Tidal influence on dissolved CO₂ at Sapelo Island, Georgia, USA

The study took place 21 July–17 August in a Sapelo Island marsh located at 31.44 lat, −81.28 long. Sapelo Island (31.48 lat, 81.24 long) is a long-term ecological research site and barrier island off the coast of Georgia, USA, in the humid subtropical climate (figure 1). There is a history of interdisciplinary scientific collaboration between researchers at Sapelo Island, Georgia and University of Wisconsin-Madison on biogeochemistry studies dating back more than 50 years (Ragotzkie and Bryson 1955, Jones 1980). Research on marine and estuarine ecosystems at Sapelo Island by students at UW-Madison continues as part of a fall course offered by the Department of Integrative Biology (Kara and Shade 2009). Salt marsh is a common habitat around the island, but this study specifically focused on the location denoted by the blue star on the map below (figure 1). Vegetation in the marsh is predominantly Spartina alterniflora and can reach as high as 2 m (Pinton et al 2020). Salicornia virginica, Batis maritima, and Juncus roemarianus can be found in smaller numbers where elevation and salinity are adequate (Sanders 2021).

Zoom In Zoom Out Reset image size

Sapelo Island hydrography is characterized by a strong tide (Ragotzkie and Bryson 1955), which varied as much as 3.23 m during the study period at Hunt Dock. Tide height, water temperature, salinity, and turbidity were recorded by NOAA (National Oceanic and Atmospheric Administration) tide gages in 15 min intervals at three different locations: Hunt Dock, Cabretta Creek, and Dean Creek (figure 1; NOAA Tides & Currents 2021, NOAA NERRS 2022). The 15 min tide height, water temperature, salinity, and turbidity, and 1 h moon phase data were linearly interpolated to match exactly the pCO₂ timestamps for the purpose of statistical testing. After interpolation, all data had a resolution of approximately 1 h. Tides lasted an average of 12 h, 25 min, and 15 s from high tide to high tide during the study period. Water temperature ranged from 27.2 °C to 33.4 °C at Hunt Dock during the study period. Geocentric moon phase data at hourly intervals in 2021 was retrieved from NASA Scientific Visualization Studio (Wright n.d.).

Atmospheric CO₂ concentration (ppm) was from the Georgia Coastal Ecosystems LTER Flux Tower (US-GCE) at 31.44 lat, −81.28 long. US-GCE is owned and maintained by the University of Georgia Marine Institute (UGAMI), located 5.5 km from US-GCE, at 31.40 lat, −81.28 long. It is a 10 m tall triangle tower installed on an elevated dock platform. CO₂ is measured by an enclosed CO₂ gas analyzer (CSAT LI7200) at 10 Hz frequency, then averaged to produce half-hourly data (Feagin et al 2020). Half-hourly data was used in this study due to less noise, then was linearly interpolated to exactly match pCO₂ timestamps. The tower is accessed from UGAMI by car, boat, and then walk through the marsh.

The Smart Rock is a low-cost submersible sensor developed by the OPEnS Lab (Openly Published Environmental Sensing Lab) and provided by the Consortium of Universities for the Advancement of Hydrologic Sciences, Inc. (∼$200). It measures turbidity and salinity but can also measure water level and temperature when in calibration mode (Veach 2019). Salinity was measured using the Gravity: Analog Total Dissolved Solids Sensor/Meter for Arduino manufactured by DF Robot, with a measurement range 0–1000 ppm (∼0–2000 uS cm⁻¹) and accuracy within 10%. Turbidity was measured using the Gravity: Analog Turbidity Sensor for Arduino, also manufactured by DF Robot, with a measurement range of 0–4.5 V and accuracy within 0.3 V. Turbidity and salinity were recorded by the Smart Rock once every 20 min. Turbidity and salinity from both the Smart Rock and NOAA tide gage are presented in the results. However, only measurements from NOAA tide gages were used for analysis.

