Derived seawater carbonate system parameters, using pairs of measured carbonate variables were estimated using the software program CO2SYS (version 3.0, Pierrot et al., 2021) considering temperature, salinity and nutrient concentrations (silicate and phosphate) and with the output from CO2SYS at in-situ temperature conditions. The pH data from the pH sensor collected at in-situ temperature were converted to pH at 25°C using the CO2SYS program and AT estimated from salinity (see below).

Carbonate variables were estimated using pairwise combination of carbonate variables (using CO2SYS, version 3.0, Pierrot et al., 2021). For estimations of AT, CT and pH (using pH&CT, pH&AT, and AT&CT, respectively), we followed the recommendation by Álvarez et al. (2014) based upon a large Mediterranean pH&AT&CT dataset: carbonic acid dissociation constants K1 and K2 from Mehrbach et al. (1973) as refitted by Dickson and Millero (1987) and the dissociation constant for HSO₄ ^- from Dickson (1990). The Root Mean Square Deviation (RMSD) between observed and estimated values of AT, CT and pH were 19.5 μmol kg^-1, 17.6 μmol kg^-1 and 0.033 pH units, respectively. For estimations of pCO₂ (using AT&CT, pH&AT, pH&CT), various CO2SYS program options were tested (as pCO₂ measurements were not performed by Álvarez et al., 2014) to identify the K1, K2 option that gave the lowest ΔpCO_{2 obs-est}. Although the present study is the first in the Mediterranean Sea measuring four variables: AT, CT, pH (bottle data) and CO₂ (sensor data), and thus could in principle be used to examine internal consistency, this was attempted but abandoned due to several reasons explained in Supplementary Material . Therefore, the same K1, K2 option for estimating CO₂ (using AT&CT, pH&AT) as above (Álvarez et al., 2014) was deemed more appropriate and internally consistent. Using this option the comparison of seawater pCO₂ from the sensor with calculated pCO₂ had RMSD ~29 and 9 μatm when calculated from AT&CT and pH&AT, respectively.

The method proposed by Takahashi et al. (2002), with terminology as adapted by González-Dávila et al. (2016), was used in order to estimate the variability of pCO₂ recorded by the buoy sensor resulting from the contribution of the thermal (pCO_2TH) and non-thermal (pCO_2NT) control using the Equations 1 and 2:

where pCO₂ and SST are the values obtained from HCB seawater sensors and pCO₂mean (410 μatm) and SSTmean (20.58°C) their respective annual mean values.

The air-sea interface CO₂ exchange flux FCO₂, (mmol m^-2 d^-1) calculation was made according to Equation 3 from Wanninkhof (2014)

where K₀ (moles L^-1 atm^-1) is the solubility of CO₂ in seawater (Weiss, 1974), and pCO₂ ^SW and pCO₂ ^ATM are the surface seawater and atmospheric partial pressure of CO₂ (μatm) and k (cm h^-1 by convention Equation 4) is the gas transfer coefficient expressed as

where U₁₀ is the wind speed (m s^-1), Sc is the CO₂ Schmidt number (dimensionless) normalized to 660 (the Sc for CO₂ in seawater at 20°C and salinity 35) and the coefficient 0.251 has units of (cm h^-1) (m s^-1)^-2.

A positive flux indicates a release of CO₂ from the ocean toward the atmosphere.

Fluxes of CO₂ were computed using pCO₂ ^SW from HCB, pCO₂ ^ATM data from nearest (54 km distance) land-based coastal atmosphere observation station of FINOKALIA and wind data from HCB (measured every 3 h). Description of FINOKALIA station (https://www.icos-cp.eu/observations/national-networks/icos-greece) and its air CO₂ time-series can be found in Gialesakis et al. (2023). Wind data from the nearest (34 km distance) sea-based meteorological station of E1-M3A were used to complete meteorological data gaps of HCB (air pressure, wind). Description of E1-M3A can be found in Petihakis et al. (2018). Wind speed measurements obtained at 3 m were adjusted to 10 m above sea surface (U₁₀) based on Hsu et al. (1994). Due to a gap of wind data from 07/06/2021 to 12/09/21 (in both HCB and E1-M3A), a constant value of wind speed (5.50 m s^-1) was used during that period, obtained by the average of summer values from the E1-M3A time-series data (ten years).

