for Space Research (CSR, University of Texas, http://www.csr.utexas.edu/grace/) scaled with the scaling factor of the NCAR-CLM4.0 model [Landerer and Swenson, 2012]. GRACE-M data are the RL06M solution from JPL. We retrieved the JPL R06M filtered from the contribution of oceans [CRI filter, Wiese et al., 2016], therefore giving the TWS time series corresponding only to the land part of a Mascon. The proposed scaling factors which correspond in fact to downscaling factors [Scanlon et al., 2016] were not used here since an alternative approach is proposed. These downscaling factors are based on the mass distribution calculated by LSM (soil-atmosphere models) while water height in the Djeffara plain is expected to be mostly controlled by groundwater mass variations. Therefore, we derive a specific downscaling algorithm in this study. We used 170 monthly solutions for the period between April 2002 and December 2016 ( Figure 2). This is the period with the minimum amount of missing GRACE-SH data and fairly close to the end of the period when water level data are available. The trend of a TWS time series can be analyzed and interpreted in terms of water amount variation (e.g., in mm·yr −1 ) over a land surface and thus water fluxes over a domain of interest.
As outlined above, TWS accounts inextricably for groundwater, surface water, and soil moisture. Therefore, and in the absence of substantial surface water in the Djeffara plain, Soil Water Storage (SWS) has to be integrated to identify Groundwater Storage (GWS). However, soil water storage variations (∆SWS) are usually small due to almost constant and low soil moisture in arid areas, and may represent a small or even negligible component of TWS [Gonçalvès et al., 2013, Nimmo et al., 1994. We can easily verify the relative importance of ∆SWS using GRACE-SH data whose resolution is the same as the SWS data and which correctly describes the Djeffara plain (see Figure 2b for GRACE-SH and GRACE-M grid cells location). For this purpose, soil storage data were obtained from Global Land Data Assimilation System (GLDAS), which provides soil moisture for the first meters by combining ground-based, satellite data and hydrological surface model results [Hogue et al., 2005, Rodell et al., 2004. We retrieved monthly data from the Variable Infiltration Capacity (VIC) and the Community Land Model (CLM) versions with a 1°× 1°degree spatial resolution (http://disc.sci.gsfc.nasa. gov/). VIC and CLM were previously recognized as the best models for the Saharan region [Gonçalvès et al., 2013]. The time series for the SWS (anomalies from their average value calculated over the period 01/2004-12/2009 as TWS) and TWS averaged over the Djeffara plain are shown in Figure 2a which provides a ∆TWS value by far greater than ∆SWS. Considering GRACE-M data, Mascons #1 and #2 extend beyond the Djeffara plain domain, i.e., over the NWSAS domain (Figure 2b), making the comparison between ∆SWS and ∆TWS more difficult. The ∆TWS values for land surfaces of Mascons #1 and #2 are −6.1 ± 0.2 and −5.5 ± 0.2 mm·yr −1 , respectively ( Figure 2c). These values are also substantially larger than the ∆SWS of less than 0.5 mm·yr −1 discussed above for the Djeffara (Figure 2a) and reported for the NWSAS [Gonçalvès et al., 2013]. Therefore, for both GRACE-SH and GRACE-M data, the contribution of ∆SWS to ∆TWS is considered negligible.

Hydrological data
Prior to the 1940s, most of the water used was supplied by springs, notably in Tunisia. With the development of pumping wells and the continuous increase of withdrawals, the water fraction supplied by springs declined until their complete drying up [OSS, 2006]. Pumping rates are available for years 1950, 1972, 1978, and 2003from OSS [2006, and for year 2011 (only global exploitation) from Alfarrah and Walraevens [2018]. The most documented study from OSS [2006] presents relatively precise values for irrigation but highly approximate pumping rates for domestic water. Since most of the domestic water consumption occurs in the Tripoli area, we have considered the more accurate data by Sadeg and Ghazali [2008]. In this study, domestic water pumping (about 6% of the withdrawals) will be considered explicitly only in the Tripoli area where a salt mass balance calculation is made. Most of the extracted groundwater, i.e., 61%, comes from the Upper aquifer, 32% from the Middle aquifer, and the remaining fraction (7%) from the Triassic aquifer. The only area where the increase in pumping rates is not linear but seems to reach a plateau is the Tripoli region. This may be explained by a combination of salinization problems and the substitution of some part of local water demand by groundwater from aquifers in southern Libya (Nubian sandstones) conveyed by a pipeline, the Great Man-Made River (GMMR), sourcing water from outside the catchment. This is clearly true for domestic water supply until 1997 produced by local wells in the Tripoli area, and since then exclusively supplied by the GMMR at a rate of 146 × 10 6 ·m 3 ·yr −1 [Alfarrah andWalraevens, 2018, Sadeg andGhazali, 2008]. Therefore, intensive groundwater pumping for domestic water mainly concerned the Tripoli area and was interrupted in 1997 with the connection to the GMMR system.
