A comprehensive dataset for global, regional and national greenhouse gas emissions by sector 1970-2019

25 To track progress towards keeping warming well below 2°C, as agreed upon in the Paris Agreement, comprehensive and reliable information on anthropogenic sources of greenhouse gas emissions (GHG) is required. Here we provide a dataset on anthropogenic GHG emissions 1970-2019 with a broad country and sector coverage. We build the dataset from recent releases of the “Emissions Database for Global Atmospheric Research” (EDGAR) for CO2 emissions from fossil fuel combustion and industry (FFI), CH4 emissions, N2O emissions, and fluorinated gases, and use a well-established fast-track method to extend 30 this dataset from 2018 to 2019. We complement this with data on net CO2 emissions from land use, land-use change and forestry (LULUCF) from three bookkeeping models. We provide an assessment of the uncertainties in each greenhouse gas at the 90% confidence interval (5-95 percentile) by combining statistical analysis and comparisons of global emissions inventories with an expert judgement informed by the relevant scientific literature. We identify important data gaps: CH4 and N2O emissions could be respectively 10-20% higher than reported in EDGAR once all emissions are accounted. F-gas 35 emissions estimates for individual species in EDGARv5 do not align well with atmospheric measurements and the F-gas total exceeds measured concentrations by about 30%. However, EDGAR and official national emission reports under the UNFCCC do not comprehensively cover all relevant F-gas species. Excluded F-gas species such as chlorofluorocarbons (CFCs) or https://doi.org/10.5194/essd-2021-228 O pe n A cc es s Earth System Science Data D icu ssio n s Preprint. Discussion started: 14 July 2021 c © Author(s) 2021. CC BY 4.0 License.

hydrochlorofluorocarbons (HCFCs) are larger than the sum of the reported species. GHG emissions in 2019 amounted to 59±6.6 GtCO2eq: CO2 emissions from FFI were 38±3.0 Gt, CO2 from LULUCF 6.6±4.6 Gt, CH4 11±3.3 GtCO2eq, N2O 40 2.4±1.5 GtCO2eq and F-gases 1.6±0.49 GtCO2eq. Our analysis of global, anthropogenic GHG emission trends over the past five decades  highlights a pattern of varied, but sustained emissions growth. There is high confidence that global anthropogenic greenhouse gas emissions have increased every decade. Emission growth has been persistent across different (groups of) gases. While CO2 has accounted for almost 75% of the emission growth since 1970 in terms of CO2eq as reported here, the combined F-gases have grown at a faster rate than other GHGs, albeit starting from low levels in 1970. Today, F-45 gases make a non-negligible contribution to global warmingeven though CFCs and HCFCs, regulated under the Montreal Protocol and not included in our estimates, have contributed more. There is further high confidence that global anthropogenic GHG emission levels were higher in 2010-2019 than in any previous decade and GHG emission levels have grown across the most recent decade. While average annual greenhouse gas emissions growth slowed between 2010-2019 compared to [2000][2001][2002][2003][2004][2005][2006][2007][2008][2009], the absolute increase in average decadal GHG emissions from the 2000s to the 2010s has been the largest since the 50 1970sand within all human history as suggested by available long-term data. We note considerably higher rates of change in GHG emissions between 2018 and 2019 than for the entire decade 2010-2019, which is numerically comparable with the period of high GHG emissions growth during the 2000s, but we place low confidence in this finding as the majority of the growth is driven by highly uncertain increases in CO2-LULUCF emissions as well as the use of preliminary data and extrapolation methodologies for these most recent years. While there is a growing number of countries today on a sustained 55 emission reduction trajectory, our analysis further reveals that there are no global sectors that show sustained reductions in GHG emissions. We conclude by highlighting that tracking progress in climate policy requires substantial investments in independent GHG emission accounting and monitoring as well as the available national and international statistical infrastructures. The data associated with this article (Minx et al., 2021)

Overview
Our dataset provides a comprehensive set of estimates for global anthropogenic GHG emissions disaggregated by 30 economic 100 sectors and 226 countries. The focus of the data is on anthropogenic GHG emissions originally regulated under the Kyoto Protocol: natural sources and sinks are not considered, and nor are ozone depleting substances regulated under the Montreal Protocol such as chlorofluorocarbons (CFCs) and hydrochlorofluorocarbons (HCFCs). We distinguish five groups of gases: (1) CO2 emissions from fossil fuel combustion and industry (CO2-FFI); (2) CO2 emissions from land use, land-use change and forestry (CO2-LULUCF); (3) methane emissions (CH4); (4) nitrous oxide emissions (N2O); (5) the group of fluorinated gases 105 (F-gases) comprising hydrofluorocarbons (HFCs), perfluorocarbons (PFCs) as well as sulphur hexafluoride (SF6). We do not cover NF3 emissions [NOTE TO REVIEWERS: Our update of F-gas emissions to EDGAR v6 will also add NF3 emissions to our data], which are also covered under the Paris Agreement. We provide and analyse the GHG emissions data both in native units as well as in CO2-equivalents (see Section 3.7) as commonly done in wide parts of the climate change mitigation community using global warming potentials with a 100 year time horizon from the IPCC Fifth Assessment Report (Myhre et 110 al., 2013). We briefly discuss the impact of alternative metric choices in tracking aggregated GHG emissions over the past few decades and compare the emissions with estimated anthropogenic warming.
We report the annual growth rate in emissions E for adjacent years (in percent per year) by calculating the difference between the two years and then normalizing to the emissions in the first year: ((E(t0+1)-Et0)/Et0)×100. We apply a leap-year adjustment 115 where relevant to ensure valid interpretations of annual growth rates. This affects the growth rate by about 0.3%yr -1 (1/366) and causes calculated growth rates to go up by approximately 0.3% if the first year is a leap year and down by 0.3% if the second year is a leap year. We calculate the relative growth rate in percent per year for multi-year periods (e.g. a decade) by fitting a linear trend to the logarithmic transformation of E across time (see Friedlingstein et al., 2020).

120
Our dataset draws from three underlying sources: (1) the full EDGARv6 release for CO2-FFI as well as non-CO2 GHGs covering the time period 1970-2018 . Note that currently F-gas data from EDGARv6 is still being prepared.
In the meantime, we use EDGARv5 data covering the time period 1970-2015 (Crippa et al., 2019); (2) EDGAR fast-track extensions for CO2-FFI, CH4 and N2O emissions for 2019 as well as 2016-2019 for F-gas emissions based on Olivier et al. (2005) and Crippa et al. (2020) [NOTE TO REVIEWERS: F-gas emissions in EDGARv6 are currently being revised and will 125 be included in the revised version of this manuscript. F-gases will then also have a fast-track extension from 2018 to 2019]; (3) CO2-LULUCF as the average of three bookkeeping models, consistent with the approach of the global carbon project (Friedlingstein et al., 2020). As shown in https://doi.org/10.5194/essd-2021-228  Changes (AFOLU). We devise a classification for assigning our 226 countries to regions, combining the standard Annex I/non-Annex I distinction with geographical location. We provide other common regional classifications from the UN and the World Bank as part of the supplementary files. The dataset including the sector and region classification can be found at https://doi.org/10.5281/zenodo.5053056.

