The estimates of global and national fossil CO₂ emissions (EFOS) include the oxidation of fossil fuels through both combustion (e.g. transport, heating) and chemical oxidation (e.g. carbon anode decomposition in aluminium refining) activities, and the decomposition of carbonates in industrial processes (e.g. the production of cement). We also include CO₂ uptake from the cement carbonation process. Several emission sources are not estimated or not fully covered: coverage of emissions from lime production is not global, and decomposition of carbonates in glass and ceramic production are included only for the “Annex 1” countries of the United Nations Framework Convention on Climate Change (UNFCCC) for lack of activity data. These omissions are considered to be minor. Short-cycle carbon emissions – for example from combustion of biomass – are not included here but are accounted for in the CO₂ emissions from land use (see Sect. 2.2).

Our estimates of fossil CO₂ emissions are derived using the standard approach of activity data and emission factors, relying on data collection by many other parties. Our goal is to produce the best estimate of this flux, and we therefore use a prioritization framework to combine data from different sources that have used different methods, while being careful to avoid double counting and undercounting of emissions sources. The CDIAC-FF emissions data set, derived largely from UN energy data, forms the foundation, and we extend emissions to year Y-1 using energy growth rates reported by the BP energy company. We then proceed to replace estimates using data from what we consider to be superior sources, for example Annex 1 countries' official submissions to the UNFCCC. All data points are potentially subject to revision, not just the latest year. For the full details, see Andrew and Peters (2021).

Other estimates of global fossil CO₂ emissions exist, and these are compared by Andrew (2020a). The most common reason for differences in estimates of global fossil CO₂ emissions is a difference in which emissions sources are included in the data sets. Data sets such as those published by the energy company BP, the US Energy Information Administration, and the International Energy Agency's “CO₂ emissions from fuel combustion” are all generally limited to emissions from combustion of fossil fuels. In contrast, data sets such as PRIMAP-hist, CEDS, EDGAR, and GCP's data set aim to include all sources of fossil CO₂ emissions. See Andrew (2020a) for detailed comparisons and discussion.

Cement absorbs CO₂ from the atmosphere over its lifetime, a process known as “cement carbonation”. We estimate this CO₂ sink from 1931 onwards as the average of two studies in the literature (Cao et al., 2020; Guo et al., 2021). Both studies use the same model, developed by Xi et al. (2016), with different parameterizations and input data, with the estimate of Guo and colleagues being a revision of Xi et al. (2016). The trends of the two studies are very similar. Since carbonation is a function of both current and previous cement production, we extend these estimates to 2022 by using the growth rate derived from the smoothed cement emissions (10-year smoothing) fitted to the carbonation data. In the present budget, we always include the cement carbonation carbon sink in the fossil CO₂ emission component (EFOS).

We use the Kaya Identity for a simple decomposition of CO₂ emissions into the key drivers (Raupach et al., 2007). While there are variations (Peters et al., 2017), we focus here on a decomposition of CO₂ emissions into population, GDP per person, energy use per GDP, and CO₂ emissions per energy. Multiplying these individual components together returns the CO₂ emissions. Using the decomposition, it is possible to attribute the change in CO₂ emissions to the change in each of the drivers. This method gives a first-order understanding of what causes CO₂ emissions to change each year.

We provide a projection of global CO₂ emissions in 2022 by combining separate projections for China, USA, EU, India, and for all other countries combined. The methods are different for each of these. For China we combine monthly fossil fuel production data from the National Bureau of Statistics, import and export data from the Customs Administration, and monthly coal consumption estimates from SX Coal (2022), giving us partial data for the growth rates to date of natural gas, petroleum, and cement, and of the consumption itself for raw coal. We then use a regression model to project full-year emissions based on historical observations. For the USA our projection is taken directly from the Energy Information Administration's (EIA) Short-Term Energy Outlook (EIA, 2022), combined with the year-to-date growth rate of cement clinker production. For the EU we use monthly energy data from Eurostat to derive estimates of monthly CO₂ emissions through July, with coal emissions extended through August using a statistical relationship with reported electricity generation from coal and other factors. Given the very high uncertainty in European energy markets in 2022, we forego our usual history-based projection techniques and instead use the year-to-date growth rate as the full-year growth rate for both coal and natural gas. EU emissions from oil are derived using the EIA's projection of oil consumption for Europe. EU cement emissions are based on available year-to-date data from three of the largest producers, Germany, Poland, and Spain. India's projected emissions are derived from estimates through July (August for oil) using the methods of Andrew (2020b) and extrapolated assuming normal seasonal patterns. Emissions for the rest of the world are derived using projected growth in economic production from the IMF (2022) combined with extrapolated changes in emissions intensity of economic production. More details on the EFOS methodology and its 2022 projection can be found in Appendix C1.

The net CO₂ flux from land use, land-use change, and forestry (ELUC, called land-use change emissions in the rest of the text) includes CO₂ fluxes from deforestation, afforestation, logging and forest degradation (including harvest activity), shifting cultivation (cycle of cutting forest for agriculture, then abandoning), and regrowth of forests (following wood harvest or agriculture abandonment). Emissions from peat burning and drainage are added from external data sets, with peat drainage being averaged from three spatially explicit independent data sets (see Appendix C2.1).

Three bookkeeping approaches, updated estimates each of BLUE (Hansis et al., 2015), OSCAR (Gasser et al., 2020), and H&N2017 (Houghton and Nassikas, 2017), were used to quantify gross sources and sinks and the resulting net ELUC. Uncertainty estimates were derived from the dynamic global vegetation models (DGVMs) ensemble for the time period prior to 1960, using for the recent decades an uncertainty range of ±0.7 GtC yr⁻¹, which is a semi-quantitative measure for annual and decadal emissions and reflects our best value judgement that there is at least 68 % chance (±1σ) that the true land-use change emission lies within the given range for the range of processes considered here. This uncertainty range had been increased from 0.5 GtC yr⁻¹ after new bookkeeping models were included that indicated a larger spread than assumed before (Le Quéré et al., 2018a). Projections for 2021 are based on fire activity from tropical deforestation and degradation and emissions from peat fires and drainage.

Our ELUC estimates follow the definition of global carbon cycle models of CO₂ fluxes related to land-use and land management and differ from IPCC definitions adopted in national greenhouse gas (GHG) inventories (NGHGI) for reporting under the UNFCCC, which additionally generally include, through adoption of the IPCC so-called managed land proxy approach, the terrestrial fluxes occurring on land defined by countries as managed. This partly includes fluxes due to environmental change (e.g. atmospheric CO₂ increase), which are part of SLAND in our definition. This causes the global emission estimates to be smaller for NGHGI than for the global carbon budget definition (Grassi et al., 2018). The same is the case for the Food Agriculture Organization (FAO) estimates of carbon fluxes on forest land, which include both anthropogenic and natural sources on managed land (Tubiello et al., 2021). We map the two definitions to each other, to provide a comparison of the anthropogenic carbon budget to the official country reporting to the climate convention.

We project the 2022 land-use emissions for BLUE, the updated H&N2017, and OSCAR, starting from their estimates for 2021 assuming unaltered peat drainage, which has low interannual variability but adjusting the highly variable emissions from peat fires, tropical deforestation, and degradation as estimated using active fire data (MCD14ML; Giglio et al., 2016). More details on the ELUC methodology can be found in Appendix C2.

The rate of growth of the atmospheric CO₂ concentration is provided for years 1959–2021 by the US National Oceanic and Atmospheric Administration Global Monitoring Laboratory (NOAA/GML; Dlugokencky and Tans, 2022), which is updated from Ballantyne et al. (2012) and includes recent revisions to the calibration scale of atmospheric CO₂ measurements (Hall et al., 2021). For the 1959–1979 period, the global growth rate is based on measurements of atmospheric CO₂ concentration averaged from the Mauna Loa and South Pole stations, as observed by the CO₂ Program at Scripps Institution of Oceanography (Keeling et al., 1976). For the 1980–2020 time period, the global growth rate is based on the average of multiple stations selected from the marine boundary layer sites with well-mixed background air (Ballantyne et al., 2012), after fitting a smooth curve through the data for each station as a function of time and averaging by latitude band (Masarie and Tans, 1995). The annual growth rate is estimated by Dlugokencky and Tans (2022) from atmospheric CO₂ concentration by taking the average of the most recent December–January months corrected for the average seasonal cycle and subtracting this same average one year earlier. The growth rate (in units of ppm yr⁻¹) is converted to units of GtC yr⁻¹ by multiplying by a factor of 2.124 GtC ppm⁻¹, assuming instantaneous mixing of CO₂ throughout the atmosphere (Ballantyne et al., 2012; Table 1).

Since 2020, NOAA/GML provides estimates of atmospheric CO₂ concentrations with respect to a new calibration scale, referred to as WMO-CO₂-X2019, in line with the recommendation of the World Meteorological Organization (WMO) Global Atmosphere Watch (GAW) community (Hall et al., 2021). The “X” in the scale name indicates that it is a mole fraction scale, how many micro-moles of CO₂ in a single mole of (dry) air. The word “concentration” only loosely reflects this. The WMO-CO₂-X2019 scale improves upon the earlier WMO-CO₂-X2007 scale by including a broader set of standards, which contain CO₂ in a wider range of concentrations that span the range 250–800 ppm (vs. 250–520 ppm for WMO-CO₂-X2007). In addition, NOAA/GML made two minor corrections to the analytical procedure used to quantify CO₂ concentrations, fixing an error in the second virial coefficient of CO₂ and accounting for loss of a small amount of CO₂ to materials in the manometer during the measurement process. The difference in concentrations measured using WMO-CO₂-X2019 vs. WMO-CO₂-X2007 is ∼ +0.18 ppm at 400 ppm and the observational record of atmospheric CO₂ concentrations have been revised accordingly. The revisions have been applied retrospectively in all cases where the calibrations were performed by NOAA/GML, thus affecting measurements made by members of the WMO-GAW programme and other regionally coordinated programmes (e.g. Integrated Carbon Observing System, ICOS). Changes to the CO₂ concentrations measured across these networks propagate to the global mean CO₂ concentrations. The recalibrated data were first used to estimate GATM in the 2021 edition of the global carbon budget (Friedlingstein et al., 2022a). Friedlingstein et al. (2022a) verified that the change of scales from WMO-CO₂-X2007 to WMO-CO₂-X2019 made a negligible difference to the value of GATM (-0.06 GtC yr⁻¹ during 2010–2019 and -0.01 GtC yr⁻¹ during 1959–2019, well within the uncertainty range reported below).

