3.1.1 Global ocean

At the global scale, the model fits the data with a mean bias close to zero, an RMSD of 20.48 µatm, and a coefficient of determination (r²) of 0.76. The temporal fluctuation of the spatial mean of the model–data mean difference over the global ocean is displayed in Fig. 4a along with the number of available gridded data. The time series of the yearly bias (black curve) starts with a large positive value (7.47±1.60 µatm) in the year 1985 (∼740 gridded data). The bias drops during the following years and fluctuates around zero from 1994 onwards (the number of grid boxes containing SOCAT observations per year is generally larger than 5000). In general, the magnitudes of the yearly model bias and model spread are correlated with the number of observation-based data, which has increased greatly since the 1990s. The importance of sustained data coverage is emphasized by Fig. S4. It illustrates the fact that large model–data mismatches are frequently associated with the interruption of voluntary observing ship (VOS) lines and thus with the tracking of CO₂ fugacity over large regions. The larger bias computed prior to the 1990s (Fig. 4a) might intuitively lead to the conclusion that model outputs are less reliable than those in the later periods. However, this global mean score is influenced by the number and distribution of data, and consequently the increased data density does not fully explain the reconstruction skill. For instance, even with a higher number of observation-based data than that in the pre-1990s, the years 2001 and 2007 stand out with strong negative biases ($-\mathrm{5.44}\pm \mathrm{1.26}$ and $-\mathrm{3.12}\pm \mathrm{0.92}$ µatm, respectively). While such a comparison between the global bias and the number of data highlights the lack of a simple relationship between the number of data and the skill of the mapping method, the ensemble spread (dark grey area) of model errors, representing the spread of the annual mean of pCO₂ estimates on the SOCAT grid with observations, is reduced with an exponential decay constant of 0.46±0.06 per 1000 gridded data (Fig. 4b).

Figure 4(a) Time series of the yearly mean model bias, i.e. the reconstructed pCO₂ data minus SOCATv2020 data, over the global ocean. The black curve and dark grey area represent the mean estimate and 1σ envelope of errors of the 100-member ensemble; the light grey curve represents the total number of gridded SOCAT data used in the FFNN model construction. (b) Exponential fits of the model uncertainty (the magnitude of the 1σ envelope in Fig. 4a) against the number of gridded data per year. The exponential function is $y=a{\exp }^{-bx}+c$. The black curve is derived from the best fit, and the grey-shaded area corresponds to the spread derived from standard errors of parameter estimates. (c) Statistical scores for 11 oceanic regions with the size of each scattered object proportional to the number of regional data (Table S2).

Download

The model scores for the open ocean over the period 1985 to 2019 are 17.87 µatm for the RMSD and 0.78 for r². The skill of this novel method, which uses only two-thirds of SOCAT data for fitting each of 100 FFNN models, ranks similarly to those from alternative statistical reconstruction approaches (Rödenbeck et al., 2013; Landschützer et al., 2014; Gregor et al., 2019) which have been used to complement model-based estimates of the ocean carbon sink (Friedlingstein et al., 2019, 2020).

The CMEMS-LSCE-FFNN reconstruction over the coastal regions for the full period is roughly 2 times less effective than over the open ocean in terms of the RMSD (35.86 µatm), while it shows a rather good fit with r²=0.70. The high RMSD reflects high local model errors along the continental shelves (Fig. S3). For the 1998–2015 period, the CMEMS-LSCE-FFNN approach scored an RMSD of 35.84 µatm while a recent coastal reconstruction by Landschützer et al. (2020) obtained an error of 26.8 µatm (see their Table 1). The latter presents a global ocean pCO₂ climatology product by unifying data over the same period from two conceptually equivalent reconstruction models: one covering the open ocean at a 1^∘ × 1^∘ resolution (Landschützer et al., 2016) and one targeting the coastal ocean at a 0.25^∘ × 0.25^∘ resolution (Laruelle et al., 2017). These previous reconstructions cover the coastal region with a broader definition (400 km distance from the seashore) than the MARCATS mask used in this study, leading to the differences in characteristics and numbers of evaluation data of pCO₂. In addition, the CMEMS-LSCE-FFNN model was designed with the leave-p-out cross-validation approach excluding many more independent data from monthly model fitting for model evaluation than the previous models. Overall model errors remain high despite the increase in the spatial resolution and in the number of observations. Coastal and shelf seas are characterized by complex physical and biological dynamics leading to high variability at small scales. For instance, pCO₂ levels over the Californian shelf can exceed 850 µatm, with a spatial gradient of pCO₂ as large as 470 µatm over a distance of less than 0.5 km (Chavez et al., 2018; Feely et al., 2008). Clearly, further model improvement is needed in order to capture such high spatial and temporal variability in surface ocean pCO₂ present in observations (see also Bakker et al., 2016; Laruelle et al., 2017, and references therein).

