The study was conducted at the Bossys perdus salt marsh situated along the French Atlantic coast on Ré Island (Fig. 1). It corresponds to a vegetated intertidal area of 52.5 ha that has been protected inside the National Natural Reserve (NNR) (Fig. 1). Between the 17th and most of the 20th century, the salt marsh experienced successive periods of intensive land use (salt harvesting and oyster farming) and returned to natural conditions before becoming a permanent part of the NNR in 1981 for the biodiversity protection without major restoration work (Julien Gernigon, personal communication, 2023). It is currently managed to restore its natural hydrodynamics while conserving the site's specific typology due to past human activities (channel networks, humps and dykes; Fig. 2). This salt marsh is linked to the Fier d'Ars tidal estuary that exchanges between 2.4 and 10.2 ×106 m3 of coastal waters with the Breton Sound continental shelf allowing a maximal tidal range of 5 m in the estuary (Bel Hassen, 2001). This communication allows (1) drainage of the intertidal zone of the estuary including mudflats (Mayen et al., 2023) and tidal salt marshes (Mayen et al., 2023) and (2) supply of coastal water to a large complex of artificial salt marshes (i.e. salt ponds) located upstream of the dyke (Fig. 1). The artificial marsh waters managed by the NNR for biodiversity protection (Mayen et al., 2023) are flushed back to the estuary downstream through the Bossys perdus channel (Fig. 1).

The Bossys perdus salt marsh, located upstream of the estuary (Mayen et al., 2023), is subjected to semi-diurnal tides from the Breton Sound continental shelf (Fig. 1) allowing the marsh immersion by two main channels differently in space, time and frequency according to the tidal periods (Fig. 2). At high tide, advected coastal waters can completely fill channels (Fig. S1b in the Supplement) and immerse the marsh through variable water heights depending on tidal amplitudes and meteorological conditions (Fig. S1c). In contrast, at low tide, the marsh vegetation at the benthic interface is emerged into the atmosphere without any coastal waters (Fig. S1a). During this time, Bossys perdus channels allow drainage of upstream artificial marsh waters to the estuary (Fig. 2). The marsh vegetation assemblage was mainly composed by three halophytic species as perennial plants (Halimione portulacoides, Spartina maritima and Suaeda vera; Fig. 2) that associated with different metabolic pathways (the C3-type photosynthesis for H. portulacoides and S. vera and the C4-type photosynthesis for S. maritima; Duarte et al., 2013, 2014). Whereas H. portulacoides and S. vera are evergreen plants throughout the year, the growing season for S. maritima was shorter (from spring) with a flowering period between August and October (plants persist only in the form of rhizomes in winter and fall; Julien Gernigon, personal communication, 2023).

The atmospheric eddy covariance technique allowed us to quantify the net CO2 fluxes at the ecosystem–atmosphere interface through micrometeorological measurements of the vertical component of atmospheric turbulent eddies (Aubinet et al., 2000; Baldocchi, 2003; Burba, 2021). The averaged vertical flux of any gas (F, µmol m-2 s-1) can be expressed as the covariance between the vertical wind speed (w, ms-1), air density (ρ, kgm-3) and the dry mole fraction (s) of the gas of interest as 1F=ρws‾≈ρ‾w′s′‾, where the overbar represents the time average of the parameter (i.e. 10 min in this study due to strong fluctuations at the tidal scale; Polsenaere et al., 2012) and the prime indicates the instantaneous turbulent fluctuations in these parameters relative to their temporal average (Reynolds, 1883). The Reynold's decomposition was used to break the instantaneous term down into its mean and deviation (e.g. w = w‾+w′) (Reynolds, 1883; Burba, 2021). This equation (Eq. 1) is obtained by assuming, on a flat and homogeneous surface, that (1) the variation in air density is negligible, (2) there is no divergence or convergence of large-scale vertical air motion and (3) atmospheric conditions are stable and stationary (Aubinet et al., 2012). A negative flux of atmospheric CO2 is directed towards the ecosystem, and is therefore characterized as a sink, and vice versa for positive fluxes qualified as sources of CO2 to the atmosphere.