The uncertainty around the atmospheric growth rate is due to four main factors. First, the long-term reproducibility of reference gas standards (around 0.03 ppm for 1σ from the 1980s; Dlugokencky and Tans, 2022). Second, small unexplained systematic analytical errors that may have a duration of several months to 2 years come and go. They have been simulated by randomizing both the duration and the magnitude (determined from the existing evidence) in a Monte Carlo procedure. Third, the network composition of the marine boundary layer with some sites coming or going, gaps in the time series at each site, and so on (Dlugokencky and Tans, 2022). The latter uncertainty was estimated by NOAA/GML with a Monte Carlo method by constructing 100 “alternative” networks (Masarie and Tans, 1995; NOAA/GML, 2019). The second and third uncertainties, summed in quadrature, add up to 0.085 ppm on average (Dlugokencky and Tans, 2022). Fourth, the uncertainty associated with using the average CO₂ concentration from a surface network to approximate the true atmospheric average CO₂ concentration (mass-weighted, in three dimensions) as needed to assess the total atmospheric CO₂ burden. In reality, CO₂ variations measured at the stations will not exactly track changes in total atmospheric burden, with offsets in magnitude and phasing due to vertical and horizontal mixing. This effect must be very small on decadal and longer timescales, when the atmosphere can be considered well mixed. The CO₂ increase in the stratosphere lags the increase (meaning lower concentrations) that we observe in the marine boundary layer, while the continental boundary layer (where most of the emissions take place) leads the marine boundary layer with higher concentrations. These effects nearly cancel each other. In addition, the growth rate is nearly the same everywhere (Ballantyne et al., 2012). We therefore maintain an uncertainty around the annual growth rate based on the multiple stations dataset ranges between 0.11 and 0.72 GtC yr⁻¹, with a mean of 0.61 GtC yr⁻¹ for 1959–1979 and 0.17 GtC yr⁻¹ for 1980–2020, when a larger set of stations were available as provided by Dlugokencky and Tans (2022). We estimate the uncertainty of the decadal averaged growth rate after 1980 at 0.02 GtC yr⁻¹ based on the calibration and the annual growth rate uncertainty but stretched over a 10-year interval. For years prior to 1980, we estimate the decadal averaged uncertainty to be 0.07 GtC yr⁻¹ based on a factor proportional to the annual uncertainty prior and after 1980 (0.02 × [0.61/0.17] GtC yr⁻¹).

We assign a high confidence to the annual estimates of GATM because they are based on direct measurements from multiple and consistent instruments and stations distributed around the world (Ballantyne et al., 2012; Hall et al., 2021).

To estimate the total carbon accumulated in the atmosphere since 1750 or 1850, we use an atmospheric CO₂ concentration of 278.3 ± 3 ppm or 285.1 ± 3 ppm, respectively (Gulev et al., 2021). For the construction of the cumulative budget shown in Fig. 3, we use the fitted estimates of CO₂ concentration from Joos and Spahni (2008) to estimate the annual atmospheric growth rate using the conversion factors shown in Table 1. The uncertainty of ±3 ppm (converted to ±1σ) is taken directly from the IPCC's AR5 assessment (Ciais et al., 2013). Typical uncertainties in the growth rate in atmospheric CO₂ concentration from ice core data are equivalent to ±0.1–0.15 GtC yr⁻¹ as evaluated from the Law Dome data (Etheridge et al., 1996) for individual 20-year intervals over the period from 1850 to 1960 (Bruno and Joos, 1997).

We provide an assessment of GATM for 2022 based on the monthly calculated global atmospheric CO₂ concentration (GLO) through August (Dlugokencky and Tans, 2022), and bias-adjusted Holt–Winters exponential smoothing with additive seasonality (Chatfield, 1978) to project to January 2023. Additional analysis suggests that the first half of the year (the boreal winter–spring–summer transition) shows more interannual variability than the second half of the year (the boreal summer–autumn–winter transition), so that the exact projection method applied to the second half of the year has a relatively smaller impact on the projection of the full year. Uncertainty is estimated from past variability using the standard deviation of the last 5 years of monthly growth rates.

The reported estimate of the global ocean anthropogenic CO₂ sink SOCEAN is derived as the average of two estimates. The first estimate is derived as the mean over an ensemble of 10 global ocean biogeochemistry models (GOBMs, Tables 4 and A2). The second estimate is obtained as the mean over an ensemble of seven observation-based data products (Tables 4 and A3). An eighth product (Watson et al., 2020) is shown but is not included in the ensemble average as it differs from the other products by adjusting the flux to a cool, salty ocean surface skin (see Appendix C3.1 for a discussion of the Watson product). The GOBMs simulate both the natural and anthropogenic CO₂ cycles in the ocean. They constrain the anthropogenic air–sea CO₂ flux (the dominant component of SOCEAN) by the transport of carbon into the ocean interior, which is also the controlling factor of present-day ocean carbon uptake in the real world. They cover the full globe and all seasons and were recently evaluated against surface ocean carbon observations, suggesting they are suitable to estimate the annual ocean carbon sink (Hauck et al., 2020). The data products are tightly linked to observations of fCO₂ (fugacity of CO₂, which equals pCO₂ corrected for the non-ideal behaviour of the gas; Pfeil et al., 2013), which carry imprints of temporal and spatial variability, but are also sensitive to uncertainties in gas exchange parameterizations and data sparsity. Their asset is the assessment of interannual and spatial variability (Hauck et al., 2020). We use two further diagnostic ocean models to estimate SOCEAN over the industrial era (1781–1958).

The global fCO₂-based flux estimates were adjusted to remove the pre-industrial ocean source of CO₂ to the atmosphere of 0.65 GtC yr⁻¹ from river input to the ocean (Regnier et al., 2022) to satisfy our definition of SOCEAN (Hauck et al., 2020). The river flux adjustment was distributed over the latitudinal bands using the regional distribution of Aumont et al. (2001; north: 0.17 GtC yr⁻¹; tropics: 0.16 GtC yr⁻¹; south: 0.32 GtC yr⁻¹), acknowledging that the boundaries of Aumont et al. (2001; namely 20° S and 20° N) are not consistent with the boundaries otherwise used in the GCB (30° S and 30° N). A recent study based on one ocean biogeochemical model (Lacroix et al., 2020) suggests that more of the riverine outgassing is located in the tropics than in the Southern Ocean, and hence this regional distribution is associated with a major uncertainty. Anthropogenic perturbations of river carbon and nutrient transport to the ocean are not considered (see Sect. 2.7 and Appendix D3).

We derive SOCEAN from GOBMs by using a simulation (sim A) with historical forcing of climate and atmospheric CO₂, accounting for model biases and drift from a control simulation (sim B) with constant atmospheric CO₂ and normal-year climate forcing. A third simulation (sim C) with historical atmospheric CO₂ increase and normal-year climate forcing is used to attribute the ocean sink to CO₂ (sim C minus sim B) and climate (sim A minus sim C) effects. A fourth simulation (sim D; historical climate forcing and constant atmospheric CO₂) is used to compare the change in anthropogenic carbon inventory in the interior ocean (sim A minus sim D) to the observational estimate of Gruber et al. (2019) with the same flux components (steady state and non-steady state anthropogenic carbon flux). Data products are adjusted to represent the full ice-free ocean area by a simple scaling approach when coverage is below 99 %. GOBMs and data products fall within the observational constraints over the 1990s (2.2 ± 0.7 GtC yr⁻¹, Ciais et al., 2013) after applying adjustments.

SOCEAN is calculated as the average of the GOBM ensemble mean and data product ensemble mean from 1990 onwards. Prior to 1990, it is calculated as the GOBM ensemble mean plus half of the offset between GOBMs and data product ensemble means over 1990–2001.

We assign an uncertainty of ± 0.4 GtC yr⁻¹ to the ocean sink based on a combination of random (ensemble standard deviation) and systematic uncertainties (GOBM bias in anthropogenic carbon accumulation, previously reported uncertainties in fCO₂-based data products; see Appendix C3.3). We assess a medium confidence level to the annual ocean CO₂ sink and its uncertainty because it is based on multiple lines of evidence, it is consistent with ocean interior carbon estimates (Gruber et al., 2019, see Sect. 3.5.5) and the interannual variability in the GOBMs, and data-based estimates are largely consistent and can be explained by climate variability. We refrain from assigning a high confidence because of the systematic deviation between the GOBM and data product trends since around 2002. More details on the SOCEAN methodology can be found in Appendix C3.

The ocean CO₂ sink forecast for the year 2022 is based on the annual historical and estimated 2022 atmospheric CO₂ concentration (Dlugokencky and Tans, 2022), the historical and estimated 2022 annual global fossil fuel emissions from this year's carbon budget, and the spring (March, April, May) Oceanic Niño Index (ONI) (NCEP, 2022). Using a non-linear regression approach, i.e. a feed-forward neural network, atmospheric CO₂, ONI, and fossil fuel emissions are used as training data to best match the annual ocean CO₂ sink (i.e. combined SOCEAN estimate from GOBMs and data products) from 1959 through 2021 from this year's carbon budget. Using this relationship, the 2022 SOCEAN can then be estimated from the projected 2021 input data using the non-linear relationship established during the network training. To avoid overfitting, the neural network was trained with a variable number of hidden neurons (varying between 2–5), and 20 % of the randomly selected training data were withheld for independent internal testing. Based on the best output performance (tested using the 20 % withheld input data), the best performing number of neurons was selected. In a second step, we trained the network 10 times using the best number of neurons identified in step 1 and different sets of randomly selected training data. The mean of the 10 training sequences is considered our best forecast, whereas the standard deviation of the 10 ensembles provides a first-order estimate of the forecast uncertainty. This uncertainty is then combined with the SOCEAN uncertainty (0.4 GtC yr⁻¹) to estimate the overall uncertainty of the 2022 projection.

