ECCO version 4 release 4 (v4r4) is an ocean state estimate that integrates the Massachusetts Institute of Technology General Circulation Model (MITgcm) (Marshall et al., 1997; Adcroft et al., 2004) with a wide array of observational data (Forget et al., 2015; ECCO Consortium et al., 2021). In this framework, observations are assimilated using an optimized least-squares method, ensuring dynamic and kinematic consistency without artificial heat or buoyancy sources (Heimbach et al., 2005; Wunsch et al., 2009; Wunsch and Heimbach, 2013). Argo temperature and salinity profiles and other CTD profiles from world ocean Database are used to constrain the ECCO v4r4 solution (Fukumori et al., 2017). At the sea surface, ECCO is constrained by forcing derived from the ERA-Interim reanalysis dataset (Dee et al., 2011). ECCO v4r4 spans the period 1992–2017, with a global domain and 50 vertical layers. Its resolution is 1° zonally and varies meridionally, from 1/3° at the equator to 1° at midlatitudes. For this study, we focus on the period from 2003 to 2017, during which the quality of observational data is significantly improved, enabling robust analysis of salinity, surface freshwater flux, and bottom pressure changes.

Although many previous studies have demonstrated that ECCO estimates reliably represent in situ measurements for salinity, sea level, and other variables (e.g., Stammer et al., 2004; Wunsch and Heimbach, 2007; Liu et al., 2019), this study seeks to further reduce uncertainties specifically associated with the Mediterranean Sea region. The salinity field is from the UK Met Office Hadley Centre EN4.2.2 (Good et al., 2013). EN4 is an objectively-analyzed monthly dataset, covering the period from 1950 to 2016, with a horizontal resolution of 1 degree and a vertical resolution of 42 levels. It has been reported that EN4 displays some spurious salty bias after 2015 (Ponte et al., 2021; Liu et al., 2024). However, it has not been observed in the Mediterranean area (Liu et al., 2020).

The precipitation product used is the latest version, 2.3, from the Global Precipitation Climatology Project (Huffman et al., 2009). The GPCP product is created by combining various satellite and gauge-based datasets to form a coherent spatial and temporal representation. The evaporation data is obtained from the OAFlux (Yu et al., 2007), which is generated using an objective analysis method that integrates satellite and atmospheric reanalysis output, and calculates global surface fluxes using the state-of-the-art bulk flux parameterizations.

The study period (2003–2017) was selected to align with the onset of widespread Argo float deployments, which largely enhanced observational coverage and data quality in the Mediterranean Sea. While Argo data are not directly utilized in this study, they underpin the ECCO solution and contribute to the EN4 dataset, ensuring greater reliability and consistency in the input data during this period.

The area-averaged timeseries from these datasets were compared with the monthly mean ECCO anomaly (Fig. 2). Overall, the surface freshwater flux term of ECCO strong align with the observational P-E, with a correlation of 0.92 for surface freshwater flux and R2 (the proportion of variance explained by the seasonal cycle) of 0.72. Discrepancies are observed, particularly during the winter months, where ECCO values are lower by 0.01–0.02 Sv compared to observational-based estimates. This discrepancy could come from the river runoff term R, which is incorporated into the total freshwater flux. In ECCO, river runoff is derived from observed seasonal climatology, applied as a mass flux over several surface grid cells near river mouths (Fekete et al., 2002; Stammer et al., 2004; Feng et al., 2021). Additionally, ECCO's freshwater flux estimates may also be influenced by precipitation biases in the ERA-Interim (Turuncoglu, 2015; Grist et al., 2016).

For salinity, ECCO shows good agreement with EN4, with a correlation of 0.70 and similar long-term trends of approximately 0.01 per year. However, differences arise at seasonal and shorter timescales. ECCO's seasonal cycle has an R2 of 52 %, which is substantially higher than EN4's (less than 20 %). To provide further context, the temporal variability and long-term trend of ECCO's basin-mean salinity time series generally align with previously reported estimates (e.g., Jordá et al., 2017b; Llasses et al., 2018). More recently, Aydogdu et al. (2023) presented an ensemble mean for the top 300 m of the water column, and despite this depth difference, the ECCO results fall within the spread of their multi-product uncertainty range.

