Decadal trends in the ocean carbon sink

Significance

The ocean and land absorb anthropogenic ${\mathrm{C}\mathrm{O}}_2$ from industrial fossil-fuel emissions and land-use changes, helping to buffer climate change. Here, we compare decadal variability of ocean ${\mathrm{C}\mathrm{O}}_2$ uptake using three independent methods and find that the ocean could be responsible for as much as 40% of the observed decadal variability of ${\mathrm{C}\mathrm{O}}_2$ accumulation in the atmosphere. The remaining variability is due to variability in the accumulation of carbon in the terrestrial biosphere. Models capture these variations, but not as strongly as the observations, implying that ${\mathrm{C}\mathrm{O}}_2$ uptake by the land and ocean is more sensitive to climate variability than currently thought. Models must capture this sensitivity to provide accurate climate predictions.

Abstract

Measurements show large decadal variability in the rate of ${\mathrm{C}\mathrm{O}}_2$ accumulation in the atmosphere that is not driven by ${\mathrm{C}\mathrm{O}}_2$ emissions. The decade of the 1990s experienced enhanced carbon accumulation in the atmosphere relative to emissions, while in the 2000s, the atmospheric growth rate slowed, even though emissions grew rapidly. These variations are driven by natural sources and sinks of ${\mathrm{C}\mathrm{O}}_2$ due to the ocean and the terrestrial biosphere. In this study, we compare three independent methods for estimating oceanic ${\mathrm{C}\mathrm{O}}_2$ uptake and find that the ocean carbon sink could be responsible for up to 40% of the observed decadal variability in atmospheric ${\mathrm{C}\mathrm{O}}_2$ accumulation. Data-based estimates of the ocean carbon sink from ${\mathrm{p}\mathrm{C}\mathrm{O}}_2$ mapping methods and decadal ocean inverse models generally agree on the magnitude and sign of decadal variability in the ocean ${\mathrm{C}\mathrm{O}}_2$ sink at both global and regional scales. Simulations with ocean biogeochemical models confirm that climate variability drove the observed decadal trends in ocean ${\mathrm{C}\mathrm{O}}_2$ uptake, but also demonstrate that the sensitivity of ocean ${\mathrm{C}\mathrm{O}}_2$ uptake to climate variability may be too weak in models. Furthermore, all estimates point toward coherent decadal variability in the oceanic and terrestrial ${\mathrm{C}\mathrm{O}}_2$ sinks, and this variability is not well-matched by current global vegetation models. Reconciling these differences will help to constrain the sensitivity of oceanic and terrestrial ${\mathrm{C}\mathrm{O}}_2$ uptake to climate variability and lead to improved climate projections and decadal climate predictions.

Anthropogenic emissions of carbon dioxide (${\mathrm{C}\mathrm{O}}_2$) are a major contributor to climate change, accounting for >80% of the radiative forcing of anthropogenic greenhouse gases over the past several decades (1). There is therefore a pressing need to understand the factors influencing the rate at which anthropogenic ${\mathrm{C}\mathrm{O}}_2$ accumulates in the atmosphere. The primary driver of atmospheric ${\mathrm{C}\mathrm{O}}_2$ accumulation is anthropogenic emissions from industrial activity and deforestation (2), which has increased by ∼60% over the past 30 y (Fig. 1A). ${\mathrm{C}\mathrm{O}}_2$ accumulation in the atmosphere, however, has not always followed the trend in ${\mathrm{C}\mathrm{O}}_2$ emissions. From 1990 to 1999, atmospheric ${\mathrm{C}\mathrm{O}}_2$ accumulated more rapidly than expected from the relatively slow growth in emissions, while in the decade from 2000 to 2009, atmospheric ${\mathrm{C}\mathrm{O}}_2$ accumulation was relatively steady, while emissions rose rapidly (Fig. 1A).

Fig. 1.