An EC system was continuously deployed at the Bossys perdus salt marsh to measure the net ecosystem CO2 exchange (NEE, µmol m-2 s-1). The set of EC sensors (Fig. 2), at a height of 3.15 m, was composed of an open-path infrared gas analyser (model EC150; Campbell Scientific) to measure the CO2 (mgm-3) and H2O (gm-3) concentrations in the air as well as the atmospheric pressure (kPa) and an ultrasonic anemometer (model CSAT3; Campbell Scientific) to measure the three-dimensional components of wind speed (U, V and W; ms-1) at a frequency of 20 Hz and averaged every 10 min (Fig. 2). The EC150 gas analyser also measured the air temperature using a thermistor probe (model 100K6A1A; BetaTherm). The EC100 electronics module (model EC100; Campbell Scientific) allowed us to synchronize high-frequency measurements and rapid communications between the CR6 datalogger (model CR6; Campbell Scientific) and EC devices including EC150 and CSAT3A (Fig. 2). The CR6 datalogger is a powerful core component for the data acquisition system. Additional meteorological data, such as relative humidity (RH, %), air temperature (Ta, °C) and photosynthetically active radiation (PAR, µmol m-2 s-1), were recorded every 10 min simultaneously and at the same height as the EC sensors, by a temperature/relative humidity sensor (HMP155A; Campbell Scientific), with RAD14 natural ventilation shelter) and a silicon quantum sensor (SKP215; Skye Instruments), respectively (Fig. 2). The vapour pressure deficit (VPD, Pa) was calculated every 10 min from saturated vapour pressure (calculated from Ta) and from actual vapour pressure (calculated from RH). A rainfall sensor (TE525MM; Rain Gauge, Texas Electronics), located 10 m away and connected to the EC station, simultaneously measured the cumulative precipitation at a height of 1 m (rainfall, mm). All high-frequency EC data were recorded on an SD micro-card (2 Go; Campbell Scientific) that was replaced every 2 weeks, whereas meteorological data were recorded and stored in the central acquisition system (CR6). The EC system was connected to two rechargeable batteries (12 volts and 260 Ah-1; AGM) powered by a monocrystalline solar panel (24 V, 200Wp module with MPPT 100 V/30 A controller; Victron Energy). The EC sensors were checked and cleaned every 2 weeks and the EC150 was calibrated each season with a zero-air calibration of 0 ppm (Campbell Scientific) and a certificated CO2 standard of 520 ppm (Gasdetect). Water height (Hw; ± 0.3 m) and water temperature (Tw; ± 0.1 °C) were also measured every 10 min along with EC data using a STPS probe (NKE Instrumentation) located 20 m away from the EC system (Fig. 2). The sensor was checked every two months at the laboratory to verify possible derivations in the measured parameters.

Footprints were estimated using the model of Kljun et al. (2015) applied to data from the year 2020 to obtain an annual averaged footprint from the constant measurement height (Zm = 3.15 m), the constant displacement height (d = 0.1 m; estimated from 0.67 times the canopy height; LI-COR, EddyPro^® 7 Software, LI-COR Environmental), mean wind velocities (u_mean, ms-1), standard deviations of the lateral velocity fluctuations after rotation (sigma_v, ms-1), the Obukhov length (L), friction velocities (u∗, ms-1) and wind directions (°) obtained from the EC measurements and the processing software (EddyPro^® v7.0.8; LI-COR) output. For verification, we performed the footprint estimations both with variable Zm from water height measurements and with constant Zm from data at emersion and we obtained the same footprint shapes and extends. For all calculations (i.e. habitat coverage, relationships with CO2 fluxes, etc.), we used the 70 % footprint contour line that corresponds to an average footprint of 13 042 m2 of the studied salt marsh area of interest (Fig. 2). A land-use map was also created (Fig. 2) from geo-referenced IGN BD orthogonal images with a resolution of 20 cm (2019) using ArcGIS 10.2 (ESRI). The spatial analysis tool of ArcGIS 10.2 was used to perform an unsupervised classification of the BD orthogonal images. We checked the resulting map by selecting 20 random locations within the footprint of the studied salt marsh and compared their land use on the ground and on the map.

In some situations, based on the tide (neap tides), due to meteorology influence (wind direction and atmospheric pressure) and the local altimetry heterogeneity, our one-location Hw measurements could not accurately account for the whole spatial emersion and immersion of the marsh in the EC footprint (Fig. 2). At incoming tide, when coastal waters begin to fill the channel and then overflow over the marsh (from 0.5 h in spring tides to 2.5 h in neap tides; data not shown), the SSW sector (Fig. 2) was first immersed and a non-zero Hw value was measured. However, although some marsh sectors were immersed at the same time, others were still emerged. Indeed, lowest marsh levels (56 % of the footprint area), mainly composed of mudflats and S. maritima (Fig. 2; Table 1), were quickly immersed from Hw > 0 m (south), whereas the whole marsh immersion (muds and plants) only occurred 0.75 h later from Hw > 1.0 m at high tide during spring tide. Thus, the highest marsh levels (44 % of the footprint area), mainly composed of H. portulacoides and S. vera (Fig. 2; Table 1), were still emerged for 0 < Hw < 1.0 m. Conversely, at neap tide, this footprint immersion versus emersion marsh heterogeneity could still be present even at high tide due to insufficient water levels. Although a digital field model for water heights could not be performed in 2020 to have a better spatial representation of the immersion/emersion footprint, all these important considerations were considered in our computations and analyses in this study.

