BGC Argo float 5906623 was deployed in December 2020 during SOLACE voyage IN2020_V08 on the RV Investigator (https://mnf.csiro.au/en/Voyages/IN2020_V08). Due to the science objectives of this voyage the float was programmed with a non-standard sampling schedule at the start of its mission. This meant that for the first 5 months, deep cycles (to ~2000m) were interspersed with a set of shallow profiles (to 500m and 12h later to 1000m) every 2 days. This reduced the deep profile sampling frequency from the usual 10 days to between 11 and 15 days, with no deep profiles between 3 Feb 2021 and 4 March 2021.

A similar sampling schedule was re-instated in October 2021 to April 2022, with 12-hourly shallow profile sampling interspersed between deep profiles and only one deep profile between 21 October 2021 and 16 December 2021.

For float pCO₂ and air-sea flux calculations we used the delayed mode quality-controlled (QC) data (Argo variable PH_IN_SITU_ADJUSTED, pH_adj) as available from the Argo Global Data Assembly Center (GDAC). Delayed-mode quality control (DMQC) adjustment of float pH was performed at the Australian BGC Argo facility and followed Maurer et al. (2021) using reference estimates calculated from LIPHRv2 equation 7 at 1500m depth. Although there were discrete timeframes for which data at this depth was not available, using an alternate reference depth throughout these periods was not advised due to the greater uncertainty in reference estimates at shallower depth (Carter et al., 2018). Additionally, the longest period for which profiles did not reach QC depth was 32 days, with each period bounded on either side by deep profiles, providing anchors for DMQC assessment. Furthermore, erratic sensor drift during short time frames would be uncharacteristic of pH sensor behavior, so use of deep profiles at the frequency available was deemed sufficient in the DMQC process.

For float TA estimates (TA^PF, where ^PF stands for profiling float and is added to all float related parameters) we used the Locally Interpolated Alkalinity Regression (LIARv2) as described in Carter et al. (2018), which uses adjusted values of temperature, salinity (psal) and dissolved oxygen to estimate total alkalinity. Quality control of these parameters followed currently available best practice as outlined in Maurer et al. (2021). If pCO₂ is calculated from pH and TA, it differs from results based on DIC and TA samples and direct pCO₂ measurements. This is due to a difference between pH measured spectrophotometrically and pH calculated from DIC and TA (Carter et al., 2013; Williams et al., 2017). Therefore, before profiling float-based pCO₂ estimates (pCO₂ ^PF) can be compared to mooring observations, a bias correction was applied to each profile, as per Williams et al. (2017):

Where pH_{1500m, 25˚C, 0dbar} is the pH_adj at 1500m (the standard QC reference depth), adjusted to 25˚C and 0dbar. Note that application of this bias correction is in line with the current operational procedure for calculating pCO₂ from floats with SOCCOM and GO-BGC arrays (Riser et al., 2023).

The Matlab script CO2SYSv3.1.1 (Lewis and Wallace, 1998; van Heuven et al., 2011; Sharp et al., 2021) was used to calculate pCO₂ ^PF from float measured pH^{adj corr} and float estimated TA^PF, with dissociation constants of carbonate by Lueker et al. (2000), of fluoride from Perez and Fraga (1987), of sulphate from Dickson (1990) and the boron to salinity ratio by Lee et al. (2010). Silicate and phosphate concentration was taken from WOA18 annual averages as 2.5 µmol kg^-1 and 0.9 µmol kg^-1, respectively. NH₄ was set to 2 µmol kg^-1 and H₂S to 0 µmol kg^-1. Orr et al. (2018) found combined uncertainties in carbon system parameter estimates to be largely driven by uncertainties in dissociation constants. We therefore tried the most commonly used K₁ and K₂ constants by Lueker et al. (2000), as well as more recently refined constants by Schockman and Byrne (2021), which are experimentally determined with spectrophotometric methods and evaluated with shipboard and laboratory data.

For air-sea flux calculations (FCO₂ ^PF) we averaged the top 20m of each profile (Gray et al., 2018).

Air-sea fluxes were calculated as follows:

where k is the gas transfer velocity, scaled according to Fay et al. (2021) to account for the fact that we use ERA5 windspeeds, α is the coefficient of CO₂ solubility according to Weiss (1974) and ΔpCO₂ is the gradient between seawater pCO₂ and atmospheric pCO₂ such that negative FCO₂ indicates CO₂ uptake by the ocean and positive FCO₂ indicates outgassing.

