First, we provide a description of BGC-Argo’s QC flag system which is based on the pre-existing flagging scheme used operationally within the Argo data management system. The QC flag system sets depth-level and profile-level quality flags for each BGC property ( Table 1 ; Argo Data Management Team, 2022). The depth-level flag system in Argo denotes the QC flags as “<PARAM>_QC” or “<PARAM_ADJUSTED>_QC” for unadjusted or adjusted measurements, respectively. These QC flags will flag “good” data as “1”, “probably good” data as “2”, “probably bad” data as “3”, and “bad” data as “4” ( Table 1 ). Data are labeled with a flag of “5” when they are changed (e.g., the correction for NPQ of chlorophyll-a fluorescence). If values are interpolated or extrapolated, a label of “8” is used. If no QC is performed, a flag of “0” is used, while missing data are flagged as “9”. Generally, data with flags “1”, “2”, “5”, and “8” are treated as high-quality data, while data with flags “3” or “4” are treated as low-quality data.

Profile-level quality flags, denoted as “PROFILE_<PARAM>_QC” in Argo, describe the percentage of high-quality data points out of all the available data points in a profile (N). For the calculation of N, the QC flags associated with the adjusted data (from either adjusted RTQC or DMQC) are used. If there are no adjusted data available, the QC flags associated with the unadjusted data are used in the calculation instead (from unadjusted RTQC processing). Depth-level data points with flags “0” or “9” are not used in the determination of the profile-level flag. The calculation for profile-level quality flags is as follows:

where n is the number of QC flags in the profile with the subscript indicating the value of the depth-level QC flag. Then, N is categorized as one of seven different flags: flag “A” means when 100% of the profile has high-quality data, flag “B” when 75 to 100% of the profile has high-quality data, and so on in 25% increments to flag “F” which means that 0% of the profile data are high quality ( Table 1 ). If no QC is performed or the only depth-level quality flags are “0” and “9”, then no profile-level flag is given, and an empty value is assigned.

For this analysis, we created two simplified systems of flags based on BGC-Argo’s quality flag scales described in the previous section. Our first flag classification system is important for those interested in scientific analysis as it focuses on the availability of high-quality data profiles.

The second flag system we created describes the prevalence of functioning sensors, as opposed to those that have malfunctioned (e.g., the sensor stops reporting data or reports oceanographically inconsistent values entirely). Data profiles from malfunctioning sensors cannot be recovered for biogeochemical analysis and may be meaningless. The purpose of the second flag classification system is to identify causes of poor-quality data that result from an unfixable sensor malfunction, and not from intermittent error or other operational limitations. This flag scale is likely of interest to those who are currently deploying floats and those who plan to. Both flag scales provide a foundation for tracking BGC-Argo data quality and sensor performance through repeated future analyses.

The density and spatial coverage of BGC data are important for determining where regional biogeochemistry may need to be better constrained. We assessed these metrics in different marine regions (Flanders Marine Institute, 2021) which include the North Atlantic, South Atlantic, North Pacific, South Pacific, Indian Ocean, Southern Ocean, Arctic Ocean, Baltic Sea, Black Sea, and Mediterranean Sea. Floats in the Black, Mediterranean, and Baltic seas were also reviewed both individually and as a single region we call ‘seas around Europe’.

We determined the density of high-quality profiles (equal to S _HQ divided by surface area) from both the total number of profiles available and on a yearly basis. The surface area of ocean was calculated for each marine region and from a global ocean grid resolved to 5° longitude by 5° latitude. For all regions, except for the seas around Europe, areas where the seabed was shallower than 200 m was excluded. We report the average number of high-quality profiles from PAR/E _d profiles since this BGC data comes from the same sensor and have similar numbers of high-quality profiles.

For the annual assessments, we compared the density of high-quality profiles with a target profile density of a BGC-Argo network of 1,000 active floats (assuming they are evenly-spaced). This comparison was made by calculating 1) the region-wide density as a percentage of target density, and 2) the percent area of grid cells where high-quality profile density met the target density in each region. A similar calculation was also made for the latter but by including areas with profile densities above 0. The target density is equal to ~1.1 profile per 10,000 km² per year (i.e., ~1 profile per 100 km x 100 km square of ocean per year) and is based on a network of 1,000 floats (presumably profiling every 10 days) over ~335 million km² of ocean each year.