Considering the recharging fluxes at the western and southern boundaries of the system, a constant value of 2.6 m 3 ·s −1 was estimated for the Triassic flux by means of a hydrogeological model [OSS, 2006]. Using an approach combining water stable isotopes, geostatistical and reservoir simulations, and a simple hydrological model based on piezometric levels, Gonçalvès et al. [2015Gonçalvès et al. [ , 2020 obtained a temporal variation for the recharging flux from the CI to the Djeffara aquifers showing a decline from 1.8 m 3 ·s −1 in 1950 to 0.6 m 3 ·s −1 in 2009 due to the overexploitation of the CI aquifer. In addition, global rainfall rates beginning in 1901 were retrieved from the CRU data center (Climatic Research Unit, http://wps-web1. ceda.ac.uk) at a resolution of 0.5°× 0.5°. These values indicate an annual domain-averaged rainfall rate of around 200 mm·yr −1 for the Djeffara plain.
Piezometric information available from OSS [2006] is mostly made of cumulative drawdown ∆h values in 2003, and drawdown (not absolute piezometric levels) time series but for a small fraction of piezometers. Moreover, the data is not homogeneously distributed over the domain with a higher density in Libya especially for the Upper aquifer (10 data in Tunisia against 33 in Libya). Therefore, ground-based ∆GWS estimates were focused on the Libyan Djeffara (cells 11-14, 19-21, Figure 2). Drawdown data from OSS [2006] do not fully cover the period of investigation here (2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016). Extending the applicability of such ∆h data requires analyzing longer time series. Piezometric levels time series until the 2010's could only be reconstructed for nine piezometers (see Figure 4a, b, and c) using punctual hydraulic head values [1972, 1978, 1993, and 2010Gejam, 2018], and drawdown records [OSS, 2006, Saleh andRashrash, 2015]. Among the available piezometers, well 1006 is of particular importance since it belongs to the reference piezometric network and is located at the center of the major piezometric depression affecting the Upper aquifer in Libya. This piezometer shows an almost continuous deepening of the piezometric depression. Continuous albeit moderate drawdown is also recorded in coastal piezometers (1020, 1057, and 1058; Figure 4a) while piezometer 1344 shows a stabilization which is inconsistent with the overall piezometric behavior. Piezometers in the Middle and Trias aquifers show a continuous lowering of the piezometric level. The comparison between recent drawdown (since 2002) and long-term values given by OSS [2006] shown in Figure 4d and expressed in meters per year for comparison suggest that the latter can be used for the recent period considered here (2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016) within an error of about 16% (regression slope uncertainty). Moreover, these drawdown values are the only independent and relevant ground-based estimates for comparison purposes with the GRACE-derived estimates. Ground-based GWS variations ∆GWS GWL can be calculated by linear interpolation of the long-term drawdown values ∆h [OSS, 2006] on a 0.1°× 0.1°grid, then an averaging on a 0.5°× 0.5°grid and by calculating ∆GWS k GWL = S k t × (∆h/∆t ) k , with S k t the storativity (specific yield or storage coefficient) and ∆t the time interval for the drawdown, for each aquifer k (Upper, Middle, and Trias) and by summing all these values. A specific yield at 7.5 × 10 −2 corresponding to well tests [Allemmoz and Olive, 1980] was used for the Upper aquifer, and values for the Middle and Trias aquifer were taken from the model calibration by OSS [2006] at 2 × 10 −3 and 1 × 10 −3 , respectively. The beginning of the drawdown was considered in 1950 for the Upper aquifer (∆t = 53 years) and 1970 (∆t = 33 years) for the Middle and Trias aquifer whose exploitation started later [Saleh and Rashrash, 2015]. The ∆GWS GWL values for the Tripoli area are given in Table 1.

Downscaling GRACE-derived groundwater mass variations
Regional-scale groundwater budgets are used in this study to estimate different hydrological terms such as recharge, aquifer salinization, and SGD. The solution of such balance equations requires the groundwater mass variations ∆GWS to be estimated. GRACE data provide a valuable integrated mapping of water storage over large domains. This would be difficult to achieve using field data such as piezometric level monitoring due to discontinuous data acquisition in time and/or space. The first step is to isolate the groundwater storage anomalies (GWS) from the TWS anomalies expressed in water height to facilitate hydrological applications. In arid areas such as the Djeffara plain, surface water can generally be neglected leading to the following simple decomposition of the GRACE data: and thus ∆TWS = ∆GWS + ∆SWS.