(A) Global ${\mathrm{C}\mathrm{O}}_2$ emissions from fossil-fuel-burning, cement production, and land-use change (FF+LUC) (red curve), compared with the measured rate of accumulation of ${\mathrm{C}\mathrm{O}}_2$ in the atmosphere (gold curve) and the inferred rate of change of ${\mathrm{C}\mathrm{O}}_2$ accumulation in the land and ocean (blue curve). Thin lines are annual means, and thick lines are 5-y running means. (B) Decadal trends in ${\mathrm{C}\mathrm{O}}_2$ emissions (FF+LUC) and the atmospheric and total land+ocean sinks. For emissions, positive values indicate an increasing source and negative values a decreasing source (left-hand arrows; sign convention as in Eq. 1). For the atmosphere and land+ocean sinks, positive values indicate a decreasing sink and negative values an increasing sink (right-hand arrows; opposite the sign convention in Eq. 1). All data are from the 2017 Global Carbon Budget (3, 4). Error bars are 1-$\sigma$.

This decadal variability in atmospheric ${\mathrm{C}\mathrm{O}}_2$ accumulation rate is linked to variability in the sources and sinks of ${\mathrm{C}\mathrm{O}}_2$ in the natural environment (5). The most important of these natural sources and sinks are terrestrial ecosystems and ocean waters. Other natural sources and sinks such as volcanoes and rock weathering are much smaller and change very slowly (6) and can be neglected on recent timescales. Thus, the global carbon budget (3) is primarily a balance between anthropogenic ${\mathrm{C}\mathrm{O}}_2$ emissions from fossil-fuel burning and cement manufacturing (FF) and land-use change (LUC; i.e., deforestation), and changes in the accumulation of ${\mathrm{C}\mathrm{O}}_2$ in the atmosphere (${\mathrm{C}}_{atm}$), ocean (${\mathrm{C}}_{oce}$), and land biosphere (${\mathrm{C}}_{land}$),

$$\left(\text{FF}+\text{LUC}\right)-\frac{d{\text{C}}_{atm}}{dt}-\frac{d{\text{C}}_{oce}}{dt}\hspace{0.17em}-\frac{d{\text{C}}_{land}}{dt}=0.$$ [1]

Global FF and LUC emissions have an uncertainty of ∼10% (3, 7, 8), and atmospheric ${\mathrm{C}\mathrm{O}}_2$ has been measured continuously since 1980 at a global network of stations, with error on the annual average accumulation of $<5$% (9). From these observations and Eq. (1), we can infer the accumulation rate of carbon in the combined land and ocean reservoirs (Fig. 1A). The total rate of land+ocean carbon accumulation has averaged 55 $\pm$ 10% of total carbon emissions over the past 30 y, but has shown significant decadal variability. The 1990s experienced a weakening of the land+ocean carbon sink, while the first decade of the 2000s was characterized by a strengthening land+ocean carbon sink (Fig. 1B).

The relative contribution of the land and ocean carbon sinks to this decadal variability cannot be directly measured, due to the heterogeneity of carbon accumulation and large natural carbon reservoirs. For this reason, dynamic global vegetation models (DGVMs) and global ocean biogeochemistry models (GOBMs) are often used to estimate the land and ocean carbon sinks, respectively (3). Methods have also been developed for estimating ${\mathrm{C}\mathrm{O}}_2$ accumulation in the ocean indirectly from observations using inverse models (10–12) and measurements of the sea-surface partial pressure of ${\mathrm{C}\mathrm{O}}_2$ (${\mathrm{p}\mathrm{C}\mathrm{O}}_2$) (13–15).

While the terrestrial biosphere is the dominant source of interannual variability in the natural ${\mathrm{C}\mathrm{O}}_2$ sinks (5, 16), observations and numerical models have highlighted substantial decadal variability in ocean ${\mathrm{C}\mathrm{O}}_2$ uptake at both regional (17–19) and global scales (20, 21). In particular, recent estimates from several data-based models (22–24) suggest that the decadal variability in the ocean ${\mathrm{C}\mathrm{O}}_2$ sink is larger than currently estimated by global carbon budgets. To assess the robustness of decadal trends in ocean ${\mathrm{C}\mathrm{O}}_2$ uptake, here, we compare decadal variability in the ocean carbon sink from three widely used independent methods: GOBMs participating in the 2017 Global Carbon Budget (3), an ocean circulation inverse model (OCIM) (12, 24), and ${\mathrm{p}\mathrm{C}\mathrm{O}}_2$-based flux mapping models from the Surface Ocean ${\mathrm{p}\mathrm{C}\mathrm{O}}_2$ Mapping Intercomparison (SOCOM) project (15). We use these methods to deduce the contribution of the ocean carbon sink to the decadal variability of atmospheric carbon accumulation, to examine the mechanisms governing this variability, and to shed light on the decadal variability of the terrestrial ${\mathrm{C}\mathrm{O}}_2$ sink.

