The time series of pCO2sw (µatm) presented in this study were measured by different Pro-Oceanus CO₂ sensors at the EMSO-E2M3A at a depth of approximately 2 m (Cardin et al., 2025a). High frequency (every 4 h) pCO2sw measurements spanned over a period from 2015 to 2024 with declared accuracy of ±0.5 %. This time series was complemented with pCO2sw measurements collected around the EMSO-E2M3A in July 2020 by an Uncrewed Surface Vehicle (USV, Saildrone Inc., USA) during the ATL2MED demonstration experiment (Martellucci et al., 2024a, 2025). Additionally, SOCAT V2026 observations in the SAd were included in the pCO2sw time series shown here. These data cover three years of observation in 2015 (August and September), in 2016 (August and September) and 2020 (June and July) with the latter corresponding to the ATL2MED demonstration experiment data. Hourly sea surface temperature (SST, °C), salinity (Sal) and dissolved oxygen (DO, µmolkg-1) were measured by a SeaBird SBE 37-ODO with declared accuracy of ±0.002 °C, ±0.003 mS cm⁻¹, ±0.1 % full scale range, respectively. DO from SBE 37-ODO was also used to calculate the apparent oxygen utilization (AOU, µmolkg-1). This value corresponds to the difference between the equilibrium saturation of oxygen concentration in seawater with the same physical and chemical properties and the measured oxygen concentration, providing an estimation of the balance between primary production and respiration. AOU has been computed using the solubility coefficients from Benson and Krause (1984) as fitted by Garcia and Gordon (1992). A detailed description of the dataset used here, and the quality control procedures applied to the data are available in Dentico et al. (2026).

To calculate air-sea FCO₂, hourly atmospheric molar fraction of CO₂ (xCO₂, ppm) from the ENEA (Italian National Agency for New Technologies, Energy and Sustainable Economic Development) Station for Climate Observations on the island of Lampedusa (Marullo et al., 2021; di Sarra et al., 2025) were used. Indeed, atmospheric pCO₂ measurements (pCO_2 atm, µatm) at EMSO-E2M3A only started in December 2023, when a ProOceanus CO₂-Pro ATM was deployed allowing continuous measurements of both pCO2sw and pCO_2 atm in alternating mode. As in situ observations of pCO_2 atm were not long enough to match with pCO2sw, Lampedusa data was chosen as it represents one of the longest oceanic time series of pCO_2 atm in the Mediterranean Sea. Moreover, these data have been used in similar studies to compute FCO₂ at Mediterranean scale (e.g., Martellucci et al., 2024a, 2025) and in the northern Adriatic Sea (e.g., Urbini et al., 2020; Cantoni et al., 2024). From Lampedusa xCO₂ (indicated as xCO_{2_LMP}), the pCO_2 atm relative to the local physical and atmospheric condition at EMSO-E2M3A (indicated as pCO_{2 atm_E2M3Ar}) was calculated according to Weiss (1974): 1pCO2atm_E2M3Ar=xCO2_LMP(P-pH2O) where P is the air pressure (mbar) measured at EMSO-E2M3A and pH₂O is the water vapour saturation (mbar) that has been calculated according to Weiss and Price (1980): 2pH2O=Aexp⁡(B-C(100/SST)-Dlog⁡(SST/100)-E(SSS)) where SST is sea surface temperature (K), SSS is sea surface salinity both measured at EMSO-E2M3A and the coefficients A, B, C and D can be found in Weiss and Price (1980). To assess the potential bias of using xCO_{2_LMP} for the calculation of CO₂ air-sea flux in the SAd, a comparison between pCO_2 atm measured by the ProOceanus CO₂-Pro ATM sensor installed at EMSO-E2M3A (pCO_{2 atm_E2M3A}) with pCO_{2 atm_E2M3Ar} was performed. This comparison was limited to a four-month period between December 2023 and March 2024, when the CO₂-Pro ATM sensor was operational. The two time series show good overall agreement, however the differences tend to be larger at higher pCO_2 atm values, while they remain smaller at lower values (Figs. S1, S2a, Supplement). A mean difference of +2.7±3 µatm between pCO_{2 atm_E2M3A} and pCO_{2 atm_E2M3Ar} indicates that pCO_{2 atm_E2M3A} is on average higher than pCO_2 atm derived from xCO_{2_LMP}. A detailed investigation of the reasons behind this difference was beyond the scope of this work. Although the analysis of the residuals indicated potential heteroscedasticity in the data (Fig. S2a, b in the Supplement), the short duration of the overlapping time series justified applying a constant bias correction of +2.7 µatm to pCO_{2 atm_E2M3Ar}. To evaluate the robustness of this correction, a comparison between pCO_2 atm derived from the CAMS reanalysis of the European Centre for Medium-Range Weather Forecasts (ECMWF; Agustí-Panareda et al., 2023), was also considered (Fig. S3, Supplement). The difference between monthly pCO_2 atm data at Lampedusa (LMP) and at EMSO-E2M3A between 2003 and 2021 was 2 ppm which is similar to the offset estimated in the four-month period.