As of December 2022, the database of Sprof files has accumulated a total of 263,715 unique BGC profiles from 1,811 BGC-Argo floats ( Table S1 ). The most frequently measured property is O₂ (251,855 raw profiles from 1,754 floats) while the least frequently measured property is pH (39,483 raw profiles from 447 floats) after PAR/E _d (46,028 raw profiles from 273 floats). Since 2002, the total number of floats deployed has steadily increased, alongside the total number of floats sustained in the array ( Figure 1 ).

How close the float array is to the annual float or profile targets is highly dependent on the BGC parameter because most floats do not measure the full suite of BGC parameters. For example, the number of radiometer-equipped floats deployed each year has not surpassed 50, while the other five BGC parameters see a minimum of 100 floats deployed per year as of 2022. These discrepancies result from pilot programs that prioritized certain parameters. Radiometry has largely been driven by European and Australian projects, which focused on regions around Europe, the North Atlantic, and the Indian sector of the Southern Ocean. Currently, only 39 floats measure all 6 BGC parameters of interest ( Figure 1 ). As of 2022, PROVOR, one of the three main models of floats used in the BGC-Argo fleet, is the only float type that can measure all six variables, restricting the full set of parameters from being collected. However, many operational float programs are currently putting effort toward expanding sensor suite capabilities on various platforms. For example, the Global Ocean Biogeochemistry (GO-BGC) array plans to begin deploying six-parameter APEX floats in the summer of 2023. A 6-parameter Navis float is currently in the later stages of development as well.

The majority of profiles for each BGC parameter have received QC flags. Almost all O₂ profiles have undergone QC (P _NOQC =<1%). The parameters missing the most QC flags are Chla (P _NOQC = 12%) and b _bp (P _NOQC = 15%). This lack of QC on some parameters will decline as the program matures. The mode of QC received by each property also varies greatly ( Figure 2 ). At least three-quarters of O₂, NO₃^-, pH, and Chla profiles have received adjusted or delayed-mode processing. In comparison, >40% of b _bp and PAR/E _d data contain unadjusted real-time processing (also see Figure S1 ). Note that the RTQC for PAR/E _d only involves a range test and that there is no procedure for real-time adjustments; if adjustments are made to the radiometry data, then only DMQC has been applied.

Due to their specialized QC procedures, which limit the quality flag of “1” to data that has been adjusted, we report only the QC metrics derived from adjusted data of O₂, NO₃^-, pH, and Chla for the rest of this document, specifically referring to these variables as ‘adjusted’. Otherwise, the QC metrics for the other parameters (b _bp and PAR/E _d) include both adjusted and unadjusted data, unless specifically stated.

Based on our first classification system, the number of high-quality profiles has generally increased each year, except for profiles of PAR/E _d ( Figure 3 ). This lack of PAR/E _d profiles is due to no sustained rise in the deployments of floats equipped with these sensors ( Figure 1 ). We find that the database-wide P _HQ was the highest for PAR/E _d (92%, on average) and the lowest for adjusted pH (48%) ( Table S2 ). Adjusted O₂ and NO₃^- profiles have a P _HQ equal to ~90% and 87%, respectively, while profiles of b _bp and adjusted Chla have P _HQ equal to 87% and 85%, respectively. For the ~21,000 adjusted PAR/E _d profiles that have received DMQC, P _HQ increases to ~97% (or 96.7% for PAR; 97.3% for E _d[380], E _d[412], and E _d[490]).

Low-quality data profiles were the least common with PAR/E _d (P _LQ = 2.5% for PAR; 1.8% for E _d[380]; 1.8% for E _d[412]; and 1.9% for E _d[490]). Low-quality adjusted pH profiles are currently the most common (P _LQ = ~52%) among the BGC parameters. Other BGC sensors tend to report low-quality profiles 7−15% of the time. Profiles with no data (i.e., unresponsive sensors) are the most common among PAR/E _d and b _bp (P _NR = ~6% for each), while they are the least common among the adjusted O₂, NO₃^-, pH, and Chla profiles (P _NR< 1.5%; Table S2 ).