Raw EC data measured at high-frequency were processed following Aubinet et al. (2000) with the EddyPro software. First, different correcting steps were applied to our raw data according to the procedures given by Vickers and Mahrt (1997) and Polsenaere et al. (2012) for intertidal systems: (1) unit conversion to check that the units for instantaneous data are appropriate and consistent to avoid any errors in the calculation and correction of CO2 fluxes; (2) despiking to remove outliers in the instantaneous data from the anemometer and gas analyser due to electronic and physical noise and replaced the detected spikes with a linear interpolation of the neighbouring values; (3) amplitude resolution to identify situations in which the signal variance is too low with respect to the instrumental resolution; (4) double coordinate rotation to align the x axis of the anemometer to the current mean streamlines, nullifying the vertical and cross-wind components; (5) time delay removal by detecting discontinuities and time shifts in the signal acquisition from the anemometer and gas analyser; (6) detrending with removal of short-term linear trends to suppress the impact of low-frequency air movements; and (7) performing the Webb–Pearman–Leuning (WPL) correction to take into account the effects of temperature and water vapour fluctuations on the measured fluctuations in the CO2 and H2O densities (Burba, 2021). The turbulent fluctuations of CO2 fluxes were calculated with EddyPro using the linear detrending method (Gash and Culf, 1996) which involves calculating deviations from around any linear trend evaluated (i.e. over the whole flux averaged period). High-frequency CO2 fluxes were processed and averaged over intervals of 10 min (shorter than in terrestrial ecosystems) to detect fast NEE variations with the tide (Polsenaere et al., 2012; Van Dam et al., 2021). During the EC data processing by EddyPro, a correction for flux spectral losses in the low frequency range was performed according to Moncrieff et al. (2004).

A strict quality control was applied on EddyPro processed CO2 flux data to remove bad data related to instrument malfunctions, processing and mathematical artefacts, ambient conditions that do not satisfy the requirements for the EC method, wind that is not from the footprint and heavy precipitation for the open-path IRGA (Burba, 2021). Processed data were screened using tests for steady state and turbulent conditions (Foken and Wichura, 1996; Foken et al., 2004; Göckede et al., 2004). In this study, we did not apply a ustar filter in our EC data processing because we measured only 11 % of night-time data corresponding to a ustar threshold below 0.1 ms-1 and above which NEE does not increase anymore with ustar values (threshold close to values found in grassland; Gu et al., 2005). Contrary to terrestrial ecosystems (Gu et al., 2005), the low canopy height of the studied marsh strongly limited the CO2 storage in the vegetation and favours the atmospheric CO2 circulation. If the signal to noise ratio of the EC150 gas analyser was less than 0.7 and/or the percentage of high-frequency missing values over 10 min exceeded 10 % (i.e. data absent in the raw data file or removed through the quality screening procedures), no flux was calculated. This choice was the best compromise between removing poor-quality data and keeping as much of measured CO2 flux data as possible (data and associated tests not shown). Then, we used the method of Papale et al. (2006) to detect and remove outliers in the 10 min flux data. The median and median absolute deviation (MAD) were calculated over a 2-week window separating daytime and night-time periods. Data above 5.2× MAD were removed. After all post-processing and quality controls, 18.3 % of the EC data were removed and gap-filled through a machine learning approach to obtain continuous flux data in 2020.

The random forest (RF) model was used to gap-fill our EC dataset. Random forest is a supervised machine learning technique proposed by Breiman (2001) that can model a non-linear relationship with no assumption about the underlying distribution of the data population. This method has been shown to be particularly suited to gap-fill EC data (Kim et al., 2020; Cui et al., 2021). Random forest builds multiple decision trees, each of which is based on a bootstrap aggregated data sample (i.e. bagging of the EC data) and a random subset of predictors (i.e. the selected environmental data; Table S1 in the Supplement). We build RF models with environmental predictors that have been identified in the literature to control CO2 fluxes in salt marshes and which were available during the gaps and with measurements recorded between 2019 and 2020 (Table S1). Each random forest model was built from a trained bagging ensemble of 400 randomly generated decision trees (Kim et al., 2020) with the “randomForest” package in the R software (Liaw and Wiener, 2022). In this study, we used the RF2 model with PAR, air temperature, water height and relative humidity as environmental predictors because its performance indicators showed a high Pearson correlation coefficient (R2 = 0.88) and low values of root mean square error (RMSE = 1.27) and model bias (0.0024) allowing us to correctly gap-fill a large EC data (Table S1). The calculated uncertainty of the RF2 model on the resulting annual C budget was 0.43 %. Each tree was trained from bagged samples including 70 % of the initial dataset. The remaining 30 % of the data were used to estimate the fit of each random forest model. The model used was then able to explain 88 % of the variability in the test data. Daytime data were better explained than night-time data (59 % vs. 38 %), with light being the main parameter of the model. However, only 20 % of the night-time EC data were gap-filled with the random forest model. Using a partial dependence analysis and an ondelette analysis, we concluded that the relationships and temporal dynamics modelled allowed us to correctly fill the gaps in our dataset. However, extreme values of some predictors (i.e. PAR > 1000 µmol m-2 s-1) can reduce the random Forest model performance for estimation of EC data. This observation is common for random forest models, as they show poor results for extreme values. Other models, such as artificial neural networks, were also tested but showed poorer results (Table S1).

For all measured variables, the 10 min data did not follow a normal distribution (Shapiro–Wilk tests, p < 0.05). Non-parametric comparisons, such as the Mann–Whitney and Kruskal–Wallis tests, were carried out with a 0.05 level of significance. To assess the influence of meteorological and hydrological drivers on NEE fluxes at different temporal scales, we performed a pairwise Spearman's correlation analysis on the 10 min values and monthly mean values (“cor function” in R).