Atmospheric pCO₂ data (pCO₂ ^PF_air) for the float location and profile times were calculated from mole fractions of CO₂ in dry air (xCO₂) from the CSIRO Oceans and Atmosphere and the Australian Bureau of Meteorology Kennaook/Cape Grim Baseline Air Pollution Station in Tasmania, Australia (https://capegrim.csiro.au/). The monthly data were interpolated on daily timesteps using piecewise cubic Hermite polynomials (Gray et al., 2018). Hourly windspeed and mean sea level pressure data from the ERA5 reanalysis product (0.25˚ gridded data, Hersbach et al., 2018) were subset to the nearest space coordinate of each float profile and interpolated on minute timesteps to the time of the float profile using piecewise cubic Hermite polynomials. Float seawater temperature and salinity together with ERA5 mean sea level pressure were used to calculate atmospheric pCO₂ from atmospheric xCO₂ following the method of Zeebe and Wolf-Gladrow (2001). For annual flux (FCO₂ ^PF) calculations, pCO₂ ^PF data were interpolated on daily timesteps using piecewise cubic Hermite polynomials (Gray et al., 2018) and then integrated for the calendar year 2021. Where more than one profile existed per day, the results were averaged for daily fluxes.

The SOFS mooring is recovered, and a new mooring deployed once a year from research voyages on RV Investigator, and mooring named sequentially (for more details refer to annual reports Wynn-Edwards et al., 2019; Wynn-Edwards et al., 2022). This study covers data from SOFS-9 (deployed 31 August 2020, recovered 25 April 2021) and SOFS-10 (deployed 20 April 2021, recovered 13 May 2022). Mooring seawater and atmospheric pCO₂ (from here on called pCO₂ ^M and pCO₂ ^M_air, respectively, where ^M stands for mooring and is added to all mooring related parameters) were measured by a Moored Autonomous pCO₂ System (MAPCO2) mounted on the surface buoy with an estimated uncertainty of< ± 2 µatm for seawater and< ± 1 µatm for air (Sutton et al., 2014a; Sutton et al., 2019, Table 1 ). The system uses an equilibration-based method and measures xCO₂ with a nondispersive infrared gas analyzer (LI-COR LI-820), which is calibrated prior to each measurement with a CO₂ reference gas standard (Sutton et al., 2014a). Seawater temperature and salinity were measured at 1m and 30m by Sea-Bird Electronics SBE37 MicroCAT sensors and quality-controlled following (Jansen et al., 2020; Jansen et al., 2023). Salinity and temperature data were gridded to UTC hour by averaging all values of equal depth within 30 minutes of the hour (Jansen et al., 2022). Hourly windspeed data were measured by a Woods Hole Oceanographic Institution ASIMET Sonic Wind Module (Schulz et al., 2012) and paired with the nearest (<10min) three-hourly pCO₂ measurements. Data gaps in the windspeed record of less than 6 hours were interpolated on hourly timesteps using piecewise cubic Hermite polynomials. Data gaps of more than 6 hours were filled with ERA5 hourly data on single levels (Hersbach et al., 2018). The ERA5 wind speed product was chosen due to its high temporal (hourly) and spatial resolution (0.25˚ x 0.25˚) and the availability of data for the time period of interest. For the calendar year 2021, the time period for which we calculated air-sea fluxes, the mean difference between mooring and ERA5 windspeeds was 0.27 ms^-1 (± 1.09).

Air-sea fluxes were calculated per equation (2) and mooring air-sea fluxes are from here on called FCO₂ ^M. For annual flux calculations, daily averaged fluxes were integrated for the calendar year of 2021.

In late April 2021, there was a four day overlap in measurements between SOFS-9 and SOFS-10 sensors, while both moorings were in the water between recovery and deployment ( Figure S1 ). Measurements of pCO₂ ^M and calculated FCO₂ ^M of SOFS-9 and SOFS-10 were averaged for this period.

The SOFS moorings are S-tether designs that allow their surface floats to move in large ‘watch circles’. When referring to the mooring location in this manuscript, we refer to the GPS tracked location of the surface buoy, as opposed to the anchor location of the mooring.

During the time period of the two moorings discussed here, five research voyages on RV Investigator visited and/or passed the SOTS site (for more information refer to Table S1 ). RV Investigator is fitted with an automated underway pCO₂ measurement system with an accuracy of ± 2 µatm (Pierrot et al., 2009, Table 1 ). These data (pCO₂ ^UW) were used for comparison with pCO₂ ^M measurements.