The terrestrial land sink (SLAND) is thought to be due to the combined effects of fertilization by rising atmospheric CO₂ and N inputs on plant growth, as well as the effects of climate change such as the lengthening of the growing season in northern temperate and boreal areas. SLAND does not include land sinks directly resulting from land use and land-use change (e.g. regrowth of vegetation) as these are part of the land-use flux (ELUC), although system boundaries make it difficult to exactly attribute CO₂ fluxes on land between SLAND and ELUC (Erb et al., 2013).

SLAND is estimated from the multi-model mean of 16 DGVMs (Table A1). As described in Appendix C.4, DGVM simulations include all climate variability and CO₂ effects over land. In addition to the carbon cycle represented in all DGVMs, 11 models also account for the nitrogen cycle and hence can include the effect of N inputs on SLAND. The DGVM estimate of SLAND does not include the export of carbon to aquatic systems or its historical perturbation, which is discussed in Appendix D3. See Appendix C4 for DGVM evaluation and uncertainty assessment for SLAND using the International Land Model Benchmarking system (ILAMB; Collier et al., 2018). More details on the SLAND methodology can be found in Appendix C4.

Like for the ocean forecast, the land CO₂ sink (SLAND) forecast is based on the annual historical and estimated 2022 atmospheric CO₂ concentration (Dlugokencky and Tans, 2021), historical and estimated 2022 annual global fossil fuel emissions from this year's carbon budget, and the summer (June, July, August) ONI (NCEP, 2022). All training data are again used to best match SLAND from 1959 through 2021 from this year's carbon budget using a feed-forward neural network. To avoid overfitting, the neural network was trained with a variable number of hidden neurons (varying between 2–15), larger than for SOCEAN prediction due to the stronger land carbon interannual variability. As done for SOCEAN, a pre-training selects the optimal number of hidden neurons based on 20 % withheld input data, and in a second step, an ensemble of 10 forecasts is produced to provide the mean forecast plus uncertainty. This uncertainty is then combined with the SLAND uncertainty for 2021 (0.9 GtC yr⁻¹) to estimate the overall uncertainty of the 2022 projection.

Cumulative fossil CO₂ emissions for 1850–2021 were 465 ± 25 GtC, including the cement carbonation sink (Fig. 3, Table 8, all cumulative numbers are rounded to the nearest 5 GtC).

In this period, 46 % of fossil CO₂ emissions came from coal, 35 % from oil, 15 % from natural gas, 3 % from decomposition of carbonates, and 1 % from flaring.

In 1850, the UK contributed 62 % of global fossil CO₂ emissions. In 1891 the combined cumulative emissions of the current members of the European Union reached and subsequently surpassed the level of the UK. Since 1917, US cumulative emissions have been the largest. Over the entire period 1850–2021, US cumulative emissions amounted to 115 GtC (24 % of world total), the EU's to 80 GtC (17 %), and China's to 70 GtC (14 %).

In addition to the estimates of fossil CO₂ emissions that we provide here (see Sect. 2), there are three additional global data sets with long time series that include all sources of fossil CO₂ emissions: CDIAC-FF (Gilfillan and Marland, 2021), CEDS version v_2021_04_21 (Hoesly et al., 2018; O'Rourke et al., 2021), and PRIMAP-hist version 2.3.1 (Gütschow et al., 2016, 2021), although these data sets are not entirely independent of each other (Andrew, 2020a). CDIAC-FF has the lowest cumulative emissions over 1750–2018 at 437 GtC, GCP has 443 GtC, CEDS 445 GtC, PRIMAP-hist TP 453 GtC, and PRIMAP-hist CR 455 GtC. CDIAC-FF excludes emissions from lime production, while neither CDIAC-FF nor GCP explicitly include emissions from international bunker fuels prior to 1950. CEDS has higher emissions from international shipping in recent years, while PRIMAP-hist has higher fugitive emissions than the other data sets. However, in general these four data sets are in relative agreement as to total historical global emissions of fossil CO₂.

Global fossil CO₂ emissions, EFOS (including the cement carbonation sink), have increased every decade from an average of 3.0 ± 0.2 GtC yr⁻¹ for the decade of the 1960s to an average of 9.6 ± 0.5 GtC yr⁻¹ during 2012–2021 (Table 6, Figs. 2 and 5). The growth rate in these emissions decreased between the 1960s and the 1990s, from 4.3 % per year in the 1960s (1960–1969), 3.2 % per year in the 1970s (1970–1979), 1.6 % per year in the 1980s (1980–1989), and 0.9 % per year in the 1990s (1990–1999). After this period, the growth rate began increasing again in the 2000s at an average growth rate of 3.0 % per year, decreasing to 0.5 % per year for the last decade (2012–2021). China's emissions increased by +1.5 % per year on average over the last 10 years, dominating the global trend, and India's emissions increased by +3.8 % per year, while emissions decreased in EU27 by -1.8 % per year and in the USA by -1.1 % per year. Figure 6 illustrates the spatial distribution of fossil fuel emissions for the 2012–2021 period.

EFOS includes the uptake of CO₂ by cement via carbonation, which has increased with increasing stocks of cement products from an average of 20 MtC yr⁻¹ (0.02 GtC yr⁻¹) in the 1960s to an average of 200 MtC yr⁻¹ (0.2 GtC yr⁻¹) during 2012–2021 (Fig. 5).

Global fossil CO₂ emissions were 5.1 % higher in 2021 than in 2020 because of the global rebound from the worst of the COVID-19 pandemic, with an increase of 0.5 GtC to reach 9.9 ± 0.5 GtC (including the cement carbonation sink) in 2021 (Fig. 5), distributed among coal (41 %), oil (32 %), natural gas (22 %), cement (5 %), and others (1 %). Compared to the previous year, 2021 emissions from coal, oil, and gas increased by 5.7 %, 5.8 %, and 4.8 %, respectively, while emissions from cement increased by 2.1 %. All growth rates presented are adjusted for the leap year unless stated otherwise.

In 2021, the largest absolute contributions to global fossil CO₂ emissions were from China (31 %), the USA (14 %), the EU27 (8 %), and India (7 %). These four regions account for 59 % of global CO₂ emissions, while the rest of the world contributed 41 %, including international aviation and marine bunker fuels (2.8 % of the total). Growth rates for these countries from 2020 to 2021 were 3.5 % (China), 6.2 % (USA), 6.8 % (EU27), and 11.1 % (India), with +4.5 % for the rest of the world. The per capita fossil CO₂ emissions in 2021 were 1.3 tC per person per year for the globe and were 4.0 (USA), 2.2 (China), 1.7 (EU27), and 0.5 (India) tC per person per year for the four highest-emitting countries (Fig. 5).

The post-COVID-19 rebound in emissions of 5.1 % in 2021 is close to the projected increase of 4.8 % published in Friedlingstein et al. (2022a) (Table 7). Of the regions, the projection for the “rest of world” region was least accurate (off by -1.3 %), largely because of poorly projected emissions from international transport (bunker fuels), which were subject to very large changes during this period.

Globally, we estimate that global fossil CO₂ emissions (including cement carbonation) will grow by 1.0 % in 2022 (0.1 % to 1.9 %) to 10.0 GtC (36.6 GtCO₂), exceeding their 2019 emission levels of 9.9 GtC (36.3 GtCO₂). Global increase in 2022 emissions per fuel types are projected to be +1 % (range 0.2 % to 1.8 %) for coal, +2.2 % (range 1.1 % to 3.3 %) for oil, -0.2 % (range -1.1 % to 0.7 %) for natural gas, and -1.6 % (range -3.7 % to -0.5 %) for cement.

For China, projected fossil emissions in 2022 are expected to decline by 0.9 % (range -2.3 % to +0.4 %) compared with 2021 emissions, bringing 2022 emissions for China to around 3.1 GtC yr⁻¹ (11.4 GtCO₂ yr⁻¹). Changes in fuel-specific projections for China are +0.1 % for coal, -2.8 % for oil, -1.1 % for natural gas, and -7.0 % for cement.

For the USA, the Energy Information Administration (EIA) emissions projection for 2022 combined with cement clinker data from USGS gives an increase of 1.5 % (range -1 % to +4 %) compared to 2021, bringing 2022 USA emissions to around 1.4 GtC yr⁻¹ (5.1 GtCO₂ yr⁻¹). This is based on separate projections for coal of -4.6 %, oil of +2 %, natural gas of +4.7 %, and cement of +1.2 %.

For the European Union, our projection for 2022 is for a decline of 0.8 % (range -2.8 % to +1.2 %) over 2021, with 2022 emissions around 0.8 GtC yr⁻¹ (2.8 GtCO₂ yr⁻¹). This is based on separate projections for coal of +6.7 %, oil of +0.9 %, and natural gas of -10.0 %, while cement remains unchanged.

For India, our projection for 2022 is an increase of 6 % (range of 3.9 % to 8 %) over 2021, with 2022 emissions around 0.8 GtC yr⁻¹ (2.9 GtCO₂ yr⁻¹). This is based on separate projections for coal of +5.0 %, oil of +10.0 %, natural gas of -4.0 %, and cement of +10.0 %.

For the rest of the world, the expected growth rate for 2022 is 1.7 % (range 0.1 % to 3.3 %). The fuel-specific projected 2022 growth rates for the rest of the world are: +1.6 % for coal, +3.1 % for oil, -0.1 % for natural gas, +3 % for cement.

Cumulative CO₂ emissions from land-use changes (ELUC) for 1850–2021 were 205 ± 60 GtC (Table 8; Fig. 3; Fig. 14). The cumulative emissions from ELUC show a large spread among individual estimates of 140 GtC (updated H&N2017), 280 GtC (BLUE), and 190 GtC (OSCAR) for the three bookkeeping models and a similar wide estimate of 185 ± 60 GtC for the DGVMs (all cumulative numbers are rounded to the nearest 5 GtC). These estimates are broadly consistent with indirect constraints from vegetation biomass observations, giving a cumulative source of 155 ± 50 GtC over the 1901–2012 period (Li et al., 2017). However, given the large spread, a best estimate is difficult to ascertain.