Using our secondary classification system, profiles with functioning sensors were the least common among pH profiles (P _FUNC = 60%; Table S2 ). This suggests that sensor malfunction is the cause of most low-quality pH profiles, rather than correctable calibration drift errors. Functioning sensors were present in a higher portion of profiles for O₂ (91%), NO₃^- (91%), Chla (94%), b _bp (88%), and PAR/E _d (P _FUNC= 93% for PAR, E _d[380], E _d[412], and E _d[490]).

We first analyzed annual P _HQ by the year a profile was collected to summarize temporal trends in the BGC data quality available to users ( Figure 4 ). Secondly, we estimated the survival rates after one year (R _Float, R _HQ, and R _FUNC) by year the float was deployed ( Figures 5 , S2 ). The trends in R _Float, R _HQ, and R _FUNC by deployment year are useful for characterizing similar batches of floats/sensors ( Figure 5 ; Table S3 ). With both data quality and survival rates, we compared pre-2017 values with the period of 2017 to 2022 to determine changes in the past six years with respect to the early stages of the program.

We find that annual P _HQ in the past six years was >80% for most BGC parameters, except for adjusted pH, which was between 40 and 60% and below pre-2017 levels ( Figure 4 ). PAR/E _d profiles have an annual P _HQ that is always >85%, even though P _HQ in the past six years is below pre-2017 levels. For the remaining parameters, the P _HQ values between 2017 and 2022 are generally close to or above pre-2017 levels.

Regarding survival rates, we find that the annual R _Float and R _HQ can appreciably vary from year-to-year ( Figure 5 ). Generally, R _Float is ~90% at 36.5 cycles, meaning that about 9-in-10 floats survive their first year. R _HQ tends to vary more and is more parameter dependent. R _HQ should always be equal to or less than R _Float because it is not possible for the sensor to record data if the float fails (the reverse is possible). The results show that R _HQ is typically <10% below R _Float. In some cases, R _Float is more than R _HQ, which we suspect results from sensors that temporarily become unresponsive or report bad data, before reporting good data in a successive profile, and the non-linear nature of float/sensor survival over time. Furthermore, R _FUNC follows the same patterns as R _HQ ( Figure S2 ). This suggests that the cause of most low-quality profiles is not intermittent error or other operational limitations but rather sensor/cable damage, degradation, and/or bio-fouling.

R _Float for floats with pH sensors is similar to the R _Float of floats measuring other parameters. However, R _HQ derived from adjusted pH profiles had a distinct drop in 2017 and 2018 that was not associated with a decline in R _Float. This gap between R _Float and R _HQ for floats with pH sensors led to the accumulation of low-quality, adjusted pH profiles seen one-to-two years later in 2018 and 2019. Since then, the R _HQ has improved, resulting in an increase in P _HQ of adjusted pH profiles in the following years ( Figure 4 ). This recovery in P _HQ and R _HQ, especially in 2022, puts pH on track to have similar data quality as the other five BGC properties.

A similar case may have occurred with b _bp from floats deployed in 2019. The R _HQ of b _bp was ~10% lower than the R _Float in 2019 and 2020 ( Figure 5 ) and was followed by a decline in the P _HQ of b _bp profiles in 2020 and 2021 ( Figure 4 ). This difference suggests that b _bp sensors failed before the floats did, which led to a disproportionate collection of low-quality profiles. However, a similar pattern in R _HQ was observed in adjusted Chla ( Figure 5 ), without a similar decline in P _HQ in 2020 or 2021 ( Figure 5 ). The decline in annual P _HQ of b _bp appears to have been corrected since 2022.

R _Float is generally close to the R _HQ for most parameters in most years. The similarity between the two rates explains why there is little variability in the P _HQ of adjusted O₂ profiles in the last six years. For the same reason, P _HQ for PAR/E _d profiles remains generally consistent from 2017 to 2022, despite both R _Float and R _HQ declining to ~70% in 2019/2020.