During the SOLACE and mooring deployment and recovery voyages DIC and TA samples from CTD casts were collected and analyzed coulometrically and with an (open cell) potentiometric titration, respectively, following standard procedures (Dickson et al., 2007). Regular collection of TA samples during SOTS voyages allowed for the establishment of a robust relationship between TA and salinity at the SOTS site. This linear relationship was derived from samples collected in the upper 100m between 2010 and 2019 and has an error associated with the fit of ~ 8 µmol kg^-1 (Shadwick et al., 2020, Table 1 ).

The SOFS mooring surface float houses a modified McLane Remote Access Sampler (RAS), which collects discrete seawater samples at regular intervals. These samples are analyzed for TA^RAS onshore once recovered. Details on quality assurance and QC of these RAS data are published in Shadwick et al. (2020). CTD DIC and TA sample results were used to calculate seawater pH (pH^CTD) with CO2SYS v3.1.1, as described for float data. The calculation of pH from DIC and TA samples introduces an uncertainty that could be avoided in future work if pH was measured directly.

RAS, CTD and mooring TA data were used for comparison to the float estimates of TA^PF. The pCO₂ ^CTD from CTD samples were calculated with CO2SYSv3.1.1 (Lewis and Wallace, 1998; van Heuven et al., 2011; Sharp et al., 2021) with dissociation constants and nutrient concentrations as outlined for float calculations.

For the most stringent comparison between mooring observations and float estimates, we subsampled the higher resolution mooring data with the following thresholds:

These temperature and potential density thresholds are indicative of well mixed water masses at SOTS (Weeding and Trull, 2014; Shadwick et al., 2015; Jansen et al., 2020; Jansen et al., 2023) and comparable to thresholds used widely for the definition of mixed layer depth (de Boyer Montégut et al., 2004); they are used here to limit the comparison to water masses of similar physical properties. For an overall average seawater pCO₂ of 383 µatm (combining mooring and float data), this temperature difference of 0.3°C would equate to a 5 µatm pCO₂ difference (Takahashi et al., 1993).

For the comparison between pCO₂ ^M and pCO₂ ^UW we used the same temperature and potential density criteria but averaged underway data 3h either side of mooring observations and restricted voyage data to within 1° latitude and longitude of mooring location.

To assess the potential influence of distance between mooring and float on the difference in the pCO₂ data (when restricted to temperature and potential density thresholds) we used a model II regression. This reflects the fact that both variables are not controlled and that there is no a priori assumption that there is a dependent variable. Model II regression calculations were performed with the Matlab script lsqfitgm.m (Edward T Peltzer, MBARI, https://www.mbari.org/products/research-software/matlab-scripts-linear-regressions/).

Two CTD casts within 24h of the first three float profiles indicate that pH^{adj corr} compares well to pH^CTD calculated from discrete water samples of DIC and TA collected from the CTD rosette ( Figures S2 , S4 ; Table 3 ). Only float profile data within 5m of each Niskin bottle sample depth were used in the comparison.

The average difference between pH^{adj corr} of the first profile (13 Dec) and pH^CTD (12 Dec) in the upper 20m was 0.001 (± 0.0002). The average difference over the full depth of the CTD cast, ~ 1000m, was 0.003 (± 0.003). The CTD cast occurred within ~6km of the first float profile. In doing the same comparison between the first profile and a CTD a day later (13 Dec), the mean difference in pH for the top 20m was 0.007 (± 0.0003) and 0.009 (± 0.006) for all samples down to ~ 1000m. The distance between the second cast and the float profile was about 7km ( Figures S2 , S3 ; Table 3 ).

The second and third float profiles occurred on 14 December, ~ roughly 12 hours apart. We compared these two float profiles to CTD data from 13 December, as described above. For the second float profile the mean difference in pH for the top 20m was 0.008 (± 0.0006) and 0.009 (± 0.007) for all samples down to ~ 850m, with a distance of ~13km between float and CTD cast. For the third profile the mean difference in pH for the top 20m was 0.009 (± 0.0005) and 0.010 (± 0.006) for all samples down to ~ 850m and a distance of ~ 18km ( Figures S4 , S5 ; Table 3 ).