The CO₂-LAMP was deployed at the base of US-GCE. The CO₂-LAMP measured pCO₂ in both water and air in the tidal marsh. A waterproofed NDIR CO₂ gas analyzer was used to measure pCO₂. Data were logged using a CO₂-LAMP recently developed by Blackstock et al (2019). In this platform, passive equilibration of dissolved CO₂ in the water exchanges with a 'headspace' volume containing the CO₂ gas analyzer enclosed by an expanded polytetrafluoroethylene (ePTFE) semi-permeable membrane. As previously reported by Johnson et al (2010), diffusion of CO₂ in water primarily limits time needed for equilibration, but where runoff rapidly introduces water of dissimilar pCO₂ values, passive equilibrators may not fully capture peak pCO₂ values in some cases (Yoon et al 2016).

The NDIR gas analyzer used was a K30 10% manufactured by Senseair AB (Delbo, Sweden). The K30 10% manufacturer reported accuracy is ±300 ppm with a resolution of 10 ppm CO₂ and capability of measuring up to 100 000 ppm. Prior to deployment, reference measurements were made using 0 ppm CO₂ (99% N₂, O₂-balance) and 2000 ± 40 ppm CO₂ (N₂ balance) reference gases. After deployment, reference measurements were remade using a 0 ppm CO₂ (N₂) and 550 ± 2 ppm CO₂. During reference measurements, the reference gas exchanges across the semi-permeable ePTFE surface and is measured by the K30 until equilibration is reached. Pre-deployment 0 ppm CO₂ reference gas measurements were measured as 0 ppm CO₂ by the K30, and the 2000 ppm CO₂ reference measurements were measured as 1980 ppm CO₂, which were within K30 analytical and reference gas concentration uncertainty. Post-deployment 0 ppm CO₂ reference gas measurements were measured as 0 ppm CO₂ by the K30, and the 550 ppm CO₂ reference measurements were measured as 530 ppm CO_2, indicating a consistent, slight underestimation (20 ppm) of gas concentrations and negligible drift relative to the range of the gas analyzer and previous ranges observed in surface waters at Sapelo Island (Wang and Cai 2004).

CO₂-LAMP data were recorded every 10 s for 20 min, followed by 45 min of sleep. All measurements except the last measurement from each measurement cycle were removed during post-processing, resulting in approximately 22 datapoints per day with a resolution of 1.09 h each and 589 datapoints overall. Due to the unique CO₂-LAMP measurement frequency, 15 min tide data, 30 min flux data, and hourly moon phase included in our analysis were interpolated to match the temporal resolution of pCO₂. Values of pCO₂ less than or equal to 0 ppm occurred infrequently during measurement cycles but were marked as non-physical values and replaced with the value of the previous measurement. Multiple sensors are preferred to compare pCO₂ data, check for drift, and determine the degree of spatial variability. However, power limitations restricted simultaneous deployment of multiple CO₂-LAMPs.

Another issue which impacted data collection was biological fouling, which is common when monitoring in coastal or other aquatic ecosystems. Minimum daily pCO₂ measurements determined by the 22-point moving minimum linearly increased throughout the study, likely due to accumulation of biofilms and algae over the CO₂-LAMP semi-permeable membrane (figure 2). Original data were detrended by fitting the daily moving minimum pCO₂ to a simple linear regression. Fitted data was then subtracted from the original data, and atmospheric background CO₂ of 420 ppm was added back in.

