The calculation of global and national CO2 emissions from fossil fuels, including gas flaring and cement production (EFF), relies primarily on energy consumption data, specifically data on hydrocarbon fuels, collated and archived by several organisations (Andres et al., 2012). These include the Carbon Dioxide Information Analysis Center (CDIAC), the International Energy Agency (IEA), the United Nations (UN), the United States Department of Energy (DoE) Energy Information Administration (EIA), and more recently also the Planbureau voor de Leefomgeving (PBL) Netherlands Environmental Assessment Agency. Where available, we use national emissions estimated by the countries themselves and reported to the UNFCCC for the period 1990–2014 (40 countries). We assume that national emissions reported to the UNFCCC are the most accurate because national experts have access to additional and country-specific information, and because these emission estimates are periodically audited for each country through an established international methodology overseen by the UNFCCC. We also use global and national emissions estimated by CDIAC (Boden and Andres, 2016). The CDIAC emission estimates are the only data set that extends back in time to 1751 with consistent and well-documented emissions from fossil fuels, cement production, and gas flaring for all countries and their uncertainty (Andres et al., 2014, 2012, 1999); this makes the data set a unique resource for research of the carbon cycle during the fossil fuel era.

The global emissions presented here are based on CDIAC's analysis, which provides an internally consistent global estimate including bunker fuels, minimising the effects of lower-quality energy trade data. Thus, the comparison of global emissions with previous annual carbon budgets is not influenced by the use of national data from UNFCCC reports.

During the period 1959–2013, the emissions from fossil fuels estimated by CDIAC are based primarily on energy data provided by the UN Statistics Division (UN, 2015a, b; Table 4). When necessary, fuel masses/volumes are converted to fuel energy content using coefficients provided by the UN and then to CO2 emissions using conversion factors that take into account the relationship between carbon content and energy (heat) content of the different fuel types (coal, oil, gas, gas flaring) and the combustion efficiency (to account, for example, for soot left in the combustor or fuel otherwise lost or discharged without oxidation). Most data on energy consumption and fuel quality (carbon content and heat content) are available at the country level (UN, 2015a). In general, CO2 emissions for equivalent primary energy consumption are about 30 % higher for coal compared to oil, and 70 % higher for coal compared to natural gas (Marland et al., 2007).

Recent revisions in energy data for China (Korsbakken et al., 2016) have not yet fully propagated to the UN energy statistics used by CDIAC but are available through the BP energy statistics (BP, 2016b). We thus use the BP energy statistics (BP, 2016b) and estimate the emissions by fuel type using the BP methodology (BP, 2016a) to be consistent with the format of the CDIAC data. Emissions in China calculated from the BP statistics differ from those provided by CDIAC emissions mostly between 1997 and 2009. The revised emissions are higher by 5 % on average between 1990 and 2015 for a total additional emissions of 2.0 GtC during that period (41.3 GtC using the BP statistics and methodology compared to 39.3 provided by CDIAC). The two estimates converge to similar values from 2011 onwards (< 2 % difference). We propagate these new estimates for China through to the global total to ensure consistency.

Our emission totals for the UNFCCC-reporting countries were recorded as in the UNFCCC submissions, which have a slightly larger system boundary than CDIAC. Additional emissions come from carbonates other than in cement manufacture, and thus UNFCCC totals will be slightly higher than CDIAC totals in general, although there are multiple sources of differences. We use the CDIAC method to report emissions by fuel type (e.g. all coal oxidation is reported under “coal”, regardless of whether oxidation results from combustion as an energy source), which differs slightly from UNFCCC.

For the most recent 1–2 years when the UNFCCC estimates (1 year) and UN statistics (2 years) used by CDIAC are not yet available, we generated preliminary estimates based on the BP annual energy review by applying the growth rates of energy consumption (coal, oil, gas) for 2015 to the national and global emissions from the UN national data in 2014, and for 2014 and 2015 to the CDIAC national and global emissions in 2013. BP's sources for energy statistics overlap with those of the UN data but are compiled more rapidly from about 70 countries covering about 96 % of global emissions. We use the BP values only for the year-to-year rate of change, because the rates of change are less uncertain than the absolute values and to avoid discontinuities in the time series when linking the UN-based data with the BP data. These preliminary estimates are replaced by the more complete UNFCCC or CDIAC data based on UN statistics when they become available. Past experience and work by others (Andres et al., 2014; Myhre et al., 2009) show that projections based on the BP rate of change are within the uncertainty provided (see Sect. 3.2 and Supplement from Peters et al., 2013).

Estimates of emissions from cement production by CDIAC are based on data on growth rates of cement production from the US Geological Survey up to year 2013 (USGS, 2016a). For 2014 and 2015 we use estimates of cement production made by the USGS for the top 18 countries (representing 85 % of global production; USGS, 2016b), while for all other countries we use the 2013 values (zero growth). Some fraction of the CaO and MgO in cement is returned to the carbonate form during cement weathering, but this is neglected here.

Estimates of emissions from gas flaring by CDIAC are calculated in a similar manner to those from solid, liquid, and gaseous fuels and rely on the UN energy statistics to supply the amount of flared or vented fuel. For the most recent 1–2 emission years, flaring is assumed constant from the most recent available year of data (2014 for countries that report to the UNFCCC, and 2013 for the remainder). The basic data on gas flaring report atmospheric losses during petroleum production and processing that have large uncertainty and do not distinguish between gas that is flared as CO2 or vented as CH4. Fugitive emissions of CH4 from the so-called upstream sector (e.g. coal mining and natural gas distribution) are not included in the accounts of CO2 emissions except to the extent that they are captured in the UN energy data and counted as gas “flared or lost”.

The published CDIAC data set includes 255 countries and regions. This list includes countries that no longer exist, such as the USSR and East Pakistan. For the carbon budget, we reduce the list to 219 countries by reallocating emissions to the currently defined territories. This involved both aggregation and disaggregation, and does not change global emissions. Examples of aggregation include merging East and West Germany to the currently defined Germany. Examples of disaggregation include reallocating the emissions from the former USSR to the resulting independent countries. For disaggregation, we use the emission shares when the current territories first appeared. The disaggregated estimates should be treated with care when examining countries' emissions trends prior to their disaggregation. For the most recent years, 2014 and 2015, the BP statistics are more aggregated, but we retain the detail of CDIAC by applying the growth rates of each aggregated region in the BP data set to its constituent individual countries in CDIAC.

Estimates of CO2 emissions show that the global total of emissions is not equal to the sum of emissions from all countries. This is largely attributable to emissions that occur in international territory, in particular the combustion of fuels used in international shipping and aviation (bunker fuels), where the emissions are included in the global totals but are not attributed to individual countries. In practice, the emissions from international bunker fuels are calculated based on where the fuels were loaded, but they are not included with national emissions estimates. Other differences occur because globally the sum of imports in all countries is not equal to the sum of exports and because of inconsistent national reporting, differing treatment of oxidation of non-fuel uses of hydrocarbons (e.g. as solvents, lubricants, feedstocks), and changes in stock (Andres et al., 2012).

The uncertainty in the annual emissions from fossil fuels and industry for the globe has been estimated at ±5 % (scaled down from the published ±10 % at ±2σ to the use of ±1σ bounds reported here; Andres et al., 2012). This is consistent with a more detailed recent analysis of uncertainty of ±8.4 % at ±2σ (Andres et al., 2014) and at the high end of the range of ±5–10 % at ±2σ reported by Ballantyne et al. (2015). This includes an assessment of uncertainties in the amounts of fuel consumed, the carbon and heat contents of fuels, and the combustion efficiency. While we consider a fixed uncertainty of ±5 % for all years, in reality the uncertainty, as a percentage of the emissions, is growing with time because of the larger share of global emissions from non-Annex B countries (emerging economies and developing countries) with less precise statistical systems (Marland et al., 2009). For example, the uncertainty in Chinese emissions has been estimated at around ±10 % (for ±1σ; Gregg et al., 2008), and important potential biases have been identified suggesting China's emissions could be overestimated in published studies (Liu et al., 2015). Generally, emissions from mature economies with good statistical bases have an uncertainty of only a few percent (Marland, 2008). Further research is needed before we can quantify the time evolution of the uncertainty, as well as its temporal error correlation structure. We note that even if they are presented as 1σ estimates, uncertainties in emissions are likely to be mainly country-specific systematic errors related to underlying biases of energy statistics and to the accounting method used by each country. We assign a medium confidence to the results presented here because they are based on indirect estimates of emissions using energy data (Durant et al., 2011). There is only limited and indirect evidence for emissions, although there is a high agreement among the available estimates within the given uncertainty (Andres et al., 2014, 2012), and emission estimates are consistent with a range of other observations (Ciais et al., 2013), even though their regional and national partitioning is more uncertain (Francey et al., 2013).