Since computing pCO₂ from float pH measurements requires a second carbon system parameter, i.e. TA, we also compared LIARv2 estimates of TA^PF with TA^CTD from the two CTD casts within 24h of the first three float profiles. We again compared CTD bottle data to float profile data within 5m of each Niskin bottle sample depth. The average difference in TA^PF between the first float profile (13 Dec) minus TA^CTD (12 December) was 0.22 (± 0.64) µmol kg^-1 for the top 20m and -1.83 (± 1.93) µmol kg^-1 for the entire cast to about 1000m. The CTD cast occurred within ~6km of the first float profile. In doing the same comparison between the first profile and CTD cast from 13^th December, the mean difference in TA was 2.16 (± 0.51) µmol kg^-1 for the top 20m and -2.24 (± 5.66) µmol kg^-1 for all samples down to ~ 1000m. The distance between the second cast and the float profile was about 7km ( Table 3 ).

The second and third float profiles occurred on the 14^th of December, roughly 12 hours apart. We compared these two float profiles to CTD data from the 13 December, as described above. For the second float profile the mean difference in TA was 2.34 (± 0.62) µmol kg^-1 for the top 20m and -2.14 (± 5.68) µmol kg^-1 for all samples down to ~ 850m, with a distance of ~13km between float and CTD cast. For the third profile the mean difference in TA was 2.32 (± 0.63) µmol kg^-1 for the top 20m and -2.15 (± 6.01) µmol kg^-1 for all samples down to ~ 850m and a distance of ~ 18km between float and CTD cast ( Table 3 ).

Sharp et al. (2021) included error estimates in their updated CO2SYS Matlab script based on the error propagation analysis by Orr et al. (2018), which calculated TA^PF associated uncertainty of about 8 µatm kg^-1 (0.36%, Table 1 ). The uncertainty of TA^CTD and TA^RAS is typically ± 2μmol kg^-1 (95% confidence interval, Shadwick et al., 2020, Table 1 ).

We also compared TA^PF with TA^M estimates and applied the same rules as outlined in the Methods section to limit mooring salinity data (1m and 30m sensors) and TA^M estimates for our comparison. Overall, the difference between TA^M and TA^PF was 0.11 ( ± 2.49) µmol kg^-1 ( Table 2 ), with a maximum difference of ~ 4 µmol kg^-1. Only two RAS samples fell within the restrictions for this comparison (23/6/2021 and 1/7/2021) and the difference between TA^RAS samples and average TA^PF estimates of the top 20m was 2.46 and 0.93 µmol kg^-1, respectively.

In late April 2021, there was a four day overlap between SOFS-9 and SOFS-10 with the moorings about 35km apart ( Figures S1 , S6 ). The overlapping seawater temperature sensor data show that the two moorings were exposed to different water masses in the top ~350m, with a temperature difference of about 4˚C and a salinity difference of 0.83. While SOFS-10 was in warmer waters than SOFS-9, it recorded on average seawater pCO₂ of about 12.35 µatm (± 3.45) lower than SOFS-9 ( Figure S6 ), indicating that the difference in pCO₂ was not due to temperature effects but likely due to different water masses, likely warmer, saltier subtropical water, with lower DIC (and thus lower pCO₂; e.g. Pardo et al., 2019), observed by SOFS-10.

The mean difference between pCO₂ ^M minus pCO₂ ^CTD was -6.33 µatm (± 4.46, n=5 CTD samples, Table 2 ). The CTD casts for this comparison were within<70km of the mooring. We also compared pCO₂ ^M with pCO₂ ^UW measurements with the same temperature and potential density criteria as described for pCO₂ and air sea flux. The mean difference between pCO₂ ^M with pCO₂ ^UW across the five voyages was 2.29 µatm (± 3.93) with a range of -8.03 µatm to 17.94 µatm (mostly within 60km, with some observations up to ~120km apart, Table 2 ). Comparisons between the mooring pCO₂ MAPCO2 sensor and the ship underway pCO₂ system in the lab have shown the two systems to be within 2 µatm of each other under ideal conditions (pers. comm. Craig Neill).

