Research Article

The oceanic sink for anthropogenic CO2 from 1994 to 2007

See allHide authors and affiliations

Science  15 Mar 2019:
Vol. 363, Issue 6432, pp. 1193-1199
DOI: 10.1126/science.aau5153

The state of ocean CO2 uptake

The ocean is an important sink for anthropogenic CO2 and has absorbed roughly 30% of our emissions between the beginning of the industrial revolution and the mid-1990s. This effect is an important moderator of climate change, but can we count on it to remain as strong in the future? Gruber et al. calculated the ocean uptake of anthropogenic CO2 for the interval from 1994 to 2007, which continued as expected. They also observed clear regional deviations from this pattern, suggesting that there is no guarantee that uptake will remain as robust with time.

Science, this issue p. 1193


We quantify the oceanic sink for anthropogenic carbon dioxide (CO2) over the period 1994 to 2007 by using observations from the global repeat hydrography program and contrasting them to observations from the 1990s. Using a linear regression–based method, we find a global increase in the anthropogenic CO2 inventory of 34 ± 4 petagrams of carbon (Pg C) between 1994 and 2007. This is equivalent to an average uptake rate of 2.6 ± 0.3 Pg C year−1 and represents 31 ± 4% of the global anthropogenic CO2 emissions over this period. Although this global ocean sink estimate is consistent with the expectation of the ocean uptake having increased in proportion to the rise in atmospheric CO2, substantial regional differences in storage rate are found, likely owing to climate variability–driven changes in ocean circulation.

Using observations from the first global survey of inorganic carbon in the ocean conducted as part of the World Ocean Circulation Experiment (WOCE)–Joint Global Ocean Flux Study (JGOFS) programs during the 1980s and early 1990s, Sabine et al. (1) estimated that the ocean has taken up 118 ± 18 Pg (1 Pg = 1015 g) of anthropogenic carbon, Cant, from the atmosphere from the beginning of the industrial revolution to the mid-1990s. Anthropogenic carbon represents the additional inorganic carbon present in the ocean-atmosphere system as a consequence of human emissions to the atmosphere (2, 3) through the burning of fossil fuel, the production of cement, and land-use change (4, 5). We distinguish this anthropogenic component from the fluxes and storage changes associated with natural CO2, i.e., the quantity of carbon in the atmosphere-ocean system that existed already in preindustrial times (6). Here, using newly developed methods and high-quality ocean observations collected since 2003 on repeat hydrographic cruises (7), we extend the analysis of Sabine et al. (1) to 2007. In particular, we reconstruct the increase in the oceanic storage of Cant between 1994 and 2007 and contrast this to the expected change given the continued increase in atmospheric CO2.

Detecting the change in anthropogenic CO2

We use the data synthesized by the Global Ocean Data Analysis Project version 2 (GLODAPv2) (8) and the recently developed eMLR(C*) method (9) to identify the change in Cant (∆tCant) between the WOCE-JGOFS (nominal 1994) and the Repeat Hydrography–GO-SHIP (nominal 2007) periods (see supplementary materials for details, figs. S3 to S5). This method builds on the extended multiple linear regression (eMLR) approach (10), which was designed to separate ∆tCant from any natural CO2-driven change in dissolved inorganic carbon. The eMLR(C*) method has been extensively tested with synthetic data from a biogeochemical model (9), demonstrating its ability to reconstruct ∆tCant with high accuracy at global and basin scales. These tests also revealed that subbasin scales are less well resolved, especially in regions characterized by high temporal variability. To quantify the uncertainties in the reconstructions, we used the spread from a Monte Carlo analysis and an ensemble of 14 sensitivity studies (see supplementary materials for details).

Global distribution

The global, vertical distribution of ∆tCant between 1994 and 2007 reveals the strong gradients that are characteristic for a passive conservative tracer invading the ocean from the surface (Fig. 1). In the upper 100 m, Cant increased, on average, by 14 μmol kg−1 over this period, close to the expected level given the rise in atmospheric CO2 and the ocean’s buffer capacity (11). Below that, ∆tCant decreases rapidly with depth, reaching half of the surface value at 375 m and ⅒ of it at 1000 m. Half of the global ∆tCant signal is found in the top 400 m, more than 75% above 1000 m, and only about 7% between 2000 and 3000 m.