Given the fact that ∆SWS ∆TWS (Section 2.2, Figure 2) in the case of the Djeffara plain, ∆GWS can be assimilated to ∆TWS allowing a direct interpretation of GRACE data. Due to the coarse resolution of the GRACE data, the Djeffara plain only corresponds to seven 1°×1°grid cells for GRACE-SH and two 3°×3°M ascons for GRACE-M (blue and green squares, respectively in Figure 2b). Similar to what was done using the domain-averaged TWS as depicted in Figure 2, the interannual trend (slope of the linear regression on TWS time series) gives the required ∆GWS = ∆TWS values in mm·yr −1 for each grid cell of GRACE-SH and GRACE-M solutions. However, to enable the integration of the CRU annual rainfall data which is at a higher resolution (0.5°× 0.5°) and to refine the water budget calculations, the ∆GWS values obtained for each of the coarse grid cells have to be downscaled. The downscaling here is based on the assumption that groundwater mass balance is mostly influenced by pumping. However, the groundwater budget within a domain described using GRACE grid cells is controlled by both internal fluxes (abstractions, recharge) and boundary net fluxes. The validity of the assumption adopted here which leads to neglect boundary fluxes is discussed in Section 5. Given that the drawdown caused by a pumping well is proportional to Q W × ln(r ) [de Marsily, 1986] with Q W (m 3 ·s −1 ) the pumping rate and r the radial distance to the well, the lateral drainage (radial component) upon application of Darcy's law may be dependent on Q w /r . Since the lateral drainage and ∆GWS are related in the groundwater mass balance equation (see Section 3.2), pumping rates divided by the distances to pumping zones can be heuristically introduced as weighting factors in a downscaling approach. As illustrated in Figure 5a, for a given cell j out of the n c 0.5°× 0.5°refined grid cells contained in a sub-domain i to be downscaled (e.g., n c = 4 for a 1°× 1°grid) with a known groundwater storage variation ∆GWS i , the weight writes: with: where Q l W is the total withdrawal (m 3 ·s −1 ) of pumping area l and r l k is the average distance between the four corners of the refined grid cell k and the centroïd of the pumping area l (Figure 5a) to avoid singularity issues (r l k → 0). Note that the centers of the main pumping zones (centroid) were used here to avoid considering thousands of wells whose location is uncertain for a large number of them. The groundwater storage variation for a given refined grid j writes finally: The closer the node j is to a pumping centroïd and the larger the pumping rate, the greater the weight. Moreover, from (3) to (5), it can be easily verified that The downscaling approach proposed here is applied in two different manners depending on the GRACE solution used. For GRACE-SH, the application is straightforward using n c = 4 for each 1°×1°grid and the pumping rates and centroïds within the Djeffara plain. For the downscaling of each Mascon land surface ∆GWS value, the weights are calculated for each 0.5°× 0.5°cell but using exclusively the pumping wells located in the subdomain (NWSAS or Djeffara) of the considered cell. Therefore, the main pumping areas and rates of the NWSAS within the retained Mascons had to be introduced [OSS, 2003, see Figure 2b].
For consistency reasons, we considered average pumping rates for the same period as . No further refinement of the grid in the north-west of the domain (first row) was made to avoid making the linear system resolution unnecessarily more complex for an uncertain gain in precision (see discussion). The mass balance grid is therefore maintained at the resolution of the AAR data. The flux F 3 x,r therefore represents the overall exchange with the sea in the area of Gabès-Djerba-Zarzis.

Mass balance approach
Groundwater storage variations ∆GWS provided by GRACE can be introduced in a regionalized mass balance equation to identify recharge and coastal exchange fluxes. For this purpose, we consider the groundwater mass balance equation for an aquifer system as: where T x and T y are the transmissivities (m 2 ·s −1 ) in the x and y direction, h is the hydraulic head (or piezometric level in m), S t is the storativity (specific yield S y or storage coefficient S for unconfined or confined aquifers), Q W is the pumping rate in m·s −1 (discharge in m 3 ·s −1 divided by a cell grid surface area), and R N and R A are natural and artificial recharges, respectively (m·s −1 ). The artificial recharge mostly corresponds to irrigation return flow to the aquifer which is accounted for using the relation R A = f R × Q W with f R standing for the fraction of irrigation return flow (to be ascertained, see Section 4). For the most populated region of Tripoli only, an additional term due to domestic water network leaks is considered (see Section 4). Generally, Equation (6) is solved for h and a single aquifer, and then the fluxes are calculated using Darcy's law.