Zoom In Zoom Out Reset image size

The CO₂-LAMP was attached to a foam board capable of varying with tide height. This deployment was based on a similar, unpublished design by Thomas L O'Halloran at Clemson University (personal communication). The CO₂-LAMP protective housing was attached to a foam swimming board' using marine epoxy and an extra-large zip tie to keep it elevated at the water surface with the waterproofed gas analyzer submerged below the water surface (figure 3). Because pCO₂ concentrations at the air–water interface are most important to wind-driven gas exchange (Wanninkhof 2014), and pCO₂ can vary within the water column (Beaubien et al 2014), consistent depth of measurement near the water surface is important. The bottom baseboard was held down with four t-shaped PVC pipes with holes drilled into the piping for drainage, which prevented the pipes from moving upwards over time. The top baseboard, sized 60.96 × 60.96 × 0.635 cm, was secured to the flux tower with zip ties threaded through holes in the wood and decking. Metal flanges with male adapter fittings were screwed into the top and bottom boards with wood screws. Three hollow five-foot copper pipes were then easily placed inside the adapter fittings on both ends and pressed or hammered down, with no further seal necessary. Only metal flanges and adapters can be used because the allowances fit with copper piping. Plastic pipe fittings did not fit with the metal pipes or flanges, though transition adapters could work in the future.

Zoom In Zoom Out Reset image size

Lateral C export is referred to as a 'potential' throughout this study because it is solely based on contributions from pCO₂ and does not account for the import or export of other forms of C. Tide velocity was calculated as change in tide height divided by change in time between pCO₂ measurements. Atmospheric CO₂ (ppm) measured at the flux tower was subtracted from pCO₂ prior to the potential tidal C export estimation (table 1) and linear regression (figure 5) to eliminate any measurements potentially taken in air. Any measurements below zero after this correction were set to zero. Lateral import was assumed to occur when tide was increasing, and export when tide was decreasing. Cumulative imported and exported C were calculated for each tide cycle. Cumulative tidal imports were subtracted from cumulative tidal exports to achieve net export (gC m⁻² s⁻¹). Net export was then multiplied by the length of each tidal cycle in seconds and averaged to find the potential average C export per tidal cycle. Net export in seconds was also summed across the duration of the study to find potential total C export for the entire study. The equation used for calculations was:

$$\begin{equation*}{\text{C}}\;{} {\text{import}}{} \;{\text{or}}\;{} {\text{export}}{} \left( {{\text{gC}}\;{{\text{m}}^{ - 2}}{{\text{s}}^{ - 1}}} \right) = {} \frac{{p{\text{C}}{{\text{O}}_2}\left( {{\text{ppm}}} \right) \times {\text{tide}}\;{} {\text{velocity}}{} \left( {{\text{m }}{{\text{s}}^{ - 1}}} \right) \times {\rho _{{\text{air}}}}\left( {{\text{g}}\;{{\text{m}}^{ - 3}}} \right) \times 12\left( {{\text{g}}\;{\text{mol}}{{} ^{ - 1}}\;{{\text{C}}^{ - 1}}} \right)}}{{28.97\left( {{\text{mo}}{{\text{l}}^{ - 1}}\;{\text{ai}}{{\text{r}}^{ - 1}}} \right) \times {{10}^6}}}.\end{equation*}$$

ρ_air was assumed to be 1225 g m⁻³. About 10⁶ is the conversion factor for pCO₂ (ppm).

Vector autoregression and Granger causality were the selected methods for hypothesis testing. These methods work together and are ideal for complex datasets with multiple predictors. Vector autoregressive models are stationary, multivariate time series models which can be used to describe random stochastic processes with limited prior knowledge of the forces influencing each variable. One equation is created for each dependent variable using a linear function of the lagged dependent variable and other information. Akaike information criterion (AIC) of each model is calculated from the number of predictor variables and model performance. The best-fit model according to the AIC is then used in the Granger causality test.

The Granger causality test works well to examine underlying relationships in a dataset which the vector autoregression cannot. Granger causality originated in econometrics, but has since been adopted by many other fields, including the geosciences (Granger 1969, Detto et al 2012, Desai 2014). Granger causality should only be computed on stationary, modeled data rather than original or nonstationary data, and cannot be computed for singular matrices made of multiple, strongly interdependent variables. This model-test combination is a computationally simple way of assessing temporal relationships between variables. Some disadvantages of Granger causality testing are that results may be skewed by infrequent sampling, too-frequent sampling, nonlinear causal relationships, and more. Interpretations of Granger causality testing should align with the physical dynamics of a system, and not be used as stand-alone statistical results due to these limitations.