Fig. 1 Vertical sections of the change in anthropogenic CO2, ∆tCant, between the JGOFS–WOCE era (~1994) and the Repeat Hydrography–GO-SHIP era (~2007).

Shown are the zonal mean sections in each ocean basin organized around the Southern Ocean in the center. The upper 500 m are expanded. Contour intervals are 2 μmol kg−1.

There are strong spatial variations in the vertical penetration of ∆tCant (Fig. 1), reflecting the differences in the efficacy with which the surface signal is transported and mixed down by the large-scale overturning circulation (1, 3, 1214). As transport occurs primarily along sloping neutral density surfaces (15), it is instructive to investigate the distribution of ∆tCant on such iso-surfaces (Fig. 2). The relatively shallow surfaces associated with mode waters (such as the neutral surface 26.60 kg m−3) (16) are ventilated on time scales of a few decades or less, such that the anthropogenic CO2 signal is transported rather effectively from the outcrops at the mid to high latitudes toward the ocean’s interior (Fig. 2A). By contrast, substantial changes in Cant on the deeper (~1000 m in the mid and low latitudes) 27.40 kg m−3 neutral surface are essentially limited to the regions close to the outcrop (Fig. 2B). This reflects the decade- to century-long ventilation ages of the water masses occupying this neutral surface, consisting, in the Southern Hemisphere, primarily of Antarctic Intermediate Water. Despite its limited horizontal reach, the downward transport of the ∆tCant signal along this neutral surface from the Southern Ocean outcrop leads to the second-deepest penetration of ∆tCant anywhere in the ocean (Fig. 1). The deepest penetration is found in the North Atlantic, where the downward spreading of newly formed North Atlantic Deep Water (NADW) causes a ∆tCant with a maximum of 5 μmol kg−1 below 1000 m. The rapid southward spreading of NADW between about 1500 and 2500 m transports this elevated ∆tCant into the South Atlantic. The corresponding process is much less vigorous and deep-reaching in the North Pacific, as there is no deep-water formation there. This leads to a shallow penetration by mode and intermediate waters.

Fig. 2 Distribution of the change in anthropogenic CO2,tCant, between 1994 and 2007 on two selected neutral surfaces.

(A) ∆tCant on the neutral surface 26.60 kg m−3, representing subtropical mode waters and located around 400-m depth in the centers of the subtropical gyres. (B) ∆tCant on the neutral surface 27.40 located at a depth of about 1000 m. In the Southern Hemisphere and in the Indian and Pacific oceans, this neutral surface represents southern-sourced Antarctic Intermediate Water, whereas in the North Atlantic it represents Intermediate Water formed in the north. Stippled areas poleward of the highest concentration of ∆tCant indicate the outcrop areas of these neutral surfaces. Hatched areas indicate regions where no estimate of ∆tCant was possible owing to data limitations.

These regional differences in the vertical penetration cause large spatial variations in the column inventory of ∆tCant, i.e., the change in Cant integrated vertically from the surface down to 3000 m (Fig. 3A). Large parts of the mid-latitude North Atlantic increased their anthropogenic CO2 loads by 16 ± 1 mol C m−2 or more between 1994 and 2007. This corresponds to an average storage rate of 1.2 ± 0.1 mol m−2 year−1; twice the global mean storage rate of 0.65 ± 0.08 mol m−2 year−1. Large changes are also found along a broad zonal band in the mid-latitudes of the Southern Hemisphere, with maximum values above 15 ± 2 mol m−2 throughout much of the South Atlantic, and progressively smaller values going eastward into the Indian and Pacific Oceans. This reflects the deep, but spatially variable, penetration of Cant into the thermoclines of these basins induced by downward transport by mode and intermediate waters (17, 18) (compare Fig. 1), i.e., the upper cell of the meridional overturning circulation in the Southern Hemisphere (19).