Subsurface temperature and salinity have been measured at the POSEIDON HCB mooring site since 2016. Between November 2020 and May 2023, the mooring captured seasonal variability in temperature and salinity over the ~2.5 year observation period ( Figure 2 ). There were some minor gaps in the salinity data (mostly <3 day periods), while a larger gap occurred from mid-July to November 2022. Over the ~2.5 year observation period, the SST ranged from 15.3°C in winter to 28.3°C in summer. According to the entire temperature and salinity data record of HCB station (available at https://poseidon.hcmr.gr) the hydrological situation over the studied period showed no extreme climatic events. The drop in salinity observed in May 2022 could be due to the intrusion of a different water mass (see Discussion). The SST satellite product (daily values) showed a consistent relationship with the in-situ data during the entire period ( Figure 2 ). The SSS satellite product (weekly values) difference with in-situ data varied generally between ~0.2 and ~0.4 units.

The variations in Chla concentration over the studied period displayed smooth seasonal patterns (especially 0-120 m mean depth-integrated values), with somewhat higher values in winter-spring (December to May) and lower in summer-autumn (June to November). The highest concentrations were observed during spring 2022 which was the only time during the study period when Chla exceeded ~0.2 µg L^-1 ( Figure 3 ). Mean depth-integrated Chla values during the summer-autumn period were generally greater than surface Chla values, reflecting the deepening of the Chla maximum. The Chla satellite product was generally higher than in-situ surface data with the difference being generally less that ~0.05 µg L^-1.

The variations in inorganic nutrients surface concentrations over the studied period ( Supplementary Material Figure 1 ) were overall low (<0.92 μM for NO₃+NO₂, <0.02 μM for PO₄, <0.88 μM for NH₄, <1.50 μM for SiO₄) without any distinct seasonal pattern.

The area acted as a sink of atmospheric CO₂ during approximately 7.5 months in winter-spring, between mid-October and May, and as source during approximately 4.5 months in summer-autumn, between June and mid-October. In the winter-spring (sink) period, the highest negative flux was -18.0 mmol CO₂ m^-2 d^-1, whereas the summer-autumn (source) period the highest positive flux was +13.1 mmol m^-2 d^-1. A strong variability was observed in both periods (+3.8 ± 2.3 and -3.5 ± 3.5 mmol m^-2 d^-1) with wind being the main driver of the flux variability ( Figure 7 ). To estimate the net flux over the entire source/sink acting periods, sea surface pCO₂ data-lacking periods were completed using pCO_2buoySSS&pH, and the remaining gaps using pCO_2satSST&Chla. Based on this composite pCO₂ dataset, the total flux integrated over the sink period was between -834 and -662 mmol CO₂ m^-2 (over 219 and 234 days, during the first- and second-year period respectively), whereas the total flux integrated over the source period was between +607 and +563 mmol CO₂ m^-2 (over 154 and 144 days, during the first- and second-year period respectively). The estimated annual flux ranged between -0.23 ± 0.02 and -0.10 ± 0.01 mol CO₂ m^-2 y^-1 (during the first- and second-year period respectively).

There are few time-series stations in the Mediterranean that monitor the carbonate system, with the vast majority being in the northwestern Mediterranean and the Northern Adriatic (Hassoun et al., 2022 and references therein; Champenois and Borges, 2021; Wimart-Rousseau et al., 2023). Among these, high-frequency (daily to hourly scale) published data in the Eastern Mediterranean are found sporadically in the Northern Adriatic (e.g. Turk et al., 2010, 2013) although initiatives are maintained (ICOS RI, 2022), and one year in the Saronikos Gulf (González-Dávila et al., 2016). The eutrophic Northern Adriatic is a particular case where the carbonate system in under strong riverine influence and strong benthic pelagic coupling due to its shallowness (Urbini et al., 2020 and references therein), thus not representative of the generally oligotrophic and deep system of the Eastern Mediterranean and therefore we will not include it in the following when mentioning “Eastern Mediterranean”. Time-series in the remaining Eastern Mediterranean are in the South-Eastern Levantine sub-basin (5- to 6-year cycles in Lebanese and Israeli coast, respectively Sisma-Ventura et al., 2017; Hassoun et al., 2019) and one in the Aegean Sea (one-year cycle in Saronikos Gulf; González-Dávila et al., 2016). Furthermore, a fourth one-year cycle in the North-Western Levantine (South of Crete) could be considered a reconstructed time-series out of three cruises and Argo floats casts (Wimart-Rousseau et al., 2021). In fact, the carbonate observing effort devoted to the Eastern Mediterranean, has been essentially with R/V cruises, mostly measuring CT, AT and pH (Wimart-Rousseau et al., 2021 and references therein; Álvarez et al., 2023 and references therein; Hassoun et al., 2022 and references therein). In addition, high frequency (hourly to daily scale) measurements are also scarce in the Mediterranean (Hassoun et al., 2022 and references therein), with the Eastern Mediterranean having before the present study only one annual high frequency cycle of pH (González-Dávila et al., 2016).