Wind speed (WSPD, m s⁻¹) was measured by a Young sensor on the meteorological station at EMSO-E2M3A (Cardin et al., 2025a; Dentico et al., 2026) with a manufacturer accuracy of ±0.3 m s⁻¹ and was used as input for CO₂ flux calculation. Given that at EMSO-E2M3A wind speed is measured at 2 m height and CO₂ air-sea flux calculation requires the standard reference height of 10 m (Wanninkhof, 2014), a conversion was performed using a logarithm wind speed profile according to: 3uLOG(z)=(ln⁡(z/z0)/ln⁡(zm/z0))u(zm) where uLOG(z) is the wind speed (m s⁻¹) at the desired height (10 m), z is the desired measurement height (10 m), zm the actual measurement height (2 m), z0 is the roughness length that for oceanic regions was set equal to 1.52×10-4 (Peixoto and Oort, 1992) and u(zm) is the wind speed (m s⁻¹) measured at EMSO-E2M3A at 2 m height. A detailed description of the data collection and quality check of the data can be found in Dentico et al. (2026).

The European Copernicus Marine Service (CMEMS) provides regular and systematic information on the ocean physics and biogeochemistry for the global ocean and the European regional seas (Le Traon et al., 2019). Data from different Mediterranean CMEMS products were used as an additional reference to investigate the variability of the in situ pCO2sw and FCO₂. In particular, the following modelling products were used: the Sea Physics Reanalysis (PHY, Escudier et al., 2021) for the physical variables of surface temperature (SST, °C), salinity (Sal) and Mixed Layer Depth (MLD, m), the Biogeochemistry Reanalysis (RD; Cossarini et al., 2021) and the Biogeochemistry Analysis and Forecast product (AF; Salon et al., 2019) for pCO2sw (µatm) and FCO₂ (mmolm-2d-1). Additionally, the Ocean Color Satellite product (Volpe et al., 2019a) for the surface chlorophyll a (Chl a, µgL-1) was used as a proxy of primary production. Detailed information on the models and variables used are summarized in Table 1. The time series presented here are spatial averages of the grid points in the 10 km radius from EMSO-E2M3A from the native spatial resolution of 4.5 and 1 km for models and satellite data, respectively.

ERA5 wind product, which is the fifth generation of ECMWF reanalysis for the global climate and weather (Hersbach et al., 2023), was also used and compared with EMSO-E2M3A wind measurements. ERA5 provides hourly estimates for a large number of atmospheric, ocean-wave and land-surface quantities from 1940 to present. The spatial resolution of atmospheric data is 0.25° × 0.25° and wind speed (WSPD_ERA5, m s⁻¹) in the subregion (30 km radius) and in the nearest point close to EMSO-E2M3A was calculated as the square root of the sum of squares of the 10 m u component (m s⁻¹), which is the eastward horizontal wind speed, and of the 10 m v component (m s⁻¹) that represents the northward horizontal wind speed.