Here, identifying h is not the purpose since several aquifers are in fact present, and our goal is to obtain mass balance equations written in fluxes, applied to a multilayer aquifer where GRACE-derived ∆GWS values can be introduced. Upon discretization over the 0.5°× 0.5°grid considered here for the Djeffara plain (Section 3.1) using ∆x and ∆y as x and y spacing, and a time step ∆t of one year (6) becomes: where Q i x,l and Q i x,r are the discharges (m 3 ·yr −1 ) in the x direction at the left and right edges of cell i positive if oriented in the x axis direction; Q i y,l and Q i y,u are the discharges (m 3 ·yr −1 ) in the y direction at the lower and upper edges of cell i positive if oriented in the y axis direction. Introducing the fluxes (in m·s −1 ) corresponding to discharges (e.g., F i x,l = Q i x,l /(∆x∆y)) in the left-hand side of (7), and noting that S t ∆h i /∆t = ∆GWS i , Equation (7) yields the mass balance equation applying to each of the 210.5°×0.5°g rid cells of the Djeffara which writes: (8), Q W i , R Ai , and ∆GWS i are known at each grid cell, the left-hand side fluxes and the natural recharge are the unknowns. In the most general case, the system formed by the 21 equations (mass balance (8) for each grid of the Djeffara) is unsolvable since 76 unknown variables (fluxes and natural recharge values) are involved. Therefore, two simplifications are proposed to solve the set of equations. The first simplification is to consider a linear relation between natural recharge and AAR in arid areas [Scanlon et al., 2006], i.e., R N i = αAAR i with α a dimensionless coefficient to identify. The time lag of several years to several decades [e.g. Scanlon et al., 2006] for the recharge to reach the water table makes the use of the annual rainfall inaccurate when the use of AAR is more adapted to account for the buffering effect of the generally thick vadose zones in arid domains. According to OSS (2006, and see Section 2.1), diffuse recharge characterized by large lag times could represent almost 75% of the overall recharge in the Djeffara plain. The second simplification consists in considering only the main flow directions, i.e., west to east in the Tunisian Djeffara and south to north in the Libyan part (Figure 5b). This assumption relies on the general natural hydraulic gradients from reliefs to coastal zones, and on the alignment of pumping wells along the coast creating hydraulic divides also globally directed from the reliefs toward the coast. These two assumptions lead to a groundwater mass balance linear system of 21 equations and 22 unknowns (spatially variable fluxes) shown in Figure 5b which is therefore still undetermined. However, the mass balance equation system can be simplified to a stream tube form by summing groundwater balance equations of individual cells for the following blocks: 1-3; 4-6; 7-10 and 15-18; 11-19; 12-20; 13-21; 14. For all blocks ("tubes") intermediate fluxes at meshes sides would simplify leaving only one unknown (besides α) per equation: the coastal flux. Note that for the coastal cells, negative values of F i x,l or F i y,u represent seawater intrusion. Conversely, positive values correspond to groundwater discharge (SGD).

Salinity model for the Tripoli refined grid cell
In order to solve the indetermination problem of the groundwater mass balance equations system of Section 3.2, a simple domain-averaged salinity model for the cell corresponding to the Tripoli region (cell 13, Figure 5b and c) is considered. Given the heavy withdrawals in this area, fresh groundwater discharge is assumed negligible. We consider that the coastal flux mostly corresponds to seawater intrusion which occurs mainly in the Upper aquifer, therefore neglecting the seawater fluxes for deeper aquifers. The time evolution of the salinity in the Tripoli area is sufficiently well documented which enables the calculation of average salinity values for the corresponding 0.5°× 0.5°grid cell using the salinity maps available for 1950for , 1960for , and 2000for from OSS [2006 for the Upper aquifer. The simple salinity mass balance equation in the Upper aquifer for cell 13 writes: where V (m 3 ) is the volume of the Upper aquifer in the Tripoli grid cell, f is the fraction of the pumping in the Tripoli area Q T W (m 3 ·s −1 ) which returns to the aquifer (irrigation return flow and/or domestic water network leaks) with a salt concentration C 1 (kg·m −3 ), Q M and Q S (m 3 ·s −1 ) are the seawater entering the aquifer and the flux coming from the south, and C M and C 0 (kg·m −3 ) are the seawater salinity and the natural salinity (water rock interactions equilibrium concentration). Note that in (9) the natural recharge does not appear because it involves the concentration of rainwater (or evaporated rainwater) and represents a negligible term compared to the others. We assume that the time variation of the total pumping rate in the Tripoli area (Figure 3) can be described by: with the fitted values A = 13.3 m 3 ·s −1 , β = 0.05 yr −1 and a time origin in 1948 when Q T W = 0. Assuming that seawater intrusion and southern fluxes closely follow the global pumping rate time evolution leads to consider similar expressions for Q M and Q S writing: A simplified resolution of the salt balance equation (9) can be achieved by considering that the return flow is mainly related to irrigation (85% of the water pumped before 2000, i.e., the calibration period). The salt is extracted from aquifers at a rate Q T w ×C but the fraction (1− f ) of the groundwater abstraction Q T w is lost by evaporation leaving the salt in the top soil. Assuming a perfect salt leaching by the excess irrigation water, the concentration in the water returning to the aquifers is C 1 = [(1− f )C + f C ]/ f = C / f , which exactly corresponds to an evaporation effect. Introducing this expression for C 1 in the mass balance equation (9), the two first terms at the righthand side vanish (total salt return). This simplification is therefore equivalent to considering that the salt balance is controlled by lateral fluxes (Q M ,Q S ) depending on pumping but is not directly affected by withdrawals due to irrigation practices. In addition, assuming small temporal variations for V , and introducing (11) into (9) yields, upon integration: (12) Equation (12) is used to reproduce the cellaveraged salinity values available at three measurement times by calibrating the initial (natural) salinity of the Upper aquifer in 1948 C 0 , and A M and A S . Once calibrated, Equation (11) can be used to calculate the seawater flux.