Observations at high frequency are important as they provide more information on changes in ocean chemistry essential for the assessment of the marine ecosystem’s responses to warming, eutrophication, and the oceanic CO₂ uptake (Flecha et al., 2015), on how diel variations in plankton production and respiration contribute to the variability of the carbonate system (Honkanen et al., 2021), as well as on developing/validating algorithms and exploring drivers connected with sub-mesoscale phenomena. Measuring from moving platforms may compromise the results as biological diurnal oscillations may not be appropriately accounted for (Honkanen et al., 2021). The present study provides the first marine pCO₂ time series in the Eastern Mediterranean measuring pCO₂ directly and at high frequency (hourly scale) and the second pH time-series cycle also at high frequency (a previous annual cycle done in Saronikos Gulf in 2013; González-Dávila et al., 2016). This allowed for the examination of the drivers of surface pCO₂ and pH temporal variability over a period of almost two years. In the area studied the strong seasonal variability and low values of net air-sea CO₂ flux demonstrate the need of high frequency in-situ observations (including measurements of pCO₂, temperature and wind speed) in order to reduce the uncertainty of the estimation.

Temperature plays a major role in driving seawater pCO₂ and pH variability whereas changes induced by non-thermal processes are considered to be mainly due to biological activity, but could also include changes due to advection, vertical diffusion, and air-sea gas exchanges (Takahashi et al., 1993). The main drivers in determining pCO₂ and pH on continental shelves include water temperature, water mixing, presence of organic matter in river plumes and biological control. At the low latitudes non-upwelling shelves, which are typically oligotrophic, the biological pump is less effective (review by Chen et al., 2013). Therefore, the importance of pCO₂ seawater variability due to biological control is revisited in the Eastern Mediterranean (Wimart-Rousseau et al., 2021). Decomposing the drivers contributing to thermal control (pCO_2TH) and non-thermal control (pCO_2NT) (which in general cancel each other out) is important for understanding which drivers are more important (Ko et al., 2021). In our study area and period, we infer that pCO_2TH was the main factor followed by pCO_2NT due to evaporation, water mixing and biological remineralization-production, as explained below. Temperature is the dominant factor controlling the diel to seasonal variability of pCO₂ (and pH), as previously described by other studies in the Eastern Mediterranean (González-Dávila et al., 2016; Wimart-Rousseau et al., 2020). The observed annual cycle of pCO_2NT was similar to the one described in the North-Western Mediterranean (Merlivat et al., 2018; Wimart-Rousseau et al., 2023), however with biological control (remineralisation-production) and evaporation being the main non-thermal drivers, as suggested for the North-Western Levantine (Wimart-Rousseau et al., 2021). Starting from March, the pCO_2NT decrease from March to May would be associated with photosynthetic uptake (decreasing CT and slightly increasing AT) as indicated by Chla higher values and progressively also associated with evaporation (mainly increasing AT). Then from June to September, the decrease of pCO_2NT would be linked to the effect of AT increase related to evaporation (indicated by the salinity increase) as well as, but to a lesser extent, to autotrophic processes. The effect of remineralization would increase progressively from summer onwards. In fact though the biomass of heterotrophs is higher than autotrophs in this oligotrophic area during summer (Siokou-Frangou et al., 2002), the low amount of organic matter reduces the importance of remineralization in summer. It is only after accumulation of organic matter, with simultaneous decrease of autotrophic and evaporation processes that the increase of pCO_2NT due to remineralization would appear in October (initiating the increase of CT). This increase of pCO_2NT (and CT) would be supported by winter water mixing after December for a period lasting 1 to 3 months (Krasakopoulou et al., 2009; Tsiaras et al., this issue). Then pCO_2NT increase would progressively be slowed down by autotrophs after February, closing the annual cycle. On the other hand, physical factors other than evaporation (i.e. land-based fluxes, precipitation, lateral advection) appear to play a minor role in the pCO_2NT in the study area and period. In fact, land-based water fluxes are generally low as there are only small rivers in the area (Poulos et al., 2009) and as precipitation data (from land-based meteorological station data located in Heraklion city, Lagouvardos et al., 2017) showed no significant rain events during the period of pCO₂ records (as confirmed by the absence of salinity decreases at HCB during rain events). Lateral advection appears to have negligible effect in pCO_2NT at HCB. At this location, depending on the prevailing circulation pattern, local high salinity surface water shares occasionally the surface layers with less saline Modified Atlantic Water (MAW), karst waters (Georgopoulos et al., 2000) or Modified Black Sea Water (MBSW) (Velaoras et al., 2019; Kassis and Korres, 2021). A surface water mass of MAW or MBSW could explain the drop in salinity observed in May 2021 (but not karst waters since AT-SSS relation was not affected). This water mass intrusion did not affect pCO₂.