In the following subsections, we present and discuss the reconstruction skills for different ocean regions, as well as for open-ocean and coastal domains (Fig. 4c). Complete results including the numbers of gridded data, RMSDs, and r² for each region are summarized in Table S2.

3.1.2 Ocean basins

Arctic

Data coverage is particularly sparse over the Arctic Ocean (Ar) with 50 to 220 grid boxes with observations per year since 2007 and an interruption in 2010 (Fig. S4). While continental shelves account for 50 % of the region's area, only one-third of the observation-based data are from coastal regions. Moreover, observations are seasonally biased towards ice-free summer months (Bakker et al., 2016). Though reconstruction standard errors are similar for open basins and coastal regions (RMSDs of 33.01 and 30.65 µatm, respectively), the coefficient of determination is higher over the open ocean (r²=0.61) compared to coastal seas (r²=0.44), suggesting a higher model skill over open basins. The close-to-zero bias of the coastal reconstruction shown in Fig. 4c results from the compensation between highly positive and negative values over the continental shelves of Alaska, the Canadian archipelagos, and the Barents and Kara seas (see Fig. S3); the yearly bias fluctuates within $[-\mathrm{50},\mathrm{30}]$ µatm (Fig. S4). Of all open-ocean regions, the Arctic reconstruction has the highest bias (3.19 µatm). Cold Arctic waters are characterized by low levels of surface ocean pCO₂ due to the temperature effect on CO₂ solubility and the seasonal drawdown of dissolved inorganic carbon (DIC) during summer months by intense biological production (Feely et al., 2001; Takahashi et al., 2009; Arrigo et al., 2010). Assuming that the Arctic predictors remain within the range of global relationships, the overestimation of pCO₂ by CMEMS-LSCE-FFNN, as seen in Fig. 4c, suggests a possible underestimation of biological productivity. While this remains conjectural, we acknowledge a large uncertainty in the contribution of biological activity (net primary production, NPP) to surface ocean pCO₂, as it is “proxied” by chlorophyll a derived from remote sensing (Maritorena et al., 2010; Babin et al., 2015). Overall, these scores point to the Arctic as a relatively poorly reconstructed region.

Atlantic

The North Atlantic stands out as a region with high data coverage (Fig. S1a) and a rapidly increasing number of data since 2000 (Fig. S4). A sustained sampling effort adds between 2000 and 4000 data each year to the database over the Subtropical Atlantic (StA) and Subpolar Atlantic (SpA) regions (including between 10 %–40 % of coastal data). The data density over the North Atlantic stands in strong contrast to the often fewer than 1000 gridded data per year collected over the Equatorial Atlantic (EA) and South Atlantic (SA) and their strong year-to-year variability.

The comparison between the reconstructed open-ocean pCO₂ and evaluation data over the four sub-regions of the open Atlantic (Fig. 4c and Table S2) reveals small mean model–data differences, which together with the two other scores, identify the Atlantic as the basin with the highest reconstruction skill. RMSDs corresponding to the StA, the EA, and the SA are below 15.50 µatm, and r² values are in the range of [0.69,0.77]. While a larger RMSD is obtained over the SpA (23.68 µatm), the r² of 0.76 falls close to the upper end of the range determined for the three other regions. As discussed in Schuster et al. (2013), large temporal and spatial gradients of pCO₂ as well as its variability driven by a diversity of physical and biological processes (e.g. surface ocean temperature gradients, biological production, vertical mixing, and horizontal advection of water masses) keep the analysis of pCO₂ over the SpA challenging.

Despite accounting for over 59 % of the total coastal data, skilful data reconstruction over the coastal Atlantic regions remains difficult. RMSDs are in general above 30 µatm, and, with the exception of the coastal SpA (r²=0.79), below 51 % of the observed variance is predicted by the model over the other regions (StA, 0.51; EA, 0.25; SA, 0.46). The large model–data mismatch along the Atlantic continental shelves (Fig. S3) reflects the poor reconstruction of pCO₂ over regions under the influence of upwelling systems (e.g. Moroccan coast, Benguela), large river discharges (e.g. Amazon, Congo, Florida, Mississippi), and the bottlenecks of gulfs or bays (e.g. Bahamas, English Channel).