The spatial patterns of the time-mean values are also compared. Overall, the time-mean patterns from ECCO align reasonably well with the observational data products. The correlation for freshwater flux and salinity patterns ranges between 0.6 and 0.7 (p < 0.01), respectively. Overall, ECCO captures the primary features of the observed oceanic variables in the Mediterranean Sea with reasonable accuracy (Fukumori et al., 2007; García-García et al., 2010; Soto-Navarro et al., 2010; Calafat et al., 2012). This gives us confidence in using ECCO for salinity and mass budget analyses in the region.

Below we provide a brief summary of the salinity and mass budget analyses with the ECCO estimates. In general, our approach is very consistent with other budget analysis using ECCO (e.g., Tesdal and Abernathey, 2021; Siddiqui et al., 2024), which benefit greatly from the provided diagnostic terms. Details on how to close the budgets are provided in Forget et al. (2015) and Piecuch (2017).

In the Mediterranean Sea, significant spatial differences are observed at the sub-basin scale (Bonaduce et al., 2016; Mohamed et al., 2019). This spatial variability is caused by the complex thermohaline changes and local circulation (Menna et al., 2012; Mauri et al., 2019; Menna et al., 2019; Poulain et al., 2021). Compared to other regional models (Escudier et al., 2021; Meli et al., 2023), ECCOv4r4's coarse resolution may limit its ability to confidently resolve these sub-basin and mesoscale processes, and it is unclear how precise the flux estimates must be to ensure accurate simulations and avoid potential model drifts. Additionally, narrow straits such as the Dardanelles are not explicitly resolved as lateral boundary fluxes in ECCO but are instead incorporated within the surface flux term, along with river runoff and other unresolved processes. This introduces some uncertainty in the precise partitioning of boundary inputs. Therefore, in this study, we focus on the Mediterranean basin as a whole, rather than attempting to resolve the sub-basin processes in detail. More discussion on the model resolution and uncertainty will be discussed in later section.

The budgets for the whole Mediterranean Basin can be described as a simple box model. The balance in the Mediterranean Sea budgets is between two boundary terms, the surface flux and the inflow through the Strait of Gibraltar. Since ECCO uses volume-conserving Boussinesq approximation (Greatbatch, 1994; Marshall et al., 1997), at each grid, the equation for mass budget of the whole water column can be estimated as: 1ρsw∂Pb∂t≈-∇(ρswu)+ρfwFfw where Pbis the bottom pressure equivalent water thickness, u is the horizontal velocity, Ffw is the surface freshwater flux (P+R-E). For the benefit of discussion, we do not separate the individual sources of freshwater inputs (i.e., P+R are treated as a single term). ρfw and ρsw are the freshwater and seawater density, respectively.

The left-hand side represents the tendency term of bottom pressure, i.e., the rate of OBP change, the first right-hand side term is the convergence of seawater, and the second term marks the contribution of surface freshwater flux.

Given the Mediterranean Sea's semi-enclosed nature, when integrating Eq. (1) over the entire basin, the integral of the convergence term naturally equals to the net influx through the strait, and the integral of the second term represents the total freshwater flux at the sea surface. Vertical exchanges between surface and deep Mediterranean layers are implicitly represented in ECCO but not explicitly analyzed in this study.

At the Strait of Gibraltar, we divided the water column into two layers to estimate the inflow and outflow: the upper layer (0–150 m) represents the transport of the AW (eastward, Fig. 3a), while the lower layer (150 to ∼ 320 m, model bottom) represents the MOW (westward, Fig. 3b).

The salinity budget in the Mediterranean Sea is balanced in a similar form. The local salinity conservation can be simplified as: 2∂S∂t=-∇Su+Ds+FfwS̃ where S̃ is the local surface salinity, and Ds represents the subgrid-scale processes parameterized as mixing (diffusive salt flux). The sum of the first two terms on the right-hand side describes the total flux of oceanic transport, which, when integrated over the entire basin, corresponds to the net influx of salinity through the strait.