National emission inventories take a territorial (production) perspective and “include greenhouse gas emissions and removals taking place within national territory and offshore areas over which the country has jurisdiction” (Rypdal et al., 2006). That is, emissions are allocated to the country where and when the emissions actually occur. The territorial emission inventory of an individual country does not include the emissions from the production of goods and services produced in other countries (e.g. food and clothes) that are used for consumption. Consumption-based emission inventories for an individual country are another attribution point of view that allocates global emissions to products that are consumed within a country; these inventories are conceptually calculated as the territorial emissions minus the “embedded” territorial emissions to produce exported products plus the emissions in other countries to produce imported products (consumption = territorial - exports + imports). The difference between the territorial- and consumption-based emission inventories is the net transfer (exports minus imports) of emissions from the production of internationally traded products. Consumption-based emission attribution results (e.g. Davis and Caldeira, 2010) provide additional information to territorial-based emissions that can be used to understand emission drivers (Hertwich and Peters, 2009), quantify emission transfers by the trade of products between countries (Peters et al., 2011b), and potentially design more effective and efficient climate policy (Peters and Hertwich, 2008).

We estimate consumption-based emissions from 1990 to 2014 by enumerating the global supply chain using a global model of the economic relationships between economic sectors within and between every country (Andrew and Peters, 2013; Peters et al., 2011a). Our analysis is based on the economic and trade data from the Global Trade and Analysis Project (GTAP; Narayanan et al., 2015), and we make detailed estimates for the years 1997 (GTAP version 5), 2001 (GTAP6), and 2004, 2007, and 2011 (GTAP9.1) (using the methodology of Peters et al., 2011b). The results cover 57 sectors and up to 141 countries and regions. The detailed results are then extended into an annual time series from 1990 to the latest year of the GDP data (2014 in this budget), using GDP data by expenditure in current exchange rate of US dollars (USD; from the UN National Accounts Main Aggregates Database; UN, 2015c) and time series of trade data from GTAP (based on the methodology in Peters et al., 2011b).

We estimate the sector-level CO2 emissions using our own calculations based on the GTAP data and methodology, include flaring and cement emissions from CDIAC, and then scale the national totals (excluding bunker fuels) to match the CDIAC estimates from the most recent carbon budget. We do not include international transportation in our estimates of national totals, but include them in the global total. The time series of trade data provided by GTAP covers the period 1995–2013 and our methodology uses the trade shares as this data set. For the period 1990–1994 we assume the trade shares of 1995, while for 2014 we assume the trade shares of 2013.

Comprehensive analysis of the uncertainty in consumption emissions accounts is still lacking in the literature, although several analyses of components of this uncertainty have been made (e.g. Dietzenbacher et al., 2012; Inomata and Owen, 2014; Karstensen et al., 2015; Moran and Wood, 2014). For this reason we do not provide an uncertainty estimate for these emissions, but based on model comparisons and sensitivity analysis, they are unlikely to be larger than for the territorial emission estimates (Peters et al., 2012a). Uncertainty is expected to increase for more detailed results, and to decrease with aggregation (Peters et al., 2011b; e.g. the results for Annex B countries will be more accurate than the sector results for an individual country).

The consumption-based emissions attribution method considers the CO2 emitted to the atmosphere in the production of products, but not the trade in fossil fuels (coal, oil, gas). It is also possible to account for the carbon trade in fossil fuels (Andrew et al., 2013), but we do not present those data here. Peters et al. (2012a) additionally considered trade in biomass.

The consumption data do not modify the global average terms in Eq. (1) but are relevant to the anthropogenic carbon cycle as they reflect the trade-driven movement of emissions across the Earth's surface in response to human activities. Furthermore, if national and international climate policies continue to develop in an un-harmonised way, then the trends reflected in these data will need to be accommodated by those developing policies.

We report the annual growth rate in emissions for adjacent years (in percent per year) by calculating the difference between the two years and then comparing to the emissions in the first year: EFFt0+1-EFFt0EFFt0×100 % yr-1. This is the simplest method to characterise a 1-year growth compared to the previous year and is widely used. We apply a leap-year adjustment to ensure valid interpretations of annual growth rates. This affects the growth rate by about 0.3 % yr-1 (1/365) and causes growth rates to go up approximately 0.3 % if the first year is a leap year and down 0.3 % if the second year is a leap year.

The relative growth rate of EFF over time periods of greater than 1 year can be re-written using its logarithm equivalent as follows: 1EFFdEFFdt=d(ln⁡EFF)dt. Here we calculate relative growth rates in emissions for multi-year periods (e.g. a decade) by fitting a linear trend to ln⁡(EFF) in Eq. (2), reported in percent per year. We fit the logarithm of EFF rather than EFF directly because this method ensures that computed growth rates satisfy Eq. (6). This method differs from previous papers (Canadell et al., 2007; Le Quéré et al., 2009; Raupach et al., 2007) that computed the fit to EFF and divided by average EFF directly, but the difference is very small (< 0.05 % yr-1) in the case of EFF.

Energy statistics from BP are normally available around June for the previous year. To gain insight on emission trends for the current year (2016), we provide an assessment of global emissions for EFF by combining individual assessments of emissions for China and the USA (the two biggest emitting countries) and the rest of the world.

We specifically estimate emissions in China because the data indicate a significant departure from the long-term trends in the carbon intensity of the economy used in emissions projections in previous global carbon budgets (e.g. Le Quéré et al., 2015a), resulting from a rapid deceleration in emissions growth against continued growth in economic output. This departure could be temporary (Jackson et al., 2016). Our 2016 estimate for China uses (1) coal consumption estimates from the China Coal Industry Association for January through September (CCIA, 2016), (2) estimated consumption of natural gas (IEW, 2016; NDRC, 2016a) and domestic production plus net imports of petroleum (NDRC, 2016b) for January through July from the National Development and Reform Commission, and (3) production of cement reported for January to September (NBS, 2016). Using these data, we estimate the change in emissions for the corresponding months in 2016 compared to 2015 assuming a 2 % increase in the energy (and thus carbon) content of coal for 2016 resulting from improvements in the quality of the coal used, in line with the trends reported by the National Bureau of Statistics for recent years. We then assume that the relative changes during the first months will persist throughout the year. The main sources of uncertainty are from the incomplete data on stock changes, the carbon content of coal, and the assumption of persistent behaviour for the rest of the year. These are discussed further in Sect. 3.2.1.

For the USA, we use the forecast of the US Energy Information Administration (EIA) for emissions from fossil fuels (EIA, 2016). This is based on an energy forecasting model which is revised monthly, and takes into account heating degree days, household expenditures by fuel type, energy markets, policies, and other effects. We combine this with our estimate of emissions from cement production using the monthly US cement data from USGS for January–July, assuming changes in cement production over the first seven months apply throughout the year. While the EIA's forecasts for current full-year emissions have on average been revised downwards, only seven such forecasts are available, so we conservatively use the full range of adjustments following revision, and additionally assume symmetrical uncertainty to give ±2.3 % around the central forecast.