Hourly moon phase and 15 min tide height, salinity, turbidity, and water temperature were linearly interpolated to match pCO₂ measurement times, which had a resolution of 1.09 h, for statistical testing. Moon phase was eliminated from statistical testing because it created a singular matrix due to its close relationship with tide height. The vector autoregression model was fitted to the dataset including tide height, salinity, turbidity, and pCO₂ with lags ranging from 1 to 5. Model estimation was initialized using the first six observations as pre-sample data. The model with the best fit, according to the lowest AIC, was then used in the Granger causality test. To evaluate the accuracy of the vector autoregression model fit, half of the data (even-numbered observations) were withheld from the fitting procedure and pCO₂ was estimated from the fitted model. The ability of the fitted model to predict the withheld values was then assessed using correlation and significance between pCO₂ predictions and observations. Adjusted R² of modeled tide height, salinity, turbidity, and pCO₂ was also calculated. Lastly, model consistency and autopower spectral densities of the vector autoregression were reported. Model consistency measures the proportion of the correlation structure in both the observed and modeled data. Autopower spectra can be used to determine overlapping peak frequencies between observed and modeled data.

The leave-one-out Granger causality test was performed using the multivariate Granger causality toolbox in MATLAB, which assesses whether each variable in a best-fit vector autoregression model forecasts another variable (Barnett and Seth 2014). Results of the Granger causality test are described using p-values of each Granger causality and causal density, which is the average of Granger causalities between each pair of variables in a dataset with respect to the variables not included in each pairwise causality test. Datasets composed of variables that behave relatively independently will have higher causal densities because their predictors produce unique and useful information. Datasets with completely unrelated variables will have scores closer to zero, as the dynamics guiding each variable are different (Seth et al 2011).

To estimate the uncertainty of our approach, we constructed 100 pCO₂ alternative timeseries datasets using a Monte Carlo style approach. The 100 datasets were created by randomly selecting numbers in the range of the original measurement ±300 ppm, to account for sensor accuracy. The potential tidal export was then recalculated using the simulated datasets to produce estimated upper and lower limits of tidal exports from the marsh.

The twice daily, or semidiurnal, tide at Sapelo Island had a clear influence on pCO₂ (figure 4(A)). Tide also influenced changes in salinity and turbidity (figures 4(B) and (C)). Tide height at Hunt Dock reached a maximum of 4.5 m during the study (figure 4(D)). Water temperature was not strongly correlated with pCO₂ (R = −0.09, p = 0.02). Surface water salinity as electrical conductivity surpassed the maximum limit of the Smart Rock sensor. Salinity peaks lined up well with peaks in tide height. Low tide was reflected in pCO₂ measurements and tide height, but not by the salinity sensor. Salinity and pCO₂ were also poorly aligned from 31 July to 3 August when both dropped to low levels. Relative turbidity indicated murky water was consistently brought in by the tide, and was especially murky on 25–27 July, and 21 August. Relative turbidity continued to decrease throughout the study, which means that surface water got darker with time.

Zoom In Zoom Out Reset image size

Values of pCO₂ ranged from 390 to 6106 ppm ± 54 (standard error), with a mean of 1421 ppm. This equated to a range of 0–5596 ppm throughout the study period after accounting for atmospheric CO₂, with a mean of 1081 ppm (0 standard error). There was a positive linear relationship (R² = 0.35) between mean tide height and median binned pCO₂ (figure 5). Most outliers, having pCO₂ values greater than 1.5 times the interquartile range, occurred at tide heights below 3 m.