During the year 2020, temporal variations in NEE fluxes were studied at the seasonal and diurnal/tidal scales. Seasons were defined based on calendar dates: the winter period from 1 January 2020 to 19 March 2020 and from 21 to 31 December 2020, the spring period from 20 March 2020 to 19 June 2020, the summer period from 20 June 2020 to 21 September 2020 and the fall period from 22 September 2020 to 20 December 2020. Daytime and night-time were separated into PAR > 10 and PAR ≤ 10 µmol m-2 s-1, respectively. For the NEE flux analysis according to environmental drivers, NEE fluxes were grouped into five PAR groups (0 < PAR ≤ 10, 10 < PAR ≤ 500, 500 < PAR ≤ 1000, 1000 < PAR ≤ 1500 and 1500 < PAR ≤ 2000 µmol m-2 s-1) to reduce NEE fluctuations due to PAR variations. Water heights (Hw) measured at one location over the marsh (Fig. 2) relative to the mean sea level were used to distinguish emersion (Hw = 0 m at low tide) and immersion (Hw > 0 m at high tide) periods (see Sect. 2.3) and thus, the influence of tides on NEE fluxes.

To study marsh metabolism related to photosynthesis and respiration processes, measured NEE fluxes were partitioned into gross primary production (GPP) and ecosystem respiration (Reco), respectively. During marsh emersion, NEE fluxes occur at the marsh–atmosphere interface involving only benthic metabolism (or marsh metabolism) resulting in NEE = GPP-Reco. During marsh immersion, NEE fluxes are the result of benthic metabolism, planktonic metabolism and lateral C exchanges by tides thereby making it more difficult to study the marsh metabolism (Polsenaere et al., 2012). Negative NEE values indicated a marsh CO2 uptake from the atmosphere and positive values indicated a marsh CO2 source into the atmosphere. GPP was expressed in negative values and Reco was expressed in positive values. In this study, NEE flux partitioning into marsh metabolic fluxes (NEEmarsh) was performed according to the following equation using the model of Kowalski et al. (2003): 2NEEmarsh=GPP-Reco=a1PARa2+PAR-Reco, where a1 is the maximal photosynthetic CO2 uptake at light saturation (µmol CO2 m-2 s-1) and a2 is the PAR at half of the maximal photosynthetic CO2 uptake (µmol photon m-2 s-1). The a1/a2 ratio corresponds to photosynthetic efficiency (Kowalski et al., 2003). Reco was calculated as follows according to Wei et al. (2020b): 3Reco=R0exp(bTa), where Reco is the night-time ecosystem respiration (µmol CO2 m-2 s-1), R0 is the ecosystem respiration rate at 0 °C (µmol CO2 m-2 s-1), Ta is the air temperature (°C) and b is a response coefficient of the temperature variation (Wei et al., 2020b).

For NEE flux partitioning, estimations of the GPP coefficients (a1 and a2; Eq. 2) and Reco coefficients (R0 and b; Eq. 3) were performed by the least squares method (“minpack.lm” package in R) at the monthly scale only during emersion periods where measured NEE fluxes corresponded to estimated NEEmarsh fluxes. First, for each month, R0 and b were estimated during night-time emersion periods where NEE = Reco following Eq. (3) (Wei et al., 2020b). Then, a1 and a2 were estimated during daytime emersion periods using night-time respiration coefficients (R0 and b) where NEE = GPP-Reco following Eqs. (2) and (3) (Kowalski et al., 2003). Finally, NEEmarsh (net marsh metabolic fluxes without tidal influence) were calculated from PAR and Ta values measured at a 10 min frequency throughout the year using the monthly coefficients calculated for the partitioning (Eq. 2). As our ecosystem had a low phenological variation (Table S2), we concluded that a monthly time step for the coefficient estimation was sufficient to answer our study objectives. During emersion periods, monthly net C balances (i.e. budgets) of measured NEE and estimated NEEmarsh, as well as the monthly mean fluxes, were very similar (Table S3), confirming the correct NEE flux partitioning calculations done in this study.

Within the EC footprint, halophyte marsh vegetation (66 %) composed of Halimione portulacoides, Spartina maritima and Suaeda vera mainly dominated, whereas muds and channels only accounted for 27 and 7 %, respectively (Fig. 2). The area occupied by S. vera, crossing the EC footprint from WNW to ESE (Table 1), corresponded to the highest marsh level that was partly immersed only during the highest tidal amplitudes (Fig. 2). H. portulacoides and S. maritima occupied mostly the NNE (70 %), SSE (69 %), WSW (68 %) and SSW (67 %) wind sectors. In contrast, mud habitats mostly covered the NNW sector, where the lowest vegetation cover was found (Fig. 2; Table 1). The highest channel area was found in the SSW sector (Fig. 2; Table 1).