The Emissions Database for Global Atmospheric Research (EDGAR) 140
EDGAR emission estimates included in our dataset are derived from two methodologies: a) full bottom-up emission inventory data; b) fast-track emission inventory data imputed from incomplete input data. As described in Janssens-Maenhout et al.
(2019), the EDGAR bottom-up emission inventory estimates are calculated from international activity data and emission factors following the 2006 IPCC Guidelines for National Greenhouse Gas Inventories (IPCC, 2006) -updated according to the latest scientific knowledge. Emissions (EMs) from a given sector i in a country C accumulated during a year t for a chemical 145 compound x are calculated with the country-specific activity data (AD), quantifying the activity in sector i, with the mix of j technologies (TECH) and with the mix of k (end-of-pipe) abatement measures (EOP) installed with the share k for each technology j, the emission rate with an uncontrolled emission factor (EF) for each sector i and technology j and relative reduction (RED) by abatement measure k, as summarised in the following formula: The activity data are sector dependent and vary from fuel combustion in energy units (TJ) of a particular fuel type, to the amount (ton) of products manufactured, or to the number of animals or the area (hectares) or yield (ton) of cultivated crops.
The technology mixes, (uncontrolled) emission factors and end-of-pipe measures are determined at different levels: countryspecific, regional, country group (e.g. Annex I/non-Annex I), or global. Technology-specific emission factors are used to 155 enable an IPCC tier-2 approach, taking into account the different management and /technology processes or infrastructures (e.g., different distribution networks) under specific "technologies", and modelling explicitly abatements/ emission reductions, e.g. the CH4 recovery from coal mine gas at country level under the "end-of-pipe measures". As with national inventories, emissions are accounted over a period of one calendar year in the country in which they took place (i.e. a territorial accounting principle) (IPCC, 2006(IPCC, , 2019. A full description of data sources and methodology for EDGARv6 is provided in Crippa et al. 160 (2021).
Extensions to 2019 are derived using a "fast-track methodology", which is designed to update full EDGAR inventories to more recent years based on less information Olivier et al., 2005;Olivier and Peters, 2020). The underlying idea is to extrapolate emissions trends based on observed activity trends in key sectors. For CO2-FFI emissions, the fast track 165 estimates were based on the latest BP coal, oil and natural gas consumption data (BP, 2019). Updates for cement, lime, ammonia and ferroalloys production are based on US Geological Survey statistics, urea production and consumption on statistics from the International Fertilizer Association, gas used from flaring on data from the Global Gas Flaring Reduction Partnership, steel production on statistics from the World Steel Association, and cement clinker production on UNFCCC data.
For methane and nitrous oxide emissions, fast-track extensions are based on detailed agricultural statistics from FAO (CH4 170 and N2O), fuel production and transmission statistics from IEA and BP (CH4) as well as UNFCCC-CRF data for Annex I countries on coal production (CH4 recovery) and the production of chemicals (N2O abatement). Finally, for F-gas emissions, a more extensive fast-track extension covering 2016-2019 was undertaken. For Annex I countries, these fast-track extensions were based on the most recent national emission inventories, submitted under the UNFCCC (up to 2018). For all remaining countries and years, simple extrapolation was used given the absence of international statistics. Available fast-track data is 175 from EDGARv5, which we link to the full EDGARv6 release by calculating the county and sector specific emissions growth between 2018 and 2019 and multiplying it with the 2018 values in our data.

Accounting for CO2 emissions Land Use, Land-Use Change and Forestry (CO2-LULUCF)
We consider all fluxes of CO2 from land use, land-use change and forestry. This includes CO2 fluxes from the clearing of (including harvest activity), shifting cultivation (cycles of forest clearing for agriculture, then abandonment), and regrowth of forests and other natural vegetation following wood harvest or abandonment of agriculture, and emissions from peat burning and drainage. Some of these activities lead to emissions of CO2 to the atmosphere, while others lead to CO2 sinks. CO2-LULUCF therefore is the net sum of emissions and removals from all human-induced land use changes and land management. 185 Note that CO2-LULUCF is referred to as (net) land-use change emissions, ELUC, in the context of the global carbon budget (Friedlingstein et al., 2020). Agriculture per se, apart from conversions between different agricultural types, does not lead to substantial CO2 emissions as compared to land-use changes such as clearing or regrowth of natural vegetation. Therefore, CO2 fluxes in the AFOLU sector refer mostly to forestry and other land use (changes), while the agricultural part of the sector is covered by CH4 and N2O emissions. 190 Since in reality anthropogenic CO2-LULUCF emissions co-occur with natural CO2 fluxes in the terrestrial biosphere, models have to be used to distinguish anthropogenic and natural fluxes (Friedlingstein et al., 2020). CO2-LULUCF as reported here is calculated via a bookkeeping approach, as originally proposed by Houghton et al. (2003), tracking carbon stored in vegetation and soils before and after land-use change. Response curves are derived from literature and observations to describe the 195 temporal evolution of the decay and regrowth of vegetation and soil carbon pools for different ecosystems and land use transitions, including product pools of different lifetimes. These dynamics distinguish bookkeeping models from the common approach of estimating "committed emissions" (assigning all present and future emissions to the time of the land use change event), which is frequently derived from remotely-sensed land use area or biomass observations (Ramankutty et al., 2007).
Most bookkeeping models also represent the long-term degradation of primary forest as lowered standing vegetation and soil 200 carbon stocks in secondary forests, and include forest management practices such as wood harvesting.
Following the approach taken by the global carbon budget (Friedlingstein et al., 2020), we take the average of three bookkeeping estimates: the bookkeeping of land use emissions model, BLUE (Hansis et al., 2015), H&N (Houghton and Nassikas, 2017), and OSCAR (Gasser et al., 2020). Key differences across these estimates, including land-use forcing, are 205 summarised in Table 4. Since bookkeeping models do not include emissions from organic soils, emissions from peat fires and peat drainage are added from external datasets: Peat burning is based on the Global Fire Emission Database (GFED4s; (van der Werf et al., 2017)) and introduces large interannual variability to the CO2-LULUCF emissions due to synergies of landuse and climate variability particularly in Southeast Asia, strongly noticeable during El-Niño events such as 1997. Peat drainage is based on estimates by Hooijer et al. (2010) for Indonesia and Malaysia in H&N, and added to BLUE and OSCAR 210 from the global FAO data on organic soils emissions from croplands and grasslands (Conchedda and Tubiello, 2020).

Uncertainties in GHG emission estimates
Estimates of historic GHG emissions -CO2, CH4, N2O and F-gasesare uncertain to different degrees. Assessing and reporting uncertainties is crucial in order to understand whether available estimates are sufficiently accurate to answer, for example, whether GHG emissions are still rising, or if a country has achieved an emission reduction goal (Marland, 2008). These 215 uncertainties can be of a scientific nature, such as when a process is not sufficiently understood. They also arise from incomplete or unknown parameter information (activity data, emission factors etc.), as well as estimation uncertainties from imperfect modelling techniques. There are at least three major ways to examine uncertainties in emission estimates (Marland et al., 2009): 1) by comparing estimates made by independent methods and observations (e.g. comparing top-down vs bottomup estimates; modelling against remote sensing data) (Petrescu et al., 2020c(Petrescu et al., , 2020aSaunois et al., 2020;; 2) 220 by comparing estimates from multiple sources and understanding sources of variation (Andres et al., 2012;Andrew, 2020a;Ciais et al., 2021;Macknick, 2011); 3) by evaluating multiple estimates from a single source (e.g. Hoesly and Smith, 2018) including approaches such as uncertainty ranges estimated through statistical sampling across parameter values, applied for example at the country or sectoral level (e.g. Andres et al., 2014;Monni et al., 2007;Solazzo et al., 2021), or to spatially distributed emissions (Tian et al., 2019). This section assesses the relevant peer-reviewed literature on uncertainties in historic 225 GHG emission estimates and places an expert judgement on the uncertainties for the different (groups of) GHGs. Uncertainty estimates can be rather different depending on the method chosen. For example, the range of estimates from multiple sources is bounded by their interdependency; they can be lower than true structural plus parameter uncertainty estimates or than estimates made by independent methods. In particular it is important to account for potential bias in estimates, 230 which can result from using common methodological or parameter assumptions across estimates, or from missing sources, which can result in a systemic bias in emission estimates (see N2O discussion below). Independent top-down observational constraints are, therefore, particularly useful to bound total emission estimates (Petrescu et al., 2020c(Petrescu et al., , 2020b. Solazzo et al. (2021) evaluated the uncertainty of the EDGAR's source categories and their totals for all the main GHGs (CO2-235 FFI, CH4, N2O). The study is based on the propagation of the uncertainty associated with input parameters (activity data and emission factors) as estimated by expert judgement (tier-1) and complied by IPCC (2006IPCC ( , 2019. A key methodological challenge is determining how well uncertain parameters are correlated between sectors, countries, and regions. The more highly correlated parameters (e.g. emission factors) are across scales, the higher the resulting overall uncertainty estimate. Solazzo et al. (2020) assume full covariance between same source categories where similar assumptions are being used, and 240 independence otherwise. For example, they assume full covariance where the same emission factor is used between countries or sectors, while assuming independence where country-specific emission factors are used. This strikes a balance between extreme assumptions (full independence or full covariance in all cases) that are likely unrealistic, but still leans towards higher uncertainty estimates. When aggregating emission sources, assuming covariance increases the resulting uncertainty estimate. Uncertainties calculated with this methodology tend to be higher than the range of values from ensemble of dependant 245 inventories (Saunois et al., 2016(Saunois et al., , 2020. The uncertainty of emission estimates derived from ensembles of gridded results from bio-physical models (Tian et al., 2018) adds an additional dimension of spatial variability, and is therefore not directly comparable with aggregate country or regional uncertainty, estimated with the methods discussed above.
We adopt a 90% confidence interval (5 th -95 th percentile) to report the uncertainties in our GHG emissions estimates, i.e., there 250 is a 90 % likelihood that the true value will be within the provided range if the errors have a Gaussian distribution, and no bias is assumed. This is in line with previous reporting in IPCC AR5 (Blanco G. et al., 2014;Ciais et al., 2014). The uncertainties reported here combine statistical analysis, comparisons of global emissions inventories and expert judgement of the likelihood of results lying outside this range, rooted in an understanding gained from the relevant literature. At times, we also use a qualitative assessment of confidence levels to characterize the annual estimates from each term based on the type, amount, 255 quality, and consistency of the evidence as defined by the IPCC (IPCC, 2014).