Figure 2 shows that between Dec 2020 and May 2021, there were times of close agreement between pCO₂ ^M and pCO₂ ^PF, but also times of significant difference. The mean combined uncertainty for pCO₂ ^PF estimates based on the error propagation routine of the CO2SYS package (Sharp et al., 2021) is 22 µatm (5.6%) for all observations considered here. This assumes a temperature uncertainty of 0.002°C, psal uncertainty of 0.01 (Gray et al., 2018), and no assigned error for nutrients and errors associated with dissociation constants after Orr et al. (2018). Alternatively, if we are to apply a general uncertainty estimate of 2.7% for pCO₂ ^PF as outlined in Williams et al. (2017), which is based on a top-down uncertainty analysis of float pCO₂ estimates as compared to shipboard underway measurements, this reduces the magnitude of uncertainty to an average of 11 µatm throughout the period of interest. For pCO₂ ^PF, both error estimate approaches are displayed in Figure 2 as uncertainty envelopes. The pCO₂ ^M fall within the bounds of uncertainty for pCO₂ ^PF for the majority of the study period, except for ~90 days in Jan to May 2022 and ~12 days in May 2022 for which they deviated beyond the uncertainty envelopes. Over the entire study period the float drifted in an area of up to ~ 4° Latitude and ~ 3° Longitude from the mooring location ( Figure 1 ). This equates to distances of up to ~ 414km, although over 83% of profiles were within 300km of the mooring. If we restrict the comparison to only the first 10 profiles, within 1˚ latitude from the mooring and use average pCO₂ ^M 6 hours either side of the float profile, then pCO₂ ^PF estimates were 6.04 µtam (± 11.82) higher than the average of pCO₂ ^M. If we extend this to the whole float record and compare mooring data within 6h of the float profiles when both are within 1° latitude, the difference in pCO₂ ^M minus pCO₂ ^PF was -3.4 µatm (± 8.39).

A recent uncertainty propagation estimate for the marine carbon dioxide system by Orr et al. (2018) showed that dissociation constants K₁ and K₂ contribute the most to the combined standard uncertainty. Therefore we also calculated pCO₂ ^PF with those recently presented by Schockman and Byrne (2021) which are experimentally determined with spectrophotometric methods and evaluated with shipboard and laboratory data. The pCO₂ ^PF using the K₁ and K₂ constants by Schockman and Byrne (2021) are ~4 µatm higher than those calculated using constants by Lueker et al. (2000), increasing the gap between pCO₂ ^M and pCO₂ ^PF further. To remain consistent with other float-based pCO₂ estimates available in this region and used within other studies (Johnson et al., 2017; Gray et al., 2018; Johnson et al., 2022), for the remainder of this work, we therefore focus on calculations based on constants by Lueker et al. (2000) but note the size of the uncertainty introduced by the different choices in K₁ and K₂ constants. We also note that the application of the Williams et al. (2017) bias correction reduces pCO₂ ^PF estimates by ~5.9 µatm (± 0.29) over the entire float record, or in other words, without the application of said bias correction the gap between pCO₂ ^M and pCO₂ ^PF would be larger.

The comparison between the overlapping SOFS-9 and SOFS-10 at close proximity (~35km) shows that limiting comparisons by distance alone may not be enough. We therefore used temperature and potential density as criteria for comparable water masses. Of the 242 float profiles that occurred during the time covered in this study, only 27 profiles passed the temperature and potential density criteria outlined in the Methods section, highlighting the great spatial variability in seawater properties in this region. Considering only the subset of data that passed these criteria, pCO₂ ^PF was always higher than pCO₂ ^M (overall mean difference of -11.06 µatm ± 4.76, Table S2 ). This was most noticeable for the winter period of 2021 (grey shading Figure 2A ). Application of the Takahashi temperature correction (Takahashi et al., 1993) before comparing pCO₂ ^PF to pCO₂ ^M for this subset of data makes little difference to the average discrepancy (– 11.17 µatm ± 7.74). Diurnal variations in pCO₂ could contribute to the discrepancy between pCO₂ ^PF and pCO₂ ^M comparisons when averaging pCO₂ ^M either side of float profile times. A comparison with mooring observations 2 days either side of the float profile averages this diurnal signal and therefore has an overall small effect on the comparisons. We also calculated the overall mean difference between pCO₂ ^M and pCO₂ ^PF estimate with the same temperature and potential density criteria, but 6 hour (-10.48 µatm ± 3.95) and 3h (-9.92 µatm ± 4.04) either side of the float profile and note that this does not make a statistically distinguishable difference.

To investigate the influence of distance between the mooring and float on the pCO₂ differences under the most stringent comparison conditions (temperature and potential density), we calculated a Model II regression between the difference in pCO₂ and the distance in km between mooring and float for those data only ( Figure 3 ). Mooring pCO₂ ^M that met the temperature, potential density and time criteria (outlined in the Methods) were averaged before subtracting pCO₂ ^PF. Float profiles included in this regression were south and north of the mooring locations ( Figure 4 ), yet the differences between pCO₂ ^M and pCO₂ ^PF were all negative, unlike the differences between pCO₂ ^M and pCO₂ ^UW described above, which were both negative and positive.