Fig. 3 Maps of the column inventories of anthropogenic CO2 in the ocean (0 to 3000 m).

(A) Change in column inventory between 1994 and 2007 based on the eMLR(C*) method. (B) Column inventory for the year 1994 (1) based on the ∆C* method (45). (C) Anomalous change in column inventory estimated from the difference between the change in the vertical column inventory [shown in (A)] and that expected on the basis of the transient steady-state model and the Cant inventory [shown in (B)], namely, ∆tCantanom = ∆tCant − α·Cant(1994), α = 0.28. The column inventories were obtained by vertically integrating the values from the surface down to 3000 m. Hatched areas indicate regions where no estimate of ∆tCant was possible owing to data limitations.

In contrast to the regions with high accumulation, the low column inventory changes in the region south of 60°S stand out. The average change in storage there is less than 4 ± 2 mol m−2, corresponding to a rate of only 0.30 ± 0.15 mol m−2 year−1, i.e., less than half of the global mean. This low storage rate is associated with the lower cell of the meridional overturning circulation (19). First, the upwelling of old waters with low concentrations of Cant (compare Fig. 1) around the Antarctic Polar Front prevents a substantial accumulation of Cant there. Second, there is little downward transport associated with the downward component of the lower cell, likely owing to the physical blocking of the air-sea exchange of CO2 by sea ice in the Antarctic zone and the relatively short residence time of these waters at the surface (20). This contrasts strongly with the upper overturning cell, which takes up a lot of anthropogenic CO2 from the atmosphere but quickly transports it northward into the band of high storage between 50°S and 30°S (14, 17, 21).

Global and regional inventory change

Integrated globally and down to a depth of 3000 m, we estimate a change in the inventory of oceanic anthropogenic CO2 of 31.2 ± 4 Pg C for the period 1994 to 2007 (Table 1 and table S3). To this we add the uptake and storage by regions outside our gridded domain, namely, the Mediterranean Sea and the Arctic Ocean, which are estimated to account for ~1.5 Pg C (2224). We add an additional ~1 Pg C to our estimate to account for the accumulation of Cant below 3000 m, estimated from the fraction of Cant found below that depth in 1994 in the observational estimates (1) [see also (25)] and from the fraction modeled for the 1994 to 2007 period (14). This yields a global oceanic storage increase of anthropogenic CO2 of 34 ± 4 Pg C, which equals a mean annual uptake rate of 2.6 ±0.3 Pg C year−1 over the 1994 to 2007 period.

Table 1 Change in the inventory of anthropogenic CO2 between 1994 and 2007 as estimated on the basis of the eMLR(C*) method.

Shown in italics are the estimated uncertainties based on the sensitivity and Monte Carlo analyses.

View this table:

The individual ocean basins and hemispheres contribute very differently to both the global inventory and its uncertainty (Table 1). Between the hemispheres, the majority (60 ± 11%) of the increase in Cant is found in the Southern Hemisphere (Table 1). Between the basins, the storage in the Pacific (39 ± 4%) and Atlantic (35 ± 4%) has increased by roughly equal amounts, whereas the contribution of the Indian Ocean is much smaller (21 ± 10%). Although this smaller storage in the Indian ocean reflects its fractional area coverage, the areal storage in the Atlantic is nearly twice that in the Pacific, primarily reflecting the differences in the uptake and downward transport in the high northern latitudes of these two basins. The largest uncertainty stems from the Indian Ocean, primarily as a result of the poor data coverage since 2000 (fig. S1B).

Adding the 34 ± 4 Pg C increase between 1994 and 2007 to the 118 ± 19 Pg C estimated for the change between the preindustrial period and 1994 (1) yields a global ocean storage for 2007 of 152 ± 20 Pg C. Extrapolating this estimate linearly to the year 2010 gives an inventory of 160 ± 20 Pg C, which is consistent with the “best” estimate provided by Khatiwala et al. (13) of 155 ± 31 Pg C, obtained by combining models and other constraints. Our estimate confirms also the results of a recent diagnostic model (14), which simulated a cumulative storage of about 155 to 161 Pg C for 2010. Our estimate stands out by its use of inorganic carbon measurements as the foundation to determine anthropogenic CO2 inventories, whereas the other referenced studies employed indirect or model-based methods. Our approach also implicitly includes the effect of a time-changing ocean circulation, whereas the indirect estimates of (14, 26) assume a steady-state ocean circulation.