Zoom In Zoom Out Reset image size

Mean pCO₂ during incoming tide was roughly half of mean pCO₂ during outgoing tide, with an average concentration of 492 ppm compared to 1569 ppm after accounting for atmospheric CO₂. This resulted in a potential net lateral C export of 0.03–0.06 gC m⁻² tide cycle⁻¹ and 1.71–3.02 gC m⁻² during the entire study period as pCO₂ at all tide gages, assuming 100% sensor accuracy (table 1). Potential tidal C export was largest when using Hunt Dock tide data. Potential lateral net C exports per tide cycle based on tide gages at Cabretta Creek and Dean Creek were similar despite different distances from the CO₂-LAMP when 100% sensor accuracy was assumed.

Flow out of the marsh was highly variable dependent on sensor accuracy. Potential lateral net C export estimated using Hunt Dock tide data was 3.20 gC m⁻² higher than estimations based on tide data at Cabretta Creek and 5.82 gC m⁻² higher than estimations based on tide data at Dean Creek when summed across the entire study period according to the Monte Carlo style simulation. Although the range produced by the upper and lower bounds of the Monte Carlo style simulations for each tide gage were relatively small on average per tide cycle (range of 0.12 gC m⁻² tide cycle⁻¹), these led to large uncertainties in total C export for the study period overall (average range of 6.22 gC m⁻² study period⁻¹). Lower bounds of the Monte Carlo style estimates were similar to observations, but upper bounds were at least two times greater than observations across all tide gages.

The vector autoregression model explained 98% of the variation in tide height, 96% of the variation in salinity, 97% of the variation in water temperature, and 58% of the variation in pCO₂ when considering the number of independent variables. The model results did not fit well with observations of turbidity, with an adjusted R² value of −0.28. Model consistency was low at only 5%. Strong peaks in the autopower spectra for tide height and pCO₂ corresponded with the frequency of the semidiurnal tide at roughly 2.3 × 10⁻⁵ Hz (figures 6(A) and (D)). A similar but weaker maximum frequency was observed in the autopower spectra for salinity (figure 6(B)). Turbidity reflected maximum autopower at the semidiurnal frequency for modeled data but not for observed data (figure 6(C)). Vector autoregression predictions of pCO₂ aligned well with observations when withholding half of the data (R = 0.84) and when using the full dataset to fit the model (R = 0.83).

Zoom In Zoom Out Reset image size

Granger causality testing of the vector autoregression model using Hunt Dock tide data revealed causal relationships between tide height and water temperature, water temperature and pCO₂, tide height and pCO₂, and tide height and salinity (figure 7). Causal relationship direction (e.g. pCO₂ –> tide ht.) was the opposite of reality in some cases. Mean causal density was 0.03.

Zoom In Zoom Out Reset image size

We demonstrated that the CO₂-LAMP could measure the influence of tides on dissolved CO₂ in a tidal marsh using a low-cost platform ($346–497 in 2022). Uncertainty in pCO₂ measurements due to sensor accuracy using the manufacturer-provided ranges had a large influence over potential net export, in addition to atmospheric CO₂ concentration and tide gage location. Nevertheless, it is important to note that the manufacturer-provided analytical error represents a maximum. Reference measurements inform that actual field measurements are likely much more accurate than the manufacturer-stated accuracy. A more thorough set of reference measurements would provide better statistics regarding individual sensor accuracy, which could then be used for refining the parameterization of the Monte Carlo style analysis. We assume the estimated uncertainty would be much lower in this case.

Large ranges in lateral flow estimates demonstrated difficulty with comprehensive measurement of water flow and the importance of co-located tide gages when measuring pCO₂. When using water flow outside of the marsh to understand lateral C imports or exports from the marsh, overland and groundwater flow corrections should be made (Wang et al 2016), though such information was not available in this study. Overland flow is when water follows a path through the marsh not captured by a tide gage. More spatiotemporally frequent measurements of water flow are essential. One way to do this could be to use a fixed acoustic Doppler current profiler or measure particle tracking velocity using video, though video may be difficult in the marsh as wind effects on apparent surface water flow can lead to erroneous flow estimations.