There is a clear signal that the farther away the float was from the mooring, the larger the difference in pCO₂. While the float moved east and west around the mooring, it predominantly moved south, with only one short excursion to the north ( Figure 4 ). Only 10 profiles (November 2021) met the temperature and potential density criteria and were within 1° latitude and 1° longitude of the moorings. The average difference in pCO₂ under those conditions was -7.01 µatm (± 2.19). For these profiles (unlike the other comparisons) the float measured higher temperatures in the top 20m compared to the mooring sensors at 1m and 30m, and therefore applying the temperature correction to pCO₂ ^PF reduces the discrepancy to -3.27 µatm (± 2.57).

To investigate whether there were any co-varying properties along with distance, we overlaid float dissolved oxygen, chlorophyll (Chl a), mixed layer depth, nitrate concentration and windspeeds at the location of the float profiles (not shown). However, none of these parameters correlated with the distance to the mooring. There was also no correlation between these parameters and the difference in seawater pCO₂ between mooring and float. Figure 4 shows the float track in relation to the mooring locations colored with the ΔpCO₂. To illustrate the difference in pCO₂ for all float profiles, we averaged pCO₂ ^M within 2d of float profiles, but not taking into account temperature and potential density differences ( Figure 4 ). This highlights that the greatest differences in pCO₂ are seen in profiles that occurred south of the mooring, which occurred during autumn months. In Figure 4 we also marked those profiles that additionally met the temperature and potential density criteria with grey hexagons, which shows that the profiles with the overall largest (negative) difference between pCO₂ ^M and pCO₂ ^PF did not meet temperature and density criteria.

If the majority of the difference in mooring and float pCO₂ was attributable to distance, then the two results could be brought in line via the above regression. We therefore applied the slope and intercept to all pCO₂ ^PF data, as per equation (4):

Where d_m-f is the distance (km) between mooring and float. This brings pCO₂ ^PF closer in line with pCO₂ ^M for parts of the time period discussed here, particularly during the autumn/winter period of 2021 ( Figure 5B ). The intention of this was not to “correct” pCO₂ ^PF estimates but to test the theory that some of the difference in pCO₂ is based on a true (predominantly latitudinal) distance gradient.

The annual FCO₂ ^M for the calendar year 2021, without limiting the data otherwise, was -1.68 (± 0.004) mol m^-2 yr^-1 compared to FCO₂ ^PF -0.89 (± 0.003) mol m^-2 yr^-1 ( Figures 6 ; Figure S8 ). Most notable are short periods of outgassing seen in the float estimates during austral autumn 2021 (April and May), when the mooring estimates show significant CO₂ uptake instead. This was also the time period of the largest difference in seawater temperature and salinity ( Figure 6 ), when a warmer, saltier water mass that extended to ~ 350m was recorded by SOFS-10 for about a month ( Figure S6 ). Both mooring and float agree, however, on periods of outgassing in austral spring 2021 (Sept and Oct).

Since squared windspeed factors into air-sea flux calculations we also calculated air-sea fluxes for the float with windspeed observations from the mooring instead of ERA5 windspeeds. For this we used the average windspeed mooring data 3h either side of the float profile time ( Figure S8 ) and equation 2, but with the same scaling factor for the gas transfer coefficient. Mooring windspeed is not assumed to be more accurate than ERA5 winds for the location of the float profile (a comparison between ERA5 winds and moorings winds is shown in Figure S7 ), rather this calculation is to visualize the effect of choosing different windspeed products on air-sea flux calculations. The mean difference between mooring windspeed minus ERA5 windspeed at the float locations for 2021 was -0.27 (± 1.79) m s^-1, with a range of 0.02 to 5.47 m s^-1 and no bias over time. Using mooring windspeed for the float calculations resulted in a near identical annual flux of -0.88 (± 0.003) mol m^-2 yr^-1. To exemplify the impact of pCO₂ ^PF uncertainties as compared to the impact of the choice of windspeed, we calculated FCO₂ ^PF with a Monte Carlo approach (1000 simulations) using a mean pCO₂ uncertainty of - 11.2 µatm (± 7.7), - 7 µatm (± 2.2) and – 3.3 µatm (± 2.6) for the calendar year 2021, which resulted in an air-sea flux of -1.92 mol m^-2 yr^-1 (± 0.004 for ERA5 winds, vs -1.89 mol m^-2 yr^-1 ± 0.004 for mooring windspeeds), - 1.54 mol m^-2 yr^-1 (± 0.004 ERA5, -1.51 mol m^-2 yr^-1 ± 0.004 mooring windspeed) and – 1.20 mol m^-2 yr^-1 (± 0.003 ERA5, - 1.18 mol m^-2 yr^-1 ± 0.003 mooring windspeed), respectively; indicating that the uncertainty in pCO₂ ^PF has a far great impact on air-sea flux estimates than the choice of windspeed product.