Decadal Variability of the Ocean Carbon Sink

Estimates of the global ocean carbon sink from the GOBMs, SOCOM products, and the OCIM are in broad agreement regarding the magnitude and temporal evolution of ocean carbon accumulation over the past 30 y (Fig. 2A). Estimates of the ocean anthropogenic carbon sink in 2010 from these methods cluster around a mean of $\sim$2.4 GtC$\cdot$y⁻¹ with an uncertainty of $\sim$25% due to differences among the various methods and models (Fig. 2A).

Fig. 2.

(A) Estimates of the ocean carbon sink from a subset of models participating in the SOCOM project (15), a subset of GOBMs participating in the 2017 Global Carbon Budget (3) and an OCIM with (24) and without (12) decadal variability in ocean circulation. Thick lines are the ensemble mean from each method, with shading representing one SD uncertainty. For the OCIM with variable circulation, the mean value at the end of each decade (1989, 1999, and 2009) is shown, with error bars representing one SD. For the OCIM with constant circulation, error bars are the ensemble range. SOCOM results have been adjusted for outgassing of riverine ${\mathrm{C}\mathrm{O}}_2$ (Materials and Methods). (B) Decadal trends in the net (land+ocean) carbon sink (blue bar; same as in Fig. 1) and four estimates of decadal trends in the ocean carbon sink from SOCOM models (red bar), GOBMs (purple bar), and OCIM with decadal variability in ocean circulation (gold bar) and without any variability in ocean circulation (dashed line).

A closer look at the decadal trends in ocean ${\mathrm{C}\mathrm{O}}_2$ uptake reveals that the various methods of estimating the oceanic ${\mathrm{C}\mathrm{O}}_2$ sink differ in the magnitude of their decadal variability (Fig. 2B). The OCIM with steady circulation simulates ${\mathrm{C}\mathrm{O}}_2$ uptake by an ocean with no variability in circulation or biology (12), and therefore the decadal trends are very similar for both the 1990s and the 2000s, with global ocean ${\mathrm{C}\mathrm{O}}_2$ accumulation accelerating at $\sim$0.4 GtC$\cdot$y⁻¹⋅decade⁻¹. All of the other methods display significantly more decadal variability, strongly suggesting decadal trends in ocean circulation and/or biology over this time period (Fig. 2B).

Decadal trends in ocean ${\mathrm{C}\mathrm{O}}_2$ uptake are strongest in the observation-based models. In the 1990s, SOCOM products (15) and the OCIM with decadally varying circulation (24) diagnose a weakening trend of 0.15 $\pm$ 0.43 and 0.28 $\pm$ 0.26 GtC$\cdot$y⁻¹⋅decade⁻¹, respectively, which in turn accounts for 8% (−10% to 83%) and 16% (1–77%) of the observed 1.8 $\pm$ 1.1 GtC$\cdot$y⁻¹⋅decade⁻¹ weakening of the net (land+ocean) carbon sink. In the 2000s, the SOCOM products estimate a strengthening of the ocean carbon sink by 0.80 $\pm$ 0.51 GtC$\cdot$y⁻¹⋅decade⁻¹ that is consistent with the 1.0 $\pm$ 0.2 GtC$\cdot$y⁻¹⋅decade⁻¹ strengthening inferred by the OCIM with variable circulation. These trends account for 35% ($9-109$%) and 43% (24–100%), respectively, of the observed 2.3 $\pm$ 1.1 GtC$\cdot$y⁻¹⋅decade⁻¹ strengthening trend of the total (land+ocean) carbon sink in the 2000s. Based on the average trends in the observation-based models over the 1990s and the first decade of the 2000s, the ocean is responsible for $\sim$10–40% of the observed decadal variability in the natural carbon sinks.