The first term on the right-hand side, which represents the advection of salinity, includes two distinct physical processes that contribute to salinity changes (Piecuch, 2017): one process represents the overall dilution/concentration due to the convergence/divergence of the mass transport (S∇u), while the other reflects the exchange of salt content carried by the advective flow (u∇S).

It is important to note that ECCO-v4 solutions are based on the MITgcm model with the z* vertical coordinate system (Adcroft and Campin, 2004), which allows vertical grid-cell thickness to stretch or compress according to changes in sea surface height. Additionally, freshwater is added or removed as an explicit surface mass flux (Campin et al., 2008). This approach ensures that changes in salinity directly reflect mass input and removal, with the model naturally accounting for dilution and concentration effects. These characteristics support the physical consistency of ECCO estimates and justify our use of diagnosed mass fluxes in interpreting salinity variability.

The focus of this study is on the interannual variability of the fluxes. The non-seasonal signal is obtained subtracting the climatology from the original timeseries instead of a fitted annual sinusoid, since the annual variation is not always sinusoidal in shape (García-García et al., 2022). This approach is consistent with previous studies using ECCO to investigate Mediterranean Sea variability and phenomena associated with Gibraltar Strait transport (Fukumori et al., 2007; Menemenlis et al., 2007; Volkov and Landerer, 2015). These studies highlight a strong correlation between OBP and sea level on non-seasonal timescales, providing a more accurate representation of the relationship between mass (OBP) variations and volume (sea level) variations. By our estimation, the seasonal cycle accounts for less than 10 % of the total variance in the mass flux, while the salinity component contributes slightly more, below 30 %.

Here we express mass fluxes in Sverdrups (Sv) for consistency with common oceanographic practice. Salinity flux is represented in g kg⁻¹ Sv, enabling direct comparison between mass and salinity fluxes.

In this section we present the estimates for each flux term of Eqs. (1) and (2) applied to the entire Mediterranean Sea. Figure 4a presents the temporal variability of the Mediterranean Sea's mass budget from 2003 to the end of 2017. In general, the net influx at the Strait of Gibraltar (i.e., inflow plus outflow, labeled “Strait”), shows large month-to-month variability: its standard deviation is 2.4 times larger than the surface freshwater flux (labeled “Surface”).

The mass tendency term, i.e., the sum of Strait and Surface terms, is labeled “Total” and mainly driven by the net influx at the Strait. The Pearson correlation coefficient between Strait and Total terms' timeseries is 0.96 (p<0.01), while there is no significant correlation between Total and Surface terms (r<0.2). The coefficient of determination (R2) of Strait to Total term is 0.92, showing that the net influx explains nearly all of the variance in the total mass tendency. Previous studies have shown that the mass exchange through Gibraltar significantly impacts the Mediterranean, dominating the mean sea-level trend in the region (Calafat et al., 2010; Pinardi et al., 2014).

The interannual variability of both the Strait and Total terms is evidently much smaller on a year-to-year scale compared to month-to-month fluctuations (Fig. 4d). The transport through the Strait of Gibraltar exceeded 0.06 Sv in 2009 and 2017, while the lowest values occurred in 2010 and 2015, around 0.03 Sv. In contrast, the Surface term exhibits very limited interannual variability, with a range of only 0.01 Sv, reaching a maximum in 2014. Similar numbers have been previously reported in the literature; García-García et al. (2022) provided annual mean estimates of net exchange at the Strait of Gibraltar at 0.04 Sv in 2008, and 0.0224 Sv in 2010.

Linear trends of the fluxes were also estimated, showing a small upward trend in both the Strait and Total terms (O(10⁻⁴ Sv per year)). However, these trends are not statistically significant as their 95 % confidence intervals include zero. This indicates that the net transport through the Strait of Gibraltar, as well as the overall mass within the Mediterranean Sea, remained highly stable over the study period. On average, the total mass tendency is close to zero (O(10⁻⁴ Sv)), and the net oceanic influx through the Strait of Gibraltar and the surface freshwater flux appear to be balanced almost simultaneously, with no noticeable lag.