Pacific

With the exception of the Subpolar Pacific (SpP), the number of observations has increased regularly over the Pacific basin. In recent years, there are from 1000 to 3500 grid boxes with observations recorded over the Subtropical Pacific (StP), the Equatorial Pacific (EP), and the South Pacific (SP) (Fig. S4). Forty percent of total open-ocean data belong to the StP and the EP in the years 1985–2019. Corresponding RMSDs are 17.15 and 16.68 µatm, with r² above 0.78. Despite a data coverage below one-third of that reported for the two previous regions, the model proved skilful in reconstructing pCO₂ over the SP (Fig. 4c) with an RMSD of 11.50 µatm and r² of 0.76.

The overall good performance of the FFNN over these three Pacific sub-regions contrasts with its lack of skill over the open SpP. The data density is poor and highly variable. From before 1994, fewer than 250 gridded data per year are available to constrain the reconstruction, followed by several years of intense effort and a maximum of about 1250 data in 2000 before decreasing again to the pre-1994 values. At first order, skill scores fluctuate in line with data density. During the first period (up to 1994), the bias varies within $[-\mathrm{25},\mathrm{25}]$ µatm (Fig. S4); it decreases close to $[-\mathrm{2},\mathrm{4}]$ µatm between 1997 and 2000 and increases again along with decreasing data density. Much like the SpA, the SpP is a region characterized by a strong spatial and temporal variability in pCO₂ (Ishii et al., 2014), challenging any reconstruction method. The difficulty is further aggravated by the paucity of data in this region compared to the SpA. Skill scores are modest over the SpP with an RMSD of 29.08 µatm and r² of 0.64 (Fig. 4c and Table S2).

The ratio between coastal and open-ocean observation-based data is 1 : 24. The paucity of data for the coastal domain is reflected by lower skill scores compared to the open ocean. Over the coastal SpP, for example, the RMSD amounts to 54.69 µatm, while it is 29.08 µatm for the corresponding open-ocean region. Comparable to the SpP, data reconstruction over the coastal regions of the StP (e.g. North American coast, Sea of Japan), as well as over the western EP (e.g. Peruvian upwelling) and the SP (e.g. offshore Chile), remains difficult (Fig. S3). Similar results have been found by Landschützer et al. (2020).

The EP is characterized by strong equatorial upwelling, making it one of the major outgassing regions of CO₂ (Feely et al., 2001). Surface ocean pCO₂ shows a strong interannual variability predominantly in response to the El Niño–Southern Oscillation (ENSO), the dominant regional climate mode (Rödenbeck et al., 2015; Landschützer et al., 2016; Denvil-Sommer et al., 2019). Before the 2000s, negative (positive) peaks of bias (Fig. S4) coincide with La Niña years, e.g. 1988–1990, 1995–1996, 1999–2001 (El Niño years, e.g. 1986–1987, 1991–1992, 1997–1998) (see the ENSO events highlighted in Fig. 9). A strong negative bias is again computed in 2010–2012, which could reflect the lack of data during that cooling phase. On the contrary, the reconstruction seems less sensitive to the strong warm anomalies associated with the 2015–2016 El Niño. The model appears to be more efficient at reconstructing surface ocean pCO₂ during the hot climate mode (El Niño) than during the cool one (La Niña) when enhanced upwelling drives surface ocean pCO₂ up and towards unusually large values. This allows us to anticipate the effect of a general decrease in data collection and processing since 2020 in response to the 2019 coronavirus disease (COVID-19) pandemic on the estimation of the ocean carbon sink. We expect a high negative bias in model estimates of pCO₂ and the consequent underestimation of CO₂ outgassing due to the combined impact of data decreasing and La Niña conditions governing since August and September 2020 (https://public.wmo.int/en/media/press-release/la-nina-has-developed, last access: December 2020). It is worthwhile to also note that monthly gridded SOCAT data in the eastern EP have declined in the last 5 years compared to the other years in the 2010s.