For the rest of the world, we use the close relationship between the growth in GDP and the growth in emissions (Raupach et al., 2007) to project emissions for the current year. This is based on the so-called Kaya identity (also called IPAT identity, the acronym standing for human impact (I) on the environment, which is equal to the product of population (P), affluence (A), and technology (T)), whereby EFF (GtC yr-1) is decomposed by the product of GDP (USD yr-1) and the fossil fuel carbon intensity of the economy (IFF; GtC USD-1) as follows: EFF=GDP×IFF. Such product-rule decomposition identities imply that the relative growth rates of the multiplied quantities are additive. Taking a time derivative of Eq. (3) gives dEFFdt=d(GDP×IFF)dt and, applying the rules of calculus, dEFFdt=dGDPdt×IFF+GDP×dIFFdt. Finally, dividing Eq. (5) by Eq. (3) gives 1EFFdEFFdt=1GDPdGDPdt+1IFFdIFFdt, where the left-hand term is the relative growth rate of EFF, and the right-hand terms are the relative growth rates of GDP and IFF, respectively, which can simply be added linearly to give overall growth rate. The growth rates are reported in percent by multiplying each term by 100. As preliminary estimates of annual change in GDP are made well before the end of a calendar year, making assumptions on the growth rate of IFF allows us to make projections of the annual change in CO2 emissions well before the end of a calendar year. The IFF is based on GDP in constant PPP (purchasing power parity) from the IEA up to 2013 (IEA/OECD, 2015) and extended using the IMF growth rates for 2014 and 2015 (IMF, 2016). Interannual variability in IFF is the largest source of uncertainty in the GDP-based emissions projections. We thus use the standard deviation of the annual IFF for the period 2006–2015 as a measure of uncertainty, reflecting a ±1σ as in the rest of the carbon budget. This is ±1.0 % yr-1 for the rest of the world (global emissions minus China and USA).

The 2016 projection for the world is made of the sum of the projections for China, USA, and the rest of the world. The uncertainty is added in quadrature among the three regions. The uncertainty here reflects the best of our expert opinion.

Land-use-change CO2 emissions are calculated by a bookkeeping method approach (Houghton, 2003) that keeps track of the carbon stored in vegetation and soils before deforestation or other land-use change, and the changes in forest age classes, or cohorts, of disturbed lands after land-use change, including possible forest regrowth after deforestation. The method tracks the CO2 emitted to the atmosphere immediately during deforestation, and over time due to the follow-up decay of soil and vegetation carbon in different pools, including wood products pools after logging and deforestation. It also tracks the regrowth of vegetation and associated build-up of soil carbon pools after land-use change. It considers transitions between forests, pastures, and cropland; shifting cultivation; degradation of forests where a fraction of the trees is removed; abandonment of agricultural land; and forest management such as wood harvest and, in the USA, fire management. In addition to tracking logging debris on the forest floor, the bookkeeping method tracks the fate of carbon contained in harvested wood products that is eventually emitted back to the atmosphere as CO2, although a detailed treatment of the lifetime in each product pool is not performed (Earles et al., 2012). Harvested wood products are partitioned into three pools with different turnover times. All fuel wood is assumed burned in the year of harvest (1.0 yr-1). Pulp and paper products are oxidised at a rate of 0.1 yr-1, timber is assumed to be oxidised at a rate of 0.01 yr-1, and elemental carbon decays at 0.001 yr-1. The general assumptions about partitioning wood products among these pools are based on national harvest data (Houghton, 2003).

The primary land-cover change and biomass data for the bookkeeping method analysis is the Forest Resource Assessment of the FAO which provides statistics on forest-cover change and management at intervals of 5 years (FAO, 2010). The data are based on countries' self-reporting, some of which include satellite data in more recent assessments (Table 4). Changes in land cover other than forest are based on annual, national changes in cropland and pasture areas reported by the FAO Statistics Division (FAOSTAT, 2010). Land-use-change country data are aggregated by regions. The carbon stocks on land (biomass and soils), and their response functions subsequent to land-use change, are based on FAO data averages per land-cover type, per biome, and per region. Similar results were obtained using forest biomass carbon density based on satellite data (Baccini et al., 2012). The bookkeeping method does not include land ecosystems' transient response to changes in climate, atmospheric CO2, and other environmental factors, and the growth/decay curves are based on contemporary data that will implicitly reflect the effects of CO2 and climate at that time. Published results from the bookkeeping method are available from 1850 to 2010, with preliminary results available to 2015.

CO2 emissions associated with land-use change calculated from satellite-based fire activity in tropical forest areas (van der Werf et al., 2010) provide information on emissions due to tropical deforestation and degradation that are complementary to the bookkeeping approach. They do not provide a direct estimate of ELUC as they do not include non-combustion processes such as respiration, wood harvest, wood products, or forest regrowth. Legacy emissions such as decomposition from on-ground debris and soils are not included in this method either. However, fire estimates provide some insight in the year-to-year variations in the sub-component of the total ELUC flux that result from immediate CO2 emissions during deforestation caused, for example, by the interactions between climate and human activity (e.g. there is more burning and clearing of forests in dry years) that are not represented by other methods. The “deforestation fire emissions” assume an important role of fire in removing biomass in the deforestation process and thus can be used to infer gross instantaneous CO2 emissions from deforestation using satellite-derived data on fire activity in regions with active deforestation. The method requires information on the fraction of total area burned associated with deforestation vs. other types of fires, and this information can be merged with information on biomass stocks and the fraction of the biomass lost in a deforestation fire to estimate CO2 emissions. The satellite-based deforestation fire emissions are limited to the tropics, where fires result mainly from human activities. Tropical deforestation is the largest and most variable single contributor to ELUC.

Fire emissions associated with deforestation and tropical peat burning are based on the Global Fire Emissions Database (GFED4; accessed July 2016) described in van der Werf et al. (2010) but with updated burned area (Giglio et al., 2013) as well as burned area from relatively small fires that are detected by satellite as thermal anomalies but not mapped by the burned-area approach (Randerson et al., 2012). The burned-area information is used as input data in a modified version of the satellite-driven Carnegie–Ames–Stanford Approach (CASA) biogeochemical model to estimate carbon emissions associated with fires, keeping track of what fraction of fire emissions was due to deforestation (see van der Werf et al., 2010). The CASA model uses different assumptions to compute decay functions compared to the bookkeeping method, and does not include historical emissions or regrowth from land-use change prior to the availability of satellite data. Comparing coincident CO emissions and their atmospheric fate with satellite-derived CO concentrations allows for some validation of this approach (e.g. van der Werf et al., 2008). Results from the fire-based method to estimate land-use-change emissions anomalies added to the bookkeeping mean ELUC estimate are available from 1997 to 2015. Our combination of land-use-change CO2 emissions where the variability in annual CO2 deforestation emissions is diagnosed from fires assumes that year-to-year variability is dominated by variability in deforestation.

Land-use-change CO2 emissions have been estimated using an ensemble of DGVM simulations. New model experiments up to year 2015 have been coordinated by the project “Trends and drivers of the regional-scale sources and sinks of carbon dioxide” (TRENDY; Sitch et al., 2015). We use only models that have estimated land-use-change CO2 emissions following the TRENDY protocol (see Sect. 2.5.2). Models use their latest configurations, summarised in Tables 5 and 6.

Two sets of simulations were performed with the DGVMs, first forced with historical changes in land-cover distribution, climate, atmospheric CO2 concentration, and N deposition, and second, as further described below with a time-invariant pre-industrial land-cover distribution, allowing for estimation of, by difference with the first simulation, the dynamic evolution of biomass and soil carbon pools in response to prescribed land-cover change. Because of the limited availability of the land-use forcing (see below), 14 DGVMs performed historical simulations with time-invariant land-cover distribution, but only 5 DGVMs managed to simulate realistic simulations with time varying land-cover change. These latter DGVMs accounted for deforestation and (to some extent) regrowth, the most important components of ELUC, but they do not represent all processes resulting directly from human activities on land (Table 5). All DGVMs represent processes of vegetation growth and mortality, as well as decomposition of dead organic matter associated with natural cycles, and include the vegetation and soil carbon response to increasing atmospheric CO2 levels and to climate variability and change. In addition, eight models explicitly simulate the coupling of C and N cycles and account for atmospheric N deposition (Table 5), with three of those models used for land-use-change simulations. The DGVMs are independent of the other budget terms except for their use of atmospheric CO2 concentration to calculate the fertilisation effect of CO2 on primary production.