In contrast to growing fossil emissions, CO₂ emissions from land use, land-use change, and forestry have remained relatively constant over the 1960–1999 period but show a slight decrease of about 0.1 GtC per decade since the 1990s, reaching 1.2 ± 0.7 GtC yr⁻¹ for the 2012–2021 period (Table 6) but with large spread across estimates (Table 5, Fig. 7). Different from the bookkeeping average, the DGVM model average grows slightly larger over the 1970–2021 period and shows no sign of decreasing emissions in the recent decades (Table 5, Fig. 7). This is, however, expected as DGVM-based estimates include the loss of additional sink capacity, which grows with time, while the bookkeeping estimates do not (Appendix D4).

ELUC is a net term of various gross fluxes, which comprise emissions and removals. Gross emissions on average over the 1850–2021 period are 2 (BLUE, OSCAR) to 3 (updated H&N2017) times larger than the net ELUC emissions. Gross emissions show a moderate increase from an average of 3.2 ± 0.9 GtC yr⁻¹ for the decade of the 1960s to an average of 3.8 ± 0.7 GtC yr⁻¹ during 2012–2021 (Fig. 7). Increases in gross removals, from 1.8 ± 0.4 GtC yr⁻¹ for the 1960s to 2.6 ± 0.4 GtC yr⁻¹ for 2012–2021, were slightly larger than the increase in gross emissions. Since the processes behind gross removals, foremost forest regrowth and soil recovery, are all slow, while gross emissions include a large instantaneous component, short-term changes in land-use dynamics, such as a temporary decrease in deforestation, influences gross emissions dynamics more than gross removal dynamics. It is these relative changes to each other that explain the small decrease in net ELUC emissions over the last 2 decades and the last few years. Gross fluxes often differ more across the three bookkeeping estimates than net fluxes, which is expected due to different process representation; in particular, treatment of shifting cultivation, which increases both gross emissions and removals, differs across models.

There is a smaller decrease in net CO₂ emissions from land-use change in the last few years (Fig. 7) than in last year's estimate (Friedlingstein et al., 2021), which places our updated estimates between last year's estimate and the estimate from the GCB2020 (Friedlingstein et al., 2020). This change is principally attributable to changes in ELUC estimates from BLUE and OSCAR, which relate to improvements in the underlying land-use forcing (see Appendix C2.2 for details). These changes address issues identified with last year's land-use forcing (see Friedlingstein et al., 2022a) and remove or attenuate several emission peaks in Brazil and the Democratic Republic of the Congo and lead to higher net emissions in Brazil in the last decades compared to last year's global carbon budget (the emissions averaged over the three bookkeeping models for Brazil for the 2011–2020 period were 168 MtC yr⁻¹ in GCB2021 as compared to 289 MtC yr⁻¹ in GCB2022). A remaining caveat is that global land-use change data for model input does not capture forest degradation, which often occurs on small scale or without forest cover changes easily detectable from remote sensing and poses a growing threat to forest area and carbon stocks that may surpass deforestation effects (e.g. Matricardi et al., 2020; Qin et al., 2021). While independent pan-tropical or global estimates of vegetation cover dynamics or carbon stock changes based on satellite remote sensing have become available in recent years, a direct comparison to our estimates is not possible, most importantly because satellite-based estimates usually do not distinguish between anthropogenic drivers and natural forest cover losses (e.g. from drought or natural wildfires) (Pongratz et al., 2021).

We additionally separate the net ELUC into four component fluxes to gain further insight into the drivers of emissions: deforestation, afforestation, reafforestation, and wood harvest (i.e. all fluxes on forest lands); emissions from organic soils (i.e. peat drainage and peat fires); and fluxes associated with all other transitions (Fig. 7; Sect. C2.1). On average over the 2012–2021 period and over the three bookkeeping estimates, fluxes from deforestation amount to 1.8 ± 0.4 GtC yr⁻¹, and from afforestation, reafforestation, and wood harvest fluxes amount to -0.9 ± 0.3 GtC yr⁻¹ (Table 5). Emissions from organic soils (0.2 ± 0.1 GtC yr⁻¹) and the net flux from other transitions (0.2 ± 0.1 GtC yr⁻¹) are substantially less important globally. Deforestation is thus the main driver of global gross sources. The relatively small deforestation flux (1.8 ± 0.4 GtC yr⁻¹) in comparison to the gross emission estimate above (3.8 ± 0.7 GtC yr⁻¹) is explained by the fact that emissions associated with wood harvesting do not count as deforestation as they do not change the land cover. This split into component fluxes clarifies the potential for emission reduction and carbon dioxide removal: the emissions from deforestation could be halted (largely) without compromising carbon uptake by forests and would contribute to emissions reduction. By contrast, reducing wood harvesting would have limited potential to reduce emissions as it would be associated with less forest regrowth; sinks and sources cannot be decoupled here. Carbon dioxide removal in forests could instead be increased by afforestation and reafforestation.

Overall, the highest land-use emissions occur in the tropical regions of all three continents. The top three emitters (both cumulatively 1959–2021 and on average over 2012–2021) are Brazil (in particular the Amazon Arc of Deforestation), Indonesia, and the Democratic Republic of the Congo, with these three countries contributing 0.7 GtC yr⁻¹ or 58 % of the global total land-use emissions (average over 2012–2021) (Fig. 6b). This is related to massive expansion of cropland, particularly in the last few decades in Latin America, Southeast Asia, and sub-Saharan Africa (Hong et al., 2021), a substantial part of which has been for export of agricultural products (Pendrill et al., 2019). Emission intensity is high in many tropical countries, particularly in Southeast Asia, due to high rates of land conversion in regions of carbon-dense and often still pristine undegraded natural forests (Hong et al., 2021). Emissions are further increased by peat fires in equatorial Asia (GFED4s, van der Werf et al., 2017). Uptake due to land-use change occurs partly due to expanding forest area as a consequence of the forest transition in the 19th and 20th centuries and the subsequent regrowth of forest, particularly in Europe (Fig. 6b) (Mather, 2001; McGrath et al., 2015).

While the mentioned patterns are supported by independent literature and robust, we acknowledge that model spread is substantially larger at regional rather than global levels, as has been shown for bookkeeping models (Bastos et al., 2021) and DGVMs (Obermeier et al., 2021). Assessments for individual regions will be performed as part of REgional Carbon Cycle Assessment and Processes (RECCAP2; Ciais et al., 2022) or already exist for selected regions (e.g. for Europe by Petrescu et al., 2020; for Brazil by Rosan et al., 2021; and for eight selected countries and regions in comparison to inventory data by Schwingshackl et al., 2022).

National GHG inventory data (NGHGI) under the LULUCF sector or data submitted by countries to FAOSTAT differ from the global models' definition of ELUC that we adopt here in that the natural fluxes (SLAND) are counted towards ELUC when they occur on managed land in the NGHGI reporting (Grassi et al., 2018). In order to compare our results to the NGHGI approach, we perform a re-mapping of our ELUC estimates by adding SLAND in managed forest from the DGVM simulations (following Grassi et al., 2021) to the bookkeeping ELUC estimate (see Appendix C2.3). For the 2012–2021 period, we estimate that 1.8 GtC yr⁻¹ of SLAND occurred in managed forests and is then reallocated to ELUC here, as has been done in the NGHGI method. By doing so, our mean estimate of ELUC is reduced from a source of 1.2 GtC to a sink of 0.6 GtC, which is very similar to the NGHGI estimate of a 0.5 GtC sink (Table 9). The re-mapping approach has been shown to also be generally applicable for country-level data (Grassi et al., 2022b; Schwingshackl et al., 2022). Country-level analysis suggests, e.g. that the bookkeeping mean estimates higher deforestation emissions than the national report in Indonesia but estimates less CO₂ removal by afforestation than the national report in China. The fraction of the natural CO₂ sinks that the NGHGI estimates include differs substantially across countries, related to varying proportions of managed vs. all forest areas (Schwingshackl et al., 2022). Comparing ELUC and NGHGI on the basis of the four component fluxes (Grassi et al., 2022b), we find that NGHGI deforestation emissions are reported to be smaller than the bookkeeping estimate (1.1 GtC yr⁻¹ averaged over 2012–2021). A reason for this lies in the fact that country reports do not (fully) capture the carbon flux consequences of shifting cultivation. Conversely, carbon uptake in forests (afforestation, reafforestation, and forestry) is substantially larger than the bookkeeping estimate (1.75 GtC yr⁻¹ averaged over 2012–2021), owing to the inclusion of natural CO₂ fluxes on managed land in the NGHGI. Emissions from organic soils and the net flux from other transitions are similar to the estimates based on the bookkeeping approach and the external peat drainage and burning data sets. Though estimates between NGHGI, FAOSTAT, individual process-based models, and the mapped budget still differ in value and need further analysis, the approach taken here provides a possibility to relate the global models' and NGHGI approach to each other routinely and thus link the anthropogenic carbon budget estimates of land CO₂ fluxes directly to the Global Stocktake as part of UNFCCC Paris Agreement.

The global CO₂ emissions from land-use change are estimated as 1.1 ± 0.7 GtC in 2021, similar to the 2020 estimate. However, confidence in the annual change remains low.

Land-use change and related emissions may have been affected by the COVID-19 pandemic (e.g. Poulter et al., 2021). During the period of the pandemic, environmental protection policies and their implementation may have been weakened in Brazil (Vale et al., 2021). In other countries monitoring capacities and legal enforcement of measures to reduce tropical deforestation have also been reduced due to budget restrictions of environmental agencies or the impairments of ground-based monitoring intended to prevent land grabs and tenure conflicts (Brancalion et al., 2020; Amador-Jiménez et al., 2020). Effects of the pandemic on trends in fire activity or forest cover changes are hard to separate from those of general political developments and environmental changes, and the long-term consequences of disruptions in agricultural and forestry economic activities (e.g. Gruère and Brooks, 2021; Golar et al., 2020; Beckman and Countryman, 2021) remain to be seen. Overall, there is limited evidence so far that COVID-19 was a key driver of changes in LULUCF emissions at a global scale. Impacts vary across countries and deforestation-curbing and enhancing factors may partly compensate each other (Wunder et al., 2021).