CO2 emissions from fossil fuels and industrial processes
Several studies have compared estimates of annual CO2-FFI emissions from different global inventories (Andres et al., 2012;Andrew, 2020a;Gütschow et al., 2016;Janssens-Maenhout et al., 2019;Macknick, 2011;Petrescu et al., 2020c). However, 260 estimates are not fully independent as they all ultimately rely on many of the same data sources. For example, all global inventories use one of four global energy datasets to estimate CO2 emissions from energy use, and these energy datasets themselves all rely on the same national energy statistics, with few exceptions (Andrew, 2020a). Divergence between these estimates (see Figure 1) are mainly related to differences in the estimation methodology, conversion factors, emission coefficients, assumptions about combustion efficiency, and calculation errors (Andrew, 2020a;Marland et al., 2009 also Table 1 for further information on the inventories). Another major source of divergence between datasets is differences in their respective system boundaries (Andres et al., 2012;Andrew, 2020a;Macknick, 2011). Hence, differences across CO2-FFI emissions estimates do not reflect full uncertainty due to source data dependencies. At the same time, the observed range across estimates from different databases exaggerates uncertainty, to the extent that they largely originate in 270 system boundary differences (Andrew, 2020a;Macknick, 2011).   Figure 1). However, this variability is almost halved when system boundaries are harmonised (Andrew, 2020a).
EDGARv6 CO2-FFI emissions as used in this report track at the top of the range as shown in Figure 1. This is partly due to the comprehensive system boundaries of EDGAR, but also due to the assumption of 100% oxidation of combusted fuels as per IPCC default assumptions. Once system boundaries are harmonised EDGAR continues to track at the upper end of the range, but no longer at the top. EDGAR CO2 FFI estimates are further well-aligned with emission inventories submitted by 285 Annex I countries to the UNFCCCeven though some variation can occur for individual countries (Andrew, 2020a). Uncertainties in CO2-FFI emissions arise from the combination of uncertainty in activity data and uncertainties in emission 290 factors including assumptions for combustion completeness and non-combustion uses. CO2-FFI emissions estimates are largely derived from energy consumption activity data, where data uncertainties are comparatively small due to well established statistical monitoring systems, although there are larger uncertainties in some countries and time periods (Andres et al., 2012;Andrew, 2020a;Ballantyne et al., 2015;Janssens-Maenhout et al., 2019;Macknick, 2011). Most of the underlying uncertainties are systematic and related to underlying biases in the energy statistics and accounting methods used 295 (Friedlingstein et al., 2020). Uncertainties are lower for fuels with relatively uniform properties such as natural gas, oil or gasoline and higher for fuels with more diverse properties, such as coal (IPCC 2006;Blanco G. et al. 2014). Uncertainties in CO2 emissions estimates from industrial processes, i.e. non-combustive oxidation of fossil fuels and decomposition of carbonates, are higher than for fossil fuel combustion. At the same time, products such as cement also take up carbon over their life cycle, which are often not fully considered in carbon balances (Guo et al., 2021;Sanjuán et al., 2020;Xi et al., 2016). 300 However, recent versions of the global carbon budget include specific estimates for the cement carbonation sink and estimate average annual CO2 uptake at 0.70 GtCO2 for 2010-2019 (Friedlingstein et al., 2020).
Uncertainties of energy consumption data (and, therefore, CO2-FFI emissions) are generally higher for the first year of their publication when less data is available to constrain estimates. In the BP energy statistics, 70% of data points are adjusted by 305 an average of 1.3% of a country's total fossil fuel use in the subsequent year with further more modest revisions later on (Hoesly and Smith, 2018). Uncertainties are also higher for developing countries, where statistical reporting systems do not have the same level of maturity as in many industrialised countries (Andres et al., 2012;Andrew, 2020b;Friedlingstein et al., 2019Friedlingstein et al., , 2020Gregg et al., 2008;Guan et al., 2012;Janssens-Maenhout et al., 2019;Korsbakken et al., 2016;Marland, 2008).
Example estimates of uncertainties for CO2 emissions from fossil fuel combustion at the 95% confidence interval are ±3-5% 310 for the U.S., ±15 -±20% for China and ±50% or more for countries with poorly developed or maintained statistical infrastructure (Andres et al., 2012;Gregg et al., 2008;Marland et al., 1999). However, these customary country groupings do not always predict the extent to which a country's energy data has undergone historical revisions (Hoesly and Smith, 2018).
Uncertainties in CO2-FFI emissions before the 1970s are higher than for more recent estimates. Over the last two to three decades uncertainties have increased again because of increased production in some developing countries with less rigorous 315 statistics and more uncertain fuel properties (Ballantyne et al., 2015;Friedlingstein et al., 2020;Marland et al., 2009).
The global carbon project (Friedlingstein et al., 2019(Friedlingstein et al., , 2020Le Quéré et al., 2018) assesses uncertainties in global anthropogenic CO2-FFI emissions estimates within one standard deviation (1σ) as ±5% (±10% at 2σ). This is broadly consistent with the ±8.4% uncertainty estimate for CDIAC (Andres et al., 2014) as well as the ±7 -±9% uncertainty estimate for 320 EDGARv4.3.2 and v5 (Janssens-Maenhout et al., 2019;Solazzo et al., 2021) at 2σ. It remains at the higher end of the ±5% -±10% range provided by Ballantyne et al. (2015). Consistent with the above uncertainty assessments, we present uncertainties for global anthropogenic CO2 emissions at ±8% for a 90% confidence interval in line with IPCC AR5 and the UN emissions gap report (Blanco G. et al., 2014;UNEP, 2020). 1 respectively. For 1990-2019 maximum average differences are 9.1 and 7.8 GtCO2 yr -1 for gross emissions and removals, respectively (Friedlingstein et al., 2020). Note that 2016-2019 is extrapolated in H&N and 2019 in OSCAR based on the anomalies of the net flux for the gross fluxes. Differences in the models underlying this observed variability are reported in Table 4. In the longer term, a consistent general upward trend since 1850 across models is reversed during the second part of the 20th century. Since the 1980s, however, differing trends across models are related to, among other things, different land-350 use forcings (Gasser et al., 2020). Further differences between BLUE and H&N can be traced in particular to: (1) differences in carbon densities between natural and managed vegetation, or between primary and secondary vegetation; (2) a higher allocation of cleared and harvested material to fast turnover pools in BLUE compared to H&N; and (3) to the inclusion subgrid scale transitions .