The GOBMs also simulate weaker-than-expected ocean ${\mathrm{C}\mathrm{O}}_2$ uptake during the 1990s followed by a strengthening trend during the 2000s, but the magnitude of decadal variability is smaller than that estimated by SOCOM and the variable-circulation OCIM. For example, in the 2000s, the growth rate of oceanic ${\mathrm{C}\mathrm{O}}_2$ uptake in the GOBMs was slightly less than simulated by the OCIM with constant circulation and biology, while the other methods estimate that oceanic uptake was accelerating roughly twice as fast as it would with constant circulation and biology (Fig. 2B). According to average trends in the GOBMs over the 1990s and the first decade of the 2000s, the ocean is responsible for $\sim$0–20% of the decadal variability in the natural carbon sinks, which is about half of the variability estimated by the observation-based approaches.

Despite the overall agreement among the methods on the sign of the decadal variability in the ocean ${\mathrm{C}\mathrm{O}}_2$ sink, there is substantial spread in the magnitude of the decadal trends both across models within a particular method and across oceanographic regions (Fig. 3). With respect to the global ocean ${\mathrm{C}\mathrm{O}}_2$ uptake, the SOCOM products range from a trend of −0.21 to 1.11 GtC$\cdot$y⁻¹⋅decade⁻¹ in the 1990s to −0.21 to −2.13 GtC$\cdot$y⁻¹⋅decade⁻¹ in the 2000s. Almost all (eight of nine) of the SOCOM products show a more rapidly strengthening ${\mathrm{C}\mathrm{O}}_2$ sink in the 2000s compared with the 1990s. Different GOBMs also exhibit substantially different decadal variability, although all of the GOBMs simulate a strengthening of the ocean ${\mathrm{C}\mathrm{O}}_2$ sink in the 2000s relative to the 1990s (Fig. 3A).

Fig. 3.

Decadal trends in ocean carbon uptake for the global ocean (A) and for different ocean regions (B–F) as defined by the biomes of refs. 25 and 26 (see SI Appendix for biome definitions and definitions of the models used here). (B) Southern Ocean. (C) North Atlantic. (D) North Pacific. (E) Low-latitude Atlantic. (F) Low-latitude Pacific + Indian. The global ocean in A is the sum of the regions in B–F and does not include coastal regions and marginal seas. Trends and color-coding are as in Fig. 2B, with symbols representing individual models. Positive trends represent a weakening oceanic ${\mathrm{C}\mathrm{O}}_2$ sink and negative trends a strengthening oceanic ${\mathrm{C}\mathrm{O}}_2$ sink.

To examine regional patterns of decadal variability in the ocean ${\mathrm{C}\mathrm{O}}_2$ sink, we integrated the air–sea ${\mathrm{C}\mathrm{O}}_2$ fluxes within different regions based on biomes defined by ref. 25 (SI Appendix). The model-average trends across different methods (SOCOM, GOBMs, and OCIM), and in different oceanographic regions, display a remarkable pattern: In every region, every method (on average) predicts that the oceanic ${\mathrm{C}\mathrm{O}}_2$ uptake increased faster in the 2000s than in the 1990s (Fig. 3 B–F). The best agreement at regional scales across methods is found between the SOCOM products and the OCIM with variable circulation. In all regions, these methods infer an oceanic ${\mathrm{C}\mathrm{O}}_2$ sink that strengthened much faster in the 2000s than in the 1990s. In the high latitudes, the SOCOM-based estimates place more of the weakening in the 1990s ${\mathrm{C}\mathrm{O}}_2$ sink in the Southern Ocean, while the OCIM-based estimates suggest that more of the weakening occurred in the North Atlantic and North Pacific (Fig. 3 B–D). In the low latitudes, the SOCOM and OCIM models agree that the Pacific and Indian Oceans were a weakening sink in the 1990s (Fig. 3F), while the OCIM simulates a weaker-trending Atlantic Ocean sink than most of the SOCOM products (Fig. 3E). The strengthening of the ocean ${\mathrm{C}\mathrm{O}}_2$ sink in the 2000s is consistent across regions in both the SOCOM and OCIM models.

Decadal trends in the GOBM-simulated oceanic ${\mathrm{C}\mathrm{O}}_2$ uptake are not as variable as those diagnosed by the SOCOM products or the variable-circulation OCIM. For example, in the Southern Ocean, the observation-based methods infer large decadal variations in the ocean ${\mathrm{C}\mathrm{O}}_2$ sink, but the GOBMs simulate only a slight strengthening trend from the 1990s to the 2000s, with the exception of the NEMO-PISCES (CNRM) model, which simulates a large strengthening (Fig. 3B). The same is true in the low-latitude Pacific and Indian, which has the largest decadal variability next to the Southern Ocean in the observation-based estimates, but displays weak decadal variability in the GOBMs (Fig. 3F).