Indian Ocean

The Indian Ocean (IO) is the third-largest oceanic region by area but also the one with the lowest data density. With the exception of the year 1995 (approximately 1900 grid boxes including observations), as few as 500 gridded data have been provided per year (Fig. S4), yielding a total number of data often below 10 per grid cell for the entire reconstruction period (Fig. S1a). There have been even fewer than 75 grid boxes with observations per year over the continental shelf. However, the reconstruction over the coastal region is comparable to the open IO with a low RMSD (<19 µatm) and a high correlation with the observation-based data (r²=0.65). The overall negative bias shown in Fig. 4c for the coastal IO points to the model underestimating coastal pCO₂ levels. Large errors are distributed along the western Arabian Sea, western Madagascar, and the tropical eastern IO (Fig. S3). These regions are under the influence of the southwest monsoon, giving rise to a seasonal upwelling regime (see Feely et al., 2001; Sabine et al., 2002; Sarma et al., 2013, and references therein). Strong seasonal upwelling results in a marked seasonal cycle of surface ocean pCO₂ with high levels during the upwelling season. The paucity of data is likely to limit the skill of the model reconstruction of the seasonal cycle over large parts of the IO with consequences for the annual mean analysed here.

Southern Ocean

Until recently, data coverage over the Southern Ocean (SO) was sparse (Fig. S1a), irregular at the grid cell scale, and biased towards austral summer months (e.g. Bushinsky et al., 2019; Gregor et al., 2019). A strong sampling effort allowed a recent increase in observations to reach up to 2000 gridded data per year (Fig. S4). Model scores for the open and the coastal ocean are RMSDs of 19.18 µatm and 35.73 µatm, respectively, and r² values of 0.62 and 0.65, respectively. The reconstruction lacks skill over the continental shelves of South America and Antarctica (see Fig. S3).

In general, the pCO₂ reconstruction over the SO has less skill compared to the Atlantic or the Pacific due to the paucity of observation-based data compared to its large area. Rödenbeck et al. (2015) reported inconsistent reconstructed interannual variability in pCO₂ between different data-based methods. The interannual variability is large due to the natural variability in the coupled ocean–atmosphere system characterized by one of the globe's strongest ocean currents, strong winds, and vertical mixing and upwelling of DIC-rich deep waters (Gregor et al., 2018; Gruber et al., 2019). Efforts to improve pCO₂ reconstruction are ongoing and include model development (e.g. Gregor et al., 2017), as well as the increase in data coverage by the addition of data from different sampling platforms (e.g. profiling floats; Bushinsky et al., 2019). For the time being, CMEMS-LSCE-FFNN stands out as one of the skilful models with respect to observation-based data in the SO (Friedlingstein et al., 2020; Hauck et al., 2020).

3.1.3 Time series stations

CMEMS-FFNN-LSCE estimates of pCO₂ are now compared with moored pCO₂ time series provided by Sutton et al. (2019). This data product comprises pCO₂ measurements collected from a wide range of oceanic regions since 2004 (Figs. S5–S8). Most of the stations were established in the North Atlantic and the North Pacific and Equatorial Pacific; one site is in the IO and another in the SO. Approximately one-third of the Sutton et al. (2019) sites belong to the coastal seas and shelves (Fig. S8). Table S3 details the information of the moored pCO₂ time series.

Observation-based data used for model–data comparison (black points in Figs. S6–S8) are monthly averages of pCO₂ measurements at each site. This interpolation results in monthly time series with a number of data N between 9 (NH10) and 98 (WHOTS). The ensemble mean μ_ensemble and ensemble spread σ_ensemble (Eq. 2) are computed from the CMEMS-LSCE-FFNN ensemble of model outputs at the four nearest model grid boxes of each location. Results confirm a reasonably good reconstruction of the proposed approach. The model best estimates (thick coloured lines) characterize pCO₂ trends and variations in in situ data well, and the model ensembles almost catch the observation-based data in their 99 % confidence interval (light shaded envelope). For over 90 % of the time series stations, the model estimation obtains a moderate to high coefficient of determination r² with a linear model–data correlation r larger than 0.5 (e.g. BTM, 0.98; CRESCENTREEF, 0.92; HOGREEF, 0.84; SOFS, 0.79; TAO110W, 0.75; WHOTS, 0.73). Mean bias μ_misfit (Eq. 3) and the RMSD (Eq. 4) are relatively low compared to mean pCO₂ values of the time series stations.