355
Uncertainties in CO2-LULUCF emissions can be more comprehensively assessed through comparisons across a suite of dynamic global vegetation models (DGVM) (Friedlingstein et al., 2020). DGVM models are not combined in the CO2-LULUCF mean estimate in our data because the typical DGVM setup includes the loss of additional sink capacity, which makes up about 40% of the DGVM estimate in recent years (Obermeier et al., 2020) and is excluded in bookkeeping estimates.
Nonetheless, a CO2-LULUCF estimate from the DGVM multi-model mean remains consistent with the average estimate from 360 the bookkeeping models, as shown in Figure 1. Variation across DGVMs is large with a standard deviation at around 1.8 GtCO2 yr -1 , but is still smaller than the average difference between bookkeeping models at 2.6 GtCO2 yr -1 as well as the current estimate of H&N (Houghton and Nassikas, 2017) and its previous model versions . DGVMs differ in methodology, input data and how comprehensively they represent land-use-related processes. In particular land management, such as crop harvesting, tillage, or grazing (all implicitly included in observation-based carbon densities of bookkeeping 365 models) can alter CO2 flux estimates substantially, but are included to varying extents in DGVMs, thus increasing model spread (Arneth et al., 2017). For all types of models, land-use forcing is a major determinant of emissions and removals, and its high uncertainty impacts CO2-LULUCF estimates . The reconstruction of land-use change of the historical past, which has to cover decades to centuries of legacy LULUCF fluxes, is based on sparse data or proxies (Hurtt et al., 2020;Klein Goldewijk et al., 2017), while satellite-based products suffer from complications in distinguishing natural from 370 anthropogenic drivers (Hansen et al., 2013;Li et al., 2018) or accounting for small-scale disturbances and degradation (Matricardi et al., 2020). Lastly, regional carbon budgets can be substantially over-or underestimated when carbon embodied in trade products is not accounted for (Ciais et al., 2021).
We base our uncertainty assessment on Friedlingstein et al. (2020) and take ±2.6 GtCO2 yr -1 as a best-value judgement for the 375 ±1σ uncertainty range (thus ±5.1 GtCO2 yr -1 for ±2σ) in CO2-LULUCF emissions, constant over the last decades. This absolute uncertainty estimate presented above corresponds roughly to a relative uncertainty of about ±50% over 1970-2019, which is much higher than for most fossil-emission terms, but reflects the large model spread and large differences between the current estimate of H&N and its previous model version . This corresponds to a relative uncertainty of about ±80% for a 90% confidence interval (5 th -95 th percentile) and is larger but still broadly in line with the upper end of the relative 380 uncertainty of ±50 -±75% considered in AR5 (Blanco G. et al., 2014). Much larger uncertainties in CO2-LULUCF emissions have been identified across the literature, but were traced back to different definitions used in various modelling frameworks (Pongratz et al., 2014) as well as inventory data (Grassi et al., 2018). Overall, we use a relative uncertainty estimate of about ±70% for a 90% confidence interval. This recognizes the choice of a constant absolute uncertainty estimate taken elsewhere (Friedlingstein et al., 2020) and in recognition of a possible trend towards higher CO2-LULUCF emissions estimates in more 385 recent years.
Finally, note that attempts to constrain the estimates of CO2-LULUCF emissions by observed biomass densities have been undertaken, but were successful only in some non-tropical regions . While providing valuable independent and observation-driven information, remote-sensing derived estimates have limited applicability for model evaluation for the total 390 CO2-LULUCF flux, since they usually only quantify vegetation biomass changes and exclude legacy emissions from the presatellite era. Further, with the exception of the (pan-tropical) estimates by Baccini et al. (2012) they either track committed instead of actual emissions (e.g. Tyukavina et al., 2015), combine a static carbon density map with forest cover changes, or include the natural land sink (e.g. Baccini et al., 2017) to infer fluxes directly from the carbon stock time seriesnone of which fully distinguishes natural from anthropogenic disturbances. 395 Literature: a (Hansis et al., 2015); b (Houghton and Nassikas, 2017); c (Gasser et al., 2020); d (Hurtt et al., 2020); e (Chini et al., 2020); f (Nations, 2015); g (Conchedda and Tubiello, 2020); h (Hooijer et al., 2010)

Anthropogenic CH4 emissions 405
About 60% of total global methane emissions come from anthropogenic sources (Saunois et al., 2020). These are linked to a range of different sectors: agriculture, fossil production and use, waste as well as biomass and biofuel burning. Methane emissions can be derived either using bottom-up (BU) estimates that rely on anthropogenic inventories such as EDGAR (Janssens-Maenhout et al., 2019), land surface models that infer part of natural emissions (Wania et al., 2013) or observationbased upscaling for some specific sources such as geological sources (e.g. Etiope et al., 2019). Alternatively, top-down (TD) 410 approaches can be used, such as atmospheric transport models that assimilate methane atmospheric observations to estimate past methane emissions (Houweling et al., 2017). Some TD systems aim to optimize certain emission sectors based on differences in their spatial and temporal distributions (e.g. Bergamaschi et al., 2013), while other only solve for net emissions at the surface. Then the partitioning of TD posterior (output) fluxes between specific source sectors (e.g. Fossil vs. BB&F) is carried out with various degrees of uncertainty depending of the methods and the degree of refinement of sectors, but often 415 rely on ratios from the prior knowledge of fluxes. Comprehensive assessments of methane sources and sinks have been provided by Saunois et al. (2016Saunois et al. ( , 2020 and Kirschke et al. (Kirschke et al., 2013).
However, they may differ in the assumptions and data used for the calculation. While the US-EPA inventory uses the reported emissions by the countries to UNFCCC, other inventories produce their own estimates using a consistent approach for all countries, and country specific activity data, emission factor and technological abatement when available. EDGAR mostly apply a Tier 1 approach to estimate methane emissions while GAINS uses a Tier 2 approach (Höglund-Isaksson et al., 2020). CEDS is based on pre-existing emission estimates from FAOSTAT and EDGAR and then scales these emissions to match country-specific inventories, largely those reported to UNFCCC.
Global anthropogenic CH4 emission estimates are compared in Figure 1. EDGARv5 has revised total global CH4 emissions 435 about 10 Mt CH4 yr -1 higher than EDGARv4.3.2 due to a higher estimate for the waste sector (see supplementary material).
Subsequent revisions of the estimation methodology in EDGARv6 in alignment with the IPCC guidelines refinement (IPCC, 2019) lead to very substantial differences in total CH4 emissions of up to 50 MtCH4yr -1 before the 1990s, but these differences are smaller ranging from 1-13 MtCH4yr -1 since the 2000s. The cause of these differences is a new procedure to separately  Table 5, uncertainties in total global methane emissions across all anthropogenic and natural sources 450 are comparatively small at ±6% -a range larger than errors in transport models only (Locatelli et al., 2015). However, uncertainty in the chemical sink was not fully considered in the TD estimates in Saunois et al (2020). Uncertainty on the global burden of OH is about 10-15%, which translates to an uncertainty of approximately ±9% on total global emissions . Based on both TD and BU ensemble, uncertainty (reported as the minimum-maximum range across estimates) on the global anthropogenic methane emissions is about ±10% to ±30% depending on the category, with larger uncertainty in the 455 fossil fuel sectors than in the agriculture and waste sector (Saunois et al., 2020). However, these uncertainties are underestimated as they do not consider the uncertainty in each individual estimate, which includes potential uncertainties in activity data, emission factors, and equations used to estimate emissions.
Uncertainties in EDGARv5 CH4 emissions using a Tier 1 approach are estimated at -33% to +46% at 2σ, but there is great 460 variability across individual sectors ranging from ±30% (agriculture) to more than ±100% (fuel combustion), with high uncertainties in oil and gas sector (±93%) and coal fugitive emissions (±65%) . Inventories at national scale, such as in the USA also show large uncertainties depending on the sector (NASEM, 2018), though the activity data uncertainty may be lower than those for less developed countries. For example, global inventories, such as EDGAR, estimate   The most recent UN emissions gap report (UNEP, 2020) gives an uncertainty range for global anthropogenic methane 480 emissions with one standard deviation of ±30% (i.e. ±60% for 2σ), which is slightly higher than recent estimates in the literature. On the other hand, IPCC AR5 provides a comparatively low estimates at ±20% for a 90% confidence interval.
Overall, we apply a best value judgment of ±30% for global anthropogenic methane emissions for a 90% confidence interval. This is justified by the large uncertainties reported in the methane budgets (Kirschke et al., 2013;Saunois et al., 2016Saunois et al., , 2020 as well as for FAO activity statistics by Tubiello et al. (Tubiello et al., 2015), is broadly in line with the uncertainties quantified 485 for EDGARv5.