The spatial variability in this region of the Southern Ocean is well known and makes direct comparison of drifting float measurements and fixed-mooring observations difficult. A comparison based purely on time is not appropriate for periods when the float was up to ~360km away and/or “up- or downstream” of the mooring site. Ideally, only measurements in the same water mass would be used. Therefore, we started by comparing observations between platforms when they were close in time and space first, e.g. CTD casts shortly after the deployment of the float.

The initial comparison between CTD samples and float measurements showed that the float pH sensor was calibrated and functioning correctly at the time of deployment. Extending the comparison to the first 10 profiles, which were within 1˚ latitude from the mooring, pCO₂ ^PF estimates were about 6.0 µtam (± 11.82) higher than mooring observations within 6h of each profile. This agrees with previous reports (using various methods) of a bias in float pCO₂ estimates of about 4 – 8 µatm higher (Gray et al., 2018; Williams et al., 2018; Bushinsky et al., 2019; Long et al., 2021; Wu and Qi, 2022). It is also important to note that Maurer et al. (2021) stated that float pH measurements adjusted according to the method described in their paper (and applied in this study) are accurate to approximately 0.007 (based on validation efforts which compared adjusted float pH to shipboard data taken alongside deployment). This also matches our comparison between float pH^{adj corr} and pH^CTD estimates at the time of deployment. An uncertainty of 0.007 pH units roughly equates to 7 µatm pCO₂, which is only ~ 1 µatm higher than the magnitude of the difference between mooring and the first 10 float profiles close in space reported in this study. Gray et al. (2018) (based on Williams, 2017), estimated pCO₂ ^PF uncertainty to be in the order of 2.9%, which equates to 11 µatm for the entire observation period presented here. Combined uncertainties after Orr et al. (2018) were about twice as large (~ 22 µatm, 5.6% relative uncertainty, Table 1 ). Therefore, we emphasize that comparing pCO₂ ^PF to pCO₂ ^M close in space and time yields differences that are well within the limits of stated pH sensor capabilities and pCO₂ estimate uncertainties. This was also true when we compared the whole float record (over a year) to mooring data, while applying the latitude and time restrictions (see below).

It is not clear how much of the observed difference in pCO₂ over the course of the year can be attributed directly to uncertainty in float estimates. As from the overlapping mooring observations, pCO₂ can be significantly different if water masses with genuinely different characteristics are encountered. We therefore used temperature and potential density as criteria for when mooring and float were in comparable water masses. The difference between pCO₂ ^M and pCO₂ ^PF when only looking at data restricted with this water mass criteria was around 11 µatm ( Table 2 ), which is within stated uncertainties. Mean uncertainty estimates of TA^PF were about 8 µmol kg^-1, which is larger than the 5 µmol kg^-1 reported by Williams et al. (2018). Comparison between TA^PF, TA^M and CTD samples were well within these uncertainties. A relative uncertainty in TA of 0.4% (based on the LIARv2 TA) would translate to a relative uncertainty of 0.4% in pCO₂ (Dickson and Riley, 1978). Thus, the difference in pCO₂ ^PF is not predominantly due to uncertainties in TA^PF, but could be attributed to uncertainties in seawater carbonate system thermodynamics as well as uncertainties in the adjusted float pH measurements, which would also include any uncertainty in measured dissolved oxygen concentration used to adjust the pH and used for LIARv2 estimates of TA (Johnson et al., 2016; Johnson et al., 2017; Williams et al., 2017; Gray et al., 2018). Orr et al. (2018) found combined uncertainties in carbon system parameter estimates to be largely driven by uncertainties in dissociation constants. We therefore tried the most commonly used K₁ K₂ constants by Lueker et al. (2000), as well as more recently refined constants by Schockman and Byrne (2021). The pCO₂ ^PF estimates were about 4 µatm higher when using constants by Schockman and Byrne (2021) and therefore increased rather than reduce the difference between pCO₂ ^M and pCO₂ ^PF.