Results
GRACE-derived ∆GWS for the period 2002-2016 were downscaled based on the distance to the centroids of the pumping areas and on the pumping rates as reconstructed in 2009 (see Section 3.1, Figures 2b  and 6a for locations and rates) to obtain the same resolution as the AAR data from the CRU (0.5°× 0.5°, Figure 6c). Downscaling produces realistic groundwater depletion especially in the region of Tripoli concentrating about one half of the Libyan withdrawals as shown in Table 1 and Figure 6b using GRACE-SH and GRACE-M data. A better agreement between downscaled GRACE-derived ∆GWS values and ground-based estimates ∆GWS GWL is obtained using GRACE-SH than GRACE-M as shown by the Root Mean Square error (RMSE) values in Figure 7. The discrepancy is mostly related to cell 13 (Tripoli) with a large underestimation. For groundwater budget calculations proposed hereafter, the average values ∆GWS A calculated as weighted mean values of ∆GWS SH and ∆GWS M using the MSE as weighting factors were adopted. The calculated domainaveraged value (using ∆GWS A in Table 1) for ∆GWS over the 0.5°square grid is about −8.0 ± 0.2 mm·yr −1 , which confirms that the depletion shown by the piezometric levels decline. GRACE data points to an overall aquifer depletion with ∆GWS values varying in space between −1.9 and −22.6 mm·yr −1 . The highest depletion rate is observed in the Libyan part (Tripoli area).
The transient salt balance for the 0.5°× 0.5°cell corresponding to Tripoli (cell 13 in Figure 5b, see Section 3.3) was calculated to obtain an independent estimate of the seawater flux at the shoreline. This flux is required to enable solving the linear system (8) of groundwater mass balance equations accounting for the downscaled ∆GWS values. Combining the mass balance equations for cells 13 and 21 (Figure 5b) yields: where R D A (mm·yr −1 ) is the leaks from the domestic water system that recharges the Upper aquifer, which is considered only for cell 13 corresponding to Tripoli (near 1.5 × 10 6 inhabitants). According to CEDARE [2014], these leaks represent 43 ± 4% of the 146 × 10 6 m 3 ·yr −1 of domestic water supplied to the Tripoli area by the GMMR. This corresponds to an W only corresponds to irrigation pumping. Solving (13) requires identifying the fraction of irrigation return flow f R . For this purpose, we consider the available estimates for the irrigation efficiency: for the Libyan Djeffara, where Q D is the plant water demand while Q W is the water effectively supplied (pumping). According to Abdudayem and Scott [2014], E irr is about 0.64 which is in agreement with other estimates observed in southern parts of Libya [Shaki and Adeloye, 2006]. Given the fact that water supplied satisfies the evapotranspiration demand of cultivated plants, the water provided in excess infiltrates except a certain fraction which is evaporated from the soil surface. Quantifying this evaporation from the soil surface is not straightforward despite its pivotal importance for water balance calculations [Dewandel et al., 2008]. Here, we consider a simple calculation derived from the method proposed by Allen et al. [1998] in an irrigation context. Irrigation water saturates a soil surface layer whose thickness is the excess irrigation (in m) divided by the porosity. A fraction of this saturated layer corresponding to the drainage porosity S y is immediately removed by vertical drainage due to gravity, the remaining part evaporates. Assuming that evaporation can bring the soil moisture of top soils to almost zero (residual porosity is evaporated), our simple model writes: where ω is the total porosity. More elaborated models are available, but require additional parameterization such as unsaturated flow coefficients [Dewandel et al., 2008, Or andLehmann, 2019]. In the Tripoli area, Allemmoz and Olive [1980] reported a value for S y of between 5 and 10%. Considering this range as the 95% confidence interval leads to S y = 7.50 ± 1.25%. Together with a value of 10% for ω [Sadeg and Karahanoglu, 2001], this S y estimate leads to f R = 27 ± 4% consistent with the value of 30% considered by Sadeg and Karahanoglu [2001] in their modeling study. The volume V of groundwater in cell 13 for the Upper aquifer was calculated at 3.87 × 10 10 m 3 using ω = 10%, a mean thickness of 150 m [OSS, 2006], and a cell surface area of 2580 km 2 . We considered an uncertainty of 20% for V in particular to account for the error associated with the assumption of a time constant volume (see discussion). This value for V was introduced in (12) to reproduce the average salinity evolution from 1948 to 2000 as shown in Figure 8. Then, upon calibration of C 0 (0.63 g·L −1 ), A S (2 m 3 ·s −1 ) and A M (1.1 m 3 ·s −1 ), an average value of 1.0 ± 0.2 m 3 ·s −1 for the seawater intrusion in mesh 13 for the period 2002-2016 was calculated using (11). This value corresponds to F 13 y,u = −13.2±2.8 mm·yr −1 (discharge divided by surface area of cell 13 and negative sign due to incoming flux orientation). Introducing the value for F 13 y,u into (13) yields α = 2.2 ± 1.4% using the average value for Q 13 W (irrigation only) of 332 × 10 6 m 3 ·yr −1 for the period 2002-2016 obtained using the exponential fit of Figure 3 (Equation (10) with A = 13.3 m 3 ·s −1 and β = 0.05 yr −1 ). Now that α is known, the full linear system of groundwater mass balance equations can be solved for the coastal fluxes (red arrows in Figure 5b). The results are given in Table 2.