Anthropogenic N2O emissions
Anthropogenic N2O emissions occur in a number of sectors, namely agriculture, fossil fuel and industry, biomass burning, and waste. The agriculture sector consists of four components: direct and indirect emissions from soil and water bodies (inland, 490 coastal, and oceanic waters), manure left on pasture, manure management, and aquaculture. Besides these main sectors, a final 'other' category represents the sum of the effects of climate, elevated atmospheric CO2, and land cover change. This is a new sector that was developed as part of the global nitrous oxide budget )a recent assessment to quantify all sources and sinks of N2O emissions, updating previous work (Kroeze et al., 1999;Mosier et al., 1998;Mosier and Kroeze, 2000;Syakila and Kroeze, 2011). Overall, anthropogenic sources contributed just over 40% to total global N2O emissions 495 .
There are a variety of approaches for estimating N2O emissions. These include inventories (Janssens-Maenhout et al., 2019;Tian et al., 2018;Tubiello et al., 2013), statistical extrapolations of flux measurements (Wang et al., 2020), andas processbased land and ocean modelling (Tian et al., 2019;Yang et al., 2020). There are at least five relevant global N2O emissions 500 inventories available: EDGAR (Crippa et al., 2019Janssens-Maenhout et al., 2019), GAINS (Höglund-Isaksson, 2012), FAO-N2O (Tubiello, 2018;Tubiello et al., 2013), CEDS (Hoesly et al., 2018;McDuffie et al., 2020;O'Rourke et al., 2020) and GFED . While EDGAR and GAINS cover all sectors except biomass burning, FAOSTAT-N2O is in EDGARv6 lead to an average increase of 1.5% yr -1 in total N2O emissions estimates between 1999 and 2018 compared to EDGARv5 (differences before 1999 were negligible at less than 1% yr -1 ). Differences across different versions of the EDGAR dataset are shown in the Supplementary Material ( Fig. SM-1). The main discrepancies across different global inventories are in agriculture, where emission estimates from the global nitrous oxide budget (also referred to as "GCP")  and FAOSTAT are on average 1.5 Mt N2O yr -1 higher than those from GAINS and EDGAR during 1990-2016, due to much 510 higher estimates of direct emissions from fertilised soils and manure left on pasture. GCP provides the largest estimate, because it synthesised from the other three inventories and further informed by additional bottom-up modelling estimatesand is as such more comprehensive in scope (Figure 1). In particular, it includes an additional sector that considers the sum of the effects of climate, elevated atmospheric CO2, and land cover change . EDGAR estimates of anthropogenic N2O emissions as used in this dataset should therefore be considered as lower bound estimates. 515 Anthropogenic N2O emissions estimates are subject to considerable uncertaintylarger than those from FFI-CO2 or CH4 emissions. N2O inventories suffer from high uncertainty on input data, including fertiliser use, livestock manure availability, industry, and waste. In the recent Emissions Gap Report (UNEP, 2020) relative uncertainties for global anthropogenic N2O emissions are estimated at ±50% for a 68% (1σ) confidence interval. This is larger than the ±60% uncertainties reported in IPCC AR5 for a 90% confidence interval (Blanco G. et al., 2014), but is comparable with the ranges for anthropogenic emissions in the global N2O budget . Overall, we assess the relative uncertainty for global anthropogenic N2O emissions at ±60% for a 90% confidence interval.     Where available, the various top-down estimates agree with each other within uncertainties. The magnitude of the difference between WMO (2018) and EDGARv5 estimates varies markedly between species; for CF4, the median annual ratio between the top-down and bottom-up estimates is close to 1.0, whereas for c-C4F8 it is more than 100. Such differences have been 575 previously noted, for example, by Mühle, et al. (2019) as well as in some earlier papers. For SF6, the relatively close agreement between a previous version of EDGAR (v4) and a top-down estimate has been discussed in Rigby, et al. (2010). They estimated uncertainties in EDGARv4 of ±10% to ±15%, depending on the year, and indeed, top-down values were consistent within these uncertainties. For CF4, these is close agreement between EDGARv4 and atmospheric observations after 1991, while for C2F6 there is closer agreement before 1991 (Mühle et al., 2010). This remains the case here for EDGARv5. However, it should 580 https://doi.org/10.5194/essd-2021-228 be noted that some assumptions within EDGAR had previously been validated against atmospheric observations, hence EDGARv4 might be considered a hybrid of top-down and bottom-up methodologies for these species, as some parameters may have been chosen based on comparison with atmospheric observations. Mühle, et al. (2010) noted a substantial gap between EDGARv4 and top-down estimates (with EDGARv4 emissions being less than 30% of the top-down values before 2008), which has apparently closed considerably in recent years in EDGARv5. However, for this species, as for many others, 585 the cause of this discrepancy is not known.
When species are aggregated into an F-gas total, weighted by their 100-year GWPs (Figure 3), the EDGARv5 estimates are around 10% lower than the WMO 2018 values in the 1980s. Subsequently, EDGARv5 estimates grow more rapidly than the top-down values and are almost 30% higher than WMO 2018 by the 2010s. Given that detailed uncertainty estimates are not 590 available for all EDGAR F-gas species, we base our uncertainty estimate solely on this single comparison with the top-down values, and therefore suggest an uncertainty in aggregated F-gas emissions of ±30% for a 90% confidence interval. For individual species, the magnitude of this discrepancy can be orders of magnitude larger.
The F-gases in EDGARv5 do not include chlorofluorocarbons (CFCs), hydrochlorofluorocarbons (HCFCs) and some 595 perfluorinated species such as NF3most of these species being regulated under the Montreal Protocol. Historically, total CO2-equivalent F-gas emissions have been dominated by the CFCs (Engel and Rigby, 2018). In particular, during the 1980s, peak annual emissions due to CFCs reached 9.1±0.4 GtCO2eqyr -1 (Figure 3), comparable to that of CH4, and substantially larger than the 2019 emissions of the gases included in EDGARv5 (1.6 GtCO2eq) (Table 7). Subsequently, following the controls of the Montreal Protocol, emissions of CFCs declined substantially, while those of HCFCs and HFCs rose, such that 600 CO2eq emissions of the HFCs, HCFCs and CFCs were approximately equal by 2016, with a smaller contribution from PFCs, SF6 and some more minor F-gases. Therefore, the GWP-weighted F-gas emissions in EDGARv5, which are dominated by the HFCs, represent less than half of the overall CO2eq F-gas emissions in 2018.

Aggregated GHG emissions
Based on our assessment of relevant uncertainties above, we apply constant, relative uncertainty estimates for GHGs at a 90% 605 confidence interval that range from relatively low for CO2 FFI (±8%), to intermediate values for CH4 and F-gases (±30%), to higher values for N2O (±60%) and CO2 from LULUCF (±70%). To aggregate these and estimate uncertainties for total greenhouse gases in terms of CO2eq emissions, we taking the square root of the squared sums of absolute uncertainties for individual (groups of) gases, using 100-year Global Warming Potentials (GWP100) to weight emissions of non-CO2 gases but excluding uncertainties in the metric itself (see Section 3.7). Overall, this is broadly in line with IPCC AR5 (Blanco G. et al., 610 2014), but provides important adjustments both in the evaluation of uncertainties (CH4, F-gases, CO2-LULUCF) as well as the approach in reporting total uncertainties across greenhouse gases.  (Allen et al., 2018;Cain et al., 2019;Collins et al., 2019;Lynch et al., 2020).