The direction and magnitude of FCO₂ is governed by the difference between pCO2sw and pCO_2 atm and the air-sea CO₂ gas transfer velocity (Takahashi et al., 2002): 4FCO2=k0k(pCO2sw-pCO2atm) where k0 is the solubility coefficient of CO₂ at in situ temperature and salinity (Weiss, 1974; Zeebe and Wolf-Gladrow, 2001) and k is the gas transfer velocity (m d⁻¹). In this study, if the FCO2<0 the flux is from air to seawater, on the contrary if FCO2>0 the flux is from the seawater to the atmosphere. Several algorithms for k calculation exist, here different parametrizations were used to compute FCO₂: 5kW=0.251u2(Sc/660)-0.56kW&McG=0.0283u3(Sc/660)-0.57kN=0.222u2+0.333u(Sc/600)-0.5 where kW is a quadratic relationship proposed by Wanninkhof (2014) and it is one of the most frequent parametrizations used in the FCO₂ estimations at both global and Mediterranean scale (Urbini et al., 2020; Cantoni et al., 2024; Martellucci et al., 2025; Pecci et al., 2026); kW&McG is a cubic wind speed dependence from Wanninkhof and McGillis (1999); and kN is a mixed (linear and quadratic) dependency from Nightingale et al. (2000) used in the literature (for example in Gulf of Biscay; Otero et al., 2013). In Eqs. (5), (6) and (7) u is the wind speed (m s⁻¹) corrected at 10 m height and Sc the Schmidt number for CO₂. Different temporal resolutions of u have been used to compute Eq. (5), (6) and (7) including hourly, 6 h averages and daily averages. From these, daily k values have been calculated to have daily FCO₂ values. Additionally, 6 h averages ERA5 wind speed data have been used to calculate the different parametrizations of k. ERA5 wind speed data were selected for two main reasons: (i) wind reanalysis products, unlike in situ time series, provide continuous spatial and temporal coverage without observational gaps; (ii) ERA5 wind data are used as the wind forcing input in the Mediterranean Sea Physical and Biogeochemistry Reanalyses. Thus, considering that CO₂ air-sea flux in the CMEMS products (Salon et al., 2019; Cossarini et al., 2021) and in the present work was computed following the same Wanninkhof (2014) formulation, the use of the same wind dataset, allowed a focus on the analyses of pCO2sw difference between in situ data and model outputs. In summary, different wind products (EMSO-E2M3A versus ERA5), different formulations of k and different temporal aggregation of wind speed data were used to compute the FCO₂ and to assess their impact on annual and seasonal estimates.

Thermal (pCO_{2 sw_T}, µatm) and non-thermal (pCO_{2 sw_NT}, µatm) components of pCO2sw were estimated according to Takahashi et al. (2002). The thermal component is associated to the effect of temperature on pCO2sw, while the non-thermal component can be attributed to other factors not strictly related to temperature, as biological activity (photosynthesis and respiration), changes in dissolved inorganic carbon (DIC), mixing or advective processes and gas exchange with the atmosphere. pCO_{2 sw_T} and pCO_{2 sw_NT} have been computed using the following equations: 8pCO2sw_T=mean(pCO2sw)exp⁡0.0423(SSTobs-SSTmean)9pCO2sw_NT=pCO2swexp⁡0.0423(SSTmean-SSTobs) where, “mean” refers to yearly mean values of sea surface temperature (SST, °C), while the other variables (pCO2sw and SST_obs) refer to the daily values measured during the study period.