We then examined the cumulative mass flux (Fig. 4b) and the detrended timeseries (Fig. 4c). The cumulative curves provide a clear visualization of long-term trends, where the slope corresponds to the time-averaged flux values. In contrast, the detrended time series isolates shorter-term variability, revealing patterns that are otherwise obscured by the dominant trend. Notably, the magnitude of the detrended series is approximately two orders of magnitude smaller than the cumulative fluxes. Surface and Strait terms exhibit nearly exact opposite variations, with the Surface term showing relatively smaller month-to-month fluctuations. This inverse relationship is expected, as the total mass of the Mediterranean Sea remains nearly conserved (Fig. 4b).

A notable turning point occurs in the winter of 2010/2011. Prior to this, from 2003 to 2010, the surface flux shows a predominantly negative anomaly, followed by a subsequent seven-year period of positive anomalies. The Strait term exhibits the exact opposite pattern, with positive anomalies before 2010 and negative anomalies afterward. Notably, both fluxes tend to contribute to the removal of freshwater from the Mediterranean basin. Before 2010, the freshwater loss from surface fluxes exceeded the net inflow through the strait by approximately 100 km³. After 2010, the Strait term has a negative anomaly and reaches values as low as -500 km³, showing a larger deficit.

Figure 5 presents the salinity fluxes for the Mediterranean Sea. Unlike the mass budget, the Surface term exhibits considerably larger month-to-month variability in salinity, with a standard deviation 2.5 times greater than that of the Strait term. Moreover, the surface freshwater flux demonstrates a strong linear relationship with the total salinity tendency: the correlation between surface flux and Total salinity tendency is exceptionally high, exceeding 0.97 (p<0.01) with an R2 value of 0.95, indicating that air-sea fluxes alone account for nearly all observed salinity changes. In contrast, the net influx through the Strait of Gibraltar explains only 4 % of the salinity variability, with a much weaker correlation of 0.25. This is because freshwater, which contains no salt, could substantially dilutes the denser, saltier Mediterranean waters when introduced, impacting the overall salinity dynamics within the basin.

A significant positive change was identified in the Strait term, with an estimated rate of 0.02 ± 0.01 g kg⁻¹ Sv per year (Fig. 5d), indicating a weakening of the net salinity flux. As illustrated in Fig. 4, both the inflowing and outflowing water masses exhibit freshening trends over the analysis period. This concurrent decline in salinity across both layers results in a reduction in the mean salinity at the Strait of Gibraltar. Consequently, the net salinity flux through the strait decreases over time. This decline in the salinity gradient reduces the efficiency of salt export from the Mediterranean, assuming the net transport via the strait remain relatively stable (Fig. 4d).

The observed freshening of the inflowing water has been attributed in some studies to Arctic ice melt and changes in North Atlantic circulation patterns (Dukhovskoy et al., 2019; Holliday et al., 2020). However, it is important to note that ECCO does not include tidal forcing, which has been shown to increase net salt export at Gibraltar by approximately 25 % (Sanchez-Roman et al., 2018). This likely contributes to ECCO's relatively low estimate of salt export compared to tidal-resolving models.

Other research also suggests that the salinity of North Atlantic waters may be increasing in recent decades (Bates and Johnson, 2020; Sukhonos et al., 2024). A recent study by Lu et al. (2024) further highlights the complexity of North Atlantic salinity trends at climate scales, noting that climate models tend to underestimate observed salinity increases in the Atlantic basin. These contrasting findings suggest that the salinity characteristics of North Atlantic are still subject to significant uncertainty and debate, warranting further investigation to clarify their role in shaping Mediterranean salinity trends.

In Fig. 5b, we present the cumulative contribution of salinity fluxes to the mean salinity level in the Mediterranean Sea. Over the study period, the mean salinity of Mediterranean seawater shows a steady increase at a rate of approximately 2.2 ± 0.2 × 10⁻³ g kg⁻¹ per year, translating to a total salinity increase of about 0.03 over 15 years. The air-sea freshwater flux, driven primarily by substantial net evaporation, contributes significantly to this trend. On its own, it would have raised the mean salinity by 0.2 over the period, with a rate of 14.0 ± 0.2 × 10⁻³ g kg⁻¹ per year. In comparison, the contribution from the inflow through the Strait of Gibraltar is estimated at -12.1 ± 0.2 × 10⁻³ g kg⁻¹ per year, accumulating to a reduction of 0.17 over 15 years.