The eMLR(C*)-based estimates of ∆tCant compare overall well with other regional data-based analyses of the change in anthropogenic CO2 conducted so far, both in terms of the vertical distributions and the column inventories (see supplementary materials). Our data-based storage rates for the period 1994 to 2007 are also in good agreement with those recently inferred from a diagnostic model of the ocean circulation (21) (fig. S6).

Comparison with the JGOFS-WOCE era reconstructions

The column inventory change between 1994 and 2007 (Fig. 3A) is spatially well correlated (r2 = 0.68) with the total inventory of anthropogenic CO2 in 1994 (Fig. 3B) (1). This relationship is consistent with the nearly exponential and multidecadal nature of the anthropogenic perturbation of the global carbon cycle, which results in the establishment of a “transient steady state” (27). For an ocean with invariant circulation and mixing, this transient steady state implies that the change of anthropogenic CO2 over any time period t1 to t2 is linearly related to the amount of anthropogenic CO2 at the initial time, t1, i.e., ∆tCant(t2 – t1) ≈ α·Cant(t1) (28). We estimate the proportionality α from theoretical considerations for the period 1994 through 2007 and obtain a value of α = 0.28 ± 0.02 (see supplementary materials). This value varies little by region, but it varies substantially over time. This theoretically estimated value is statistically indistinguishable from the ratio of the global change in inventory between 1994 and 2007 (34 ± 4 Pg C) and the inventory in 1994 (118 ± 19 Pg C). This implies that, to first order, the global ocean has continued to take up anthropogenic CO2 from the atmosphere at a rate expected from the increase in atmospheric CO2, i.e., there is no indication of a major change in the uptake over the 1994 to 2007 period relative to the long-term mean.

On a regional basis, however, the reconstructed distribution of ∆tCant differs from that inferred from the Cant distribution in 1994 and the assumption of a transient steady state (Fig. 3C). We interpret these differences, i.e., the anomalous accumulation of Cant (∆tCantanom = ∆tCant – α·Cant) to be primarily the result of variations in ocean circulation. But care must be taken when interpreting ∆tCantanom, because the associated uncertainties accumulate the errors from all terms, i.e., ∆tCant, α, and Cant. As shown in the supplementary materials, the most important source of error is the reconstruction of ∆tCant itself. In particular, tests with synthetic data from an ocean biogeochemical model showed that although the eMLR(C*) method generally works well, changes in ocean circulation tend to lead to biases in the reconstructed ∆tCant, which directly affect the inferred ∆tCantanom (9). However, these tests also revealed that the method is still able to recover the most important signals associated with ∆tCantanom, especially in the North Atlantic. The error is largest in the North Pacific, yet across the globe, more than 50% of the modeled variance in ∆tCantanom at basin scales is correctly recovered. Globally, the bias of the recovered ∆tCantanom is essentially zero.

The most prominent anomaly is found in the North Atlantic, where the reconstructed change in inventory over the entire basin is 20% smaller than that predicted by the transient steady state (Fig. 3C). This anomaly, likely robust given that it was well recovered in the tests with synthetic data, characterizes nearly the entire water column (Fig. 4A). The strongest ∆tCantanom signals are found in the subpolar gyre and within the Subpolar Mode Water and in the Intermediate Water, both belonging to the water masses with the highest burden of anthropogenic CO2 (10, 29). This lower-than-expected increase in storage in the North Atlantic during the 1990s and early 2000s has been described previously (2931) and was attributed to the slow-down and reorganization of the North Atlantic overturning circulation at that time, which led to a reduction in the downward transport of Cant (29, 31). This slowdown was probably temporary, as indicated by the more recent rapid increase in Cant in the Irminger Sea (32).

Fig. 4 Zonal mean sections of the anomalous change in Cant, i.e., ∆tCantanom = ∆tCant − α·Cant(1994), α = 0.28.