Discussion
Several simplifying hypotheses supporting our approach must first be discussed. The downscaling approach as developed here requires boundary fluxes for a given domain described by GRACE cells to be low in comparison with the internal withdrawals. This requirement is satisfied using GRACE-SH data, the resolution of which allows obtaining a domain close to the real geographical extent of the Djeffara characterized by low boundary fluxes. Using coarser GRACE-M grids which extend beyond the Djeffara  (8) are positive when oriented in the same direction than x or y axis (see Figure 5). * Value between parentheses is the domestic water leaks considered only for the Tripoli area. Coastal cells are indicated in bold character. The boundary fluxes F CI (cell 1) and F T (cells 18-21, 14; see Figure 5) are 18.9 × 10 6 and 16.4 × 10 6 m 3 ·yr −1 , respectively.  1950, 1960, and 2000[OSS, 2006.
limits and more precisely in the NWSAS domain raises the question of the boundary fluxes. According to the piezometric maps for the NWSAS [Baba-Sy, 2005, Besbes andHorriche, 2008], the net fluxes at the southern and western boundaries of Mas-con#1 ( Figure 2b) are almost zero for the CT (orthogonal hydraulic head contours). Regarding the CI, the piezometric map points to an entering flux at the western boundary of Mascon#1 corresponding to the outflow of the CI in the Djeffara plain estimated at 0.6 m 3 ·s −1 (Section 2.3). Therefore, the neglected net boundary flow for Mascon#1 represents no more than 0.2 mm·yr −1 considering its surface area. For Mascon#2, the Nafusa represents a recharge area for both the CI and CT with a cumulated value at 1.7 m 3 ·s −1 calculated by Baba-Sy [2005]. This would 313 represent a neglected net boundary outflow of less than −0.5 mm·yr −1 for Mascon #2. For both GRACE version (M and SH), considering small boundary fluxes is thus a reasonable assumption. In case of non-negligible boundary fluxes, the application of the proposed approach is more complex and requires additional steps: (i) identification of the net boundary flux, (ii) subtraction of this flux from the GRACEderived ∆GWS signal to isolate the purely internal ∆GWS, (iii) application of the proposed downscaling approach and, (iv) addition of the net boundary flux to downscaled ∆GWS values.
A salt balance provides coastal flux estimates, assuming a constant volume of groundwater in the phreatic aquifer for cell 13 (Tripoli) along the simulation extending over the period 1948-2000. As active recharge and withdrawals are at work, this may represent a rough approximation. For the calculation period , the average pumping rate is 267×10 6 m 3 ·yr −1 , of which about 35% (average value between domestic leaks and irrigation return flow, f ) returns to the aquifer (93 × 10 6 m 3 ·yr −1 , about 80% comes from irrigation and 20% from domestic water leaks). Recharge is approximately 15 × 10 6 m 3 ·yr −1 , average seawater intrusion 22 × 10 6 m 3 ·yr −1 , and the flux out of cell 21 is about 40 × 10 6 m 3 ·yr −1 (temporal average values of Q M and Q S using (11)). The net annual average water withdrawal from cell 13 is therefore 97 × 10 6 m 3 ·yr −1 which leads to a net volume extracted from the phreatic aquifer over the 52 years of simulation of 5.0 km 3 representing a groundwater storage variation of about 13%. This is consistent with the 20% uncertainty considered for V in the calculation of seawater intrusion flux. In this study, we considered only principal flow directions, i.e., east to west in Tunisia and south to north in Libya, which makes it possible to solve the discrete linear system of mass balance equations. This clear simplification is however relatively well supported by piezometric data [e.g. OSS, 2006, Sadeg and Karahanoglu, 2001, Trabelsi et al., 2012. Although adapted to the overall objective of fluxes assessment using parsimonious modeling approaches and to the resolution of the rainfall data, the mesh may appear coarse. However, noting that the nine pumping zones are mostly aligned along the 450 km of the shoreline, the average spacing of hydraulic divides associated with pumping is about 50 km which corresponds to the grid resolution. This resolution enables to compensate for the fact that the pumping wells are actually located around the centroids which were conveniently used here for downscaling and mass balance calculations. Even though small portions of the Djeffara plain in the south of the domain covering no more than half a grid cell were excluded, we believe that the spatial resolution does not significantly affect the results. As mentioned in Section 3.1, additional downscaling and thus grid refinement would have been especially beneficial for the area of Gabès (Figure 1). However, Chulli et al. [2007] estimated the SGD in the gulf of Gabès at 0.2 m 3 ·s −1 (0.1 m 3 ·s −1 for half of the coast length) using Darcy's calculations for the Upper aquifer. This relatively low value also obtained in a recent modeling study by Vernoux et al. [2020] justifies maintaining the grid resolution as adopted here.