The time series of the daily surface pCO2sw between 2015 and 2024 is presented in Fig. 2a. The incomplete coverage of the full seasonal cycle in several years, prevented a detailed assessment of interannual variability. Therefore, the discussion presented in this section mainly focuses on the seasonal patterns and their main drivers, and, when possible, on interannual differences of the seasonal cycle. Mean winter values were around 380 µatm and mean summer values reached around 484 µatm, resulting in a seasonal amplitude of about 100 µatm. This pronounced seasonality reflected the combined influence of thermal and non-thermal processes (Fig. 2b). During winter, low sea surface temperature (SST <15 °C; Fig. 2c) increases CO₂ solubility, leading to lower pCO2sw (Fig. 2b). However, during winter convection the upward transport of DIC-rich intermediate and deep waters (Touratier et al., 2016; Ingrosso et al., 2017) led to an increase of surface pCO2sw as shown by the pCO_{2_NT} component (Fig. 2b). This was particularly evident between late December and late February of 2016–2017, 2017–2018, 2020–2021, and 2023–2024 (Fig. 2a). Positive AOU values observed during these periods (Fig. 2d) further support this interpretation, indicating the upward displacement of oxygen-depleted intermediate waters during vertical mixing, also described in Martellucci et al. (2024b). The post-convection periods (Fig. 2e; usually from March to May–June) were characterised by a phytoplankton bloom (Fig. 2f). Although direct pCO2sw observations were limited during this period, the decrease in AOU and the increase in DO concentration at the surface (Fig. 2d) provided indirect but consistent evidence of primary production. These observations suggest that biological CO₂ uptake likely contributes to lowering pCO2sw following winter mixing, acting as an additional non-thermal driver of the seasonal variability. In spring, surface warming began, leading, in summer, to maximum SST values (>25 °C; Fig. 2c) and strong stratification. During this period, pCO2sw reached its annual maximum, closely following the pCO_{2 sw_T} component and indicating a dominant thermal effect (Fig. 2b). The comparison between the Saildrone and EMSO-E2M3A observations further confirmed this pattern, showing coherent short-term variations with the EMSO-E2M3A time series (r=0.77, p value <0.01) and a nearly constant offset of 16 µatm (Fig. 2a). Similarly, SOCAT V2026 observations showed very good agreement with the EMSO-E2M3A measurements during August–September 2015 (mean difference of approx. 8.4 µatm; Fig. 2a). In contrast, larger differences (mean difference approx. 40 µatm) were observed during the same period in 2016. The SOCAT observations in 2016 are located along the eastern side of the SAd, therefore, the discrepancies may reflect differences in the hydrographic conditions and water mass distribution between the two surveys, resulting in the sampling of waters with different physical and biogeochemical characteristics. However, further analyses would be required to confirm this hypothesis. Interestingly, higher pCO2sw values were recorded in summer 2022, with a peak in July 2022, coinciding with the occurrence of a Mediterranean Marine Heatwave. In the SAP, Pirro et al. (2024) detected positive temperature anomalies across the whole water column, with the strongest warming at the surface. Thus, the elevated pCO2sw observed in this period might be linked to the increased stratification that limited vertical exchange favouring the accumulation of respired carbon in surface waters. This behaviour was consistent with recent observations in the northwestern Mediterranean Sea (e.g., Metzl et al., 2025), which highlighted that the exceptional warming observed in summer 2022 impacted not only thermodynamic processes but also stratification, wind field (Pecci et al., 2026) and biological activity, with consequences for the variability of pCO2sw and air–sea CO₂ flux.

The annual cycle of surface pCO2sw in the SAd was comparable to other regions of the Mediterranean Sea with generally lower values than those observed in the SAd (Dentico et al., 2026). Similar seasonal variations especially in winter were described for the northern Adriatic shelf (Ingrosso et al., 2016; Urbini et al., 2020; Cantoni et al., 2024), where intense Bora wind events and vertical mixing periodically disrupt water column stratification, promoting an efficient exchange between bottom and surface pCO2sw. Seasonal pCO2sw variability was reported between 200 µatm in winter up to 500 µatm in summer (e.g., Cantoni et al., 2024). However, unlike the SAd, the riverine inputs in this area also play a key role in modulating the carbonate system dynamics (e.g., Giani et al., 2023). In the convective region of the northwestern Mediterranean Sea, several studies have emphasized the importance of vertical mixing, horizontal advection and biological productivity in driving pCO2sw variability (e.g., Merlivat et al., 2018; Fourrier et al., 2022; Ulses et al., 2023). Here winter values were comparable to the SAd (around 350 µatm) and summer values around 500 µatm (Merlivat et al., 2018; García-Ibáñez et al., 2024) or higher, as reported for the DYFAMED site in the Ligurian Sea (Coppola et al., 2020) where values exceeded 550 µatm. In the SAd the effect of biological activity was less clear because of data gaps. Nevertheless, further investigations could be essential to clarify the role of biological processes in shaping the local carbon cycle as highlighted in previous studies (Turchetto et al., 2012; Cerino et al., 2012). Non-convective regions such as the eastern Mediterranean Sea, exhibited a stronger thermal control on seasonal pCO2sw variability. The first pCO2sw time series from the Cretan Sea showed that temperature was the dominant driver of pCO2sw, while the non-thermal component was mainly associated with evaporation, water mass mixing, and biological remineralisation-production processes (Frangoulis et al., 2024). Similarly, in the southeast Levantine, seasonal pCO2sw was found to be tightly correlated with sea surface temperature indicating that its seasonal cycle is largely controlled by thermodynamic processes (Sisma-Ventura et al., 2017). In contrast to the Levantine Basin, where pCO2sw variability is strongly influenced by warming and evaporation, the Southern Adriatic is affected by recurrent deep convection and higher biological activity (e.g., Pérez et al., 2024; Tsiaras et al., 2024). Nevertheless, seasonal variations were similar ranging between 350 µatm in winter and around 500 µatm in summer (Frangoulis et al., 2024). The analysis of hydrological data showed the highly dynamic characteristics of the region. First, the intensity and occurrence of convection in the SAd have varied substantially among years (Fig. 2e). Depending on atmospheric forcing and water column preconditioning, the depth of vertical mixing ranged from relatively shallow mixing in the intermediate layers (<400 m depth) to deeper convective events (>800 m depth). Weaker events were recorded in 2015, 2016, 2019, and 2020, whereas deeper mixing (>600 m) occurred in more recent years (especially after 2022; Fig. 2e). This interannual variability resulted from strong/weak air-sea interactions, where stronger heat losses could occur especially under the influence of the cold and dry Bora wind (Le Meur et al., 2025). Variations in the regional hydrographic structure, particularly changes in salinity (Amorim et al., 2024; Le Meur et al., 2025), also affected the stability of the water column and the depth of convection. This variability in salinity (Fig. 2g) was associated with the inflow of Levantine/Ionian surface water, which is regulated by changes in the Northern Ionian Gyre (NIG) vorticity (Gačić et al., 2010; Rubino et al., 2020; Menna et al., 2022; Civitarese et al., 2023). The influence of NIG in the SAd was evident in the salinity measurements (Fig. 2g), which showed periods of higher salinity (>39 in 2017, 2021, 2022) and lower salinity (<38.5 in 2015, 2018). Through its impact on stratification and preconditioning, this large-scale circulation variability likely modulates the intensity of winter mixing and, indirectly, pCO2sw variability.