In Indonesia, peat fire emissions are very low, potentially related to a relatively wet dry season (GFED4.1s, van der Werf et al., 2017). In South America, the trajectory of tropical deforestation and degradation fires resembles the long-term average; global emissions from tropical deforestation and degradation fires were estimated to be 206 TgC by 14 October 2020. (GFED4.1s, van der Werf et al., 2017). Our preliminary estimate of ELUC for 2022 is substantially lower than the 2012–2021 average, which saw years of anomalously dry conditions in Indonesia and high deforestation fires in South America (Friedlingstein et al., 2022a). Based on the fire emissions until 14 October, we expect ELUC emissions of around 1.1 GtC in 2022. Note that although our extrapolation is based on tropical deforestation and degradation fires, degradation attributable to selective logging, edge effects, or fragmentation will not be captured. Further, deforestation and fires in deforestation zones may become more disconnected, partly due changes in legislation in some regions. For example, Van Wees et al. (2021) found that the contribution from fires to forest loss decreased in the Amazon and in Indonesia over the period of 2003–2018. More recent years, however, saw an uptick in the Amazon again (Tyukavina et al., 2022 with update), and more work is needed to understand fire–deforestation relations.

The fires in Mediterranean Europe in summer 2022 and in the US in spring 2022, though above average for those regions, only contribute a small amount to global emissions. However, they were unrelated to land-use change and are thus not attributed to ELUC but would be part of the natural land sink.

Land-use dynamics may be influenced by the disruption to the global food market associated with the war in Ukraine, but scientific evidence so far is very limited. High food prices, which preceded (but were exacerbated by) the war (Torero and FAO, 2022), are generally linked to higher deforestation (Angelsen and Kaimowitz, 1999), while high prices on agricultural inputs such as fertilizers and fuel, which are also under pressure from embargoes, may impair yields.

Atmospheric CO₂ concentration was approximately 278 ppm in 1750, 300 ppm in the 1910s, 350 ppm in the late 1980s, and 414.71 ± 0.1 ppm in 2021 (Dlugokencky and Tans, 2022); Fig. 1). The mass of carbon in the atmosphere increased by 48 % from 590 GtC in 1750 to 879 GtC in 2021. Current CO₂ concentrations in the atmosphere are unprecedented in the last 2 million years, and the current rate of atmospheric CO₂ increase is at least 10 times faster than at any other time during the last 800 000 years (Canadell et al., 2021).

The growth rate in atmospheric CO₂ level increased from 1.7 ± 0.07 GtC yr⁻¹ in the 1960s to 5.2 ± 0.02 GtC yr⁻¹ during 2012–2022 with important decadal variations (Table 6, Figs. 3 and 4). During the last decade (2012–2021), the growth rate in atmospheric CO₂ concentration continued to increase, albeit with large interannual variability (Fig. 4).

The airborne fraction (AF), defined as the ratio of atmospheric CO₂ growth rate to total anthropogenic emissions, i.e. 2AF=GATM/(EFOS+ELUC), provides a diagnostic of the relative strength of the land and ocean carbon sinks in removing part of the anthropogenic CO₂ perturbation. The evolution of AF over the last 60 years shows no significant trend, remaining at around 44 %, albeit showing a large interannual and decadal variability driven by the year-to-year variability in GATM (Fig. 9). The observed stability of the airborne fraction over the 1960–2020 period indicates that the ocean and land CO₂ sinks have on average been removing about 55 % of the anthropogenic emissions (see Sect. 3.5 and 3.6).

The growth rate in atmospheric CO₂ concentration was 5.2 ± 0.2 GtC (2.46 ± 0.08 ppm) in 2021 (Fig. 4; Dlugokencky and Tans, 2022), slightly above the 2020 growth rate (5.0 GtC) but similar to the 2011–2020 average (5.2 GtC).

The 2022 growth in atmospheric CO₂ concentration (GATM) is projected to be about 5.3 GtC (2.5 ppm) based on global observations until October 2022, bringing the atmospheric CO₂ concentration to an expected level of 417.2 ppm averaged over the year, 51 % over the preindustrial level.

Cumulated since 1850, the ocean sink adds up to 175 ± 35 GtC, with more than two-thirds of this amount (120 GtC) being taken up by the global ocean since 1960. Over the historical period, the ocean sink increased in pace with the exponential anthropogenic emissions increase (Fig. 3b). Since 1850, the ocean has removed 26 % of total anthropogenic emissions.

The ocean CO₂ sink increased from 1.1 ± 0.4 GtC yr⁻¹ in the 1960s to 2.9 ± 0.4 GtC yr⁻¹ during 2012–2021 (Table 6), with interannual variations of the order of a few tenths of a gigatonne of carbon per year (Fig. 10). The ocean-borne fraction (SOCEAN/(EFOS+ELUC) has been remarkably constant at around 25 % on average (Fig. 9). Variations around this mean illustrate decadal variability of the ocean carbon sink. So far there is no indication of a decrease in the ocean-borne fraction from 1960 to 2021. The increase in the ocean sink is primarily driven by the increased atmospheric CO₂ concentration, with the strongest CO₂-induced signal in the North Atlantic Ocean and the Southern Ocean (Fig. 11a). The effect of climate change is much weaker, reducing the ocean sink globally by 0.11 ± 0.09 GtC yr⁻¹ (-4.2 %) during 2012–2021 (nine models simulate a weakening of the ocean sink by climate change with a range of -3.2 to -8.9 %, and only one model simulates a strengthening by 4.8 %), and it does not show clear spatial patterns across the GOBM ensemble (Fig. 11b). This is the combined effect of change and variability in all atmospheric forcing fields, previously attributed to wind and temperature changes in one model (Le Quéré et al., 2010).

The global net air–sea CO₂ flux is a residual of large natural and anthropogenic CO₂ fluxes into and out of the ocean with distinct regional and seasonal variations (Figs. 6 and B1). Natural fluxes dominate on regional scales but largely cancel out when integrated globally (Gruber et al., 2009). Mid-latitudes in all basins and the high-latitude North Atlantic dominate the ocean CO₂ uptake where low temperatures and high wind speeds facilitate CO₂ uptake at the surface (Takahashi et al., 2009). In these regions, formation of mode, intermediate, and deep-water masses transport anthropogenic carbon into the ocean interior, thus allowing for continued CO₂ uptake at the surface. Outgassing of natural CO₂ occurs mostly in the tropics, especially in the equatorial upwelling region, and to a lesser extent in the North Pacific and polar Southern Ocean, mirroring a well-established understanding of regional patterns of air–sea CO₂ exchange (e.g. Takahashi et al., 2009; Gruber et al., 2009). These patterns are also noticeable in the Surface Ocean CO₂ Atlas (SOCAT) data set, where an ocean fCO₂ value above the atmospheric level indicates outgassing (Fig. B1). This map further illustrates the data sparsity in the Indian Ocean and the Southern Hemisphere in general.

Interannual variability of the ocean carbon sink is driven by climate variability with a first-order effect from a stronger ocean sink during large El Niño events (e.g. 1997–1998) (Fig. 10; Rödenbeck et al., 2014; Hauck et al., 2020). The GOBMs show the same patterns of decadal variability as the mean of the fCO₂-based data products, with a stagnation of the ocean sink in the 1990s and a strengthening since the early 2000s (Fig. 10, Le Quéré et al., 2007; Landschützer et al., 2015, 2016; DeVries et al., 2017; Hauck et al., 2020; McKinley et al., 2020). Different explanations have been proposed for this decadal variability, ranging from the ocean's response to changes in atmospheric wind and pressure systems (e.g. Le Quéré et al., 2007; Keppler and Landschützer, 2019), including variations in upper-ocean overturning circulation (DeVries et al., 2017), to the eruption of Mount Pinatubo and its effects on sea surface temperature and slowed atmospheric CO₂ growth rate in the 1990s (McKinley et al., 2020). The main origin of the decadal variability is a matter of debate, with a number of studies initially pointing to the Southern Ocean (see review in Canadell et al., 2021), but contributions from the North Atlantic and North Pacific oceans (Landschützer et al., 2016; DeVries et al., 2019) or a global signal (McKinley et al., 2020) were also proposed.

Although all individual GOBMs and data products fall within the observational constraint, the ensemble means of GOBMs and data products adjusted for the riverine flux diverge over time with a mean offset increasing from 0.28 GtC yr⁻¹ in the 1990s to 0.61 GtC yr⁻¹ in the decade 2012–2021 and reaching 0.79 GtC yr⁻¹ in 2021. The SOCEAN positive trend over time has diverged by a factor of 2 since 2002 (GOBMs: 0.28 ± 0.07 GtC yr⁻¹ per decade; data products: 0.61 ± 0.17 GtC yr⁻¹ per decade; SOCEAN: 0.45 GtC yr⁻¹ per decade) and by a factor of 3 since 2010 (GOBMs: 0.21 ± 0.14 GtC yr⁻¹ per decade; data products: 0.66 ± 0.38 GtC yr⁻¹ per decade; SOCEAN: 0.44 GtC yr⁻¹ per decade). The GOBM estimate is slightly higher (< 0.1 GtC yr⁻¹) than in the previous global carbon budget (Friedlingstein et al., 2022a) because two new models are included (CESM2, MRI) and four models revised their estimates upwards (CESM-ETHZ, CNRM, FESOM2-REcoM, PlankTOM). The data product estimate is higher by about 0.1 GtC yr⁻¹ compared to Friedlingstein et al. (2022a) as a result of an upward correction in three products (Jena-MLS, MPI-SOMFFN, OS-ETHZ-Gracer), the submission of LDEO-HPD (which is above average), the non-availability of the CSIR product, and the small upward correction of the river flux adjustment.

The discrepancy between the two types of estimates stems mostly from a larger Southern Ocean sink in the data products prior to 2001 and from a larger SOCEAN trend in the northern and southern extratropics since then (Fig. 13). Note that the location of the mean offset (but not its trend) depends strongly on the choice of regional river flux adjustment and would occur in the tropics rather than in the Southern Ocean when using the data set of Lacroix et al. (2020) instead of Aumont et al. (2001). Other possible explanations for the discrepancy in the Southern Ocean could be missing winter observations and data sparsity in general (Bushinsky et al., 2019, Gloege et al., 2021) or model biases (as indicated by the large model spread in the Southern Hemisphere, as shown in Fig. 13, and the larger model–data mismatch, as shown in Fig. B2).