Throughout the year 2020, the full seasonal range in solar radiation was measured (Fig. 3a) with an increase in daytime PAR from winter (lowest light season) to summer (brightest season). A similar seasonal pattern was recorded for air temperatures (Ta) with values ranging from 1.5 °C in winter (coldest season) to 33.6 °C in summer (warmest season; Fig. 3b). On average, the winter and fall seasons were the wettest (RH > 82 %), associated with the lowest vapour pressure deficit (VPD) values, whereas spring and summer were the driest ones (RH < 75 %), associated with the highest VPD values (Fig. 3b). Indeed, the highest and lowest cumulative rainfalls were recorded in fall (342 mm) and summer (62 mm), respectively. The highest mean seasonal wind speed was measured in winter (4.9 ± 2.3 ms-1) with maximal speeds up to 13 ms-1 (Fig. 3c). Winds came mostly from the SSW–WSW sectors both in winter (55 %) and summer (41 %) and from the NNE–ENE sectors both in spring (51 %) and fall (31 %) (Fig. 2). Tidal activities reflected the typical hydrological conditions of the Atlantic coasts with a bi-monthly succession of spring tides and neap tides (Fig. 3d). Water heights (Hw) strongly varied according to tidal amplitudes with a maximal Hw of 1.4 m during neap tides and 2.0 m during spring tides (overall annual mean of 0.6 ± 0.4 m; Fig. 3d). Throughout the year, 25.5 % of the EC data were measured when the salt marsh was immersed through variable immersion durations and water heights (Table 2). On average, the daily immersion durations ranged between 5.7 hd-1 in winter (23.7 % of the EC data) and 6.5 hd-1 in fall (28 % of the EC data). In winter, the EC data during immersion were split into 19 % for 0 < Hw < 1 m and 4.7 % for 1 < Hw < 2 m, whereas in fall, these latter were split into 20 % for 0 < Hw < 1 m and 8 % for 1 < Hw < 2 m. In summer, the lowest marsh immersion was measured with no Hw value higher than 1.5 m (Table 2).

The annual mean NEE value was -1.27 ± 3.48 µmol m-2 s-1 with strong temporal variabilities recorded over both long and short timescales (Fig. 3e). Significant NEE variations were highlighted between each season (Kruskal–Wallis test, p < 0.001) where, on average, the highest and lowest atmospheric CO2 sinks were recorded in spring (-1.93 ± 3.84 µmol m-2 s-1) and fall (-0.59 ± 2.83 µmol m-2 s-1), respectively (Fig. 4). NEE flux partitioning gave an annual mean NEEmarsh value of -1.28 ± 3.16 µmol m-2 s-1, ranging from -2.00 ± 3.49 µmol m-2 s-1 in spring to -0.53 ± 2.51 µmol m-2 s-1 in fall. On average, in winter and fall, the measured NEE values were more negative than the estimated NEEmarsh values, whereas in spring and summer, the opposite trend was recorded (Fig. 4). Contrary to NEE and NEEmarsh, the highest seasonal values of GPP and Reco were estimated in summer, whereas the lowest seasonal values were estimated in winter (Fig. 4). The highest and lowest photosynthetic efficiencies (a1/a2 ratio) were found in winter (-2.08 × 10-2) and summer (-1.36 × 10-2), respectively.

At each season, significant diurnal differences in NEE fluxes were highlighted (Mann–Whitney tests, p < 0.05) with, on average, an atmospheric CO2 sink during daytime and an atmospheric CO2 source during night-time, irrespective of emersion or immersion periods (Table 3). For instance, in spring, NEE means were -3.93 ± 3.72 and 1.06 ± 1.09 µmol m-2 s-1 during daytime and night-time, respectively (Fig. 5b). Over all seasons, similar diurnal variations in measured NEE and estimated NEEmarsh were recorded with, on average, a rapid increase in CO2 uptake during the morning up to the middle of the day (low Ta and VPD values) and then, a decrease in CO2 uptake during the afternoon (high Ta and VPD values) to become a CO2 source during night-time (Figs. 5 and S2). On average, during the afternoon, the GPP decreases and Reco increases explained the measured decrease in CO2 uptake (Fig. 5). For each season, the highest marsh CO2 uptakes were measured during daytime emersion periods between 12:00 and 13:00 UT (maximal PAR levels), with the latter increasing from winter (-4.84 ± 2.87 µmol m-2 s-1) to spring–summer (-6.94 ± 2.80 µmol m-2 s-1; Fig. 5).

At each season, the tidal rhythm strongly disrupted NEE fluxes with, in general, no change in the marsh metabolism status (sink/source). During daytime, significantly lower CO2 uptakes were recorded during immersion than during emersion (Mann–Whitney tests, p < 0.05) when marsh plants were mostly immersed in tidal waters, and during night-time, a similar tidal pattern was recorded for CO2 emissions (Mann–Whitney tests, p < 0.05; Table 3). For instance, in spring, NEE means were -4.39 ± 3.76 and -2.59 ± 3.24 µmol m-2 s-1 during daytime emersion and daytime immersion, respectively, and were 1.25 ± 0.98 and 0.51 ± 1.22 µmol m-2 s-1 during night-time emersion and night-time immersion, respectively. In winter, during some night-time periods, weak CO2 sinks were recorded both during emersion (-0.79 ± 0.84 µmol m-2 s-1; 137 h over 71 d) and immersion (-0.82 ± 0.91 µmol m-2 s-1; 143 h over 55 d associated with a mean Hw of 0.80 m; Fig. S2). The maximal CO2 uptakes were -4.80 and -5.31 µmol m-2 s-1 during night-time emersion and night-time immersion, respectively (Table 3).