Daily FCO₂ in the SAd shown in Fig. 3 were computed using a reference configuration based on Eqs. (4) and (5) using as input pCO2sw, pCO_{2 atm_E2M3Ar} (Sect. 3.1) and 6 h averaged wind speed data from EMSO-E2M3A corrected at 10 m height (Sect. 3.3). The gas transfer velocity was calculated using the Wanninkhof (2014) parameterisation to allow a more consistent comparison with previous studies and model data (Sect. 4.4). FCO₂ estimates obtained using different wind products, wind temporal resolutions, and gas transfer velocity parameterisations are discussed in Sect. 4.3. Based on the reference configuration, CO₂ air-sea flux in the SAd presents two dominant phases: a source period, generally associated with post-convection/summer months (positive values in Fig. 3a), and a sink period, typically occurring during winter and during episodes of winter convection (negative values in Fig. 3a). While the intensity of daily values during sink and source phases exhibited high interannual variability, the timing of the transition months, i.e., the periods when the SAd shifts from a source (May) to a sink (October), occurred regularly (Fig. 3b).

Monthly averages were computed for those months with at least 14 valid daily values. The mean over the full period was subsequently obtained from these monthly values (Fig. 3b). Given the observed variability (i.e., high daily variability with a regular annual cycle) and the presence of substantial data gaps in the FCO₂ time series, an additional computational approach was applied to derive robust estimates of mean CO₂ fluxes in the SAP over the past decade (Table S1 in the Supplement). Notwithstanding the potential limitations derived by data availability, over the last decade, the SAd has functioned predominantly as a weak-to-moderate carbon sink with mean annual FCO₂ values of about -0.97 mmolm-2d-1 for the period 2015–2024. Across the Mediterranean, air-sea carbon flux is spatially uneven: the western basin takes up more CO₂ than the eastern basin, while some southeastern sectors show weak release or near-neutral behaviour (Cossarini et al., 2021). Although the values reported here for the SAd were modest in magnitude compared to other Mediterranean regions, they pointed to a net CO₂ uptake, underscoring the potential role of the SAd in regional carbon sequestration, as suggested by Ingrosso et al. (2017). While the present analysis does not directly address the carbon export below the mixed layer, the occurrence of recurrent convection events suggests that part of the CO₂ absorbed at the surface may be transferred to intermediate and deep waters. In this context, our results are consistent with observations from other Mediterranean deep-water formation regions. For instance, the northern Adriatic Sea has been reported to act as an effective CO₂ sink during winter, spring, and autumn, with mean daily fluxes of about -2.9 mmolm-2d-1 (Catalano et al., 2014; Cantoni et al., 2024) reaching however the highest CO₂ flux during winter (-14.2 mmolm-2d-1; Urbini et al., 2020). Dense waters formed in the northern Adriatic can eventually flow towards the SAP, suggesting that the Southern Adriatic may act as a regional collector of carbon-enriched waters from upstream regions in addition to locally vertical transfer of CO₂ during convection events. Similarly, in the Gulf of Lions, Ulses et al. (2023) showed that the area functions as a net CO₂ sink on an annual scale, with an estimated carbon sequestration rate of approximately -1.29 mmolm-2d-1 (estimated mean was -0.47 mol C m⁻² yr⁻¹). Here higher uptake is also linked to Atlantic inflow and spring primary production that lowers surface pCO₂ (Martellucci et al., 2025). Finally, comparable FCO₂ values could be also found between the SAd and the Central Mediterranean (Pecci et al., 2026) where estimates at the Lampedusa observatory, indicate a net CO₂ sink activity corresponding to -1.76±0.11 mmolm-2d-1 for a 1-year period (i.e., 2022).