In GCB releases until 2021, the ocean sink 1959–1989 was only estimated by GOBMs due to the absence of fCO₂ observations. Now, the first data-based estimates extending back to 1957/58 are becoming available (Jena-MLS, Rödenbeck et al., 2022, LDEO-HPD, Bennington et al., 2022; Gloege et al., 2022). These are based on a multi-linear regression of pCO₂ with environmental predictors (Rödenbeck et al., 2022, included here) or on model–data pCO₂ misfits and their relation to environmental predictors (Bennington et al., 2022). The Jena-MLS estimate falls well within the range of GOBM estimates and has a correlation of 0.98 with SOCEAN (1959–2021 and 1959–1989). It agrees well on the mean SOCEAN estimate since 1977 with a slightly higher amplitude of variability (Fig. 10). Until 1976, Jena-MLS is 0.2–0.3 GtC yr⁻¹ below the central SOCEAN estimate. The agreement, especially on phasing of variability, is impressive, and the discrepancies in the mean flux 1959–1976 could be explained by an overestimated trend of Jena-MLS (Rödenbeck et al., 2022). Bennington et al. (2022) report a larger flux into the pre-1990 ocean than in Jena-MLS.

The reported SOCEAN estimate from GOBMs and data products is 2.1 ± 0.4 GtC yr⁻¹ over the period 1994 to 2007, which is in agreement with the ocean interior estimate of 2.2 ± 0.4 GtC yr⁻¹, which accounts for the climate effect on the natural CO₂ flux of -0.4 ± 0.24 GtC yr⁻¹ (Gruber et al., 2019) to match the definition of SOCEAN used here (Hauck et al., 2020). This comparison depends critically on the estimate of the climate effect on the natural CO₂ flux, which is smaller from the GOBMs (-0.1 GtC yr⁻¹) than in Gruber et al. (2019). Uncertainties in these two estimates would also overlap when using the GOBM estimate of the climate effect on the natural CO₂ flux.

During 2010–2016, the ocean CO₂ sink appears to have intensified in line with the expected increase from atmospheric CO₂ (McKinley et al., 2020). This effect is stronger in the fCO₂-based data products (Fig. 10, ocean sink 2016 minus 2010, GOBMs: +0.42 ± 0.09 GtC yr⁻¹; data products: +0.52 ± 0.22 GtC yr⁻¹). The reduction of -0.09 GtC yr⁻¹ (range: -0.39 to +0.01 GtC yr⁻¹) in the ocean CO₂ sink in 2017 is consistent with the return to normal conditions after the El Niño in 2015/16, which caused an enhanced sink in previous years. After 2017, the GOBM ensemble mean suggests the ocean sink levelling off at about 2.6 GtC yr⁻¹, whereas the data product estimate increases by 0.24 ± 0.17 GtC yr⁻¹ over the same period.

The estimated ocean CO₂ sink was 2.9 ± 0.4 GtC in 2021. This is a decrease of 0.12 GtC compared to 2020, in line with the expected sink weakening from persistent La Niña conditions. GOBM and data product estimates consistently result in a stagnation of SOCEAN (GOBMs: -0.09 ± 0.15 GtC; data products: -0.15 ± 0.24 GtC). Seven models and six data products show a decrease in SOCEAN (GOBMs down to -0.31 GtC, data products down to -0.58 GtC), while three models and two data products show an increase in SOCEAN (GOBMs up to 0.15 GtC, data products up to 0.12 GtC; Fig. 10). The data products have a larger uncertainty at the tails of the reconstructed time series (e.g. Watson et al., 2020). Specifically, the data products' estimate of the last year is regularly adjusted in the following release owing to the tail effect and an incrementally increasing data availability with a 1–5-year lag (Fig. 10 inset).

Using a feed-forward neural network method (see Sect. 2.4) we project an ocean sink of 2.9 GtC for 2022. This is similar to the year 2021 as the La Niña conditions persist in 2022.

The additional simulation D allows us to separate the anthropogenic carbon component (steady state and non-steady state, sim D - sim A) and compare the model flux and dissolved inorganic carbon (DIC) inventory change directly to the interior ocean estimate of Gruber et al. (2019) without further assumptions. The GOBM ensemble average of anthropogenic carbon inventory changes 1994–2007 amounts to 2.2 GtC yr⁻¹ and is thus lower than the 2.6 ± 0.3 GtC yr⁻¹ estimated by Gruber et al. (2019). Only four models with the highest sink estimate fall within the range reported by Gruber et al. (2019). This suggests that the majority of the GOBMs underestimate anthropogenic carbon uptake by 10 %–20 %. Analysis of Earth system models indicate that an underestimation by about 10 % may be due to biases in ocean carbon transport and mixing from the surface mixed layer to the ocean interior (Goris et al., 2018; Terhaar et al., 2021; Bourgeois et al., 2022; Terhaar et al., 2022), biases in the chemical buffer capacity (Revelle factor) of the ocean (Vaittinada Ayar et al., 2022; Terhaar et al., 2022), and partly due to the late starting date of the simulations (mirrored in atmospheric CO₂ chosen for the pre-industrial control simulation, Table A2, Bronselaer et al., 2017; Terhaar et al., 2022). Interestingly, and in contrast to the uncertainties in the surface CO₂ flux, we find the largest mismatch in interior ocean carbon accumulation in the tropics (93 % of the mismatch), with minor contribution from the north (1 %) and the south (6 %). This highlights the role of interior ocean carbon redistribution for those inventories (Khatiwala et al., 2009).

The evaluation of the ocean estimates (Fig. B2) shows a root-mean-squared error (RMSE) from annually detrended data of 0.4 to 2.6 µatm for the seven fCO₂-based data products over the globe, relative to the fCO₂ observations from the SOCAT v2022 data set for the period 1990–2021. The GOBM RMSEs are larger and range from 3.0 to 4.8 µatm. The RMSEs are generally larger at high latitudes compared to the tropics, for both the data products and the GOBMs. The data products have RMSEs of 0.4 to 3.2 µatm in the tropics, 0.8 to 2.8 µatm in the northern extratropics (> 30° N), and 0.8 to 3.6 µatm in the southern extratropics (< 30° S). Note that the data products are based on the SOCAT v2022 database; hence, the SOCAT is not an independent data set for the evaluation of the data products. The GOBM RMSEs are more spread across regions, ranging from 2.5 to 3.9 µatm in the tropics, 3.1 to 6.5 µatm in the north, and 5.4 to 7.9 µatm in the south. The higher RMSEs occur in regions with stronger climate variability, such as the northern and southern high latitudes (poleward of the subtropical gyres). The upper ranges of the model RMSEs have decreased somewhat relative to Friedlingstein et al. (2022a).

Cumulated since 1850, the terrestrial CO₂ sink amounts to 210 ± 45 GtC, 31 % of total anthropogenic emissions. Over the historical period, the sink increased in pace with the exponential anthropogenic emissions increase (Fig. 3b).

The terrestrial CO₂ sink increased from 1.2 ± 0.4 GtC yr⁻¹ in the 1960s to 3.1 ± 0.6 GtC yr⁻¹ during 2012–2021, with important interannual variations of up to 2 GtC yr⁻¹ generally showing a decreased land sink during El Niño events (Fig. 8), responsible for the corresponding enhanced growth rate in atmospheric CO₂ concentration. The larger land CO₂ sink during 2012–2021 compared to the 1960s is reproduced by all the DGVMs in response to the increase in both atmospheric CO₂ and nitrogen deposition and the changes in climate and is consistent with constraints from the other budget terms (Table 5).

Over the period 1960 to present the increase in the global terrestrial CO₂ sink is largely attributed to the CO₂ fertilization effect (Prentice et al., 2001; Piao et al., 2009), directly stimulating plant photosynthesis and increased plant water use in water-limited systems, with a small negative contribution of climate change (Fig. 11). There is a range of evidence to support a positive terrestrial carbon sink in response to increasing atmospheric CO₂, albeit with uncertain magnitude (Walker et al., 2021). As expected from theory, the greatest CO₂ effect is simulated in the tropical forest regions, associated with warm temperatures and long growing seasons (Hickler et al., 2008) (Fig. 11a). However, evidence from tropical intact forest plots indicate an overall decline in the land sink across Amazonia (1985–2011), attributed to enhanced mortality offsetting productivity gains (Brienen et al., 2005, Hubau et al., 2020). During 2012–2021 the land sink is positive in all regions (Fig. 6) with the exception of eastern Brazil, the southwestern US, southeastern Europe, central Asia, northern and southern Africa, and eastern Australia, where the negative effects of climate variability and change (i.e. reduced rainfall) counterbalance CO₂ effects. This is clearly visible in Fig. 11 where the effects of CO₂ (Fig. 11a) and climate (Fig. 11b) as simulated by the DGVMs are isolated. The negative effect of climate is the strongest in most of South America, Central America, the southwestern US, central Europe, western Sahel, southern Africa, Southeast Asia and southern China, and eastern Australia (Fig. 11b). Globally, climate change reduces the land sink by 0.63 ± 0.52 GtC yr⁻¹ or 17 % (2012–2021).

Since 2020 the globe has experienced La Niña conditions, which would be expected to lead to an increased land carbon sink. A clear peak in the global land sink is not evident in SLAND, and we find that a La Niña-driven increase in tropical land sink is offset by a reduced high latitude extratropical land sink, which may be linked to the land response to recent climate extremes. In the past years several regions experienced record-setting fire events. While global burned area has declined over the past decades, mostly due to declining fire activity in savannas (Andela et al., 2017), forest fire emissions are rising and have the potential to counter the negative fire trend in savannas (Zheng et al., 2021). Noteworthy events include the Black Summer event in Australia in 2019–2020 (emissions of roughly 0.2 GtC; van der Velde et al., 2021) and events in Siberia in 2021 where emissions approached 0.4 GtC or 3 times the 1997–2020 average according to GFED4s. While other regions, including the western US and Mediterranean Europe, also experienced intense fire seasons in 2021, their emissions are substantially lower.