For this global carbon budget, the DGVMs used the HYDE land-use-change data set (Klein Goldewijk et al., 2011), which provides annual, half-degree, fractional data on cropland and pasture. These data are based on annual FAO statistics of change in agricultural area available to 2012 (FAOSTAT, 2010). For the years 2013 to 2015, the HYDE data were extrapolated by country for pastures and cropland separately based on the trend in agricultural area over the previous 5 years. The more comprehensive harmonised land-use data set (Hurtt et al., 2011), which also includes fractional data on primary vegetation and secondary vegetation, as well as all underlying transitions between land-use states, has not been made available yet for this year. Hence, the reduced ensemble of DGVMs that can simulate the LUC flux from the HYDE data set only. The HYDE data are independent of the data set used in the bookkeeping method (Houghton, 2003, and updates), which is based primarily on forest area change statistics (FAO, 2010). The HYDE land-use-change data set does not indicate whether land-use changes occur on forested or non-forested land; it only provides the changes in agricultural areas. Hence, it is implemented differently within each model (e.g. an increased cropland fraction in a grid cell can be at the expense of either grassland or forest, the latter resulting in deforestation; land-cover fractions of the non-agricultural land differ between models). Thus, the DGVM forest area and forest area change over time is not consistent with the Forest Resource Assessment of the FAO forest area data used for the bookkeeping model to calculate ELUC. Similarly, model-specific assumptions are applied to convert deforested biomass or deforested area, and other forest product pools, into carbon in some models (Table 5).

The DGVM runs were forced by either 6-hourly CRU-NCEP or by monthly CRU temperature, precipitation, and cloud cover fields (transformed into incoming surface radiation) based on observations and provided on a 0.5∘ × 0.5∘ grid and updated to 2015 (Harris et al., 2014; Viovy, 2016). The forcing data include both gridded observations of climate and global atmospheric CO2, which change over time (Dlugokencky and Tans, 2016), and N deposition (as used in some models; Table 5). As mentioned before, ELUC is diagnosed in each model by the difference between a model simulation with prescribed historical land-cover change and a simulation with constant, pre-industrial land-cover distribution. Both simulations were driven by changing atmospheric CO2, climate, and, in some models, N deposition over the period 1860–2015. Using the difference between these two DGVM simulations to diagnose ELUC is not fully consistent with the definition of ELUC in the bookkeeping method (Gasser and Ciais, 2013; Stocker and Joos, 2015). The DGVM approach to diagnose land-use-change CO2 emissions would be expected to produce systematically higher ELUC emissions than the bookkeeping approach if all the parameters of the two approaches were the same, which is not the case (see Sect. 2.5.2).

Differences between the bookkeeping, the addition of fire-based interannual variability to the bookkeeping, and DGVM methods originate from three main sources: the land-cover-change data set, the different approaches used in models, and the different processes represented (Table 5). We examine the results from the DGVMs and of the bookkeeping method to assess the uncertainty in ELUC.

The uncertainties in annual ELUC estimates are examined using the standard deviation across models, which averages 0.3 GtC yr-1 from 1959 to 2015 (Table 7). The mean of the multi-model ELUC estimates is consistent with a combination of the bookkeeping method and fire-based emissions (Table 7), with the multi-model mean and bookkeeping method differing by less than 0.5 GtC yr-1 over 85 % of the time. Based on this comparison, we determine that an uncertainty of ±0.5 GtC yr-1 provides a semi-quantitative measure of uncertainty for annual 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 is consistent with the uncertainty analysis of Houghton et al. (2012), which partly reflects improvements in data on forest area change using data and partly more complete understanding and representation of processes in models.

The uncertainties in the decadal ELUC estimates are also examined using the DGVM ensemble, although they are likely correlated between decades. The correlations between decades come from (1) common biases in system boundaries (e.g. not counting forest degradation in some models), (2) common definition for the calculation of ELUC from the difference of simulations with and without land-use change (a source of bias vs. the unknown truth), (3) common and uncertain land-cover change input data which also cause a bias (though if a different input data set is used each decade, decadal fluxes from DGVMs may be partly decorrelated), and (4) model structural errors (e.g. systematic errors in biomass stocks). In addition, errors arising from uncertain DGVM parameter values would be random but they are not accounted for in this study, since no DGVM provided an ensemble of runs with perturbed parameters.

Prior to 1959, the uncertainty in ELUC is taken as ±33 %, which is the ratio of uncertainty to mean from the 1960s in the bookkeeping method (Table 7), the first decade available. This ratio is consistent with the mean standard deviation of DGVMs land-use-change emissions over 1870–1958 (0.32 GtC) over the multi-model mean (0.9 GtC).

The rate of growth of the atmospheric CO2 concentration is provided by the US National Oceanic and Atmospheric Administration Earth System Research Laboratory (NOAA/ESRL; Dlugokencky and Tans, 2016), which is updated from Ballantyne et al. (2012). For the 1959–1980 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 1980–2015 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 each station with a smoothed curve as a function of time, and averaging by latitude band (Masarie and Tans, 1995). The annual growth rate is estimated by Dlugokencky and Tans (2016) 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 1 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 2.12 GtC ppm-1 (Ballantyne et al., 2012) for consistency with the other components.

The uncertainty around the annual growth rate based on the multiple stations data set ranges between 0.11 and 0.72 GtC yr-1, with a mean of 0.61 GtC yr-1 for 1959–1979 and 0.19 GtC yr-1 for 1980–2015, when a larger set of stations were available (Dlugokencky and Tans, 2016). It is based on the number of available stations, and thus takes into account both the measurement errors and data gaps at each station. This uncertainty is larger than the uncertainty of ±0.1 GtC yr-1 reported for decadal mean growth rate by the IPCC because errors in annual growth rate are strongly anti-correlated in consecutive years leading to smaller errors for longer timescales. The decadal change is computed from the difference in concentration 10 years apart based on a measurement error of 0.35 ppm. This error is based on offsets between NOAA/ESRL measurements and those of the World Meteorological Organization World Data Centre for Greenhouse Gases (NOAA/ESRL, 2015) for the start and end points (the decadal change uncertainty is the 20.35ppm2(10yr)-1 assuming that each yearly measurement error is independent). This uncertainty is also used in Table 8.

The contribution of anthropogenic CO and CH4 is neglected from the global carbon budget (see Sect. 2.7.1). 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).

In order to estimate the total carbon accumulated in the atmosphere since 1750 or 1870, we use an atmospheric CO2 concentration of 277 ± 3 or 288 ± 3 ppm, respectively, based on a cubic spline fit to ice core data (Joos and Spahni, 2008). The uncertainty of ±3 ppm (converted to ±1σ) is taken directly from the IPCC's assessment (Ciais et al., 2013). Typical uncertainties in the growth rate in atmospheric CO2 concentration from ice core data are ±1–1.5 GtC per decade as evaluated from the Law Dome data (Etheridge et al., 1996) for individual 20-year intervals over the period from 1870 to 1960 (Bruno and Joos, 1997).

A mean ocean CO2 sink of 2.2 ± 0.4 GtC yr-1 for the 1990s was estimated by the IPCC (Denman et al., 2007) based on indirect observations and their spread: ocean–land CO2 sink partitioning from observed atmospheric O2 / N2 concentration trends (Manning and Keeling, 2006), an oceanic inversion method constrained by ocean biogeochemistry data (Mikaloff Fletcher et al., 2006), and a method based on penetration timescale for CFCs (McNeil et al., 2003). This is comparable with the sink of 2.0 ± 0.5 GtC yr-1 estimated by Khatiwala et al. (2013) for the 1990s, and with the sink of 1.9 to 2.5 GtC yr-1 estimated from a range of methods for the period 1990–2009 (Wanninkhof et al., 2013), with uncertainties ranging from ±0.3 to ±0.7 GtC yr-1. The most direct way for estimating the observation-based ocean sink is from the product of (sea–air pCO2 difference) × (gas transfer coefficient). Estimates based on sea–air pCO2 are fully consistent with indirect observations (Wanninkhof et al., 2013), but their uncertainty is larger mainly due to difficulty in capturing complex turbulent processes in the gas transfer coefficient (Sweeney et al., 2007) and because of uncertainties in the pre-industrial river outgas of CO2 (Jacobson et al., 2007).

The two observation-based estimates (Landschützer et al., 2015; Rödenbeck et al., 2014) used here compute the ocean CO2 sink and its variability using interpolated measurements of surface ocean fugacity of CO2 (pCO2 corrected for the non-ideal behaviour of the gas; Pfeil et al., 2013). The measurements were from the Surface Ocean CO2 Atlas version 4, which is an update of version 3 (Bakker et al., 2016) and contains data to 2015 (see data attribution Table 1a). In contrast to last year's global carbon budget, where preliminary data were used for the past year, data used here are fully quality-controlled following standard SOCAT procedures. The SOCAT v4 were mapped using a data-driven diagnostic method (Rödenbeck et al., 2013) and a combined self-organising map and feed-forward neural network (Landschützer et al., 2014). The global observation-based estimates were adjusted to remove a background (not part of the anthropogenic ocean flux) ocean source of CO2 to the atmosphere of 0.45 GtC yr-1 from river input to the ocean (Jacobson et al., 2007) in order to make them comparable to SOCEAN, which only represents the annual uptake of anthropogenic CO2 by the ocean. Several other data-based products are available, but they show large discrepancies with observed variability that need to be resolved. Here we used the two data products that had the best fit to observations for their representation of tropical and global variability (Rödenbeck et al., 2015).