Despite challenges faced with maintenance and operation, the CO₂-LAMP performed well in detecting tidal export of C as pCO₂. The sensor platform is simple to operate, lightweight, low-cost, and captures moderately frequent measurements necessary to understand influence of a full tidal cycle on marsh biogeochemistry. Mismatched peaks in salinity and pCO₂ indicated there was likely residual moisture with relatively high salinity still on the sensor between some tidal cycles. Decreasing turbidity and pCO₂ drift could have been caused by fouling on sensor surfaces or increased inputs of DIC from terrestrial freshwaters. Data from nearby streams from smaller catchments would be needed to validate this claim. Additionally, more frequent cleaning may resolve this issue. In future versions of the Smart Rock, the upper limit of the salinity sensor should be increased. Adding higher quality internal clocks to both the Smart Rock and CO₂-LAMP will provide more reliable timestamps. Decreasing the power draw of the CO₂-LAMP would also enable simultaneous data collection by multiple sensors at stations having smaller solar arrays.

The cost of the CO₂-LAMP in 2022 ranges $346–497 with an added $180 minimum for the tide-varying station (i.e. wooden boards, copper pipes, etc), plus shipping and tax. Prices can change with individual product choices (e.g. type of K30, size of waterproof case), access to construction tools or lab supplies, and the number of supplies bought in bulk or reused. An itemized list of parts and costs are included in SI1. The CO₂-LAMP uses the same measurement method as other low-cost pCO₂ sensor platforms such as the SIPCO₂, which had a cost of $500 USD in 2017 (Hunt et al 2017). A smaller, lower-cost, and submersible pCO₂ sensor platform was developed by Hill (2018) for approximately $305 USD, though battery life was shorter. At its current price, the CO₂-LAMP is still less costly than some other alternatives (Li 2022). The Smart Rock had a low cost of around $200 but was built most easily while taking an online course workshop, which adds to the overall cost. However, the internal battery and long-lasting low-power mode make the Smart Rock simple to use and implement.

The autoregression model demonstrated that variations in observations of turbidity are difficult to explain with this suite of variables, and the tidal influence on modeled salinity and turbidity are present, but weak. Neither modeled nor observed water temperature demonstrated a tidal signal. Granger causality testing revealed that although pCO₂, turbidity, salinity, and water temperature fluctuated with tide height, any correlation between pCO₂ and salinity is likely due to the influence of tide height on salinity. Causal density and vector autoregression model consistency were low. Low model consistency in this case could be the result of a missing key variable, such as tide direction. Nonzero causal density indicates that there was some dynamical complexity, but higher causal density might be achieved with more variables or a higher frequency and duration dataset, which would enable a greater lag capability and higher confidence in results. Alternative statistical techniques such as multiple linear regression may be better suited to describing tidal data in cases where the dataset is highly interdependent, or singular. Granger causality testing did not accurately predict the direction of some relationships potentially due to edge effects from overlapping signals, or distance between the tide gage and pCO₂ sensor. More specifically, a sine wave of pCO₂ may be viewed as 90° ahead of a sine wave of water temperature or 270° behind.

Our modeling approach could be used to increase spatial estimations of C imports and exports in salt marshes given certain assumptions. With an adequate training set and validation data collected through time, C imports and exports could be estimated. Investigators could expand C exchange monitoring using instrumentation like the CO₂-LAMP to develop regression-based models where tide height and other water quality parameters are monitored. Intermittent CO₂-LAMP (or similar system) deployments at potential sites would allow validation of the regression-based model and re-evaluation of model coefficients over time should cyclical or long-term changes be present in response to C exchange dynamics.