Throughout the year, NEE fluxes were significantly controlled by solar radiations and air temperatures at the multiple timescales studied, thereby favouring marsh CO2 uptake. During daytime (PAR > 10 µmol m-2 s-1), PAR and Ta displayed the strongest negative correlations with NEE at both the monthly scale (-0.87 and -0.65, respectively; n = 12, p < 0.05) and the 10 min scale (-0.77 and -0.21, respectively; n = 27 160, p < 0.05). The highest and lowest correlations between NEE and PAR were recorded for 10 < PAR ≤ 500 and for 1500 < PAR ≤ 2000 µmol m-2 s-1, respectively, confirming the rapid increase or decrease in CO2 uptake for low daytime PAR values (Fig. 6a). During daytime, vapour pressure deficit (VPD) was negatively correlated with NEE (-0.31; n = 27 160, p < 0.05) producing a large reduction in CO2 uptake for all PAR levels and even led to a switch from sink to source of atmospheric CO2 from VPD > 1200 Pa for low PAR levels (PAR ≤ 500 µmol m-2 s-1; Fig. 6b). During night-time and daytime, air temperature (Ta) was positively (0.54; n = 27 190, p < 0.05) and negatively (-0.21; n = 25 544, p < 0.05) correlated with NEE, respectively. However, from PAR > 500 µmol m-2 s-1, high Ta values (> 20 °C) decreased CO2 uptake for all PAR levels (Fig. 6c). Water temperature (Tw) did not influence NEE during immersion (Fig. 6d). Indeed, for PAR > 500 µmol m-2 s-1 and Hw > 0.5 m, no significant relationship was found between NEE and Tw (n = 1215; p = 0.26). For low PAR levels (PAR ≤ 500 µmol m-2 s-1), wind speeds quickly increased CO2 uptake, whereas for high PAR levels (PAR > 500 µmol m-2 s-1), CO2 uptake was increased only for wind speeds higher than 7 ms-1 (Fig. 6e). For wind directions, a spatial heterogeneity of NEE was recorded according to wind sectors both during daytime and night-time (Fig. 6f). Within the footprint area composed of an assemblage of plants and muds (Fig. 2), the highest CO2 uptakes were generally recorded from the southern sectors (high vegetation:mud ratios) whereas, the lowest CO2 uptakes were generally recorded from the northern sectors (low vegetation:mud ratios; Fig. 7). For instance, our sectorial NEE analysis during daytime emersion showed that the SSE sector (vegetation:mud ratio of 2.4; Table 1) did uptake 32 % (winter), 25 % (spring) and 50 % (fall) times more atmospheric CO2 than the NNW sector (vegetation:mud ratio of 0.8; Table 1). Moreover, in winter and fall, we highlighted that CO2 uptake rates of H. portulacoides (C3 species) were significantly higher than S. maritima (C4 species) ones by comparing the SSE (60 % of H. portulacoides and 9 % of S. maritima) and WSW (33 % of H. portulacoides and 35 % of S. maritima) sectors during daytime emersion (Mann–Whitney tests, p < 0.0001). In contrast, in summer, no significant difference in NEE fluxes was recorded between these two sectors (Mann–Whitney test, p = 0.06; Fig. 7) and, more generally, between the different wind sectors (Fig. 7; Table 1). For all seasons, during night-time emersion, we recorded that southern sectors (ESE, SSE and SSW) emitted higher atmospheric CO2 than northern sectors (NNE and ENE), especially in winter and fall (Fig. 7; Table 1).

The tidal rhythm strongly influenced NEE fluxes during immersion depending on water heights (Hw) and PAR levels (Figs. 8 and S3). Throughout the year, NEE were positively correlated with Hw during the day but negatively correlated during the night (Fig. 8). More precisely, night-time immersion strongly reduced CO2 emissions and even led to a switch from source to sink of atmospheric CO2 from Hw > 0.4 m in winter (Fig. 8a), Hw > 0.7 m in spring (Fig. 8b), Hw > 1.4 m in summer (Fig. 8c) and Hw > 1 m in fall (Fig. 8d), on average. For low daytime PAR levels (PAR ≤ 500 µmol m-2 s-1), immersion only slightly reduced CO2 uptake (Fig. 8c). On the contrary, for higher daytime PAR levels (PAR > 500 µmol m-2 s-1), immersion strongly reduced CO2 uptake, especially from Hw > 0.5 m, to reach the lowest CO2 sinks from Hw > 1.0 m, irrespective of the PAR levels (Fig. 8c).