The detrended cumulative salinity timeseries are presented in Fig. 5c. Prior to 2010, both terms exhibit positive salinity anomalies, while negative anomalies are prevalent in the years that follow. In contrast, the Strait term maintains a persistent negative anomaly throughout the period, reaching its lowest point in 2011. It is also out of phase with the other flux terms, highlighting its distinct behavior relative to surface-driven salinity changes.

In this section, we examine the influence of large-scale atmospheric teleconnection patterns on Mediterranean salinity variability and water mass exchange at the Strait of Gibraltar. Specifically, we consider the most prominent modes of atmospheric variability in the North Atlantic/Europe region, i.e., the North Atlantic Oscillation (NAO), the East Atlantic (EA) pattern, and the East Atlantic/Western Russia (EA/WR) pattern, which are known to modulate regional atmospheric pressure systems, wind forcing, air-sea fluxes, and hydrographic properties in the Mediterranean basin.

The NAO has a significant impact on both weather and oceanic circulation (e.g., Hurrell, 1995; Vigo et al., 2011). Particularly relevant to this study, it modulates wind patterns near the Strait of Gibraltar, affecting non-seasonal water mass transport and variations in Mediterranean sea levels (Fukumori et al., 2007; Menemenlis et al., 2007; Landerer and Volkov, 2013; Piecuch and Ponte, 2014).The NAO also influences the Mediterranean region across multiple timescales, from seasonal to interannual and decadal (Gomis et al., 2008; Tsimplis et al., 2008; Calafat et al., 2010; Calafat et al., 2012). These processes offer a framework for understanding the relationship between the NAO and the interannual variability of the observed timeseries of mass and salinity fluxes in the Mediterranean.

The NAO index reveals that the NAO remained in a positive-to-neutral phase from approximately 2003 to 2008, shifted to a negative phase from 2008 to 2011, and then returned to a positive phase afterward. We also calculated annual and winter seasonal averages for the NAO timeseries, as previous studies have shown that the NAO is particularly influential on Mediterranean weather patterns during the winter season (e.g., Castro-Díez et al., 2002).

Mariotti et al. (2002) found significant correlations between the NAO and both annual and winter averages of precipitation and net precipitation across the broader Mediterranean region. In our study, we also observed a significant correlation between the winter NAO and the air-sea mass flux shown in Fig. 4d (blue), with a correlation coefficient of -0.51 (p < 0.05). Regarding salinity fluxes in Fig. 5d, all annual fluxes show significant correlations with the winter NAO. Both the Surface and Strait terms are correlated with the winter NAO between 0.5 and 0.6 (p < 0.05), while the Total term, i.e., the sum of the Surface and Strait terms, correlates at 0.62 (p < 0.01).

The EA pattern, often considered the second mode of atmospheric variability over the North Atlantic, is characterized by a north-south dipole of pressure anomalies similar to the NAO but displaced southeastward (Barnston and Livezey, 1987; Wallace and Gutzler, 1981). It influences storm tracks and surface pressure fields across western and central Europe and the North Atlantic. Although the EA pattern has been associated with regional climate variability (e.g., evaporation and surface pressure, Zveryaev and Hannachi, 2012), our analysis reveals no significant correlation between the EA index and the Mediterranean salinity or water mass fluxes, whether based on annual mean or winter mean values.

In contrast, the EA/WR pattern exhibits a quasi-stationary wave-train pattern extending from the North Atlantic into western Russia, with centers of action over northern Europe and western Russia (Barnston and Livezey, 1987; Lim, 2015). This mode strongly affects the Mediterranean climate through its modulation of high-pressure anomalies and upper-level atmospheric flow (Krichak et al., 2002; Josey et al., 2011). In our results, the winter mean EA/WR index shows a strong and statistically significant correlation with both salinity flux (r= 0.71, p<0.05) and mass flux (r= 0.72, p<0.05) at the surface, suggesting that it may play a critical role in shaping the interannual variability of water mass transformation and transport in the basin.