Half of the open-ocean reconstructions have model errors of less than 20 µatm and are even less than 10 µatm at KEO, PAPA, SOLS, STRATUS, and WHOTS (Figs. S6 and S7). Despite having less skill than the open-ocean reconstructions, the coastal-ocean reconstructions are quite compatible with the in situ data (Fig. S8). Most of the RMSDs remain lower than 20 % of the mean pCO₂ values of coastal time series (e.g. CCE2, 36.53 µatm; ICELAND, 12.26 µatm; M2, 36.58 µatm). For some other stations on the US west coast and in the oceanic regimes of coral reef, the estimates differ from the observation-based data in terms of the magnitude of pCO₂ (e.g. CRIMP2, LA PARGUERA) and/or of its seasonal cycle (e.g. CHABA, CHEECAROCKS, SEAK).

The reconstructed time series cover the full period 1985–2019, while observation-based data are still sparse and almost all distributed over the past 2 decades (Figs. S6–S8). The CMEMS-LSCE-FFNN time series would be useful for estimating and assessing long-term means, trends, and variations in CO₂ surface partial pressure and the corresponding air–sea fluxes.

3.2.1 Arctic

The Arctic Ocean stands out as the region with the strongest CO₂ uptake per unit area with 2.336±0.104 molC m⁻² yr⁻¹ for the open sea and 1.522±0.108 molC m⁻² yr⁻¹ for the continental shelf margins (Figs. 5b and 6b). At the scale of grid cells, air–sea gradients of pCO₂ are large, but the downward fluxes are relatively modest over the shelves of eastern Greenland, the Barents and Kara seas, and the Siberia seas (Figs. 5 or S9). During the sea-ice-covered seasons, these coastal regions are neutral while the open-ocean Arctic sectors (e.g. the Norwegian Sea, the Barents Sea, the Kara Sea) are CO₂ sinks with moderate influx densities (Fig. 8). The open-ocean influx density exceeds 3 molC m⁻² yr⁻¹ in the Arctic summer. This substantial amount of CO₂ uptake is driven by low surface ocean temperature, seasonal changes in sea-ice cover, and intense biological production. Increasing light availability and input of nutrients through meltwaters and river discharges sustain high levels of primary production and CO₂ drawdown (Bates and Mathis, 2009; Arrigo et al., 2010; Yasunaka et al., 2016, 2018). Notwithstanding, the Arctic Ocean represents roughly 0.58 % of the total surface ocean area (Table 1), and the yearly mean CO₂ uptake integrated over the Arctic for the full period amounts to only 1.64 % of the global ocean sink (Table 2 and Fig. 6a).

3.2.2 Atlantic

The open-ocean Subpolar Atlantic (SpA) sink contributes approximately 78 % to the total SpA annual C uptake (0.259±0.011 PgC yr⁻¹), as well as 12.29 % to the total ocean sink (1.643±0.125 PgC yr⁻¹, Table 2). Per unit area, the open-ocean influx amounts to 2.012±0.092 molC m⁻² yr⁻¹ and the coastal-ocean influx is 30.51 % less than its open-ocean counterpart and slightly lower than the coastal Arctic sink (Fig. 6b). However, when integrated over the region, the yearly uptake of 0.057±0.004 PgC yr⁻¹ makes the coastal SpA the strongest sink among the 11 coastal regions (Fig. 6a). The interplay between temperature- and biology-driven effects results in changes in the seasonal and spatial distributions of surface ocean pCO₂ and ultimately air–sea CO₂ fluxes. During boreal winter–spring, high wind speeds enhance gas transfer velocities and contribute to a strong cooling and an increase in CO₂ solubility (Takahashi et al., 2009; Feely et al., 2001), both enhancing uptake of CO₂ over the Labrador Sea, the North Atlantic and Norwegian currents, and the Barents and Kara seas (Fig. 8). High wind speeds also strengthen vertical mixing, a process supplying dissolved inorganic carbon (DIC) and nutrients to the surface ocean. During the spring and summer months, vigorous biological activity (Sigman and Hain, 2012) counteracts the warming-induced decrease in CO₂ solubility and increase in pCO₂ by drawing down DIC (Feely et al., 2001). Along the coast of the Barents and Kara seas, inputs of fresh water (decrease in salinity and increase in CO₂ solubility) and nutrients (biological activity and DIC drawdown) combine to strengthen CO₂ uptake (Arrigo et al., 2010; Yasunaka et al., 2016, 2018; Olafsson et al., 2021). This contrasts with other coastal regions (e.g. southern North Sea and Baltic Sea) where the respiration of terrestrial particulate organic carbon from river run-off contributes to making these areas a strong seasonal source of CO₂ (Borgesa and Gypensb, 2010; Becker et al., 2021).