Throughout the year, the annual NEE value was -483.6 gCm-2yr-1, associated with immersion duration of 6.1 hd-1, on average. Simultaneously, estimated GPP and Reco (marsh metabolic fluxes without tidal influence) absorbed and emitted 1019.4 and 533.2 gCm-2yr-1, respectively, resulting in an annual estimated NEEmarsh value similar to the measured NEE value (Fig. 9). At the seasonal scale, the highest CO2 uptakes occurred in spring and summer, associated with the lowest marsh immersion levels, and the lowest CO2 uptakes occurred in winter and fall, associated with the highest marsh immersion levels (Tables 2 and 4). In winter and fall, when the daytime immersion periods were the shortest, net C balances from measured NEE gave higher values than net C balances from estimated NEEmarsh (+7.9 and +6.2 gCm-2, respectively; Table 4). Conversely, in spring and summer when the daytime immersion periods were the longest, the opposite pattern was observed between measured NEE values and estimated NEEmarsh values (-7.3 and -9.9 gCm-2, respectively; Table 4).

In the present EC study, the salt marsh absorbed 483 gCm-2yr-1 from the atmosphere. This net C balance (i.e. budget) was lower than the values estimated for global tidal wetlands (1125 gCm-2yr-1; Bauer et al., 2013) and for tidal marshes on the Atlantic coast of the United States (775 gCm-2yr-1; Wang et al., 2016) but similar to the C balance estimated by Alongi (2020) for global salt marshes (382 gCm-2yr-1).

Currently, an increasing number of EC measurements are being taken in salt marshes in order to obtain continuous NEE data series as well as to increase knowledge about the associated metabolic processes and fluxes for these tidal systems (Table 5) (Schäfer et al., 2014; Forbrich et al., 2018; Knox et al., 2018). These EC studies confirmed the estimates of CO2 sinks in salt marshes (Wang et al., 2016; Alongi, 2020) but also revealed strong NEE flux heterogeneities according to climatic conditions and anthropogenic influences (Herbst et al., 2013; Schäfer et al., 2019). For instance, NEE measured in a natural salt marsh (S. alterniflora, S. maritima and D. spicata) showed a net C uptake from the atmosphere with high interannual variations in C balances (Table 5) mainly due to rainfall during the growing season for marsh plants (Forbrich et al., 2018). By comparison, in an urban tidal marsh, Schäfer et al. (2014) reported a higher interannual variability from 984 gCm-2 in 2009 to -310 gCm-2 in 2012 due to management practices and plant species (P. australis and S. alterniflora in 2009 and total elimination of P. australis in 2012; Table 5). In the same area, in another restored salt marsh in which the P. australis monoculture was replaced by a high diversity of emergent marsh plants (S. patens, S. cynosuroides, S. alterniflora and D. spicata), a net CO2 uptake was recorded (Table 5) which once again confirms the importance of land management practices in marsh C balances (Artigas et al., 2015). In our studied salt marsh, the natural management for several decades has allowed for a return to the natural site hydrodynamics and the development of productive marsh halophytes, mainly composed of H. portulacoides and S. maritima (59 % of the footprint area). However, past human activities and water management practices for salt farming have shaped the marsh typology (channel network, humps and dykes), producing a time-delayed immersion of plants and muds between high and low marsh areas during spring tides. Thus, due to this emersion/immersion heterogeneity, mud and S. maritima were quickly immersed by coastal waters, whereas the whole immersion of marsh habitats only occurred during the highest tidal amplitudes favouring a higher atmospheric CO2 uptake by H. portulacoides and S. vera. During the year 2020, our rewilded salt marsh took up more C from the atmosphere mainly due to strong plant photosynthesis than the other salt, brackish and freshwater marshes reported in the literature (Table 5). However, the net C balances calculated with the EC method are still too scarce to be able to take all temporal and spatial variabilities of salt marshes into account. Based on biomass production measurements in salt marshes, Sousa et al. (2010) estimated that the NPP of H. portulacoides was 505 gCm-2yr-1, whereas the NPP of S. maritima varied between 367 and 959 gCm-2yr-1 depending on the chemical–physical characteristics and marsh maturity. Thus, the net metabolism of these halophytic plants could play an important role in our net C balance but, according to “the marsh CO2 pump” (Wang et al., 2016), a significant proportion of marsh NPP was respirated by heterotrophic processes and then (1) emitted as atmospheric CO2 (38 ± 11 %) and (2) exported by tides as DIC (37 ± 15 %; Song et al., 2023).