Marginal seas are often an important sink for atmospheric CO₂ and overall play a significant role in the global carbon cycle, as they transfer carbon to deeper layers, however, not all marginal seas are CO₂ sinks and some function as CO₂ sources (Borges, 2005; Chen et al., 2013; Dai et al., 2013). The current assessment of the whole Mediterranean as source or sink of CO₂ lies mainly on modelling studies that describe it as close to equilibrium with the atmosphere, or at most functioning as a weak sink for atmospheric CO₂ (D’Ortenzio et al., 2008; Palmiéri et al., 2015). In fact, the Mediterranean in-situ studies on pCO₂ variability and air-sea fluxes are based upon a sparse time-series of pCO₂ observations mainly located in the North-Western Mediterranean and the Adriatic Sea. There is lack of surface ocean pCO₂ data with adequate spatial and/or temporal coverage in large areas of the Mediterranean, especially in the Eastern basin and the whole southern part of the basin (Hassoun et al., 2022; Álvarez et al., 2023). In the present study, the pCO₂ time series measurements allowed for estimation of the air-sea fluxes, indicating a transition from a winter-spring CO₂ sink (between mid-October and May) to a summer-autumn CO₂ source (June to mid-October). These periods are in agreement with other field studies providing direct measurements of pCO₂ or pCO₂ calculated from AT-CT in the Aegean Sea in February and May (Krasakopoulou et al., 2006, 2009, respectively), estimates of pCO₂ (by AT-pH) during an annual cycle in the Aegean (González-Dávila et al., 2016) and from reconstructed time-series out of seasonal cruises in the South-East (Sisma-Ventura et al., 2017) and the North-West Levantine Sea (Wimart-Rousseau et al., 2021). Among these studies, those that estimated annual air-sea CO₂ fluxes (González-Dávila et al., 2016; Sisma-Ventura et al., 2017), suggest their area to be a weak CO₂ source, which supports a previous hypothesis for the whole Eastern Mediterranean Sea (Durieu de Madron et al., 2009; Taillandier et al., 2012). The estimated annual flux obtained in the present study between -0.15 and -0.18 mol CO₂ m^-2 y^-1 is close to the one obtained by modelling (Tsiaras et al., this issue) for the Levantine Basin (-0.25 mol CO₂ m^-2 y^-1 for 2010-2022). However, it should be stressed that the area examined in the present study cannot be characterized as a CO₂ sink due to a) the flux being low and b) the significant variability in the flux mostly related to wind speed. This is the case also for a general characterization of the Eastern Mediterranean as source or sink. The few field studies up to now, with direct CO₂ measurements, suggest the flux being weak (either sink or source) while they are limited temporally to the present time-series and spatially to few cruises (Krasakopoulou et al., 2006; Krasakopoulou et al., 2009; Bakker et al., 2016) with most CO₂ data being estimates from other carbonate system parameters (review by Álvarez et al., 2023). We hypothesize that most areas in the oligotrophic Eastern Mediterranean Sea currently act as either weak source or sink of CO₂. This could be a transition period and with the rapid warming and increasing of atmospheric CO₂, the balance may change to a strong sink as predicted by modelling studies (Cossarini et al., 2021; Tsiaras et al., this issue). Additional field studies, especially time series measuring CO₂ (ideally at high frequency), are needed to support model and algorithm validation, to allow specifying the spatiotemporal extent as CO₂ source or sink of this area.