Climate-Driven Trends in Ocean Carbon Uptake

To separate the impacts of ${\mathrm{C}\mathrm{O}}_2$- and climate-forced variability on ocean ${\mathrm{C}\mathrm{O}}_2$ uptake in the GOBMs, we performed additional model simulations in which the climate forcing was held constant and in which the atmospheric ${\mathrm{C}\mathrm{O}}_2$ concentration was held constant (Materials and Methods). Based on these simulations, we isolated the decadal trends of oceanic ${\mathrm{C}\mathrm{O}}_2$ uptake due to atmospheric ${\mathrm{C}\mathrm{O}}_2$ increase and due to climate variability (Fig. 4). These simulations reveal that trends in ocean ${\mathrm{C}\mathrm{O}}_2$ uptake in the 1990s and 2000s are nearly indistinguishable for the ${\mathrm{C}\mathrm{O}}_2$-only forcing case (both between decades and among models) and that decadal variability in the ${\mathrm{C}\mathrm{O}}_2$ sink is driven exclusively by climate variability. Eight of nine of the GOBMs predict that climate variability drove a weakening of the global ocean ${\mathrm{C}\mathrm{O}}_2$ sink in the 1990s, and five of nine predict that climate variability drove a strengthening trend in the 2000s (Fig. 4A).

Fig. 4.

Decadal trends in ocean carbon uptake simulated by GOBMs for the regions in Fig. 3. (A) Global ocean. (B) Southern Ocean. (C) North Atlantic. (D) North Pacific. (E) Low-latitude Atlantic. (F) Low-latitude Pacific + Indian. Shown separately are the trends due to both ${\mathrm{C}\mathrm{O}}_2$ and climate variability (blue bar; same as purple bar in Fig. 3), trends due to ${\mathrm{C}\mathrm{O}}_2$ variability only (red bar), and trends due to climate variability only (gold bar). Error bars are one SD of the model ensemble mean. Symbols represent results from individual models as defined in Fig. 3.

The regions with the strongest climate-driven decadal variability in the GOBMs are the Southern Ocean (Fig 4B) and the low-latitude Pacific and Indian Oceans (Fig. 4F). Within these regions, however, the different models diverge substantially. In the Southern Ocean, the NEMO-PISCES (CNRM) model displays the largest climate-driven decadal variability, with decreasing ${\mathrm{C}\mathrm{O}}_2$ uptake in the 1990s and increasing ${\mathrm{C}\mathrm{O}}_2$ uptake in the 2000s, consistent with the observation-based estimates. But some models display the opposite trend, such as the CSIRO model, which simulates a weakening Southern Ocean ${\mathrm{C}\mathrm{O}}_2$ sink in the 2000s compared with the 1990s. In the low-latitude Pacific and Indian Oceans, it is the CSIRO model that displays the strongest climate-driven variability, in a direction consistent with the observation-based estimates.

Overall, climate variability drove a weakening of oceanic ${\mathrm{C}\mathrm{O}}_2$ uptake in the 1990s and a strengthening in the 2000s across multiple models and geographic regions. The geographical consistency of these trends suggests that this is a response to a global climatic pattern, likely large-scale changes in wind-driven ocean circulation (24, 27). These trends could be due to modes of internal variability in the climate system (22) or to external forcing [e.g., the eruption of Mount Pinatubo in 1991 (28, 29)], which can alter the states of internal climate modes (30), and thus the global winds. External drivers could be amplified by atmospheric (31) or oceanic (32) teleconnections to enhance decadal variability in ocean circulation.