Fluctuation of pCO₂ with the tide supported other studies which show tidal amplitude is a key variable in porewater discharge (Bouillon et al 2008, Maher et al 2013, Linto et al 2014, Call et al 2015, Seyfferth et al 2020) and groundwater discharge (Wang and Cai 2004, Tamborski et al 2021). Porewater and groundwater also contribute to lateral export of dissolved inorganic C, which includes CO₂, bicarbonate (HCO₃ ⁻), carbonate (CO₃ ⁻), and carbonic acid (H₂CO₃). As a result, low tides may result in higher concentrations of CO₂ partial pressure (pCO₂) in surface water than high tides at some sites as high pCO₂ porewater and groundwater become more dominant (Zablocki et al 2011, Burgos et al 2018, Taillardat et al 2018, Mayen 2020), especially during periods of low flow (Marescaux et al 2018). There was no strong evidence for that relationship in bin-averaged pCO₂ and tide height in the marsh, though most pCO₂ outliers existed at low tide heights. The marsh surface was also exposed during low tide, leading to a lack of underwater pCO₂ measurements during low tide to compare with high tide. Tidal oscillation of pCO₂ at Sapelo contrasted with coastal ecosystems of the English Channel, which have a much stronger diurnal variability despite a semidiurnal tide (Yang et al 2019). Measurements of groundwater and surface water flow from within the marsh were a limitation of this study but are crucial for understanding coastal C dynamics.

Potential net tidal export based on tide velocity at all stations was expected, as tidal marshes typically export more C than they import through lateral flow. A salt marsh in North Carolina demonstrated summertime lateral dissolved inorganic C export of approximately 0.5 gC m⁻² tidal cycle⁻¹, which was larger but similar in magnitude to the upper limits of C export based on tide velocity at Hunt Dock and Dean Creek using the Monte Carlo style simulation (Czapla et al 2020). Potential net C export for the entire month of study was lower than dissolved inorganic C export of more than 30 gC m⁻² for the month of July in a Massachusetts salt marsh in all cases (Wang et al 2016). The cumulative annual C sink for the marsh in Sapelo ranges from approximately 130–300 gC m⁻²yr⁻¹ (Nahrawi 2019) and can range from 380 to 890 gC m⁻²yr⁻¹ in other subtropical marshes (Gomez-Casanovas et al 2020, Liu et al 2020). Total lateral C export during the month of study alone would account for approximately 0.5%–9% of the cumulative annual C sink of the marsh. The estimates presented here were based on measurements taken during the peak of the growing season when photosynthetic productivity and respiration rates are high, and they should not be extrapolated throughout the rest of the year. Granger causality of tide height means that small changes in tide height across the marsh surface could have a large impact on seasonal and long-term fluctuations in C export due to pCO₂.

Additionally, subtropical salt marshes can display 'hot' moments of high C export or import, such as the coastal marsh in Codden et al (2022) where as much as 12% of annual organic C sequestration was exported as DOC in one summer month across a 16 month study. Longer term monitoring would be necessary to determine whether our measurements reflect a 'hot' moment. Our lateral C export estimate likely underestimates actual lateral C export from the marsh because it does not consider the import or export of other forms of C, such as particulate organic C, dissolved organic C, or C in CH₄. Subtropical coastal salt marshes such as Sapelo Island are among the most productive ecosystems on Earth and given high spatiotemporal variability of these types of measurements, continual monitoring is important.

A weak, positive linear relationship between marsh surface water pCO₂ and tide height was in line with results from previous studies at this site. Specifically, aquatic pCO₂ concentrations aligned well with previous measurements of ∼3250 and ∼2500 ppm during low tide and high tide, respectively in June 2001 at Barn Creek (Wang and Cai 2004). Maximum pCO₂ concentrations were still lower than those of a tidal creek in Delaware, USA of around 8400 ppm (Trifunovic et al 2020). Surface water pCO₂ aligned with the observations of ∼1086 ppm and ∼493ppm during low tide and high tide respectively at the mouth of Doboy Sound in June 2003, when accounting for atmospheric CO₂ (Jiang et al 2008).