Moreover, despite a lower benthic metabolism (photosynthesis and respiration) of muds than evergreen plants (Fig. 7), the microphytobenthos which can develop on mudflats (27 % of the footprint area) may also contribute to marsh production during daytime emersion, as highlighted in our studied marsh where static chamber measurements performed in March 2023 at midday showed a net CO2 uptake to a non-vegetated mudflat (NEE mean of -2.92 µmol m-2 s-1; unpublished results) and confirmed in an estuarine wetland in China (Xi et al., 2019). On an intertidal flat (France), EC measurements even showed a higher daily benthic metabolism with microphytobenthos (1.72 gCm-2d-1; September/October 2007) than with Zostera noltei (1.25 gCm-2d-1; July and September 2008), confirming the high biological productivity of mudflats (Polsenaere et al., 2012). However, due to the specific assemblage of our studied marsh (Fig. 2), it remains complex to accurately study these habitat effects (plants vs. microphytobenthos) on NEE fluxes at the marsh scale and draw more general conclusions. Thus, the microphytobenthos could play a significant role in the atmospheric CO2 uptake of salt marshes but also, more generally, in the carbon cycle of the coastal ocean because the resuspension of the microphytobenthos primary production during tidal immersion induce a large export of organic carbon from muds to coastal waters (up to 60 % of the benthic primary production in a nearby tidal flat; Savelli et al., 2019). These fast-growing primary producers with high labile organic carbon could also be quickly degraded locally by microbial remineralization (Ruttenberg, 1992; De Brouwer and Stal, 2001; Morelle et al., 2022), contrary to evergreen plants contributing to long-term “blue carbon” burial in sediments (Mcleod et al., 2011).

At the annual scale in 2020, the tidal rhythm did not significantly affect the net C balance of the studied salt marsh since similar annual measured NEE and estimated NEEmarsh values were recorded (Fig. 9). The loss of CO2 uptake measured during daytime immersion due to a GPP decrease could be compensated by night-time immersion where CO2 emissions and Reco were inhibited. However, strong temporal variabilities were measured, especially between the growing and non-growing seasons. In winter and fall, the salt marsh did uptake more C from the atmosphere with the tidal influence (measured NEE) than without (estimated NEEmarsh), especially in December (+35.7 %), November (+19.7 %) and January (+15.4 %), associated with the highest photosynthetic efficiencies. An opposite trend was observed in spring and summer with a reduction in net C uptake under tidal influence, especially in August (-16.9 %) and September (-9.8 %). This significant difference in the seasonal C balances could be mainly related to the photoperiod of immersion periods. We demonstrated that daytime immersion decreased CO2 uptake, whereas night-time immersion decreased CO2 emissions up to a change in metabolic status for the highest immersion levels. Thus, during seasons where daytime immersion primarily occurs, such as spring and summer, the salt marsh did uptake less atmospheric CO2 with tidal influence, whereas seasons that mostly have night-time immersion did uptake more atmospheric CO2 with tidal influence (Table 4). However, this unpublished result was only possible provided that the salt marsh switched from a source to a sink of CO2 during night-time immersion due to water undersaturation with respect to the atmosphere. In a salt marsh on Sapelo Island (USA), Nahrawi et al. (2020) highlighted tidal CO2 flux reductions all year round by distinguishing neap tide and spring tide periods. Their results showed that the highest and lowest reductions in C uptake occurred in spring (-34 %) and summer (-13 %), respectively, with a similar but greater tidal influence on the C uptake values compared with our study.

To better constrain the tidal influence on the metabolism of the salt marsh, further investigations have been carried out in 2021 in parallel with our EC measurements, with the construction of a digital field model for water heights that can be used to spatially determine, over the whole EC footprint, the exact areas of immersion and emersion (especially for the low water levels) of the marsh in each sector at a 10 min step. Similarly, during marsh immersion, EC measurements do not directly capture CO2 fluxes from benthic metabolism because of the physical barrier of the water and the lower CO2 diffusion rates in water than in air. Consequently, at the same time as when the NEE measurements were taken, water pCO2, inorganic and organic carbon concentrations associated with planktonic metabolism were determined each season through 24 h cycles to provide essential information on the contribution of planktonic communities and plants to CO2 fluxes during immersion (Mayen et al., 2024). The lateral carbon export from salt marshes by tides plays a significant role in the coastal ocean carbon cycle (Guo et al., 2009; Wang et al., 2016). Plant respiration and microbial mineralization of marsh NPP could generate Dissolved Inorganic Carbon (DIC) in waters associated with a strong benthic–pelagic coupling. Thus, our 2021 measurements of the carbon parameters, planktonic metabolism (production and respiration) and other relevant biogeochemical variables over 24 h diurnal cycles, along with measurements of the soil compartment (root OM production vs. mineralization; Arnaud et al., 2024) carried out in the EC footprint, would allow for a more integrative calculation of the studied marsh carbon budget (Mayen et al., 2024). One advantage of the EC measurements is the aggregation of CO2 fluxes from all compartments (waterbodies, soil, plants and atmosphere) in salt marshes. Yet, through this flux aggregation, we cannot mechanistically understand each marsh compartment, and therefore it can be challenging to predict CO2 fluxes under multiple global changes. Therefore, future contributions should try to simultaneously quantify all these compartments, especially soil as it is where most of the carbon is stored in salt marshes (Arnaud et al., 2024). Ongoing atmospheric CO2 exchange measurements are actually carried out since January 2023 up north over Aiguillon (intertidal) Bay in France where we precisely deployed an EC station at the edge between the tidal mud flat on the west side and salt marsh habitats on the east side of the footprint along with benthic chamber flux and water, sediment, soil carbon measurements and satellite analysis at each season to specially address questions on relative habitat (mudflat vs. salt marshes) influence on atmospheric CO2 exchanges (Pierre Polsenaere, personal communication, 2023).