Although the GOBMs display a consistent response to climate forcing, their climate-driven variability of ocean ${\mathrm{C}\mathrm{O}}_2$ uptake appears to be too weak compared with the data-based methods. Indeed, the GOBMs that perform best compared with the most accurate ${\mathrm{p}\mathrm{C}\mathrm{O}}_2$-based flux reconstructions are also the models that exhibit the largest decadal variability at the regional scale (SI Appendix, Figs. S1 and S2). The weak climate-forced variability of GOBMs might stem from either a weak ocean circulation response to atmospheric forcing or to changes in biologically driven carbon uptake that counteracts circulation-driven ${\mathrm{C}\mathrm{O}}_2$ uptake. To examine the latter possibility, we examined decadal trends in the biologically driven export of carbon below the surface ocean in the climate-forced GOBMs (SI Appendix, Fig. S3). Models with strong decadal variability in biological carbon export generally have weak decadal variability in climate-forced ${\mathrm{C}\mathrm{O}}_2$ uptake, while the opposite is true of models with weak variability in biological carbon export. Thus, the compensation between circulation-driven and biologically driven ${\mathrm{C}\mathrm{O}}_2$ uptake is one factor that reduces the sensitivity of the GOBMs to climate variability. The relative roles of biology and physics for determining decadal variability in ocean ${\mathrm{C}\mathrm{O}}_2$ uptake is poorly known and should be a priority for future study.

Discussion and Conclusions

The agreement among the various methods of determining ocean ${\mathrm{C}\mathrm{O}}_2$ uptake demonstrates a broad consensus in the magnitude of the ocean carbon sink over the past several decades and in the timing of the decadal variability (Fig. 2). This agreement is especially encouraging, considering that the three methods considered here are entirely independent. The observation-based methods (SOCOM and OCIM) predict greater decadal variability of the ocean ${\mathrm{C}\mathrm{O}}_2$ sink than ocean biogeochemistry models and suggest that ∼10–40% of the decadal variability in the natural ${\mathrm{C}\mathrm{O}}_2$ sinks can be attributed to the ocean. Ocean biogeochemistry models simulate less decadal variability of the ocean ${\mathrm{C}\mathrm{O}}_2$ sink, which could partly explain why current global carbon budgets (which rely mainly on GOBMs to estimate the oceanic ${\mathrm{C}\mathrm{O}}_2$ sink) have a declining budget imbalance in the 1990s, followed by an increasing imbalance in the 2000s (3). A muted variability of GOBMs compared with observations has also been observed for oxygen (33), suggesting that it is not unique to the carbon cycle.

These results also have important implications for decadal trends in the other major natural sink of anthropogenic ${\mathrm{C}\mathrm{O}}_2$, the terrestrial biosphere. The decadal trends in the ocean ${\mathrm{C}\mathrm{O}}_2$ sink from the three methods considered here (SOCOM, OCIM, and GOBMs) can be compared with the total land+ocean ${\mathrm{C}\mathrm{O}}_2$ sink (Fig. 1B) to deduce the decadal trends in the terrestrial ${\mathrm{C}\mathrm{O}}_2$ sink (Materials and Methods). The decadal trends in the terrestrial ${\mathrm{C}\mathrm{O}}_2$ sink so calculated demonstrate that the terrestrial biosphere was a decreasing sink of ${\mathrm{C}\mathrm{O}}_2$ in the 1990s and an increasing sink of ${\mathrm{C}\mathrm{O}}_2$ in the first decade of the 2000s (the residual land sink in Fig. 5).

Fig. 5.

Trends in the terrestrial ${\mathrm{C}\mathrm{O}}_2$ sink calculated as a residual from the global carbon budget (Eq. 1) using the estimates of the ocean ${\mathrm{C}\mathrm{O}}_2$ sink from three methods considered here (GOBMs, SOCOM, and OCIM with variable circulation) and from the DGVMs participating in the 2017 Global Carbon Budget (3). See SI Appendix for definitions of DGVMs used here.

These decadal trends are in the same direction as those of the oceanic ${\mathrm{C}\mathrm{O}}_2$ sink, but even larger in magnitude, and can place important constraints on the DGVMs that are used to estimate the terrestrial ${\mathrm{C}\mathrm{O}}_2$ sink in the Global Carbon Budget (3). The DGVMs are in good agreement with the residual land sink regarding the strengthening of the terrestrial ${\mathrm{C}\mathrm{O}}_2$ sink in the 2000s, indicating consistency between the emissions data, the ocean ${\mathrm{C}\mathrm{O}}_2$ sink estimates, and the predictions of DGVMs during this period (Fig. 5). But during the 1990s, the DGVMs show less consistency, with one group of DGVMs simulating a neutral to weakening ${\mathrm{C}\mathrm{O}}_2$ sink (in agreement with the residual land sink) and another group simulating a strengthening ${\mathrm{C}\mathrm{O}}_2$ sink.