This section provides a summary of the estimated mass and salinity budgets, broken down by their components as represented in Eqs. (1) and (2), along with associated uncertainties for the Mediterranean Sea over the period 2003–2017 (Table 1). Since these estimates are derived using ECCO, the budget calculations are inherently balanced. The budget diagram shown in Fig. 7 is on the mass balance (top) and the salinity imbalance within the Mediterranean basin (bottom). Over this 15-year period, the long-term mean water mass change is calculated at 0.001 ± 0.018 Sv, while the salinity change is estimated at 0.29 ± 0.09 Sv. This reflects a balanced state in terms of mass, contrasted with a slight imbalance in salinity. These differences arise from the distinct roles of surface processes, which exchange freshwater, and oceanic exchange of saltwater through the Strait of Gibraltar.

Starting with the mass budget, which is relatively straightforward: as outlined in the Data and Methods section, at each interface (air-sea and the Strait of Gibraltar) we can break the fluxes down into two primary processes. At the surface, these processes are evaporation and precipitation (including runoff). Evaporation results in a mass loss of -0.11 ± 0.01 Sv, while precipitation adds a mass gain of 0.06 ± 0.01 Sv. This leads to a net mass loss of -0.05 ± 0.01 Sv from all air-sea freshwater fluxes.

At the Strait of Gibraltar, the two processes are the inflow of AW and the outflow of MOW. The inflow through the upper layer brings in 0.68 ± 0.02 Sv of AW, while the outflow in the lower layer exports -0.63 ± 0.02 Sv of MOW. The net result is a modest positive mass gain, on the order of one magnitude smaller than either of the individual flows. Overall, the Mediterranean Sea gains water mass through net inflow at an average rate of 0.05 ± 0.02 Sv, which closely aligns with values reported in the literature. For instance, Jordà et al. (2017a) derived a similar number at 0.065 ± 0.033 Sv from nearly 20 independent observational estimates.

Regarding the salinity budget, as noted previously, salt and salinity transport are only meaningful for understanding freshwater fluxes when the total mass transport within a closed region is zero (Tsubouchi et al., 2012; Schauer and Losch, 2019). This is an ideal condition in our case, as we have already established a mass balance within the semi-enclosed Mediterranean basin. Meanwhile, evaporation increases the salinity level by removing freshwater, resulting in a concentration effect equivalent to 4.27 ± 0.11 Sv. On the other hand, precipitation dilutes the seawater, leading to a salinity reduction of -2.46 ± 0.11 Sv. Together, these air-sea fluxes contribute a net positive of 1.80 ± 0.10 Sv.

In contrast to the surface processes, the salinity exchange through the Strait of Gibraltar operates in a more complex manner. As outlined in the methods section, salinity flux through the strait can be split into two main components: one term represents the advection of salinity due to mass transport (u∇S), while the other captures changes in salinity caused by the differing salinity levels of the inflowing and outflowing water masses (S∇u), which lead to salinification or dilution. Starting with the salt content transport u∇S, ECCO estimate a net salt flux into the Mediterranean through this exchange at 0.30 ± 0.20 Sv. This is driven by the exchange between the AW and the relatively saltier MOW at the Strait of Gibraltar. Specifically, the AW brings in 25.21 ± 0.18 Sv of salt content, while the MOW carries out -24.91 ± 0.19 Sv.

We then consider the dilution and concentration effects resulting from the introduction of AW and MOW at the Strait of Gibraltar, which is driven by the density differences between AW, MOW, and the average Mediterranean seawater. In this context, assuming the total salt content within the Mediterranean Sea remains constant, the addition of AW (with relatively lower salinity) and the removal of MOW (with higher salinity) lead to a net dilution effect due to changes in total mass and density; specifically, the water exchange at the Strait of Gibraltar would reduce the salinity by approximately -1.78 ± 0.03 Sv due to a net dilution effect. After the adjustment, the net influence of the water exchange at the Strait of Gibraltar shifts from a net gain in salt content to an overall reduction in salinity, estimated at -1.48 ± 0.20 Sv. This hilights the critical role of mass/density changes in the Mediterranean Sea's salinity budget, which significantly impact the balance of salinity in this semi-enclosed basin.