The Subtropical Atlantic (StA) is characterized by weak to moderate mean flux densities per unit area (open, 0.733±0.036 molC m⁻² yr⁻¹; coastal, 0.457±0.064 molC m⁻² yr⁻¹). The total integrated C uptake amounts to 0.214±0.011 PgC yr⁻¹, with 0.204±0.010 PgC yr⁻¹ contributed by the open ocean. As for the SpA, the net uptake reflects the combined effect of cooling, mixing, and biological activity. Figures 5 and S9 show the regional distribution of sources and sinks. Regions of intense CO₂ uptake are associated with the warm Gulf Stream and its northeastward extension, the North Atlantic Drift. Strong uptake is also found over the western continental shelf where strong river discharges sustain high levels of biological productivity, in particular during spring (Jamet et al., 2007; Kealoha et al., 2020). Weaker sinks or sources of CO₂ in the southwestern StA and the eastern subtropical gyre are primarily driven by high surface temperature and enhanced stratification (Schuster et al., 2013). The latter restricts the vertical supply of nutrients and limits biological production. Finally, a relatively strong source of CO₂ is found over the Canary upwelling system in summer (Fig. 8).

The Equatorial Atlantic (EA) stands out as the second-strongest source region of CO₂ after the Equatorial Pacific (EP) with a yearly outgassing of $-\mathrm{0.117}\pm \mathrm{0.009}$ PgC yr⁻¹ (Fig. 6a). Most CO₂ is released from the open ocean with an average efflux of $-\mathrm{0.407}\pm \mathrm{0.031}$ $\mathrm{molC}{\mathrm{m}}^{-\mathrm{2}}{\mathrm{yr}}^{-\mathrm{1}}$ (Figs. 5b and 6b). This intense source of CO₂ stems from upwelling of cool and CO₂-rich waters in the eastern EA. A westward increase in outgassing is observed along with the advection of CO₂-rich waters (Schuster et al., 2013). The coastal EA regions release an average of $-\mathrm{0.288}\pm \mathrm{0.064}$ molC m⁻² yr⁻¹ of CO₂. Over large areas, the opposing effects of primary production and high surface temperature combine to weaken the coastal sink or seasonally switch it from a weak to a moderate source (e.g. the northeast EA, Caribbean Sea, Venezuelan and Guiana basins, Gulf of Guinea) (Fig. 8). The Amazon River is a notable exception. Its large discharges of fresh water and nutrients, as well as of dissolved and particulate carbon, turn the coastal and adjacent shelf seas into a net sink of CO₂ (Medeiros et al., 2015; Ibánhez et al., 2015).

The South Atlantic (SA) uptake amounts to 0.192±0.016 PgC yr⁻¹. Regions north of 30^∘ S act as weak sources or are neutral with respect to air–sea exchanges of CO₂, as opposed to regions to the south which are significant sinks of CO₂ (Fig. 5b). For the full period, densities over the open and coastal regions are 0.862±0.072 and 0.776±0.125 molC m⁻² yr⁻¹, respectively. Coastal regions are changing from moderate sources to sinks with increasing latitude (Fig. S9). The SA has similar seasonal dynamics to the StA with CO₂ uptake in winter and outgassing in summer (Takahashi et al., 2009; Schuster et al., 2013). During the austral winter, deep mixed layers result in cold surface waters which absorb CO₂ from the atmosphere. By contrast, warming during the summer reduces the solubility of CO₂, leading to a weak sink or even a source (Fig. 8). As explained before, biological production counteracts the effect of warming, and the vigorous spring bloom contributes to the uptake south of 30^∘ S (Sigman and Hain, 2012; Carvalho et al., 2020).