We use the data-based product of Khatiwala et al. (2009) updated by Khatiwala et al. (2013) to estimate the anthropogenic carbon accumulated in the ocean during 1765–1958 (60.2 GtC) and 1870–1958 (47.5 GtC), and assume an oceanic uptake of 0.4 GtC for 1750–1765 (for which time no data are available) based on the mean uptake during 1765–1770. The estimate of Khatiwala et al. (2009) is based on regional disequilibrium between surface pCO2 and atmospheric CO2, and a Green's function utilising transient ocean tracers like CFCs and 14C to ascribe changes through time. It does not include changes associated with changes in ocean circulation, temperature, and climate, but these are thought to be small over the time period considered here (Ciais et al., 2013). The uncertainty in cumulative uptake of ±20 GtC (converted to ±1σ) is taken directly from the IPCC's review of the literature (Rhein et al., 2013), or about ±30 % for the annual values (Khatiwala et al., 2009).

The trend in the ocean CO2 sink for 1959–2015 is computed using a combination of seven global ocean biogeochemistry models (Table 6). The models represent the physical, chemical, and biological processes that influence the surface ocean concentration of CO2 and thus the air-sea CO2 flux. The models are forced by meteorological reanalysis and atmospheric CO2 concentration data available for the entire time period. Models do not include the effects of anthropogenic changes in nutrient supply, which could lead to an increase in the ocean sink of up to about 0.3 GtC yr-1 over the industrial period (Duce et al., 2008). They compute the air-sea flux of CO2 over grid boxes of 1 to 4∘ in latitude and longitude. The ocean CO2 sink for each model is normalised to the observations by dividing the annual model values by their modelled average over 1990–1999 and multiplying this by the observation-based estimate of 2.2 GtC yr-1 (obtained from Manning and Keeling, 2006; McNeil et al., 2003; Mikaloff Fletcher et al., 2006). The ocean CO2 sink for each year (t) in GtC yr-1 is therefore SOCEAN(t)=1n∑m=1m=nSOCEANm(t)SOCEANm(1990–1999)×2.2, where n is the number of models. This normalisation ensures that the ocean CO2 sink for the global carbon budget is based on observations, whereas the trends and annual values in CO2 sinks are from model estimates. The normalisation based on a ratio assumes that if models over or underestimate the sink in the 1990s, it is primarily due to the process of diffusion, which depends on the gradient of CO2. Thus, a ratio is more appropriate than an offset as it takes into account the time dependence of CO2 gradients in the ocean. The mean uncorrected ocean CO2 sink from the models for 1990–1999 ranges between 1.7 and 2.4 GtC yr-1, with a multi-model mean of 2.0 GtC yr-1.

The uncertainty around the mean ocean sink of anthropogenic CO2 was quantified by Denman et al. (2007) for the 1990s (see Sect. 2.4.1). To quantify the uncertainty around annual values, we examine the standard deviation of the normalised model ensemble. We use further information from the two data-based products to assess the confidence level. The average standard deviation of the normalised ocean model ensemble is 0.16 GtC yr-1 during 1980–2010 (with a maximum of 0.33), but it increases as the model ensemble goes back in time, with a standard deviation of 0.22 GtC yr-1 across models in the 1960s. We estimate that the uncertainty in the annual ocean CO2 sink is about ±0.5 GtC yr-1 from the fractional uncertainty in the data uncertainty of ±0.4 GtC yr-1 and standard deviation across models of up to ±0.33 GtC yr-1, reflecting both the uncertainty in the mean sink from observations during the 1990s (Denman et al., 2007; Sect. 2.4.1) and in the interannual variability as assessed by models.

We examine the consistency between the variability in the model-based and the data-based products to assess confidence in SOCEAN. The interannual variability of the ocean fluxes (quantified as the standard deviation) of the two data-based estimates for 1986–2015 (where they overlap) is ± 0.34 GtC yr-1 (Rödenbeck et al., 2014) and ± 0.41 GtC yr-1 (Landschützer et al., 2015), compared to ±0.29 GtC yr-1 for the normalised model ensemble. The standard deviation includes a component of trend and decadal variability in addition to interannual variability, and their relative influence differs across estimates. The phase is generally consistent between estimates, with a higher ocean CO2 sink during El Niño events. The annual data-based estimates correlate with the ocean CO2 sink estimated here with a correlation of r = 0.71 (0.51 to 0.77 for individual models), and r = 0.81 (0.66 to 0.79) for the data-based estimates of Rödenbeck et al. (2014) and Landschützer et al. (2015), respectively (simple linear regression), with their mutual correlation at 0.65. The agreement is better for decadal variability than for interannual variability. The use of annual data for the correlation may reduce the strength of the relationship because the dominant source of variability associated with El Niño events is less than one year. We assess a medium confidence level to the annual ocean CO2 sink and its uncertainty because they are based on multiple lines of evidence, and the results are consistent in that the interannual variability in the model and data-based estimates are all generally small compared to the variability in the growth rate of atmospheric CO2 concentration.

For 1959–2015, the terrestrial carbon sink was estimated from the residual of the other budget terms by rearranging Eq. (1): SLAND=EFF+ELUC-(GATM+SOCEAN). The uncertainty in SLAND is estimated annually from the root sum of squares of the uncertainty in the right-hand terms assuming the errors are not correlated. The uncertainty averages to ±0.8 GtC yr-1 over 1959–2015 (Table 7). SLAND estimated from the residual of the budget includes, by definition, all the missing processes and potential biases in the other components of Eq. (8).

A comparison of the residual calculation of SLAND in Eq. (8) with estimates from DGVMs as used to estimate ELUC in Sect. 2.2.3, but here excluding the effects of changes in land cover (using a constant pre-industrial land-cover distribution), provides an independent estimate of the consistency of SLAND with our understanding of the functioning of the terrestrial vegetation in response to CO2 and climate variability (Table 7). As described in Sect. 2.2.3, the DGVM runs that exclude the effects of changes in land cover include all climate variability and CO2 effects over land, but they do not include reductions in CO2 sink capacity associated with human activity directly affecting changes in vegetation cover and management, which by design is allocated to ELUC. This effect has been estimated to have led to a reduction in the terrestrial sink by 0.5 GtC yr-1 since 1750 (Gitz and Ciais, 2003). The models in this configuration estimate the mean and variability of SLAND based on atmospheric CO2 and climate, and thus both terms can be compared to the budget residual. We apply three criteria for minimum model realism by including only those models with (1) steady state after spin-up; (2) where available, net land fluxes (SLAND-ELUC) that are a carbon sink over the 1990s as constrained by global atmospheric and oceanic observations (Keeling and Manning, 2014; Wanninkhof et al., 2013); and (3) where available, global ELUC that is a carbon source over the 1990s. Fourteen models met criteria (1), and five of the models that provided ELUC met all three criteria.

The standard deviation of the annual CO2 sink across the DGVMs' averages to ±0.8 GtC yr-1 for the period 1959 to 2015. The model mean, over different decades, correlates with the budget residual with r=0.68 (0.51 to r=0.77 for individual models). The standard deviation is similar to that of the five model ensembles presented in Le Quéré et al. (2009), but the correlation is improved compared to r=0.54 obtained in the earlier study. The DGVM results suggest that the sum of our knowledge on annual CO2 emissions and their partitioning is plausible (see Sect. 4), and provide insight on the underlying processes and regional breakdown. However, as the standard deviation across the DGVMs (0.8 GtC yr-1 on average) is of the same magnitude as the combined uncertainty due to the other components (EFF, ELUC, GATM, SOCEAN; Table 7), the DGVMs do not provide further reduction of uncertainty in the annual terrestrial CO2 sink compared to the residual of the budget (Eq. 8). Yet, DGVM results are largely independent of the residual of the budget, and it is worth noting that the residual method and ensemble mean DGVM results are consistent within their respective uncertainties. We attach a medium confidence level to the annual land CO2 sink and its uncertainty because the estimates from the residual budget and averaged DGVMs match well within their respective uncertainties, and the estimates based on the residual budget are primarily dependent on EFF and GATM, both of which are well constrained.