630
The contribution of SLCF emissions to total GHG emissions expressed in CO2eq thus depends critically on the choice of GHG metric and its time horizon. However, even for a given choice, the metric value for each gas is also subject to uncertainties.
For example, the GWP-100 for methane has changed from 21 based on the IPCC Second Assessment Report in 1995 to 28 or 34 based on the IPCC AR5 (including feedbacks). These changes and remaining uncertainties arise from parametric uncertainties, differences in methodological choices, and changes in metric values over time, due to changing background 635 conditions.  Parametric uncertainties arise from uncertainties in climate sensitivity, radiative efficacy and atmospheric lifetimes of CO2 and non-CO2 gases, etc. The IPCC AR5 assessed the parametric uncertainty of GWP for methane as ±30% and ±40% for time horizons of 20 and 100 years, and ±20% and ±30% for gases with atmospheric lifetimes of a 640 century or more. The uncertainty of GTP-100 for methane was estimated at ±75% (Myhre et al., 2013), which is larger than the uncertainty in a forcing-based metric due to due to uncertainties in climate responses to forcing (e.g., climate sensitivity). Further changes in metric values for methane and other gases within this uncertainty range are likely, given recent re-evaluations of the direct forcing of methane (Etminan et al., 2016) and adjustment of effective radiative forcing . photochemical decay products (mainly tropospheric ozone and stratospheric water vapour) contributes almost 40% of the total forcing from methane emissions. More than half of the changes in GWP-100 values for methane in successive IPCC assessments from 1995 to 2013 are due to re-evaluations of these indirect forcings. These 650 uncertainties are incorporated in the above uncertainty estimates. In addition, warming due to the emission of non-CO2 gases extends the lifetime of CO2 already in the atmosphere through climate-carbon cycle feedbacks (Friedlingstein et al., 2013). Including these feedbacks results in higher metric values for all non-CO2 gases, but the magnitude of this effect is uncertain; e.g. the IPCC AR5 found the GWP-100 value for methane without climatecarbon cycle feedbacks to be 28, whereas including this feedback would raise the value to between 31 and 34 (Gasser 655 et al., 2016;Myhre et al., 2013;Sterner and Johansson, 2017). Overall, we estimate the uncertainty in GWP-100 metric values, especially if applied to extended emission time series, as ±50% for methane and other SLCFs, and ±40% for non-CO2 gases with longer atmospheric lifetimes (specifically, those with 670 lifetimes longer than 20 years). If uncertainties in GHG metrics are considered, the overall uncertainty of total GHG emissions in 2018 increases from ±11% to ±24%. (However, in the following sections we do not include GWP uncertainties in our global, regional or sectoral estimates).
For the purpose of this paper, we use GWP-100 metric values from the IPCC AR5 (Myhre et al., 2013) without climate-carbon 675 cycle feedbacks. Even though climate-carbon cycle feedbacks are considered a robust feature of the climate system, the issue was only emerging during the IPCC AR5 and the methodology used to include this in metric calculations was indicative only.
Subsequent studies (Gasser et al., 2016;Sterner and Johansson, 2017) suggest that revisions to the simple estimation method in IPCC AR5 are necessary. As mentioned above, the most appropriate metric to aggregate GHG emissions depends on the objective. One such objective can be to understand the contribution of emissions in any given year to warming, while another can be to understand the contribution of cumulative emissions over an extended time period to warming. Sustained emissions of SLCFs such as methane do not cause the same temperature response as sustained emissions of CO2. Showing superimposed emission trends of different gases over multiple decades using GWP-100 as equivalence metric therefore does not necessarily represent the overall 685 contribution to warming from each gas over that period. In Figure 4 we therefore also show the modelled warming from emissions of each gas or group of gases -calculated using the reduced-complexity climate model FAIRv1.6 and calibrated to reproduce the pulse-response functions for each gas consistent with the IPCC AR5 (Myhre et al., 2013). Despite some differences compared to the contribution of each gas, based on GHG emissions expressed in CO2eq using GWP-100 (see

Table 7 -Average annual anthropogenic GHG emissions and emissions growth by decade and (groups of) gases for 1970-2019: CO2
715 from fossil fuel combustion and industrial processes (FFI); CO2 from land use, land-use change and forestry (LULUCF); methane (CH4); nitrous oxide (N2O); fluorinated gases (F-gases). Aggregate GHG emission trends by groups of gases reported in GtCO2eq converted based on global warming potentials with a 100-year time horizon (GWP-100) from the IPCC Fifth Assessment Report (Myhre et al., 2013). Uncertainties are reported for a 90 % confidence interval (see Section 3). Levels and growth are average values over the indicated time period. Additional supplementary tables show similar average annual GHG emissions by decade also for major sectors (Table SM- There is high confidence that emission growth has been varied, but persistent across different groups of gases. Decade-bydecade increases in global average annual emissions have been observed consistently across all (groups of) greenhouse gases (Table 7), apart from CO2-LULUCF emissions, which have been more stable, albeit uncertain, and only recently started to 725 show an upward trend. The pace and scale of emission growth has varied across groups of gases. While average annual 2.8% of total GHG emissions measured as GWP-100. Increases in average annual emissions levels from the 1970s (1970)(1971)(1972)(1973)(1974)(1975)(1976)(1977)(1978)(1979) to the 2010s (2010-2019) were lower for CO2-LULUCF (24%), CH4 (42%) as well as N2O (44%) (see Table 7).
However, there is low confidence that the reported increases in CO2-LULUCF emissions by decade actually constitute a statistically robust trend given the large uncertainties involved. In fact, two bookkeeping models underlying the AFOLU data 735 show opposing positive and negative trends (BLUE, H&N, respectively), while the third model (OSCAR), averaging over simulations that use either the same land-use forcing as BLUE (LUH2) or H&N (FAO), tracks the approximate mean of these (see also Section 3.2). Dynamic global vegetation models, which also use the LUH2 forcing, show higher estimates recently, explained by them considering the loss in sink capacity, while the bookkeeping models do not (see Figure 1). Overall, the different lines of evidence are inconclusive with regard to an upward trend in CO2-LULUCF emissions. 740 Global anthropogenic greenhouse gas emissions grew continuously slower than world GDP across all (groups of) individual gases resulting in a sustained decline in the GHG intensity of GDP as shown in Figure 5. The only exception is the group of   (Myhre et al., 2013). Coloured shadings show the associated uncertainties at a 90 % confidence interval without considering uncertainties in GDP and population data (see below). First column shows emission trends in absolute levels (GtCO2eq). Second column shows per capita emissions trends (tCO2eq/cap) using UN population data for normalization (World Bank, 2021). Third column shows emissions trends per unit of GDP (kgCO2eq/$) using GDP data in constant 2010 $ from the World Bank for normalization (World 760 Bank, 2021).
The continuous growth in global anthropogenic GHG emissions since the 1970s was mainly driven by activity growth in three major sectors: energy supply, industry and transportation (see Table SM respectively (see Table SM  In Fig. 7 we compare historic GHG emission trends with different scenarios, to explore how emissions are developing relative to the range of projected future outcomes. The Integrated Assessment Modelling (IAM) community quantified five Shared 795 Socioeconomic Pathways (SSPs) for different levels of radiative forcing in 2100 using six different IAMs (Riahi et al., 2017;Rogelj et al., 2018b). The SSPs are grouped according to their radiative forcing ranging from 1.9 Wm -2 to 8.5 Wm -2 , aimed at spanning the full range of potential outcomes. The Coupled Model Intercomparison Project Phase 6 (CMIP6) took a subset of these quantified SSPs as the basis for future climate projections (Gidden et al., 2019;O'Neill et al., 2016). In recent years, the use of the very high forcing scenariosparticularly SSP5-8.5 -is being debated in the scientific community (e.g. Hausfather 800 andPeters, 2020b, 2020a;Pedersen et al., 2020;Schwalm et al., 2020).
Historical GHG emissions from our database are consistent with the levels and trends in the scenario data, despite the scenarios being calibrated on older data sources (Gidden et al., 2019) mainly CEDS (Hoesly et al., 2018). The observed differences are larger for the GHGs with the highest uncertainty, notably CO2-AFOLU, N2O and F-gas emissions (Sections 3.2, 3.4 and 3.5). Across the different GHGs, historical emissions track on aggregate with the higher forcing scenarios such as the SSP3-805 7.0 and SSP5-8.5 markers, in terms of both levels and growth rates. CO2-FFI emissions still tend towards the higher end of the scenario range shown here, but there are signs that CO2-FFI emissions are slowing to more moderate forcing levels (e.g., SSP4-6.0 and SSP2-4.5) when considering recent trends (Hausfather and Peters, 2020a). CH4 and N2O emissions sit more in the middle and at the lower-end of the scenario rangethe latter driven by the lower levels of N2O emissions in EDGARand F-gases are consistent with the scenarios. Total GHG emissions track the higher end scenarios. 810 Figure 7 highlights the very different future emission trajectories envisioned by IAMs for individual gasesparticularly at radiative forcing levels that are consistent with the goal of the Paris Agreement such as SSP1-2.6 and SSP1-1.9. In contrast to CO2 emission, non-CO2 forcers such as anthropogenic CH4 and N2O emissions are not reduced to zero. However, in many scenarios, F-gases reach zero emissions. N2O emissions remain at similar levels to today in some of the scenarios with a 1.9 815 Wm -2 forcing at the end of the century, while they are about halved in others. Reductions in methane emissions are a bit more pronounced ranging from about 100 to 200 MtCH4yr -1 in 2100 compared to almost 400 MtCH4yr -1 in 2019. CO2-AFOLU emission trajectories overlap for different forcing levels, partly reflecting the complexities of modelling land-use change, but overall show a tendency towards a net carbon sink even in SSPs with little or even without climate policy. Given recent trends in land-use change emissions, it could be questioned whether the scenarios adequately explore the uncertainty in future land-820 use change emissions (Hausfather and Peters, 2020b).