Despite these regional negative effects of climate change on SLAND, the efficiency of land to remove anthropogenic CO₂ emissions has remained broadly constant over the last 6 decades, with a land-borne fraction (SLAND/(EFOS+ELUC)) of ∼ 30 % (Fig. 9).

The terrestrial CO₂ sink from the DGVMs ensemble was 3.5 ± 0.9 GtC in 2021, slightly above the decadal average of 3.1 ± 0.6 GtC yr⁻¹ (Fig. 4, Table 6). We note that the DGVM estimate for 2021 is larger than, but within the uncertainty of, the 2.8 ± 0.9 GtC yr⁻¹ estimate from the residual sink from the global budget (EFOS+ELUC-GATM-SOCEAN) (Table 5).

Using a feed-forward neural network method we project a land sink of 3.4 GtC for 2022, very similar to the 2021 estimate. As for the ocean sink, we attribute this to the persistence of La Niña conditions in 2022.

The evaluation of the DGVMs (Fig. B3) shows generally high skill scores across models for runoff and to a lesser extent for vegetation biomass, gross primary production (or productivity; GPP), and ecosystem respiration (Fig. B3, left panel). Skill score was lowest for leaf area index and net ecosystem exchange, with the widest disparity among models for soil carbon. These conclusions are supported by a more comprehensive analysis of DGVM performance in comparison with benchmark data (Seiler et al., 2022). Furthermore, results show how DGVM differences are often of similar magnitude compared with the range across observational data sets.

In the period 2012–2021, the bottom-up view of total global carbon sinks provided by the GCB, SOCEAN for the ocean and SLAND–ELUC for the land (to be comparable to inversions), agrees closely with the top-down global carbon sinks delivered by the atmospheric inversions. Figure 12 shows both total sink estimates of the last decade split by ocean and land (including ELUC), which match the difference between GATM and EFOS to within 0.01–0.12 GtC yr⁻¹ for inverse systems, and to 0.34 GtC yr⁻¹ for the GCB mean. The latter represents the BIM discussed in Sect. 3.8, which by design is minimal for the inverse systems.

The distributions based on the individual models and data products reveal substantial spread but converge near the decadal means quoted in Tables 5 and 6. Sink estimates for SOCEAN and from inverse systems are mostly non-Gaussian, while the ensemble of DGVMs appears more normally distributed, justifying the use of a multi-model mean and standard deviation for their errors in the budget. Noteworthy is that the tails of the distributions provided by the land and ocean bottom-up estimates would not agree with the global constraint provided by the fossil fuel emissions and the observed atmospheric CO₂ growth rate (EFOS-GATM). This illustrates the power of the atmospheric joint constraint from GATM and the global CO₂ observation network it derives from.

The total atmosphere-to-land fluxes (SLAND-ELUC), calculated here as the difference between SLAND from the DGVMs and ELUC from the bookkeeping models, amounts to a 1.9 ± 0.9 GtC yr⁻¹ sink during 2012–2021 (Table 5). Estimates of total atmosphere-to-land fluxes (SLAND-ELUC) from the DGVMs alone (1.5 ± 0.5 GtC yr⁻¹) are consistent with this estimate and also with the global carbon budget constraint (EFOS-GATM-SOCEAN, 1.5 ± 0.6 GtC yr⁻¹ Table 5). For the last decade (2012–2021), the inversions estimate the net atmosphere-to-land uptake to lie within a range of 1.1 to 1.7 GtC yr⁻¹, consistent with the GCB and DGVM estimates of SLAND-ELUC (Fig. 13 top row).

For the 2012–2021 period, the GOBMs (2.6 ± 0.5 GtC yr⁻¹) produce a lower estimate for the ocean sink than the fCO₂-based data products (3.2 ± 0.6 GtC yr⁻¹), which shows up in Fig. 12 as a separate peak in the distribution from the GOBMs (triangle symbols pointing right) and from the fCO₂-based products (triangle symbols pointing left). Atmospheric inversions (2.7 to 3.3 GtC yr⁻¹) also suggest higher ocean uptake in the last decade (Fig. 13 top row). In interpreting these differences, we caution that the riverine transport of carbon taken up on land and outgassing from the ocean is a substantial (0.65 GtC yr⁻¹) and uncertain term that separates the various methods. A recent estimate of decadal ocean uptake from observed O2/N2 ratios (Tohjima et al., 2019) also points towards a larger ocean sink, albeit with large uncertainty (2012–2016: 3.1 ± 1.5 GtC yr⁻¹).

Figure 13 also shows the latitudinal partitioning of the total atmosphere-to-surface fluxes excluding fossil CO₂ emissions (SOCEAN+SLAND-ELUC) according to the multi-model average estimates from GOBMs and ocean fCO₂-based products (SOCEAN) and DGVMs (SLAND-ELUC) and from atmospheric inversions (SOCEAN and SLAND-ELUC).

Despite being one of the most densely observed and studied regions of our globe, annual mean carbon sink estimates in the northern extratropics (north of 30° N) continue to differ. The atmospheric inversions suggest an atmosphere-to-surface sink (SOCEAN+SLAND-ELUC) for 2012–2021 of 2.0 to 3.2 GtC yr⁻¹, which is higher than the process models' estimate of 2.2 ± 0.4 GtC yr⁻¹ (Fig. 13). The GOBMs (1.2 ± 0.2 GtC yr⁻¹), fCO₂-based data products (1.4 ± 0.1 GtC yr⁻¹), and inversion systems (0.9 to 1.4 GtC yr⁻¹) produce consistent estimates of the ocean sink. Thus, the difference mainly arises from the total land flux (SLAND-ELUC) estimate, which is 1.0 ± 0.4 GtC yr⁻¹ in the DGVMs compared to 0.6 to 2.0 GtC yr⁻¹ in the atmospheric inversions (Fig. 13, second row).

Discrepancies in the northern land fluxes conform with persistent issues surrounding the quantification of the drivers of the global net land CO₂ flux (Arneth et al., 2017; Huntzinger et al., 2017; O'Sullivan et al., 2022) and the distribution of atmosphere-to-land fluxes between the tropics and high northern latitudes (Baccini et al., 2017; Schimel et al., 2015; Stephens et al., 2007; Ciais et al., 2019; Gaubert et al., 2019).

In the northern extratropics, the process models, inversions, and fCO₂-based data products consistently suggest that most of the variability stems from the land (Fig. 13). Inversions generally estimate similar interannual variations (IAVs) over land to DGVMs (0.30–0.37 vs. 0.17–0.69 GtC yr⁻¹, averaged over 1990–2021), and they have higher IAV in ocean fluxes (0.05–0.09 GtC yr⁻¹) relative to GOBMs (0.02–0.06 GtC yr⁻¹, Fig. B2) and fCO₂-based data products (0.03–0.09 GtC yr⁻¹).

In the tropics (30° S–30° N), both the atmospheric inversions and process models estimate a total carbon balance (SOCEAN+SLAND-ELUC) that is close to neutral over the past decade. The GOBMs (0.06 ± 0.34 GtC yr⁻¹), fCO₂-based data products (0.00 ± 0.06 GtC yr⁻¹), and inversion systems (-0.2 to 0.5 GtC yr⁻¹) all indicate an approximately neutral tropical ocean flux (see Fig. B1 for spatial patterns). DGVMs indicate a net land sink (SLAND-ELUC) of 0.5 ± 0.3 GtC yr⁻¹, whereas the inversion systems indicate a net land flux between -0.9 and 0.7 GtC yr⁻¹, albeit with high uncertainty (Fig. 13, third row).

The tropical lands are the origin of most of the atmospheric CO₂ interannual variability (Ahlström et al., 2015), and this is consistent among the process models and inversions (Fig. 13). The interannual variability in the tropics is similar among the ocean data products (0.07–0.16 GtC yr⁻¹) and the GOBMs (0.07–0.16 GtC yr⁻¹, Fig. B2), which is the highest ocean sink variability of all regions. The DGVMs and inversions indicate that atmosphere-to-land CO₂ fluxes are more variable than atmosphere-to-ocean CO₂ fluxes in the tropics, with interannual variability of 0.5 to 1.1 and 0.8 to 1.0 GtC yr⁻¹ for DGVMs and inversions, respectively.

In the southern extratropics (south of 30° S), the atmospheric inversions suggest a total atmosphere-to-surface sink (SOCEAN+SLAND-ELUC) for 2012–2021 of 1.6 to 1.9 GtC yr⁻¹, slightly higher than the process models' estimate of 1.4 ± 0.3 GtC yr⁻¹ (Fig. 13). An approximately neutral total land flux (SLAND-ELUC) for the southern extratropics is estimated by both the DGVMs (0.02 ± 0.06 GtC yr⁻¹) and the inversion systems (sink of -0.2 to 0.2 GtC yr⁻¹). This means nearly all carbon uptake is due to oceanic sinks south of 30° S. The Southern Ocean flux in the fCO₂-based data products (1.8 ± 0.1 GtC yr⁻¹) and inversion estimates (1.6 to 1.9 GtC yr⁻¹) is higher than in the GOBMs (1.4 ± 0.3 GtC yr⁻¹) (Fig. 13, bottom row). This discrepancy in the mean flux is likely explained by the uncertainty in the regional distribution of the river flux adjustment (Aumont et al., 2001; Lacroix et al., 2020) applied to fCO₂-based data products and inverse systems to isolate the anthropogenic SOCEAN flux. Other possibly contributing factors are that the data products potentially underestimate the winter CO₂ outgassing south of the Polar Front (Bushinsky et al., 2019) and potential model biases. CO₂ fluxes from this region are more sparsely sampled by all methods, especially in wintertime (Fig. B1). Dominant biases in Earth system models are related to mode water formation, stratification, and the chemical buffer capacity (Terhaar et al., 2021; Bourgeois et al., 2022; Terhaar et al., 2022).