The three inversion systems used in this release are the CarbonTracker (Peters et al., 2010), the Jena CarboScope (Rödenbeck, 2005), and CAMS (Chevallier et al., 2005). See Table 6 for version numbers. They are based on the same Bayesian inversion principles that interpret the same, for the most part, observed time series (or subsets thereof), but use different methodologies that represent some of the many approaches used in the field. This mainly concerns the time resolution of the estimates (i.e. weekly or monthly), spatial breakdown (i.e. grid size), assumed correlation structures, and mathematical approach. The details of these approaches are documented extensively in the references provided. Each system uses a different transport model, which was demonstrated to be a driving factor behind differences in atmospheric-based flux estimates, and specifically their global distribution (Stephens et al., 2007).

The three inversions use atmospheric CO2 observations from various flask and in situ networks. They prescribe spatial and global EFF that can vary from that presented here. The CarbonTracker and CAMS inversions prescribed the same global EFF as in Sect. 2.1.1, during 2010–2015 for CarbonTracker and during 1979–2015 in CAMS. The Jena CarboScope inversion uses EFF from EDGAR (2011) v4.2. Different spatial and temporal distributions of EFF were prescribed in each inversion.

Given their prescribed EFF, each inversion estimates natural fluxes from a similar set of surface CO2 measurement stations, and CarbonTracker additionally uses two sites of aircraft CO2 vertical profiles over the Amazon and Siberia, regions where surface observations are sparse. The atmospheric transport models of each inversion are TM5 for CarbonTracker, TM3 for Jena CarboScope, and LMDZ for CAMS. These three models are based on the same ECMWF wind fields. The three inversions use different prior natural fluxes, which partly influences their optimised fluxes. CAMS assumes that the prior land flux is zero on the annual mean in each grid cell of the transport model, so that any sink or source on land is entirely reflecting the information brought by atmospheric measurements. CarbonTracker simulates a small prior sink on land from the SIBCASA model that results from regrowth following fire disturbances of an otherwise net zero biosphere. Jena CarboScope assumes a prior sink on land as well from the LPJ model. Inversion results for the sum of natural ocean and land fluxes (Fig. 8) are more constrained in the Northern Hemisphere (NH) than in the tropics, because of the higher measurement stations density in the NH.

Finally, results from atmospheric inversions include the natural CO2 fluxes from rivers (which need to be taken into account to allow comparison to other sources) and chemical oxidation of reactive carbon-containing gases (which are neglected here). These inverse estimates are not truly independent of the other estimates presented here as the atmospheric observations include a set of observations used to estimate the global growth rate in atmospheric CO2 concentration (Sect. 2.3). However, they provide new information on the regional distribution of fluxes.

We focus the analysis on two known strengths of the inverse approach: the derivation of the year-to-year changes in total land fluxes (SLAND-ELUC) consistent with the whole network of atmospheric observations, and the spatial breakdown of land and ocean fluxes (SLAND-ELUC+SOCEAN) across large regions of the globe. The spatial breakdown is discussed in Sect. 3.1.3.

Anthropogenic emissions of CO and CH4 to the atmosphere are eventually oxidised to CO2 and thus are part of the global carbon budget. These contributions are omitted in Eq. (1), but an attempt is made in this section to estimate their magnitude and identify the sources of uncertainty. Anthropogenic CO emissions are from incomplete fossil fuel and biofuel burning and deforestation fires. The main anthropogenic emissions of fossil CH4 that matter for the global carbon budget are the fugitive emissions of coal, oil, and gas upstream sectors (see below). These emissions of CO and CH4 contribute a net addition of fossil carbon to the atmosphere.

In our estimate of EFF we assumed (Sect. 2.1.1) that all the fuel burned is emitted as CO2; thus, CO anthropogenic emissions and their atmospheric oxidation into CO2 within a few months are already counted implicitly in EFF and should not be counted twice (same for ELUC and anthropogenic CO emissions by deforestation fires). Anthropogenic emissions of fossil CH4 are not included in EFF, because these fugitive emissions are not included in the fuel inventories. Yet they contribute to the annual CO2 growth rate after CH4 gets oxidised into CO2. Anthropogenic emissions of fossil CH4 represent 15 % of total CH4 emissions (Kirschke et al., 2013) that is 0.061 GtC yr-1 for the past decade. Assuming steady state, these emissions are all converted to CO2 by OH oxidation and thus explain 0.06 GtC yr-1 of the global CO2 growth rate in the past decade, or 0.07–0.1 GtC yr-1 using the higher CH4 emissions reported recently (Schwietzke et al., 2016).

Other anthropogenic changes in the sources of CO and CH4 from wildfires, biomass, wetlands, ruminants, or permafrost changes are similarly assumed to have a small effect on the CO2 growth rate.

The approach used to determine the global carbon budget considers only anthropogenic CO2 emissions and their partitioning among the atmosphere, ocean, and land. In this analysis, the land and ocean reservoirs that take up anthropogenic CO2 from the atmosphere are conceived as independent carbon storage repositories. This approach omits the fact that carbon is continuously displaced from the land to the ocean through the land–ocean aquatic continuum (LOAC) comprising freshwaters, estuaries, and coastal areas (Bauer et al., 2013; Regnier et al., 2013). A significant fraction of this lateral carbon flux is entirely “natural” and is thus a steady-state component of the pre-industrial carbon cycle. However, changes in environmental conditions and land-use change have caused an increase in the lateral transport of carbon into the LOAC – a perturbation that is relevant for the global carbon budget presented here.

The results of the analysis of Regnier et al. (2013) can be summarised in two points of relevance for the anthropogenic CO2 budget. First, the anthropogenic perturbation has increased the organic carbon export from terrestrial ecosystems to the hydrosphere at a rate of 1.0 ± 0.5 GtC yr-1, mainly owing to enhanced carbon export from soils. Second, this exported anthropogenic carbon is partly respired through the LOAC, partly sequestered in sediments along the LOAC and, to a lesser extent, transferred in the open ocean where it may accumulate. The increase in storage of land-derived organic carbon in the LOAC and open ocean combined is estimated by Regnier et al. (2013) at 0.65 ± 0.35 GtC yr-1. The implication of a substantial LOAC carbon accumulation is that SLAND corresponds to carbon sequestered both in land ecosystems and in LOAC. We do not attempt to separate these two storage components in our study focused on SLAND.

Global CO2 emissions from fossil fuels and industry have increased every decade from an average of 3.1 ± 0.2 GtC yr-1 in the 1960s to an average of 9.3 ± 0.5 GtC yr-1 during 2006–2015 (Table 8 and Fig. 5). The growth rate in these emissions decreased between the 1960s and the 1990s, with 4.5 % yr-1 in the 1960s (1960–1969), 2.8 % yr-1 in the 1970s (1970–1979), 1.9 % yr-1 in the 1980s (1980–1989), and 1.1 % yr-1 in the 1990s (1990–1999). After this period, the growth rate began increasing again in the 2000s at an average growth rate of 3.5 % yr-1, decreasing to 1.8 % yr-1 for the last decade (2006–2015). In contrast, CO2 emissions from land-use change have remained relatively constant at around 1.3 ± 0.5 GtC yr-1 during 1960–2015. A decrease in emissions from land-use change is suggested between the 1990s and 2000s by the combination of bookkeeping and fire-based emissions used here (Table 7), but it is highly uncertain due to uncertainty in the underlying land-cover-change data. This decrease is not found in the current ensemble of the DGVMs (Fig. 6), which are otherwise consistent with the bookkeeping method within their respective uncertainty (Table 7). The decrease is also not found in the study of tropical deforestation of Achard et al. (2014), where the fluxes in the 1990s were similar to those of the 2000s and outside our uncertainty range. A new study based on FAO data to 2015 (Federici et al., 2015) suggests that ELUC decreased during 2011–2015 compared to 2001–2010.