(A) Zonal mean section in the Atlantic; (B) as in (A), but for the Pacific; and (C), as in (A), but for the Indian Ocean. Selected isolines of zonally averaged neutral density are shown as contour lines in all plots.

The anomalously low accumulation in the North Atlantic between 1994 and 2007 co-occurred with an anomalously high accumulation in the South Atlantic, such that the storage change of the entire Atlantic remained very close to the expected one. Even though we are somewhat less confident about the reconstructed ∆tCantanom in the South Atlantic given our tests with synthetic data (9), we point out that such a shift in storage from the Northern to the Southern Hemisphere was seen in previous ∆tCant reconstructions, albeit based on a single meridional section (30).

A second set of major regions of anomalous change in inventory are the Indian and Pacific sectors of the Southern Ocean, where the changes in storage are about 20% lower than expected (Fig. 3C). The tests with synthetic data showed that changes in these regions should be captured by the eMLR(C*)-based reconstruction, although likely underestimated (9). We have some confidence in the robustness of this signal because it is caused by a negative ∆tCantanom that is confined to the neutral density layers of Antarctic Intermediate Water and found in all ocean basins (Fig. 4). We interpret this signal to be the result of the southward movement and strengthening of the westerly wind belt (33), which has caused large-scale coherent changes in the meridional overturning circulation in the Southern Ocean and the ventilation of the water masses (34, 35). Further, the wind changes have caused increased upwelling over the Southern Ocean and enhanced Ekman transport to the north, leading to shorter residence times of waters at the surface. This could have led to the weaker uptake and consequently weaker downward transport of anthropogenic CO2 into the thermocline by the upper cell of the meridional overturning circulation, leading to the observed negative ∆tCantanom in the Antarctic Intermediate Water.

Additional support of our interpretation that the reconstructed ∆tCantanom are real stems from the overall agreement of our result with those inferred recently from a diagnostic ocean model (21). Although our changes are somewhat larger, the patterns agree well and support the conclusion that the changes primarily reflect changes in ocean circulation.


Our reconstructed changes in the oceanic inventory of anthropogenic CO2 imply a continuing strong role of the ocean in the recent global carbon budget (Table 2). From 1994 to 2007, anthropogenic emissions added 110 ± 8 Pg C to the atmosphere (36), most of which stemmed from the burning of fossil fuels (94 ± 5 Pg C) (5). Of these emissions, 50 ± 1 Pg C (45%) remained in the atmosphere (37). Our uptake estimate of 34 ± 4 Pg C implies that the ocean accounts for the removal from the atmosphere of 31 ± 4% of the total anthropogenic CO2 emissions over this time period (Table 2). This anthropogenic CO2 uptake fraction does not differ from that for the period from preindustrial times up to 1994 (1).

Table 2 Global CO2 budget for both the period 1800 to 1994 and the decadal period from 1994 to 2007.

In comparison to previous budgets (1), we include explicitly also the potential loss of natural CO2 from the ocean as a component of the budget. The potential contribution of changes in the land-to-ocean carbon fluxes through aquatic systems (46) is not considered here.

View this table:

Our ocean uptake estimate for anthropogenic CO2 permits us also to provide a constraint on the net land uptake for the 1994 through 2007 period. This requires us to consider also the potential contribution of a small anomalous (non–steady state) net flux of natural CO2, emerging from processes such as ocean warming and climate variability–driven changes in ocean circulation and biology (38). We obtain a rough estimate of this contribution by using the surface ocean partial pressure of CO2 (pCO2)–based air-sea flux estimate of (39) for the 1994 through 2007 period and subtracting from it the expected anthropogenic uptake flux consistent with (17), while accounting also for the steady-state outgassing of natural CO2. This yields a cumulative anomalous net release of natural CO2 of about 5 ± 3 Pg C. The resulting net (natural and anthropogenic CO2) ocean uptake estimate of 29 ± 5 Pg C for the 1994 through 2007 period implies a net terrestrial biosphere sink over this period of 31 ± 9 Pg, or about 2.4 ± 0.7 Pg C year−1 (Table 2). Our data-based net ocean uptake estimate confirms the model-based 26 ± 4 Pg C ocean sink reported by the Global Carbon Project over the 1994 through 2007 period (36).