This section examines the influence of different wind products (in situ observations and ERA5 reanalysis), different temporal aggregations of the wind data (hourly, 6-hourly, and daily averages), and alternative gas transfer velocity parameterisations (Eqs. 5, 6 and 7) on the estimation of FCO₂. Flux estimates obtained using these different methodological choices are compared on both annual basis and for two extended seasonal periods representing the main hydrographic phases of the Southern Adriatic: the winter/convective period, from October to March, and the summer/post-convective period, from April to September. For each new estimate, the annual and seasonal averages were derived using the same approach of the nominal estimate (i.e., considering months with at least 14 valid daily observations). Annual and seasonal averages were then calculated from these monthly values. The resulting annual and seasonal FCO₂ estimates are summarized in Table 2. All configurations based on in situ observations consistently indicated a net annual uptake of atmospheric CO₂ by the Southern Adriatic, although the magnitude of the estimated flux varied depending on the methodological choices adopted (Table 2). Annual mean FCO₂ ranged from -0.90 to -3.92 mmolm-2d-1. The choice of the gas transfer velocity parameterisation had a larger influence on the annual flux estimate than the temporal aggregation of the wind data. While the results using kW or kN with any temporal aggregation produced comparable results, the use of kW&McG produced higher results. For example, using the pair 6-hourly in situ wind speed and kW&McG produced an annual uptake four times larger than estimates obtained with the pair 6-hourly in situ wind speed and kW or kN. The influence of wind temporal aggregation was generally limited for kW or kN but became more pronounced for kW&McG, which is inherently more sensitive to high wind speeds that can be smoothed out due to the daily aggregation of the data. At the seasonal scale, winter CO₂ uptake was consistently larger than summer outgassing regardless of the selected parameterisation or wind temporal aggregation. Winter mean FCO₂ ranged from -14.55 to -19.26 mmolm-2d-1 using in situ winds, whereas summer values ranged between 8.75 and 13.96 mmolm-2d-1. The relative standard deviation of the summer estimates was, however, systematically higher than that of the winter period. This behaviour mainly reflects the inclusion of May in the summer period. As shown in Fig. 3b, May represents a transition month during which the Southern Adriatic shifts from winter sink conditions to summer source conditions. Although October also marked a seasonal transition, its influence on the winter mean was less pronounced because it was followed by several months characterized by persistent CO₂ uptake. In contrast, the summer period included only four months (from June to September) that consistently behave as a CO₂ source. These results suggested that the annual net CO₂ uptake in the Southern Adriatic is favoured not only by the higher winter CO₂ flux but also by the longer duration of the sink period relative to the source period. A different behaviour emerged when ERA5 wind data were used. Using ERA5 wind fields together with kW and kN parameterisations resulted in a weak annual CO₂ source, whereas the use of the nearest ERA5 grid point to the EMSO-E2M3A observatory instead of the spatial average resulted in a weak annual sink. This sign reversal highlighted the strong dependence of FCO₂ estimates on the representation of the local wind field. The difference was particularly evident during winter, when ERA5 spatially averaged winds produced CO₂ uptake estimates about 25 %–40 % lower than those obtained using the nearest ERA5 grid point. This suggests that spatial averaging may smooth high wind events, which strongly influence k because of its nonlinear dependence on wind speed (Takahashi et al., 2002). A more detailed analysis on the performance of ERA5 wind in the region revealed that the largest discrepancy between ERA5 and in situ observations (Figs. S4 and S5) occurred during strong wind events (high Beaufort conditions).