The interannual variability in the southern extratropics is low because of the dominance of ocean areas with low variability compared to land areas. The split between land (SLAND-ELUC) and ocean (SOCEAN) shows a substantial contribution to variability in the south coming from the land, with no consistency between the DGVMs and the inversions or among inversions. This is expected due to the difficulty of exactly separating the land and oceanic fluxes when viewed from atmospheric observations alone. The SOCEAN interannual variability was found to be higher in the fCO₂-based data products (0.09 to 0.19 GtC yr⁻¹) compared to GOBMs (0.03 to 0.06 GtC yr⁻¹) in 1990–2021 (Fig. B2). Model subsampling experiments recently illustrated that observation-based products may overestimate decadal variability in the Southern Ocean carbon sink by 30 % due to data sparsity, based on one data product with the highest decadal variability (Gloege et al., 2021).

A continuing conundrum is the partitioning of the global atmosphere–land flux between the Northern Hemisphere land and the tropical land (Stephens et al., 2017; Pan et al., 2011; Gaubert et al., 2019). It is of importance because each region has its own history of land-use change, climate drivers, and the impact of increasing atmospheric CO₂ and nitrogen deposition. Quantifying the magnitude of each sink is a prerequisite to understanding how each individual driver impacts the tropical and mid- and high-latitude carbon balance.

We define the north–south (N–S) difference as net atmosphere–land flux north of 30° N minus the net atmosphere–land flux south of 30° N. For the inversions, the N–S difference ranges from 0.1 to 2.9 GtC yr⁻¹ across this year's inversion ensemble with a preference across models for either a smaller northern land sink with a near-neutral tropical land flux (medium N–S difference) or a large northern land sink and a tropical land source (large N–S difference).

In the ensemble of DGVMs the N–S difference is 0.6 ± 0.5 GtC yr⁻¹, a much narrower range than the one from inversions. Only two DGVMs have a N–S difference larger than 1.0 GtC yr⁻¹. The larger agreement across DGVMs than across inversions is to be expected as there is no correlation between northern and tropical land sinks in the DGVMs, as opposed to the inversions where the sum of the two regions being well-constrained leads to an anti-correlation between these two regions. The much smaller spread in the N–S difference between the DGVMs could help to scrutinize the inverse systems further. For example, a large northern land sink and a tropical land source in an inversion would suggest a large sensitivity to CO₂ fertilization (the dominant factor driving the land sinks) for northern ecosystems, which would be not mirrored by tropical ecosystems. Such a combination could be hard to reconcile with the process understanding gained from the DGVM ensembles and independent measurements (e.g. free-air CO₂ enrichment experiments). Such investigations will be further pursued in the upcoming assessment from REgional Carbon Cycle Assessment and Processes (RECCAP2; Ciais et al., 2022).

The global carbon budget over the historical period (1850–2021) is shown in Fig. 3.

Emissions during the period 1850–2021 amounted to 670 ± 65 GtC and were partitioned among the atmosphere (275 ± 5 GtC; 41 %), ocean (175 ± 35 GtC; 26 %), and land (210 ± 45 GtC; 31 %). The cumulative land sink is almost equal to the cumulative land-use emissions (200 ± 60 GtC), making the global land nearly neutral over the whole 1850–2021 period.

The use of nearly independent estimates for the individual terms of the global carbon budget shows a cumulative budget imbalance of 15 GtC (2 % of total emissions) during 1850–2021 (Fig. 3, Table 8), which, if correct, suggests that emissions could be slightly too high by the same proportion (2 %) or that the combined land and ocean sinks are slightly underestimated (by about 3 %), although these are well within the uncertainty range of each component of the budget. Nevertheless, part of the imbalance could originate from the estimation of significant increase in EFOS and ELUC between the mid-1920s and the mid-1960s that is unmatched by a similar growth in atmospheric CO₂ concentration as recorded in ice cores (Fig. 3). However, the known loss of additional sink capacity of 30–40 GtC (over the 1850–2020 period) due to reduced forest cover has not been accounted for in our method and would exacerbate the budget imbalance (see Appendix D4).

For the more recent 1960–2021 period where direct atmospheric CO₂ measurements are available, total emissions (EFOS+ELUC) amounted to 470 ± 50 GtC, of which 390 ± 20 GtC (82 %) were caused by fossil CO₂ emissions and 85 ± 45 GtC (18 %) by land-use change (Table 8). The total emissions were partitioned among the atmosphere (210 ± 5 GtC; 45 %), ocean (120 ± 25 GtC; 26 %), and land (145 ± 30 GtC; 30 %), with a near-zero (-5 GtC) unattributed budget imbalance. All components except land-use change emissions have significantly grown since 1960, with important interannual variability in the growth rate in atmospheric CO₂ concentration and in the land CO₂ sink (Fig. 4) and some decadal variability in all terms (Table 6). Differences with previous budget releases are documented in Fig. B5.

The global carbon budget averaged over the last decade (2012–2021) is shown in Figs. 2 and 14 (right panel) and Table 6. For this period, 89 % of the total emissions (EFOS+ELUC) were from fossil CO₂ emissions (EFOS), and 11 % were from land-use change (ELUC). The total emissions were partitioned among the atmosphere (48 %), ocean (26 %), and land (29 %), with a near-zero unattributed budget imbalance (∼ 3 %). For single years, the budget imbalance can be larger (Fig. 4). For 2021, the combination of our estimated sources (10.9 ± 0.9 GtC yr⁻¹) and sinks (11.6 ± 1.0 GtC yr⁻¹) leads to a BIM of -0.6 GtC, suggesting a slight underestimation of the anthropogenic sources and/or an overestimation of the combined land and ocean sinks.

The carbon budget imbalance (BIM; Eq. 1, Fig. 4) quantifies the mismatch between the estimated total emissions and the estimated changes in the atmosphere, land, and ocean reservoirs. The mean budget imbalance from 1960 to 2021 is very small (4.6 GtC over the period, i.e. average of 0.07 GtC yr⁻¹) and shows no trend over the full time series (Fig. 4). The process models (GOBMs and DGVMs) and data products have been selected to match observational constraints in the 1990s, but no further constraints have been applied to their representation of trend and variability. Therefore, the near-zero mean and trend in the budget imbalance is seen as evidence of a coherent community understanding of the emissions and their partitioning on those timescales (Fig. 4). However, the budget imbalance shows substantial variability on the order of ±1 GtC yr⁻¹, particularly over semi-decadal timescales, although most of the variability is within the uncertainty of the estimates. The positive carbon imbalance during the 1960s and early 1990s indicates that either the emissions were overestimated or the sinks were underestimated during these periods. The reverse is true for the 1970s and to a lesser extent for the 1980s and the 2012–2021 period (Fig. 4, Table 6).

We cannot attribute the cause of the variability in the budget imbalance with our analysis, we only note that the budget imbalance is unlikely to be explained by errors or biases in the emissions alone because of its large semi-decadal variability component, a variability that is untypical of emissions and which has not changed in the past 60 years despite a near tripling of emissions (Fig. 4). Errors in SLAND and SOCEAN are more likely to be the main cause for the budget imbalance, especially on interannual to semi-decadal timescales. For example, underestimation of the SLAND by DGVMs has been reported following the eruption of Mount Pinatubo in 1991, possibly due to missing responses to changes in diffuse radiation (Mercado et al., 2009). Although since GCB2021 we accounted for aerosol effects on solar radiation quantity and quality (diffuse vs. direct), most DGVMs only used the former as input (i.e. total solar radiation) (Table A1). Thus, the ensemble mean may not capture the full effects of volcanic eruptions, i.e. associated with high light-scattering sulfate aerosols, on the land carbon sink (O'Sullivan et al., 2021). DGVMs are suspected to overestimate the land sink in response to the wet decade of the 1970s (Sitch et al., 2008). Quasi-decadal variability in the ocean sink has also been reported, with all methods agreeing on a smaller than expected ocean CO₂ sink in the 1990s and a larger than expected sink in the 2000s (Fig. 10; Landschützer et al., 2016; DeVries et al., 2019; Hauck et al., 2020; McKinley et al., 2020). Errors in sink estimates could also be driven by errors in the climatic forcing data, particularly precipitation for SLAND and wind for SOCEAN. Also, the BIM shows substantial departure from zero on yearly timescales (Fig. 4e), highlighting unresolved variability of the carbon cycle, likely in the land sink (SLAND), given its large year-to-year variability (Figs. 4d and 8).

Both the budget imbalance (BIM, Table 6) and the residual land sink from the global budget (EFOS+ELUC-GATM-SOCEAN, Table 5) include an error term due to the inconsistencies that arise from using ELUC from bookkeeping models and SLAND from DGVMs, most notably the loss of additional sink capacity (see Sect. 2.7 and Appendix D4). Other differences include a better accounting of land-use change practices and processes in bookkeeping models than in DGVMs or the error in bookkeeping models of having present-day observed carbon densities fixed in the past. That the budget imbalance shows no clear trend towards larger values over time is an indication that these inconsistencies probably play a minor role compared to other errors in SLAND or SOCEAN.

Although the budget imbalance is near zero for the recent decades, it could be due to compensation of errors. We cannot exclude an overestimation of CO₂ emissions, particularly from land-use change, given their large uncertainty, as has been suggested elsewhere (Piao et al., 2018), combined with an underestimate of the sinks. A larger DGVM (SLAND-ELUC) over the extratropics would reconcile model results with inversion estimates for fluxes in the total land during the past decade (Fig. 13; Table 5). Likewise, a larger SOCEAN is also possible given the higher estimates from the data products (see Sect. 3.1.2, Figs. 10 and 13), the underestimation of interior ocean anthropogenic carbon accumulation in the GOBMs (Sect. 3.5.5), and the recently suggested upward adjustments of the ocean carbon sink in Earth system models (Terhaar et al., 2022) and in data products, here related to a potential temperature bias and skin effects (Watson et al., 2020; Dong et al., 2022, Fig. 10). If SOCEAN were to be based on data products alone, with all data products including this adjustment, this would result in a 2012–2021 SOCEAN of 3.8 GtC yr⁻¹ (Dong et al., 2022) or > 4 GtC yr⁻¹ (Watson et al., 2020), i.e. outside of the range supported by the atmospheric inversions and with an implied negative BIM of more than -1 GtC yr⁻¹, indicating that a closure of the budget could only be achieved with either anthropogenic emissions being significantly larger and/or the net land sink being substantially smaller than estimated here. More integrated use of observations in the global carbon budget, either on their own or for further constraining model results, should help resolve some of the budget imbalance (Peters et al., 2017).