Regionalized groundwater mass balance proposed here and based on GRACE and ground data provides valuable estimates for recharge and coastal fluxes. Our approach leads to a natural recharge representing 2.2 ± 1.4% of the rainfall which corresponds to a mean value of 4.3 ± 2.7 mm·yr −1 considering the average rainfall in the Djeffara plain at 196 mm·yr −1 . This value is relatively consistent with the estimates from hydrogeological modeling at 6.8 mm·yr −1 [OSS, 2006]. In the context of circum-Saharan arid domains, this rate of 2.2 ± 1.4% is also in good agreement with some estimates for the NWSAS [2.2 ± 1.4%, Gonçalvès et al., 2013], and the Nubian Sandstones Aquifers System [1.4 ± 1.7%, Mohamed et al., 2017] using global groundwater mass balance based on GRACE data. The fluxes obtained here enable to propose a global groundwater budget for the Djeffara plain as shown in Table 3. This groundwater budget is characterized by an outflow exceeding the inflow in accordance with the overall groundwater storage depletion shown by GRACE data. In addition to the Tripoli area, seawater intrusion also occurs in the Zarzis-Djerba area (see Figure 1) which is consistent with geochemical evidences from Trabelsi et al. [2012]. In Libyan Djeffara, a lower seawater intrusion is calculated for cell 13 in comparison to cell 12 due to the domestic water leaks considered only for the most populated area (cell 13 Tripoli; see (13) and Table 2) while the remaining terms of the groundwater budget (including pumping) are rather similar. The overall seawater intrusion amounts to 193 × 10 6 m 3 ·yr −1 which is larger than the 34.7 × 10 6 m 3 ·yr −1 estimated from Uncertainty propagation using the error on ∆GWS A , α and f R .
the OSS model (2006). Total submarine groundwater discharge (SGD) is 339.6 × 10 6 m 3 ·yr −1 which is also higher than the value at 186×10 6 m 3 ·yr −1 found with the OSS model (2006). These discrepancies can be explained by a more important recharge deduced here (10.2 mm·yr −1 including an artificial recharge) compared to the model by OSS (2006; 6.8 mm·yr −1 only comprising the natural contribution). This higher recharge, if taken into account in the hydrogeological model, would lead to larger lateral drainage (SGD) and impact the model calibration probably affecting the calculated seawater intrusion fluxes. Distributed along the 450 km long coastline, this corresponds to a shore-length-normalized SGD flux 7.5 × 10 5 m 3 ·km −1 ·yr −1 . This is one of the first semi-experimentally derived SGD estimates for the north African coastline, a region that presents a notorious data gap in coastal hydrological data [e.g., El-Gamal et al., 2012; see Figure 1 in Rodellas et al., 2015]. This value compares well with model-derived estimates for freshwater SGD on the scale of the entire Mediterranean basin of 8.1 × 10 5 m 3 ·km −1 ·yr −1 [Zektser and Everett, 2004] and 1.1 × 10 6 m 3 ·km −1 ·yr −1 [Zektser et al., 2007], perhaps indicating that the aquifer system under investigation may be representative for the semi-arid Mediterranean basin. Indeed, the few studies carried out in the southern/eastern Mediterranean regions with similar semi-arid conditions also yield comparable results, e.g., 1.8 × 10 6 m 3 ·km −1 ·yr −1 [Weinstein et al., 2007]. This estimate obtained for Israel's Carmel Coast using a combination of radon and salinity as tracers is a factor 2 of the flux estimates obtained here for the Djeffara. Larger scale estimates obtained from radium isotope mass balances such as those reported by Rodellas et al. [2015] ranging from 4.0×10 5 m 3 ·km −1 ·yr −1 to 3.7×10 8 m 3 ·km −1 ·yr −1 (average 3.7 × 10 8 m 3 ·km −1 ·yr −1 ) cannot be compared in a meaningful way to the above freshwater SGD fluxes because radium-derived estimates include both freshwater SGD and (saline) seawater recirculation flux, and are correspondingly significantly larger. Although a direct comparison between our results and the radium-derived estimates is difficult, several objective reasons can however support the hypothesis of lower SGD in the southern Mediterranean domain: more extensive groundwater exploitation, lower recharge rate (recharge to AAR ratio), and lower annual precipitation. All these factors contribute to a relatively modest lateral drainage in the aquifers and therefore to lower SGD values. The groundwater exploitation results in water being removed from the lateral natural groundwater flow to the coast. For the Djeffara plain, the net water withdrawn corresponds to the pumping minus the irrigation return flow, i.e., about 70% of the exploited volumes or about 0.7 km 3 ·yr −1 . This represents 1.56×10 6 m 3 ·yr −1 ·km −1 suggesting a still low "natural" SGD value. A recharge ratio of between 23 and 52% (average 35%) for karstified carbonates [Allocca et al., 2014, Arfib and Charlier, 2016, Baudement et al., 2017, Martos-Rosillo et al., 2015, Polemio, 2016, Zagana et al., 2007, and between 10 and 21% (average 16%) for detritic aquifer formations [Séraphin et al., 2016, Yagbasan, 2016, Zagana et al., 2007 were reported in the north Mediterranean area (Spain to Turkey). Given that karstic aquifers represent almost 20% of the coastal areas in Mediterranean Europe [Goldscheider et al., 2020], an average value of about 20% for the recharge ratio, therefore tenfold higher than in the southern part of the Mediterranean domain (2 to 3%), can be considered. In addition, the average annual rainfall is generally two to three times higher on average in the coastal areas of north Mediterranean Sea. Together with the recharge ratio, this could readily explain lateral groundwater drainage toward the coasts and thus SGD values one order of magnitude higher in the north Mediterranean domain in comparison to southern parts. To date, SGD studies are primary carried out on a local scale (beach/embayment). There is an increasing interest in upscaling these coastal fluxes to, e.g., basin scale [e.g. Rodellas et al., 2015]. These large-scale estimates have their own limitations, e.g., the choice of approach which in the case of Rodellas et al. [2015] does not allow a determination of the freshwater fraction of SGD which is of interest to hydrological studies. Based on the results of this study, we propose that the correct unit to report SGD fluxes may be the catchment/aquifer scale as investigated here. It is important to note that in our GRACE-derived estimate, no SGD flux occurs from 3 out of 7 coastal cells of the Djeffara aquifer investigated in this study (Table 2). This spatial variability has to be considered when upscaling, and we suggest that GRACE data could be used in combination with local "point measurements" with other approaches to provide more accurate upscaling estimates in future studies elsewhere.

Conclusions
In this study, we proposed a downscaling approach for the GRACE-derived groundwater mass storage variations based on the pumping wells spatial distribution and rates of exploitation. In arid zones, where groundwater is the main or the only water resource, pumping is often the dominant term of the hydrological budget. This approach leads to a realistic spatial distribution of the groundwater storage variations averaged over the period 2002-2016 with lower values in the areas where large drawdown of the piezometric levels are reported, such as the Tripoli area. Downscaled groundwater storage variations were then introduced in a set of 21 discrete groundwater balance equations over the finer 0.5°× 0.5°(2580 km 2 ) grid to obtain new estimates of the recharge and coastal fluxes (outflowing or incoming) independently from any hydrogeological model. Due to indetermination issues, the saline intrusion rate in the Tripoli region was identified separately using a local salt mass balance calculation. According to our calculations, the Tripoli and Djerba-Zarzis areas are subjected to a strong seawater intrusion. The remaining coastal zones produce a cumulated groundwater outflow (SGD) of 339.6 × 10 6 m 3 ·yr −1 for the 2002-2016 period. This geophysically based value fills an important gap in SGD estimations for the southern Mediterranean coastal area which remains unquantified to date, including by radiochemical tracers widely used elsewhere. The average recharge over the Djeffara plain was estimated at 4.3 ± 2.7 mm·yr −1 or 2.2 ± 1.4% which is consistent with the recharge to average annual rainfall ratio found for other Saharan aquifers. The GRACE-derived domain-averaged water budget which is obtained here, i.e., overall groundwater depletion at −8.0 ± 0.2 mm·yr −1 , not surprisingly confirms the unsustainable mining of groundwater in the Djeffara plain. Available options to stop this unsustainable mining include pumping limitations to prevent additional salinization effects, use of underexploited resources such as the Jurassic aquifer at the border of the Djeffara, exploitation of remote but unsustainable resources (GMMR), irrigation efficiency increase, and seawater desalinization. These are all ongoing (or future) options under consideration to improve water uses and access to water. Here, we have taken advantage of the alignment of the main groundwater flow directions (toward the coast) with the GRACE grids. This favorable situation can be found for other Mediterranean coastal aquifers enabling a similar approach of distributed mass balance using satellite gravity.