Differences between the residual land sink and the DGVM land sink during the 1990s could be due to biases in the ocean ${\mathrm{C}\mathrm{O}}_2$ sink estimates, in the ${\mathrm{C}\mathrm{O}}_2$ emissions, or in the DGVMs. Given the agreement between the three independent estimates of the oceanic ${\mathrm{C}\mathrm{O}}_2$ sink, this is unlikely to be a source of bias. Errors in fossil-fuel ${\mathrm{C}\mathrm{O}}_2$ emissions (34) and LUC emissions (35) could be larger than reported and partly responsible for some of the discrepancy. The remaining discrepancies can be attributed to biases in the DGVMs, and as such could indicate a greater climate sensitivity of the terrestrial ${\mathrm{C}\mathrm{O}}_2$ sink than currently thought. In particular, the model discrepancies in the 1990s trends could partly reflect the different degrees to which the DGVMs are sensitive to the eruption of Mt. Pinatubo in 1991 (36) and the strong El Niño event of 1998 (16).

The findings of this study imply that both oceanic and terrestrial carbon-cycle models underestimate decadal variability in ${\mathrm{C}\mathrm{O}}_2$ uptake, which hinders the ability of these models to predict climate change on decadal timescales and likely contributes to decadal imbalances in current global carbon budgets (37). As the community moves toward decadal climate prediction (38, 39), it will be important to correctly resolve the climate sensitivity of oceanic and terrestrial carbon uptake. Continued development of observation-based methods for tracking ocean ${\mathrm{C}\mathrm{O}}_2$ uptake should alleviate their remaining structural errors (SI Appendix), leading to improved constraints on the magnitude and variability of the ocean ${\mathrm{C}\mathrm{O}}_2$ sink and reducing imbalances in global carbon budgets (37). This in turn will facilitate calibration of ocean biogeochemical models and terrestrial dynamic vegetation models, leading to improved climate projections and decadal predictions.

Materials and Methods

${\mathrm{p}\mathrm{C}\mathrm{O}}_2$-Based Flux Mapping Products.

The SOCOM products are based on historical observations of surface-ocean ${\mathrm{p}\mathrm{C}\mathrm{O}}_2$ compiled in the Surface Ocean ${\mathrm{C}\mathrm{O}}_2$ Atlas (SOCAT) (40) and the Lamont–Doherty Earth Observatory (41) datasets. The SOCOM models use various interpolation schemes to fill in the gaps in the data records to create continuous maps of ${\mathrm{p}\mathrm{C}\mathrm{O}}_2$ at monthly resolution, from which air–sea fluxes are calculated (15). See SI Appendix for additional information.

Inverse Models.

We use two versions of the OCIM. The first diagnoses the uptake of anthropogenic ${\mathrm{C}\mathrm{O}}_2$ in the absence of any changes to ocean circulation, solubility, or biology (12). Uncertainties are derived from the 10 different versions of the model described in ref. 12. The second version of the OCIM diagnoses the decadal-mean ocean ${\mathrm{C}\mathrm{O}}_2$ sink given decadal variations in ocean circulation along with mean state biology (24). Uncertainties are derived from 160 different versions of the model described in ref. 24. See SI Appendix for additional information.

GOBMs.

We use a subset of the GOBMs used in the 2017 Global Carbon Budget (3): NEMO-PISCES (CNRM), CSIRO, NorESM, MPIOM-HAMOCC, NEMO-PlankTOM5, MITgcm-REcoM2, and CCSM-BEC. Each model performs three simulations: Simulation A uses reanalysis climate forcing and observed atmospheric ${\mathrm{C}\mathrm{O}}_2$ concentrations from 1959–2017. Simulation B uses constant climate forcing and atmospheric ${\mathrm{C}\mathrm{O}}_2$. Simulation C uses constant climate forcing and observed atmospheric ${\mathrm{C}\mathrm{O}}_2$ concentrations from 1959–2017. In Fig. 4, ${\mathrm{“}\mathrm{C}\mathrm{O}}_2$+climate” is from simulation A, ${\mathrm{“}\mathrm{C}\mathrm{O}}_2$ only” is from simulation C–simulation B, and “climate only” is from simulation A–simulation C. Models differ in their spin-up procedure and climate forcing, as detailed in SI Appendix and SI Appendix, Table S1.