The comparison of pCO₂ ^M and pCO₂ ^PF when restricted to temperature and potential density criteria revealed a relationship of increasing difference in pCO₂ with increasing distance between mooring and float. While application of a calculated distance regression reduced the gap between pCO₂ ^PF and pCO₂ ^M, there were times when the opposite was true, particularly during the autumn of 2022 ( Figure 5A , in late March 2022), when the float drifted up to ~ 400km away from the mooring yet the distance corrected float pCO₂ disagreed more with pCO₂ ^M than with the uncorrected pCO₂ ^PF. This indicates that distance-associated variability is not the only factor contributing to the observed difference between pCO₂ ^PF and pCO₂ ^M.

For most of the study period, mooring seawater temperature in the top 30m was higher than float averages for the top 20m. Thus, the expected change in pCO₂ (Takahashi et al., 1993) for the difference in seawater temperature would not explain the disagreement, but in fact emphasize it (note the above mentioned exception for the November 2021 period). A latitudinal gradient in seawater pCO₂ in this part of the SAZ has recently been shown from ship-based repeat measurements along a transect from Tasmania to Antarctica (Brandon et al., 2022). Data presented by Brandon et al. (2022) show a southward increase in seawater pCO₂ of up to 25 µatm for the latitudinal range between mooring and float encountered in the study period. They attributed this gradient in part to an increasing influence of subtropical waters in the SOTS region of the SAZ. Temperature profiles from both mooring and float for the time period where their pCO₂ data disagree the most (February to October 2021) also indicate that there were water masses coming past the moorings that were not encountered by the float (e.g. May 2021, Figure 7 ). As discussed above, the difference in pCO₂ between SOFS-9 and SOFS-10 when they encountered different water masses, was ~12 µatm, highlighting the importance of such events for surface pCO₂ variability.

Bushinsky et al. (2019) used high-resolution ocean biogeochemical models to investigate the underlying cause of the differences between float and ship-based estimates and found that some of the difference can be explained by spatial and temporal sampling differences. This has more recently been confirmed, particularly for the SAZ, by Djeutchouang et al. (2022). By comparing the surface pCO₂ of a high-resolution coupled physical and biogeochemical model with reconstructed pCO₂ after sampling the model in time and space the way a BGC float and ship would, they show that reducing the uncertainty in the reconstructed pCO₂ depended on how well the spatial and meridional gradients and the temporal, seasonal and intra-seasonal pCO₂ variability was resolved. High temporal sampling resolution (ideally at least daily) and a meridional gradient resolving sampling strategy were best equipped to achieve that (Djeutchouang et al., 2022). In other words, a moored observation platform with high frequency sampling, resolves the seasonal and intra-seasonal variability well, and also captures mesoscale variability but it does not resolve the meridional gradient, which is particularly important in the SAZ and could explain the discrepancies between pCO₂ ^M and pCO₂ ^PFas the float drifted away.

Between 10 and 25% of the uncertainty in float based carbon flux estimate was previously attributed to the low temporal resolution of 10-day cycling which may frequently miss short-lived but significant changes in mixed-layer depth due to storms and wind-stress and the associated changes in up-welling of carbon-rich deep waters and this was found to be particularly true for the SAZ (Monteiro et al., 2015; Nicholson et al., 2022). In our study, positive float-based air-sea fluxes (i.e. outgassing) between April and June 2021 - and to a lesser degree in September 2021 – suggest that the float measured regional events which the mooring did not, and that these outgassing events may be possibly genuine oceanographic features.

The 3 – hourly resolution mooring observations show how variable seawater surface pCO₂ is in this region of the SAZ ( Figures 5 , 6 ), with subtropical water masses reaching the site intermittently. If we limit the comparison of the whole pCO₂ ^PF data to mooring data within 2d either side of float profiles, 0.3°C temperature and 0.03 kg m^-3 potential density restriction and a latitudinal and longitudinal difference of less than 1˚, we find a mean difference in seawater pCO₂ of -7.0 (± 2.19) µatm, indicating again a generally higher pCO₂ estimate from the float, but well within stated pCO₂ uncertainty estimates. Based on the comparison with previous findings and the potentially large uncertainty of pCO₂ ^PF, we find that the difference between pCO₂ ^PF and pCO₂ ^M observed in this study appears to be a combination of a potential small positive bias in pCO₂ ^PF estimates, a zonal gradient in seawater surface pCO₂ in this sector of the SAZ, and relatively fine scale spatial and temporal variability.