-Global greenhouse gas emissions for the last decade 2010-2019
There is high confidence that global anthropogenic GHG emission levels were higher in 2010-2019 than in any previous decade 830 and GHG emissions levels have grown across the most recent decade. Average annual GHG emissions for 2010-2019 were 56±6.0 GtCO2eqyr -1 compared to 47±5.4 and 39±4.9 GtCO2eqyr -1 for 2000-2009 and 1990-1999, respectively. In 2019 GHG emissions were about 6.8±1.0 GtCO2eqyr -1 or 13% higher than in 2010. F-gas and CO2-LULUCF emissions were 50% and 24% higher in 2019 than in 2010 compared to 12%, 11% and 9% for N2O, CO2-FFI and CH4 emissions, respectively. GtCO2eqyr -1 from the 2000s to the 2010s has been the largest since the 1970sand probably within all human history as suggested by available long-term data (e.g. Friedlingstein et al., 2020;Hoesly et al., 2018). 840 About 50% of the recent growth in global GHG emissions between 2010 and 2019 came from China (2.7 GtCO2eqyr -1 ) and India (0.94 GtCO2eqyr -1 ) (Figure 8). Among the major emitters, fastest GHG emissions growth was observed for Vietnam with average annual rates of 5.1% yr -1 between 2010 and 2019 followed by Turkey (4.6% yr -1 ), Indonesia (3.8% yr -1 ), Pakistan (3.4% yr -1 ), India (3.2% yr -1 ), Saudia Arabia (2.8% yr -1 ) and China (2.4% yr -1 ). GHG emission reductions achieved by countries 845 over the last decade are comparatively small even though there is a growing number of countries on sustained emissions reductions trajectories (Lamb et al., 2021b;Le Quéré et al., 2019b). The US showed the largest net anthropogenic GHG emissions reductions of 0.21 GtCO2eqyr -1 between 2010 and 2019 even though more significant reductions in CO2 emissions of 0.46 GtCO2yr -1 from a switch from coal to gas in the context of the shale gas expansion was partially compensated by additional CH4 (0.12 GtCO2eqyr -1 ) and F-gas (0.13 GtCO2eqyr -1 ) emissions (Figure 8). Other countries with decreasing GHG 850 emission levels were Germany (0.13 GtCO2eqyr -1 ) and the United Kingdom (0.14 GtCO2eqyr -1 ), where the latter shows the fastest average annual reductions at a rate of 2.6% yr -1 in the sample -in line with some GHG emission reduction scenarios that limit global warming to well below 2°C, but those ones that tend to rely more heavily on carbon dioxide removal technologies (Hilaire et al., 2019;Strefler et al., 2018). Further information on country contributions to GHG emission changes national inventories 5.6% lower than the data presented here for the same countries. The growth rates over the last decade (2010-2019) reported in the national inventories was on average 0.3 percentage points lower than the growth rates for the same set of countries in our dataset (see Figure 8). Additional analysis comparing our data with UNFCCC-CRF inventories for individual (groups of) gases and countries is provided in supplementary Fig. SM-3. 865 Sectoral GHG emissions were either stable or increased between 2010 and 2019. There is high confidence that no substantive GHG emissions reductions were observable for entire sectors at the global level (Fig. 8 d and e). The largest sectoral contribution to the 6.8±1.0 GtCO2eqyr -1 increase in GHG emissions levels between 2010 and 2019 was from CO2-AFOLU with about 1.3 GtCO2yr -1 , but this estimate is much more uncertain compared to other sectors. The continued expansion of 870 fossil-fuel based electricity production increased CO2 emissions by about 1.2 GtCO2yr -1 closely followed by CO2 emissions from road transport (0.9 GtCO2yr -1 ) and metal production (0.7 GtCO2yr -1 )the latter being the fastest large emission source in relative terms with 2.1%. Domestic and international aviation are the most rapidly growing sectors (3.8% and 3.7%, respectively), but remain globally small sources of emissions growth (0.1 and 0.17 GtCO2yr -1 ). Emissions from chemical production and waste treatment are also sizable and comparatively fast growing, contributing 0.47 GtCO2yr -1 at 1.9%yr -1 and 875 0.31 GtCO2yr -1 at 1.6%yr -1 , respectively.   Table 2).

Global greenhouse gas emissions in 2019
Global net anthropogenic greenhouse gas emissions continued to grow and reached 59±6.6 GtCO2eq in 2019 (Figure 8). In 2019, CO2 emissions from FFI were 38±3.0 Gt, CO2 from LULUCF 6.6±4.6 Gt, CH4 11±3.3 GtCO2eq, N2O 2.4±1.5 GtCO2eq and F-gases 1.6±0.49 GtCO2eq. Of the 59±6.6 GtCO2eq emissions in 2019, 33% (20 GtCO2eqyr -1 ) were from energy supply, 24% (15 GtCO2eqyr -1 ) from industry, 22% (20 GtCO2eqyr -1 ) from AFOLU, 15% (8.7 GtCO2eqyr -1 ) from transport, and 5.6% 895 (3.3 GtCO2eqyr -1 ) from buildings. In 2019, the largest absolute contributions in GHG emissions were from Asia and Developing Pacific (43%), Developed countries (25%) and Latin America and the Caribbean (10%). China (14 GtCO2eqyr -1 ), USA (6.5 GtCO2eqyr -1 ) and India (3.7 GtCO2eqyr -1 ) and the Russian Federation (2.5 GtCO2eqyr -1 ) remained the largest country contributors to global GHG emissions, excluding CO2-LULUCF as we do have not sufficient confidence to report this data at the country level. 900 In 2019, emissions were 1.4 GtCO2eqyr -1 or 2.4% higher than the 58±6.1 GtCO2eq in 2018. Most of this growth (~0.9±0.6 GtCO2eqyr -1 ) is related to increases in CO2-LULUCF, which results in particular from the high peat and tropical deforestation/degradation fires as outlined in Friedlingstein et al. (2020). Growth in CO2-FFI was very modest at 0.28±0.023 GtCO2yr -1 (Δ0.8%), while F-gas, N2O and methane grew more rapidly by 3.8%, 1.2% and 1.0% -but at much lower absolute 905 levels. While the rate of GHG emissions change between 2018 and 2019 is numerically comparable with the period of high GHG emissions growth during the 2000s, there is low confidence in the reported value due to the high share of CO2-LULUCF emissions, which are highly uncertain, and the preliminary nature of the underlying land-use data for 2019 and temporal extrapolation of two of the three bookkeeping estimates. Moreover, given prevailing uncertainties there is low confidence that GHG emissions have never been higher than in 2019 as suggested by the data, but high confidence that average annual GHGs 910 emissions have never been higher for a decade than in 2010-2019 (see Friedlingstein et al., 2020;Hoesly et al., 2018).