Emissions are partitioned among the atmosphere, ocean, and land (Eq. 1). The growth rate in atmospheric CO2 level increased from 1.7 ± 0.1 GtC yr-1 in the 1960s to 4.5 ± 0.1 GtC yr-1 during 2006–2015 with important decadal variations (Table 8). Both ocean and land CO2 sinks increased roughly in line with the atmospheric increase, but with significant decadal variability on land (Table 8). The ocean CO2 sink increased from 1.2 ± 0.5 GtC yr-1 in the 1960s to 2.6 ± 0.5 GtC yr-1 during 2006–2015, with interannual variations of the order of a few tenths of GtC yr-1 generally showing an increased ocean sink during El Niño events (i.e. 1982–1983, 1997–1998, 2015–2016) (Fig. 7; Rödenbeck et al., 2014). Although there is some coherence between the ocean models and data products and among data products regarding the mean, decadal variability, and trend, the ocean models and data products show poor agreement for interannual variability (Sect. 2.4.3 and Fig. 7). As shown in Fig. 7, the two data products and most model estimates produce a mean CO2 sink for the 1990s that is below the mean assessed by the IPCC from indirect (but arguably more reliable) observations (Denman et al., 2007; Sect. 2.4.1). This discrepancy suggests we may need to reassess estimates of the mean ocean carbon sinks, with some implications for the cumulative carbon budget (Landschützer et al., 2016).

The residual terrestrial CO2 sink increased from 1.7 ± 0.7 GtC yr-1 in the 1960s to 3.1 ± 0.9 GtC yr-1 during 2006–2015, with important interannual variations of up to 2 GtC yr-1 generally showing a decreased land sink during El Niño events, overcompensating for the increase in ocean sink and accounting for the enhanced growth rate in atmospheric CO2 concentration during El Niño events. The high uptake anomaly around year 1991 is thought to be caused by the effect of the volcanic eruption of Mount Pinatubo on climate and is not generally reproduced by the DGVMs, but it is assigned to the land by the two inverse systems that include this period (Fig. 6). The larger land CO2 sink during 2006–2015 compared to the 1960s is reproduced by all the DGVMs in response to combined atmospheric CO2 increase, climate, and variability, consistent with the budget residual and reflecting a common knowledge of the processes (Table 7). The DGVM ensemble mean of 2.8 ± 0.7 GtC yr-1 also reproduces the observed mean for the period 2006–2015 calculated from the budget residual (Table 7).

The total CO2 fluxes on land (SLAND-ELUC) constrained by the atmospheric inversions show in general very good agreement with the global budget estimate, as expected given the strong constrains of GATM and the small relative uncertainty assumed on SOCEAN and EFF by inversions. The total land flux is of similar magnitude for the decadal average, with estimates for 2006–2015 from the three inversions of 2.2, 2.3, and 3.4 GtC yr-1 compared to 2.1 ± 0.7 GtC yr-1 for the total flux computed with the carbon budget from other terms in Eq. (1) (Table 7). The total land sink from the three inversions is 1.8, 1.8, and 3.0 GtC yr-1 when including a mean river flux adjustment of 0.45 GtC yr-1, though the exact adjustment is in fact smaller because the anthropogenic contribution to river fluxes is only a fraction of the total river flux (Sect. 2.7.2). The interannual variability in the inversions also matched the residual-based SLAND closely (Fig. 6). The total land flux from the DGVM multi-model mean also compares well with the estimate from the carbon budget and atmospheric inversions, with a decadal mean of 1.7 ± 0.5 GtC yr-1 (Table 7; 2006–2015), although individual models differ by several GtC for some years (Fig. 6).

Figure 8 shows the partitioning of the total surface fluxes excluding emissions from fossil fuels and industry (SLAND+SOCEAN-ELUC) according to the process models in the ocean and on land, as well as to the three atmospheric inversions. The total surface fluxes provide information on the regional distribution of those fluxes by latitude bands (Fig. 8). The global mean CO2 fluxes from process models for 2006–2015 is 4.2 ± 0.6 GtC yr-1. This is comparable to the fluxes of 4.8 ± 0.5 GtC yr-1 inferred from the remainder of the carbon budget (EFF-GATM in Eq. 1; Table 8) within their respective uncertainties. The total CO2 fluxes from the three inversions range between 4.6 and 4.9 GtC yr-1, consistent with the carbon budget as expected from the constraints on the inversions.

In the south (south of 30∘ S), the atmospheric inversions and process models all suggest a CO2 sink for 2006–2015 of between 1.2 and 1.6 GtC yr-1 (Fig. 8), although the details of the interannual variability are not fully consistent across methods. The interannual variability in the south is low because of the dominance of ocean area with low variability compared to land areas.

In the tropics (30∘ S–30∘ N), both the atmospheric inversions and process models suggest the carbon balance in this region is close to neutral over the past decade, with fluxes for 2006–2015 ranging between -0.5 and +0.6 GtC yr-1. Both the process-based models and the inversions consistently allocate more year-to-year variability in CO2 fluxes to the tropics compared to the north (north of 30∘ N; Fig. 8), this variability being dominated by land fluxes.

In the north (north of 30∘ N), the inversions and process models are not in agreement on the magnitude of the CO2 sink with the ensemble mean of the process models suggesting a total Northern Hemisphere sink for 2006–2015 of 2.3 ± 0.4 GtC yr-1, while the three inversions estimate a sink of 2.7, 3.8, and 3.8 GtC yr-1. The mean difference can only partly be explained by the influence of river fluxes, which is seen by the inversions but not included in the process models, as this flux in the Northern Hemisphere would be less than 0.45 GtC yr-1, particularly when only the anthropogenic contribution to river fluxes is accounted for. The CarbonTracker inversion is close to the one standard deviation of the process models for the mean sink during their overlap period. CAMS and Jena CarboScope give a higher sink in the north than the process models, and a correspondingly higher source in the tropics. Differences between CarbonTracker and CAMS, Jena CarboScope may be related to differences in inter-hemispheric mixing time of their transport models, and other inversion settings. Differences between the mean fluxes of CAMS, Jena CarboScope, and the ensemble of process models cannot be explained easily. They could reflect a bias either in these two inversions or in missing processes or biases in the process models, such as the lack of adequate parameterisations for forest management in the north and for forest degradation emissions in tropics for the DGVMs.

The estimated contribution of the north and its uncertainty from process models is sensitive both to the ensemble of process models used and to the specifics of each inversion. All three inversions show substantial differences in variability and/or trend, and one inversion shows a substantial difference in the mean Northern Hemisphere sink.

Global CO2 emissions from fossil fuels and industry remained nearly constant at 9.9 ± 0.5 GtC in 2015 (Fig. 5), distributed among coal (41 %), oil (34 %), gas (19 %), cement (5.6 %), and gas flaring (0.7 %). Compared to the previous year, emissions from coal and cement decreased by -1.8 and -1.9 %, respectively, while emissions from oil and gas increased by 1.9 and 1.7 %, respectively. Due to lack of data, gas flaring in 2014 and 2015 is assumed the same as 2013.

Growth in emissions in 2015 was not statistically different from zero, at 0.06 % higher than in 2014, in stark contrast with the decadal average of 1.8 % yr-1 (2006–2015). Growth in 2015 is in the range of our projection change of -0.6 [-1.6 to +0.5] % made last year (Le Quéré et al., 2015a) based on national emissions projections for China and the USA, and projections of gross domestic product corrected for IFF improvements for the rest of the world. However, the specific projection for 2015 for China made last year (likely range of -4.6 to -1.1 %) was for a larger decrease in emissions than realised (-0.7 %). This is due to lower decline in coal production in the last 4 months of the year compared to January–August and to improvements in energy content of coal through improvements in the quality of the coal used which were at the top of the range of improvements considered in our projection.

In 2015, the largest contributions to global CO2 emissions were from China (29 %), the USA (15 %), the EU (28 member states; 10 %), and India (6.3 %). The percentages are the fraction of the global emissions including bunker fuels (3.2 %). These four regions account for 59 % of global emissions. Growth rates for these countries from 2014 to 2015 were -0.7 % (China), -2.6 % (USA), +1.4 % (EU28), and +5.2 % (India). The per capita CO2 emissions in 2015 were 1.3 tC person-1 yr-1 for the globe, and were 4.6 (USA), 2.1 (China), 1.9 (EU28), and 0.5 tC person-1 yr-1 (India) for the four highest emitting countries (Fig. 5e).