3.2.3 Pacific

The Subpolar Pacific (SpP) is the second-smallest region by area (2.78 % of the total surface ocean area) and with 0.040±0.010 PgC yr⁻¹ (net coastal and open-ocean sinks) provides the smallest contribution to the total yearly ocean C uptake (Table 2 and Fig. 6a). The coastal ocean contributes about 0.032±0.004 PgC yr⁻¹ to the total yearly C uptake, making the SpP the only region for which coastal fluxes exceed open-ocean fluxes. The strength of its coastal C sink ranks second among all coastal regions (Fig. 6a). Seasonal features of CO₂ fluxes are shown in Fig. 8. The SpP is ice-covered during the winter months, which results in close-to-zero air–sea fluxes per unit area north of 60^∘ N (e.g. Beaufort, Siberia, and Chukchi seas). Besides, vertical convection during winter brings up old DIC-rich waters, leading to CO₂ outgassing exceeding −3 molC m⁻² yr⁻¹ in the south of the region (Bates and Mathis, 2009; Arrigo et al., 2010; Ishii et al., 2014; Yasunaka et al., 2016). Intense biological production during the boreal summer drives intense uptake of CO₂ over the entire SpP (Feely et al., 2001; Sigman and Hain, 2012; Ishii et al., 2014). The interplay of these two seasonal mechanisms and their opposing effects make the open SpP a weak yearly net sink (Fig. 6). The average flux density per unit area is 0.044±0.123 molC m⁻² yr⁻¹ over the open ocean, much smaller than the value determined for the coastal ocean of 0.775±0.127 molC m⁻² yr⁻¹ (Fig. 6b). As shown in Bates (2006), Arrigo et al. (2010), and Ishii et al. (2014), surface DIC concentration is higher over the open, deep basins than the shallow coastal seas of the SpP, particularly induced by deep mixing during winter and spring. Over the same period, seasonal sea ice also restricts gas exchange; the coastal sector thus acts as a neutral region of CO₂ fluxes (Fig. 8). During spring and summer, a substantial amount of CO₂ is also absorbed in the coastal shelf seas influenced by high biological production in large ice-free areas (e.g. Chukchi and Gulf of Alaska) and/or by dilution of seawaters from river fresh water with low salinity and DIC concentration (e.g. Beaufort, Laptev, and East Siberian seas) (Arrigo et al., 2010; Yasunaka et al., 2016, 2018).

A total mean uptake of 0.523±0.016 PgC yr⁻¹ makes the Subtropical Pacific (StP) the largest sink region. The open-ocean contribution dominates the regional sink with 0.495±0.015 PgC yr⁻¹ (Table 2 and Fig. 6a). The corresponding mean flux density per unit area is 1.136±0.036 molC m⁻² yr⁻¹ (Fig. 6b) and makes the StP rank third after the open-ocean Arctic and SpA regions. As discussed for the StA, during winter months cooling and high wind intensities along the Kuroshio and North Pacific currents enhance the uptake of CO₂ (Takahashi et al., 2009; Ishii et al., 2014). By contrast, summer warming drives the StP towards close-to-neutral conditions or a weak source (Fig. 8). With a yearly mean uptake of 0.028±0.003 PgC yr⁻¹, the coastal StP sink ranks third in terms of intensity among the coastal sinks (Fig. 6a). The influx density is 1.444±0.130 molC m⁻² yr⁻¹. Western coastal systems and shelf seas are under the influence of the delivery of fresh water and nutrients by large river systems (Liu et al., 2014). The resulting intense biological production contributes to influx densities per unit area that are higher over the western continental shelf and seas (e.g. East China Sea, Sea of Japan) than over the California upwelling system (Figs. 5b, S9b, and 8).

The Equatorial Pacific (EP) is the strongest source region of CO₂ to the atmosphere with a yearly average efflux of $-\mathrm{0.490}\pm \mathrm{0.021}$ PgC yr⁻¹ from the open ocean and $-\mathrm{0.013}\pm \mathrm{0.003}$ PgC yr⁻¹ from the continental shelves. On average per unit area, the open sea emits $-\mathrm{0.616}\pm \mathrm{0.027}$ molC m⁻² yr⁻¹ of CO₂. This high rate of outgassing is a distinct feature of the EP (e.g. Feely et al., 2001; Takahashi et al., 2009; Rödenbeck et al., 2015; Landschützer et al., 2016; Denvil-Sommer et al., 2019; Landschützer et al., 2019) and is primarily due to the upwelling of DIC-rich deep waters. The magnitude of CO₂ release decreases westwards – from eastern boundary upwelling (e.g. Peru, Panama) to the International Date Line – in line with decreasing upwelling intensity, warmer sea surface temperature, and lower salinity (Ishii et al., 2014). Compared to the open EP, the efflux density of the coastal regions ($-\mathrm{0.334}\pm \mathrm{0.071}$ molC m⁻² yr⁻¹) is roughly half that of the open ocean.