Overall, the above described (section 4.2) dependencies of pCO₂ mainly with temperature, evaporation (shown by salinity) and biological production (shown by Chla) support that remote sensing could be used to estimate pCO₂, as discussed further below. In the Mediterranean Sea, only a few algorithms have been proposed for the estimation of pCO₂ using temperature (or other non-carbonate variables) despite that temperature appears to be a dominant factor in determining surface pCO₂ in the Mediterranean Sea. In the Eastern Mediterranean Sea, to our knowledge, only two studies have provided such relationships (González-Dávila et al., 2016; Sisma-Ventura et al., 2017). The present study provides the first pCO₂ algorithm developed from in-situ CO₂ measurements and remotely sensed SST data ( Figure 5 ) confirming the potential for the prediction of pCO₂ based on satellite data, as used in other marine regions (e.g. Eastern North Pacific, Hales et al., 2012). Moreover, it is suggested as a tool for extrapolation of CO₂ dynamics over wider spatial scales in the Mediterranean Sea (Sisma-Ventura et al., 2017). Satellite Chla data were also accounted for, in the developed algorithm similarly to other studies (Ford et al., 2022 and references therein) together with the SST satellite data. However, the outcome showed no improvement of the agreement with the in-situ data. The weak role of biological non-thermal control of pCO₂ was shown and discussed in the previous section. This could be explained by the generally low Chla concentration, both on average and in the amplitude of the seasonal variation, which is consistent with the oligotrophic character of the entire Cretan Sea, both in terms of Chla and primary production (Psarra et al., 2000). In addition, the fact that pCO₂ measurements did not fully cover the late winter-spring bloom periods could weaken the correlation. It should be noted that extrapolation of in-situ observations of pCO₂ using satellite observations like SST and Chla (Chen and Hu, 2019 and references therein; Ford et al., 2022 and references therein) should be done cautiously as application outside of the time–space domain in which it was developed is less reliable, essentially because of its empirical nature (D’Ortenzio et al., 2008). To examine the present algorithms’ spatiotemporal domain, we applied it (Equation 8) to a CO₂ dataset from a wider area of the Eastern Mediterranean ( Figure 1 ) spanning from 1999 to present. Data in the North Aegean with low salinities (<35) were excluded as the algorithm displayed high bias, which can be attributed to the carbonate system of this area being mostly related to Black Sea Water inflow and river runoff (Krasakopoulou et al., 2017). A linear term of +1.7 units per year accounted for long-term increase of pCO₂ was included (related to atmospheric CO₂ increase), as suggested by other studies (Chen and Hu, 2019). A +1.7 μatm pCO₂ yr^-1 (from 1983 to 2014), was also reported in the BATS and HOT stations (Bates et al., 2014), whereas in the Mediterranean Sea, reported values varied between +0.6 and +1.5 μatm pCO₂ yr^-1 (from 2010 to 2022, Tsiaras et al., this issue) and +3.5 µatm pCO₂ yr^-1 (from 2007 to 2015, Kapsenberg et al., 2017). Based upon the algorithm applicability analysis, it appeared that within the area and time window tested most (91%) of Δ_obs-est were between -20 and +20 μatm which is one order of magnitude below the amplitude of seasonal range (~150 μatm in the present study, ~200 μatm review by Álvarez et al., 2023). We thus consider that the algorithm could be useful when used, within the spatiotemporal window it was constructed, as an additional data quality control test, or to fill out data gaps of limited spatiotemporal extent. This analysis supports that the HCB station is indicative of a wider area of the Eastern Mediterranean Sea with respect to pCO₂, as previously suggested for E1-M3A for temperature, Chla, primary production, pH, and oxygen (Henson et al., 2016).

Contrary to the few published relationships used to estimate pCO₂ from temperature, many relationships used to estimate AT from salinity have been proposed for the different Mediterranean Sea sub-basins (Copin-Montégut, 1993; Schneider et al., 2007; Cossarini et al., 2015; Hassoun et al., 2015 and references therein; González-Dávila et al., 2016) and widely applied, for example, to reconstruct AT time-series (Wimart-Rousseau et al., 2021). Among these published AT-SSS relationships, the ones that performed better with the present study’s dataset were the ones of Hassoun et al. (2015) for Cretan Straits waters, followed by the one of González-Dávila et al. (2016). Despite the rather extended use of these AT-SSS relationships, satellite SSS data have been used only in preliminary estimates of AT in the Mediterranean (Sabia et al., 2020), probably as SSS satellite products have become mature only in recent years (Land et al., 2015). A comparison with latest satellite SSS products shows a difference of 0.3 salinity units demonstrating the upcoming potential to estimate AT from satellite SSS ( Figures 2 , 4 ).