Accounting for Riverine Carbon.

The OCIM and GOBMs do not account for a degassing of 0.45–0.78 GtC$\cdot$y⁻¹ (42, 43) of riverine ${\mathrm{C}\mathrm{O}}_2$, but the SOCOM products do. To make the ${\mathrm{C}\mathrm{O}}_2$ fluxes comparable across all methods, we add a flux of 0.6 GtC$\cdot$y⁻¹ to the globally integrated SOCOM CO₂ sink in Fig. 2.

Calculating Decadal Trends.

Air–sea ${\mathrm{C}\mathrm{O}}_2$ fluxes from the SOCOM products, the GOBMs, and the steady-circulation OCIM are annually averaged, then used to compute the linear trend in ocean ${\mathrm{C}\mathrm{O}}_2$ uptake for the 1990s (1990–1999) and the first decade of the 2000s (2000–2009). Uncertainties on the decadal trends for each method include ensemble uncertainty, as well as an uncertainty of $\pm$1 y for the beginning and ending years of the trend calculations (i.e., 1990 $\pm$1 to 1999 $\pm$1 and 2000 $\pm$1 to 2009 $\pm$1). For the variable-circulation OCIM, decadal trends are calculated as the average air–sea flux within a given decade minus the average air–sea flux in the preceding decade. This method minimizes the effects of discontinuities in the air–sea ${\mathrm{C}\mathrm{O}}_2$ flux introduced by abrupt changes in the ocean circulation at the demarcations of different decades (1990 and 2000) and gives trends similar to those using the final year of each decade (i.e., 2009–1999) to calculate trends. For regional decadal trends in Figs. 3 and 4, we integrate the air–sea ${\mathrm{C}\mathrm{O}}_2$ fluxes over distinct oceanographic regions based on the time-mean open-ocean biomes defined by ref. 25. To avoid differences in the model domains near the coast, the global ocean ${\mathrm{C}\mathrm{O}}_2$ uptake in all figures is the summation over all of the individual open-ocean regions and thus ignores a small contribution from coastal regions as well as the polar ice-covered regions. See SI Appendix for more information.

Calculation of Decadal Trends in the Terrestrial ${\mathrm{C}\mathrm{O}}_2$ Sink.

To calculate decadal trends in the terrestrial ${\mathrm{C}\mathrm{O}}_2$ sink, we first calculate decadal trends in the ocean carbon sink using all of the methods considered here that resolve decadal variability in the ocean ${\mathrm{C}\mathrm{O}}_2$ sink (SOCOM, GOBMs, and OCIM-variable, as displayed in Fig. 2B). We then subtract these ocean-only trends from the trend in the total (land+ocean) ${\mathrm{C}\mathrm{O}}_2$ sink (Fig. 1B) to obtain the trends in the “residual land sink” (Fig. 5). Reported uncertainties include uncertainty in the ${\mathrm{C}\mathrm{O}}_2$ emissions, uncertainty in the atmospheric ${\mathrm{C}\mathrm{O}}_2$ concentration, uncertainty in the ocean ${\mathrm{C}\mathrm{O}}_2$ sink (treating all methods of estimating the ocean ${\mathrm{C}\mathrm{O}}_2$ sink as equally probable), and uncertainty due to varying the beginning and ending years for the trend calculation by $\pm$1 y. Trends in the terrestrial ${\mathrm{C}\mathrm{O}}_2$ sink in the DGVMs are calculated in exactly the same way as those for the GOBMs, varying the starting and ending points of the trend calculation for each DGVM by $\pm$1 y. See SI Appendix for a full list of the DGVMs used here.

Data Availability.

OCIM data are available at https://tdevries.eri.ucsb.edu/models-and-data-products/. Timeseries of the SOCOM data following ref. 15 can be obtained from http://www.bgc-jena.mpg.de/SOCOM/. Timeseries of the GOBM data are available at https://doi.org/10.6084/m9.figshare.8091161.