The density of accumulated BGC profiles is important to construct global climatologies (e.g., Sauzède et al., 2016; Cossarini et al., 2021). The number of float profiles, including those of high quality, are available in most marine regions ( Figure 6 ) but at greatly varying densities ( Table S4 ). The density of high-quality profiles varies among parameters as pilot programs prioritized certain types of sensors early on (e.g., radiometry in the North Atlantic and the Mediterranean Sea).

Generally, the marine regions that have accumulated the highest density of high-quality profiles are the seas around Europe (Mediterranean Sea, Black Sea, and Baltic Sea), the Arctic Ocean, and the Southern Ocean ( Figure 6 ; Table S4 ), although these densities are highly parameter dependent. In the Arctic Ocean, sampling also has not occurred evenly throughout the region. There is a lack of floats between the Chukchi Sea and Fram Strait, leaving a large portion of the Arctic Ocean under-sampled.

The Southern Ocean has accumulated high profile densities for each BGC property due in large part to the success of the Southern Ocean Carbon and Climate Observations and Modelling (SOCCOM) project (Maurer et al., 2021; Sarmiento et al., 2023). However, floats deployed as part of SOCCOM do not include radiometers, explaining the low level of PAR/E _d data in this region. Additionally, pH data coverage is also lower than the other parameters in this region, due to sensor quality issues highlighted in the previous section.

For the major basins (Pacific, Atlantic, and Indian Oceans), the densities of high-quality profiles are generally lower than polar regions and the seas around Europe ( Table S4 ). Additionally, the high-quality, adjusted O₂ profiles are generally high compared to other BGC parameters collected in the same major basins.

With an annual global target density of ~1.1 profiles per 10,000 km² per year, the density of high-quality profiles is assessed on an annual basis to help evaluate the success of BGC-Argo ( Figures 7 , 8 ; Figures S3, S4 ). The annual, regional density of high-quality profiles, expressed as a percentage of the target density, highly varies between parameters and regions, although there is a general interannual trend of increasing profile densities. The percent area of each region with sufficient sampling also highly varies between parameters and regions. While there are significant year-to-year fluctuations, the general long-term trend is that profile densities and percent area coverage have been increasing since 2012.

BGC profiles in the seas around Europe far surpassed annual region-wide target densities at >100% in recent years ( Figure 7 ) and meet target profile densities in ~70% of the area ( Figures 8 , S4 ), except for profiles of adjusted pH and adjusted NO₃^-. For the parameters present, profile densities surpassed annual targets in the Mediterranean, Black, and Baltic seas individually ( Figure S5 ). Considering that the marginal seas around Europe are more variable than much of the open ocean, a higher sampling effort may be needed to better constrain uncertainties. The Arctic Ocean has met target densities for most parameters in the last two years as well, however, this sampling is largely concentrated in an area that consists of 30% or less of the region. The high, region-wide densities (in both the Arctic Ocean and the seas around Europe) are partly driven by floats profiling more frequently than every 10 days, which leads to more observations than what is expected from an annual target based on 10-d sampling intervals.

The Atlantic, Indian, and Pacific oceans generally have reached 10 to 30% of target densities, region-wide, for most parameters. In the same regions, about 5 to 30% of the area has hit target densities. The ocean surface area in these regions with any amount of BGC data is much higher, generally reaching 30 to 60% of a region’s area ( Figure S4 ). For O₂, profile density and percent area coverage tend to be higher than other parameters. O₂ profile densities are between 50 to 70% of the target in the North Atlantic, North Pacific, and South Atlantic oceans in recent years, while the percent area coverage is <30%.

As is the case for all regions, region-wide profile densities are closer to the target density than when evaluating target density in terms of surface area covered. This discrepancy means that the sampling effort is concentrated over a small portion of a region’s area. Future float deployments could target these under-sampled regions, although, depending on scientific interest, concentrating sampling effort in certain areas may be needed for constraining highly variable regions (e.g., the Labrador Sea).