The in situ pCO2sw and physical data proved to be a reliable asset for model validation, thus a comparison between the RD and AF CMEMS products for pCO2sw and FCO₂ was computed. The comparison between daily pCO2sw from CMEMS (Sect. 3.4) and pCO2sw measured by the probe for a decade-long period highlighted a coherent temporal evolution (Fig. 4). A RMSE of 35.96 µatm and a significant correlation coefficient of 0.82 (p value <0.01) between the two datasets was found. Nevertheless, during summer, model data were on average higher than pCO2sw from the probe, with a mean difference between RD and the probe data of +34.22 µatm. The only exception was in 2022, when pCO2sw from the probe was higher than CMEMS data (both RD and AF). As discussed in Sect. 4.1 in 2022 a marine heatwave was detected in the region (Pirro et al., 2024). The strong difference in pCO2sw between the probe and the model was likely driven by the combined effects of warming, enhanced stratification, and biogeochemical processes which are difficult to represent in current biogeochemical models. On the contrary, in winter RD data were on average lower than the probe, with a mean difference between the two datasets of -26.87 µatm. This discrepancy was mainly associated with the increase in the pCO2sw recorded by the probe between late December and late February, which was not fully resolved by the model and was underestimated by approximately 45 µatm. This mismatch may be partly attributed to the limited vertical resolution of the model and to the parameterization of vertical mixing processes, which may smooth or underestimate the upward transport of DIC to the surface during convection events (as discussed in Sect. 4.1). By contrast, with lower intensity, the AF product was able to capture this increase (Fig. 4). Similarly to RD, also AF was not able to consistently reproduce the higher pCO2sw values measured by the probe in summer 2022. Finally, in autumn and spring the mean differences were smaller (+5.13 µatm and -16.12 µatm, respectively).

The comparison between FCO₂ derived from in situ data and from the model is presented in Fig. 5. Both datasets consistently capture the two main seasonal phases of the region, with a source period during summer and a sink period during winter. However, notable differences in absolute values emerged, particularly in summer. A main factor in these differences was related to the pCO2sw measurements as discussed in the previous section. Additionally, in both seasons, the different wind forcing used by the model (ERA5) could also be a contributing factor of the FCO₂ difference (see Sect. 4.3 and Fig. S5, Supplement). On the contrary, the difference between model and in situ temperature and/or salinity was not considered a driving factor of the in situ and RD FCO₂ difference (Fig. S6, Supplement). Nevertheless, an RMSE of 16.89 mmolm-2d-1 with a correlation coefficient of 0.89 (p value <0.01) was calculated between FCO₂ from RD and FCO₂ derived from observational data.

Considering the relatively good performance of the RD, and the temporal bias in the in situ FCO₂ time series (e.g., the winter periods were by far under sampled with respect to summer periods), it was used to clarify the role of the SAd as a source or sink of CO₂. The mean annual FCO₂ using exclusively CMEMS RD resulted in -4.3 mmolm-2d-1. The spatial distribution of the mean annual FCO₂ further highlighted that spatial differences in the basin existed (Fig. 6). The strongest uptake is found in the central and deep part of the basin, where FCO₂ reaches values close to -5 mmolm-2d-1 while weaker sink activity was shown particularly along the eastern part of the basin with values being between 0 and -2 mmolm-2d-1 (Fig. 6). Nevertheless, despite the spatial differences in the magnitude of the annual CO₂ fluxes, the overall basin scale behaviour derived from the CMEMS reanalysis is consistent with the in situ observations. Although local fluxes vary across the basin, the EMSO-E2M3A can be considered a representative site for capturing the main carbon dynamics of the Southern Adriatic. Secondly, the use of the CMEMS RD product to compute the overall FCO₂ average filling the gaps of the in situ data was tested. A bias (B) using the CMEMS RD was calculated as the difference between the mean flux over the entire period (2015–2024) and the mean flux considering only the days with observations being equal to -3.7 mmolm-2d-1. B was then added to the mean flux calculated with in situ observations, resulting in a mean annual flux of -4.5 mmolm-2d-1. Even if these estimates should be considered with caution, they further confirm that the Southern Adriatic acted as a moderate carbon sink over the last decade.