Discussion
In this article we provide a comprehensive, detailed dataset for global, regional, national and sectoral GHG emissions from 915 anthropogenic sources covering the last five decades  built from the EDGARv6 GHG emissions inventory, a fasttrack update/projection as well as data on CO2-LULUCF emissions from global bookkeeping models. We assess uncertainties in our estimates by combining statistical analysis of the underlying data and expert judgement based on an in-depth review of the literature by each gas. We report uncertainties at a 90% confidence interval (5 th -95 th percentile range). This differs to the uncertainty reported by the Global Carbon Project for the global carbon, methane or nitrous oxide budgets (Friedlingstein et 920 al., 2020;Saunois et al., 2020;Wang et al., 2020), because uncertainties in our dataset are comparatively well characterized (Janssens-Maenhout et al., 2019;Solazzo et al., 2021).
Our uncertainty assessment is broadly consistent with previous assessments focussing on all GHGs (Blanco G. et al., 2014; UNEP, 2020), but we provide some important updates. Our evidence-informed uncertainty judgements are higher for CO2-925 LULUCF (±70% rather than ±50%) and CH4 (±30% rather than ±20%) drawing from work on global carbon (Friedlingstein et al., 2020) and methane (Saunois et al., 2020) budgets. We recognize the vast divergence between bottom-up inventory estimates and top-down atmospheric measurements for individual F-gases. Our revised uncertainty estimate for aggregate Fgas emissions of ±30% (rather than ±20%) reflects the smaller aggregate deviation when all individual species are considered together. 930 Our analysis involves aggregating GHG emissions into a single unit using GWP-100 values from IPCC AR5 (without carbon cycle feedbacks). By doing so we follow the practice taken in UNFCCC climate diplomacy and large parts of the literature on climate change mitigation. However, we recognise intense scientific and academic debates about the aggregation of GHGs into a single unit and alternative choices of metrics (Myhre et al., 2013) (see Section 3.7). We therefore also use a simple 935 climate model to assess the warming contribution by the individual groups of gases and find that for the historical period when emissions are growing, the GWP-100 gives a reasonable approximation to the warming contributions, but this is not expected to hold when emissions change trajectory under mitigation. In the absence of comprehensive uncertainty analysis that covers CO2-LULUCF as well as F-gas emissions, we estimate the overall uncertainties of aggregated GHG emissions by simply adding the individual uncertainties judgements by (groups of) gases in quadrature under the assumption of their independence. 940 Comprehensive uncertainty analysis of EDGAR data covering all greenhouse gases should be performed in the future, building on Solazzo et al. (2021). For the first time, we also provide an initial estimate of metric uncertainty arising from the aggregation of individual greenhouse gases into a single unit (see Section 3.7).
Our assessment highlights the comprehensive nature of our dataset covering anthropogenic sources of greenhouse gas 945 emissions. However, there are still some important data gaps. Most recent and comprehensive assessments of the methane https://doi.org/10.5194/essd-2021-228  (Saunois et al., 2020) and nitrous oxide  budgets suggest that anthropogenic CH4 and N2O emissions could be 10-20% higher than reported in EDGAR, respectively. F-gas emissions estimates for individual species in EDGARv5 do not align well with atmospheric measurements and the F-gas aggregate over-reports the measured concentrations by about 30%. However, EDGAR and official national emission reports under the UNFCCC do not comprehensively cover all relevant 950 F-gases species. We also note that our data does not cover species such as chlorofluorocarbons (CFCs), hydrochlorofluorocarbons (HCFCs) or NF3 and show that those species, which are regulated under the Montreal Protocol (except NF3), have contributed more to CO2eq emissions as well as observed warming. There is an urgent need to dedicate more resources and attention to the independent improvement of F-gas emission statistics, recognizing these current shortcomings and their increasingly important role as a driver of warming. 955 Our analysis of global, anthropogenic GHG emission trends over the past five decades (1970-2019) highlights a pattern of varied, but sustained emissions growth. There is high confidence that global anthropogenic greenhouse gas emissions have increased every decade. Emission growth has been varied, but persistent across different (groups of) gases. While CO2 has accounted for almost 75% of the emission growth since 1970 in terms of CO2eq as reported here, the combined F-gases have 960 grown much faster than other GHGs, albeit starting from very low levels. Today, they make a non-negligible contribution to global warmingrecognizing that important species such as CFCs and HCFCs are even not considered. There is further high confidence that global anthropogenic GHG emissions levels were higher in 2010-2019 than in any previous decade and GHG emissions levels have grown across the most recent decade. While average annual greenhouse gas emissions growth slowed between 2010-2019 compared to 2000-2009, the absolute increase in average decadal GHG emissions from the 2000s to the 965 2010s has been the largest since the 1970sand within all human history as suggested by available long-term data (e.g. Friedlingstein et al., 2020;Hoesly et al., 2018). We note considerably higher rates of change in GHG emissions between 2018 and 2019 than for the entire decade 2010-2019, which is numerically comparable with the period of high GHG emissions growth during the 2000s, but we place low confidence in this value as the majority is driven by highly uncertain increases in CO2-LULUCF emissions as well as the use of preliminary data and extrapolation methodologies for these most recent years. 970 While there is a growing number of countries today on a sustained emission reduction trajectory (Lamb et al., 2021b;Le Quéré et al., 2019a), it is important to study the drivers of these reductions as well as patterns of emission growth in other parts of the world (Lamb et al., 2021a). Our analysis further reveals that there are no global sectors that show sustained reductions in GHG emissions.

975
There is a growing availability of global datasets on anthropogenic emissions sources over the last 10-20 years. However, such global emission inventories have to rely on relatively simple Tier-1 estimation methods and few use more complex Tier-2 methods. Comparison of our estimates with Tier-2 and Tier-3 UNFCCC-CRFs by Annex I countries shows considerable discrepancies for some gases. On aggregate, there is a clear trend towards smaller values for GHG emission reductions and larger values for GHG emission increases in our dataset. Further work needs to be done to fully appreciate underlying 980 https://doi.org/10.5194/essd-2021-228 differences (Andrew, 2020a;Petrescu et al., 2020cPetrescu et al., , 2020b. Figure 9 further highlights the lack of recent official GHG emissions inventories for many non-Annex 1 countries outside those global emission inventories. Despite the importance of high-quality emission statistics for climate change research and tracking progress in climate policy, our analysis here emphasises considerable prevailing uncertainties and the need for improvement in emission reporting. In sectors where production efficiencies are changing rapidly, as is often the case in developing countries, using emission estimates based on 985 Tier-1 methodologies is likely to mischaracterise trends as both activity data and emission factors change over time (Wilkes et al., 2017). Moving confidently towards net-zero emissions requires high quality emissions statistics for tracking countries' progress based at least on Tier-2, if not on complex Tier-3 estimation models using comprehensive, country-specific activity data and emissions factors (IPCC, 2019). This would also support the formulation of more nuanced climate policy goals that reflect changes in emissions intensity as entry points for more comprehensive and ambitious targets to reduce absolute 990 emissions. However, underpinning such approaches with robust evidence requires the collection of a range of country-specific activity data and development of adequate statistical infrastructure for all countries of the world (FAO and GRA, 2020).
Making progress in the implementation of the Paris Agreement and keeping warming well below 2°C requires dedication and cooperation between countries: working together on a robust evidence base in GHG emissions reporting provides one important and often underappreciated step.

Data availability
The emissions dataset used for this study (Minx et al. 2021) is available at https://doi.org/10.5281/zenodo.5053056 [NOTE TO REVIEWERS: Data on CO2 emissions from fossil fuel combustion and industry, methane emissions and nitrous oxide emissions are from the most recent EDGARv6 data. As EDGARv6 data is still being compiled for F-gases, this 1010 manuscript contains EDGARv5 estimates for these, but we will update to EDGARv6 during the revision process. This procedures has been agreed upon with David Carlsonone of the chief editors of the journalbefore manuscript submission]

Funding acknowledgements
The authors would like to thank Yang Ou for helpful comments on the manuscript and Eduardo Posada as well as Lucy Banisch for their help with compiling the information for Table 1 and Fig. 9