In addition to refining quality flags through time and initiating data adjustments, there are further parameter-specific procedures that should be adopted by the data management team to further improve the absolute accuracy of BGC parameters. While high-quality profiles already represent more than 90% of adjusted O₂ profiles collected thus far, O₂ optode sensors suffer from slow response times, which are not routinely corrected within the data system. Slow sensor response times lead to a lag in true O₂ concentrations and can create errors on the order of ~10 μmol kg^-1 (Bittig et al., 2014; Gordon et al., 2020). Fortunately, various methods exist for calculating the in situ response time of O₂ optodes, given the existence of certain ancillary data and/or sampling conditions (Bittig et al., 2014; Gordon et al., 2020). Such response time corrections will help reduce uncertainties around the oxycline, although, it would not likely affect the assigned QC flags.

In situations where salinity sensor data are labeled with a flag of “4” (unrecoverable), O₂ data are given a label of “3” because salinity compensation to the O₂ measurements cannot be performed. In these cases, utilizing salinity profiles from a high-quality derived product may help recover more O₂ profiles. This procedure has been explored by the Argo community and will likely be implemented in the near future.

The pH sensors tended to fail far more than the other BGC sensors ( Figures 4 , 5 ; Table S2 ). Since a significant portion of the processed pH data was given a QC flag of “4” or characterized as “malfunctioned”, improvements to the P _HQ appear more limited by the technology. Accordingly, the community has continued to refine and improve the design of the pH sensor, leading to a steady increase in reliability over the past five years. Future, repeated analysis of the pH QC time series presented here should be performed to ensure that the upward trending pattern in P _HQ continues. It is important to note that even with relatively high failure rates, BGC-floats now collect substantially more pH profiles than shipboard measurements, playing a critical role in understanding the ocean carbon cycle.

Float fluorometers currently rely on a lab-based calibration of fluorescence-to-Chla. The conversion of fluorescence to in situ Chla concentration depends on both the composition of the phytoplankton community and their physiological status (Cullen, 1982). Ideally, the relationship between fluorescence and Chla is determined using water samples collected at the deployment site. However, this approach is not the most cost-effective, and any relationships determined at the time of deployment will likely not be the same once seasons change or the float drifts into an ecologically distinct region. Currently, each fluorometer is calibrated in a laboratory setting, using a reference fluorometer that was initially calibrated to a single culture of Thalassiosira weissflogii. This calibration has introduced a bias in the determination of Chla in situ. Roesler et al. (2017) quantified this bias with WET Labs ECO sensors and Chla concentrations determined by high pressure liquid chromatography (HPLC). Using their analysis, BGC-Argo adjusts Chla data by dividing unadjusted Chla by two to reduce this global bias. While this adjustment does not account for regional or seasonal changes in the fluorescence-Chla relationship nor greatly affects how QC flags are applied, it is a step forward for gaining more accurate estimates.

An important effort is made by the Argo Data Management Team to calibrate and qualify fluorescence data measured from BGC-Argo floats in order to deliver Chla with the best possible accuracy. It has been shown that using radiometric data associated with fluorescence measurements is an effective way to improve the fluorescence calibration (Xing et al., 2011; Xing et al., 2018). However, not all BGC-Argo floats are equipped with radiometers, making it impossible to calibrate the entire fleet homogeneously. Recently, new methods based on machine learning have made it possible to derive radiometric profiles, from merged satellite ocean color observations and hydrological data, for any BGC-Argo float, whether equipped with a radiometer or not. These synthetic radiometric profiles can thus be assigned to the fluorescence profiles to consistently calibrate the whole BGC-Argo fleet. This method is a promising alternative to other possible approaches that obtain seasonal/regional adjustments tuned to space-based estimates at the surface (Boss et al., 2008; Lavigne et al., 2015), or to HPLC Chla data.