Although the ~30% increase in the oceanic burden of anthropogenic CO2 between 1994 and 2007 constitutes a great service for humanity by slowing down the accumulation of CO2 in the atmosphere, this service comes with the cost of increased ocean acidification (40). Our reconstructed increases in Cant imply a deep reach of ocean acidification into the ocean’s interior (41, 42), causing a further shoaling of the ocean’s saturation horizons for biogenic carbonate minerals (43), and a further squeezing of the available habitats for the organisms sensitive to changes in the ocean’s CO2 chemistry (40, 44).

Documenting and quantifying these changes in anthropogenic CO2 and ocean acidification require the continuation of whole-ocean observations of inorganic carbon and of the many ancillary variables needed to produce the internally consistent data products that are so crucial for analyzing long-term changes in ocean carbon storage. In addition, the continuing documentation of the biogeochemical and physical changes in the ocean will be essential for understanding any forthcoming changes in ocean biology.

Supplementary Materials

Supplementary Text

Figs. S1 to S10

Tables S1 to S3

References (4971)

References and Notes

Acknowledgments: We are deeply indebted to the large number of principal investigators, scientists, and technicians who acquired the high-quality oceanographic data that made our study possible. This global synthesis is the outcome of an effort led by the joint IMBER-SOLAS Ocean carbon cycle working group 2. We acknowledge the support from the two parent organizations IMBER and SOLAS. We also thank the anonymous reviewers for their insightful and supportive input. Funding: We are grateful to the many funding agencies in the various countries that financially supported the global ship-based surveys that underpin much of this work. In particular, we acknowledge funding from the U.S. National Science Foundation and from the National Oceanic and Atmospheric Administration (NOAA), and the GO-SHIP program together with the International ocean carbon coordination project (IOCCP), for their efforts to initiate and coordinate the repeat hydrography program. The work of N.G. and D.C. was supported by ETH and the FP7 projects CarboChange (264879) and Geocarbon (283080). S.v.H. also received support from CarboChange (264879). R.W., R.A.F., and B.R.C. acknowledge the support of Oceanic and Atmospheric Research NOAA and the U.S. Department of Commerce, including resources from the NOAA Global Ocean Monitoring and Observations Division (fund reference 100007298). This is contribution no. 4796 from the NOAA Pacific Marine Environmental Laboratory and JISAO contribution 2018-0185. M.I. acknowledges support from the Japan Meteorological Agency and MEXT KAKENHI grant no. 24121003 “NEOPS” and JP16H01594 “OMIX”. S.K.L. acknowledges support from the Research Council of Norway (214513). F.F.P. was supported by Ministerio de Economía y Competitividad through the ARIOS (CTM2016-76146-C3-1-R) project cofunded by the Fondo Europeo de Desarrollo Regional 2014-2020 (FEDER) and EU Horizon 2020 through the AtlantOS project (grant agreement 633211). Author contributions: N.G. led the global synthesis project together with R.W., R.A.F., T.T., M.I., and J.M. N.G. conceived and developed the eMLR(C*) method together with D.C., analyzed the results, and wrote the paper with input from all coauthors. D.C. undertook additional quality controls on the data with input from S.v.H., conducted the estimate of changes in anthropogenic CO2 together with N.G., and supported all analyses. S.v.H. performed additional inversion of the data. A.O. led the team that conducted the secondary quality control of the GO-SHIP and JGOFS–WOC-era cruises. All authors provided expertise at all stages, were involved in the analysis and interpretation of the data, and contributed to the writing of the manuscript. Competing interests: The authors declare that they have no competing interests. Data and materials availability: The inorganic carbon observations from GLODAPv2 underlying this study are available from the GLODAP website: The anthropogenic CO2 estimates reported in this paper can be obtained through NCEI’s Ocean Carbon Data System:

Stay Connected to Science

Navigate This Article