Territorial emissions in Annex B countries decreased by -0.2 % yr-1 on average during 1990–2014. Trends observed for consumption emissions were less monotonic, with 0.8 % yr-1 growth over 1990–2007 and a -1.5 % yr-1 decrease over 2007–2014 (Fig. 5c). In non-Annex B countries during 1990–2014, territorial emissions have grown at 4.7 % yr-1, while consumption emissions have grown at 4.4 % yr-1. In 1990, 65 % of global territorial emissions were emitted in Annex B countries (33 % in non-Annex B, and 2 % in bunker fuels used for international shipping and aviation), while in 2014 this had reduced to 37 % (60 % in non-Annex B, and 3 % in bunker fuels). In terms of consumption emissions this split was 67 % in 1990 and 41 % in 2014 (33 to 59 % in non-Annex B). The difference between territorial and consumption emissions (the net emission transfer via international trade) from non-Annex B to Annex B countries has increased from near zero in 1990 to 0.3 GtC yr-1 around 2005 and remained relatively stable afterwards until the last year available (2014; Fig. 5). The increase in net emission transfers of 0.30 GtC yr-1 between 1990 and 2014 compares with the emission reduction of 0.4 GtC yr-1 in Annex B countries. These results show the importance of net emission transfer via international trade from non-Annex B to Annex B countries, as well as the stabilisation of emissions transfers when averaged over Annex B countries during the past decade. In 2014, the biggest emitters from a consumption perspective were China (25 % of the global total), USA (16 %), EU28 (12 %), and India (5 %).

Based on fire activity, the global CO2 emissions from land-use change are estimated as 1.3 ± 0.5 GtC in 2015, slightly above the 2006–2015 average of 1.0 ± 0.5 GtC yr-1. The slight rise in ELUC in 2015 is consistent with estimates of peat fires in Asia based on atmospheric data (Yin et al., 2016). However, the estimated annual variability is not generally consistent between methods, except that all methods estimate that variability in ELUC is small relative to the variability from SLAND (Fig. 6a). This could be partly due to the design of the DGVM experiments, which use flux differences between simulations with and without land-cover change, and thus their variability may differ due to, for example, fires in forest regions where the contemporary forest cover is smaller than pre-industrial cover used in the “without land-cover change” runs.

The growth rate in atmospheric CO2 concentration was 6.3 ± 0.2 GtC in 2015 (2.97 ± 0.09 ppm; Fig. 4; Dlugokencky and Tans, 2016). This is well above the 2006–2015 average of 4.5 ± 0.1 GtC yr-1 and reflects the large interannual variability in the growth rate of atmospheric CO2 concentration associated with El Niño and La Niña events.

The ocean CO2 sink was 3.0 ± 0.5 GtC yr-1 in 2015, an increase of 0.15 GtC yr-1 over 2015 according to ocean models. Five of the seven ocean models produce an increase in the ocean CO2 sink in 2015 compared to 2014, with near-zero changes in the last two models (Fig. 7). However, the two data products disagree over changes in the last year, with a decrease of -0.4 GtC yr-1 found in Rödenbeck et al. (2014) and an increase of 0.3 GtC yr-1 found in Landschützer et al. (2015). Thus, there is no overall consistency in the annual change in the ocean CO2 sink, although there is an indication of increasing convergence among products for the assessment of multi-year changes, as suggested by the time-series correlations reported in Sect. 2.4.3 (see also Landschützer et al., 2015). An increase in the ocean CO2 sink in 2015 would be consistent with the observed El Niño conditions and continued rising atmospheric CO2. All estimates suggest an ocean CO2 sink for 2015 that is larger than their 2006–2015 average.

The terrestrial CO2 sink calculated as the residual from the carbon budget was 1.9 ± 0.9 GtC in 2015, well below the 3.1 ± 0.9 GtC yr-1 averaged over 2006–2015 (Fig. 4), and reflecting the onset of the El Niño conditions in the second half of 2015. The DGVM mean produces a sink of 1.0 ± 1.4 GtC in 2015, also well below the 2006–2015 average (Table 7). Both models and inversions suggest that the lower sink in 2015 primarily originated in the tropics (Fig. 8).

Using separate projections for China, the USA, and the rest of the world as described in Sect. 2.1.4, we project a continued low growth in global CO2 emissions from fossil fuels and cement production in 2016 of +0.5 % (range of -0.7 to +2.1 %) from 2015 levels (Table 9), or +0.2 % (-1.0 to +1.8 %) after adjusting for leap year (see Sect. 2.1.3). Our method is imprecise and contains several assumptions that could influence the results beyond the given range, and as such is indicative only. Within the given assumptions, global emissions remain nearly constant at 9.9 ± 0.5 GtC (36.4 ± 1.8 GtCO2) in 2016, but they are still 63% above emissions in 1990. The drivers of the trends in EFF are discussed elsewhere (Peters et al., 2016).

For China, the expected change based on available data during January to July or September (see Sect. 2.1.4) is for an increase in emissions of -0.2 % (range of -3.5 to +1.6 %) in 2016 compared to 2015, or -0.5 % (range of -3.8 to +1.3 %) after adjusting for leap year (see Sect. 2.1.3). This is based on estimated decreases in coal consumption (-1.8 %) and estimated growth in apparent oil (+4.0 %) and natural gas (+7.2 %) consumption and in cement production (+2.6 %). The uncertainty range considers the spread between different data sources, as well as differences between July/August and end-of-year data observed in 2014 and 2015. The estimated reduction in coal consumption also incorporates an assumed 2 % increase in the energy density of coal – based on increases in the last 2 years, which are assumed to continue given production limits in 2016 that are likely to affect production of low-quality coal more – and the uncertainty range also reflects uncertainty in this figure.

For the USA, the EIA emissions projection for 2016 combined with cement data from USGS gives a decrease of -1.4 % (range of -3.7 to +0.9 %) compared to 2015, or -1.7 % (range of -4.0 to +0.6 %) after adjusting for leap year (see Sect. 2.1.3).

For the rest of the world, the expected growth for 2016 of +1.3 % (range of -0.2 to +2.8 %), or +1.0 % (range of -0.4 to +2.5 %) after adjusting for leap year (see Sect. 2.1.3). This is computed using the GDP projection for the world excluding China and the USA of 2.5 % made by the IMF (IMF, 2016) and a decrease in IFF of -1.2 % yr-1, which is the average from 2006 to 2015. The uncertainty range is based on the standard deviation of the interannual variability in IFF during 2006–2015 of ±1.0 % yr-1 and our estimate of uncertainty in the IMF's GDP forecast of ±0.5 %.

The growth in atmospheric CO2 concentration (GATM) was projected to be high again in 2016, at 6.7 ± 1.1 GtC (3.15 ± 0.53 ppm) for the Mauna Loa station (Betts et al., 2016). Growth at Mauna Loa is closely correlated with the global growth (r=0.95 for 1959–2015). Therefore, the global growth rate in atmospheric CO2 concentration is also expected to be high in 2016. In the 8-month period between December 2015 and August 2016, the observed global growth in atmospheric CO2 concentration was already 2.3 ppm (Dlugokencky and Tans, 2016) after seasonal adjustment, supporting the projection of Betts et al. (2016). Even with a return to El Niño neutral or possible emerging La Niña conditions for the second half of 2016, positive growth in atmospheric CO2 would still be expected during the last 4 months of the year because of the continuing persistent emissions. For example, during the transitions from El Niño to La Niña of 1986–1987, 1998–1999, and 2010–2011, atmospheric CO2 growth of 0.3, 0.6, and 0.9 ppm, respectively, was observed in the last 4 months of the year.

Combining projected EFF and GATM suggests a total for the combined land and ocean (SLAND+SOCEAN-ELUC) of about 3 GtC only. SOCEAN was 3.0 GtC in 2015 and is expected to slightly increase in 2016 from a delayed response to El Niño conditions (Feely et al., 1999). ELUC was 1.3 GtC in 2015, above the decadal mean average of 1.0 GtC yr-1, and is expected to return to average or below average in 2016 based on fire activity related to land management so far (up to August). Hence, for 2016, the residual land sink SLAND is expected to be well below its 2006–2015 average and approximately balance ELUC. This is consistent with our understanding of the response of the terrestrial vegetation to El Niño conditions and increasing atmospheric CO2 concentrations, though the uncertainties in GATM and of the partitioning among SLAND and SOCEAN are substantial.