Since the global adjustment factor is also only applicable for WET Labs (now Seabird) ECO sensors, other sensors may have to be calibrated to Chla following a different procedure and may behave differently when compared to WET Labs sensors. Subsequently, new sensors will most likely require different adjustment factors. One way to assess sensor/model-dependent adjustment factors could be to deploy some floats with multiple models of fluorometers to cross-calibrate between sensors. To catch the possibly wide range of variability in these adjustments, float deployment should be undertaken in various open ocean areas covering a wide range of biogeochemical and trophic conditions (e.g., from oligotrophic subtropical gyres to high latitude environments with strong seasonal changes in biomass). This cross-calibration could then be used to convert all Chla measurements to a common unit, assuming that differences between the same models of fluorometers are negligible. Similarly, a standard calibration procedure could be developed for BGC-Argo floats to ensure all Chla derived from fluorescence measurements from various models are intercomparable (e.g., with diffuse attenuation coefficient estimates from irradiance profiles). Regardless of which technique is applied, great care will need to be taken to ensure the interoperability of the resulting Chla concentration data so that the researchers can leverage the full database at once.

Measurements of b _bp from floats currently have a limited QC procedure that includes processing for range filters, data spikes, and bad offsets. Additional real-time QC procedures proposed by Dall’Olmo et al. (2023) will be integrated into BGC-Argo data processing soon and will likely change the proportion of QC flags in the system. Briefly, this new procedure identifies issues such as noisy b _bp data, profiles with a large fraction of missing data, and profiles with negative b _bp. Potential sensor malfunction or biofouling is also identified by setting a threshold for high deep values of b _bp, where b _bp is expected to be very low (but not equal to zero). Dall’Olmo et al. (2023) also describe a procedure that identifies biased measurements of b _bp resulting from particles that have accumulated on the float while the float is parked at depth between casts. For those interested in data spikes (e.g., Briggs et al., 2011) or high b _bp at depth, both of which could result from real oceanographic phenomena, users will likely want to apply custom QC procedures. Once implemented, these added real-time QC procedures will reduce the number of falsely identified high-quality profiles available, but also erroneously identify good data as bad. Efforts should be made in developing a DMQC procedure to improve data quality flagging.

We found that high-quality PAR/E _d profiles represent ~90% of the entire database ( Table S2 ). The data from radiometers are generally good without much QC. (Radiometry RTQC currently involves a global range test, and all radiometry data are given a QC flag of “2” unless the test is failed.) Previously tested improvements to the float radiometry processing include a dark offset correction, temperature correction, spike removal, and sensor drift detection (Organelli et al., 2016; Organelli et al., 2017; Jutard et al., 2021; O’Brien and Boss, 2022). About half of the database has DMQC processing correcting for temperature and dark values (Jutard et al., 2021) but has not yet been implemented on all floats. Future float deployments should prioritize collecting nighttime profiles of PAR/E _d, at least once a year by each float, for these temperature and dark offset corrections (Jutard et al., 2021). In a similar sense, collecting these nighttime profiles could also prove beneficial for assessing NPQ corrections. To reduce space-time biases in such an assessment without changing the standard profiling interval, average daily cycles could be extracted (Johnson and Bif, 2021; Stoer and Fennel, 2023).

Unstable sea states, clouds, and wave focusing can also momentarily alter light conditions which can lead to large fluctuations in the downwelling light profile. While there is nothing inherently wrong with the sensor during such events (as it represents natural light variability), stable light conditions are needed to better characterize the optical properties of the water column (e.g., estimation of the light diffuse attenuation coefficient). Procedures to identify these effects have been successfully applied to floats before (Organelli et al., 2016; Organelli et al., 2017) and could be further developed to create a reference dataset for use in the ocean optics community.

A few pre-deployment strategies can also be taken to help improve BGC data quality in the future. For one, testing the float from a dock could help identify issues with sensors, communications, and mechanics before deployment in the open ocean. Ideally, dock testing would prevent the deployment of floats that would otherwise quickly fail in the field. This strategy has proven to be very effective for the SOCCOM and GO-BGC projects. Secondly, while not always logistically feasible, reducing the time between when the float is manufactured and deployed may improve float survival and functionality by limiting exposure to possible damage. Finally, larger batteries can also be purchased for some float models to increase the float’s lifetime, allowing the float to collect more high-quality profiles (assuming all sensors continue to function).