The South Pacific (SP) ranks second as a sink region for CO₂ with a yearly net flux of 0.358±0.029 PgC yr⁻¹, mostly contributed by the open ocean (Fig. 6a). Uptake rates per unit area are very similar to those obtained for the SA (Fig. 6b). A detailed assessment reveals the open-ocean influx density to be slightly lower (0.791±0.066 molC m⁻² yr⁻¹) and the coastal one to be slightly higher (0.987±0.063 molC m⁻² yr⁻¹) over the SP compared to the SA. Due to the larger area of the SP (Table 1), its integrated sink is approximately twice that of the SA. Similarly to the processes discussed above for the SA, vertical mixing drives the uptake of CO₂ during austral winter (Takahashi et al., 2009; Ishii et al., 2014), and the effect of warming on CO₂ solubility during spring and summer is offset by biological production. The biological production leads to moderate to high uptake of CO₂ over the coasts and the southwest open sea (e.g. East Australian Current, southern Australia, New Zealand) (Fig. 8). The influx density decreases eastwards under the influence of the strong upwelling of DIC driven by the Peru Current.

3.2.4 Indian Ocean

The total integrated Indian Ocean (IO) sink is evaluated to 0.300±0.033 PgC yr⁻¹, with 0.305±0.033 PgC yr⁻¹ contributed by the open ocean and a weak coastal source of $-\mathrm{0.004}\pm \mathrm{0.002}$ PgC yr⁻¹. The spatial distribution of flux densities (Fig. 5b) reveals the northwestern IO to be a net source of CO₂ to the atmosphere, while the northeastern IO is close to neutral and latitudes south of 18 ^∘S act as a strong sink. This regional compensation leads to a small open-ocean influx density per unit area of 0.482±0.052 molC m⁻² yr⁻¹ and a small coastal efflux per unit area of $-\mathrm{0.131}\pm \mathrm{0.061}$ molC m⁻² yr⁻¹ (Fig. 6b). The northern IO is a strong source of CO₂ sustained by the monsoon-driven seasonal upwelling along the Arabian and Somali coasts (Behrenfeld et al., 2006; Sarma et al., 2013). The northeastern IO regions including the Bay of Bengal and its continental shelves receive fresh waters discharged from the river Ganges and lateral inputs from Indonesian outflows (see Sarma et al., 2013, and references therein) and switch between mild sources and sinks (Fig. 8). A subtropical front (40^∘ S) divides the region south of 18^∘ S into a weak sink to the north and over an oligotrophic gyre and a band of vigorous uptake to its south over the Subantarctic Zone (SAZ) between 40 and 44^∘ S (Fig. 5b). Similarly to the SA and SP, this entire region is identified as a significant net sink of CO₂ in winter (Fig. 8), possibly driven by enhanced solubility in response to cooling and mixing. While biological production maintains the sink over the SAZ during austral spring and summer months, warming reduces CO₂ uptake over the oligotrophic gyre.

3.2.5 Southern Ocean

The total Southern Ocean (SO) sink amounts to 0.349±0.070 PgC yr⁻¹, including coastal uptake of 0.018±0.002 PgC yr⁻¹. The mean influx per unit area over the open SO is 0.468±0.104 molC m⁻² yr⁻¹ and close to the one obtained for the open IO (Fig. 6b). The area-averaged CO₂ drawdown over the coastal SO is 0.599±0.089 molC m⁻² yr⁻¹ with strong coastal sinks distributed over the South American and Antarctic shelves (60^∘ westwards as seen in Figs. 5b or S4b). During the austral spring and summer, intense phytoplankton blooms enhance the consumption of CO₂ over the Subantarctic Zone and the Polar Frontal Zone between 44 and 58^∘ S (Sigman and Hain, 2012; Lenton et al., 2013), leading to a large sink with a flux density exceeding 1.667 molC m⁻² yr⁻¹ (Fig. 8). South of 58^∘S, sea-ice retreat and vertical stratification contribute to a mild sink over the Antarctic Zone. During winter, vertical mixing brings DIC-rich deep waters to the surface, triggering strong outgassing of CO₂ along the Antarctic Circumpolar Current.