A comprehensive and synthetic dataset for global, regional and national greenhouse gas emissions by sector 1970-2018 with an extension to 2019

To track progress towards keeping global warming well below 2°C or even 1.5°C, as agreed in the Paris Agreement, comprehensive up-to-date and reliable information on anthropogenic emissions and removals of greenhouse gas emissions (GHG) is required. Here we compile a new synthetic dataset on anthropogenic GHG emissions for 1970-2018 with a fast-track extension to 2019. Our dataset is global in coverage and includes CO 2 emissions, CH 4 emissions, N 2 O emissions as well as 30 those from fluorinated gases (F-gases: HFCs, PFCs, SF 6 , NF 3 ) and provides country and sector detail. We build this dataset from the version 6 release of the “Emissions Database for Global Atmospheric Research” (EDGAR v6) and three bookkeeping models for CO 2 emissions from land use, land-use change and forestry (LULUCF). We assess the uncertainties of global greenhouse gases at the 90% confidence interval (5 th -95 th percentile range) by combining statistical analysis and comparisons of global emissions inventories and top-down atmospheric measurements with an expert judgement informed by the relevant 35 scientific literature. We identify important data gaps for F-gas emissions. The agreement between EDGAR and atmospheric-based emissions estimates is relatively close for some F-gas species (~10% or less) but the estimates can differ by an order of magnitude or more for others. When aggregated, the EDGAR v6 F-gas total agrees with top-down estimates to within around

10% in recent years. However, emissions from excluded F-gas species such as chlorofluorocarbons (CFCs) or hydrochlorofluorocarbons (HCFCs) are cumulatively larger than the sum of the reported species. Using global warming 40 potential values with a 100 year time horizon from the Sixth Assessment Report by the Intergovernmental Panel on Climate Change (IPCC), global GHG emissions in 2018 amounted to 58±6.1 GtCO2eq consisting of: CO2-FFI 38±3.0 GtCO2, CO2-LULUCF 5.7±4.0 GtCO2, CH4 10±3.1 GtCO2eq, N2O 2.5±1.5 GtCO2eq and F-gases 1.3±0.40 GtCO2eq. Initial estimates suggest further growth in GHG emissions by 1.3 GtCO2eq to reach 59±6.6 GtCO2eq in 2019. Our analysis of global trends in anthropogenic GHG emissions over the past five decades  highlights a pattern of varied, but sustained emissions 45 growth. There is high confidence that global anthropogenic GHG emissions have increased every decade and emissions growth has been persistent across different (groups of) gases. There is also high confidence that global anthropogenic GHG emission levels were higher in 2009-2018 than in any previous decade and GHG emission levels grew throughout the most recent decade. While the average annual GHG emissions growth rate slowed between 2009-2018 (1.2% yr -1 ) compared to [2000][2001][2002][2003][2004][2005][2006][2007][2008][2009] (2.5% yr -1 ), the absolute increase in average decadal GHG emissions from the 2000s to the 2010s was the largest since the 50 1970s, and within all human history. Our analysis further reveals that there are no global sectors that show sustained reductions in GHG emissions. There are a number of countries that have reduced GHG emissions over the past decade, but these reductions are comparatively modest and outgrown by much larger emissions growth in some developing countries such as China, India and Indonesia. There is a need to further develop independent, robust, and timely emission estimates across all gases. As such, tracking progress in climate policy requires substantial investments in independent GHG emission accounting 55 and monitoring, as well as in national and international statistical infrastructures. The data associated with this article (Minx et al., 2021) can be found at , https://doi.org/10.5281/zenodo.5053055.

Introduction
By signing the Paris Agreement, countries acknowledged the necessity to keep the most severe climate change risks in check by limiting warming to well below 2°C, and to pursue efforts to limit warming to 1.5°C (UNFCCC, 2015). This requires rapid and sustained greenhouse gas (GHG) emission reductions towards net zero carbon dioxide (CO2) emissions well within the 65 21 st century along with deep reductions in non-CO2 emissions (Rogelj et al., 2015(Rogelj et al., , 2018a. Transparent, comprehensive, consistent, accurate and up-to-date inventories of anthropogenic GHG emissions are crucial to track progress by countries, regions and sectors in moving towards these goals. However, it is challenging to accurately track the recent GHG performance of countries and sectors. While there is a growing 70 number of global emissions inventories, only a few of them provide a wide coverage of gases, sectors, activities, and countries or regions that are sufficiently up-to-date to comprehensively track progress and thereby aid discussions in science and policy. Table 1 provides an overview of global emission inventories. Many inventories focus on individual gases and subsets of activities. Few provide sectoral detail and particularly for non-CO2 GHG emissions there is often a considerable time-lag in reporting. GHG emissions reporting under the United Nations Framework Convention on Climate Change (UNFCCC) 75 provides reliable, comprehensive and up-to-date statistics for Annex I countries across all major GHGs. Non-Annex I countries except least developed countries and small island state for which this is not mandatoryprovide GHG emissions inventory information through biennial update reports (BURs), but with much less stringent reporting requirements in terms of sector, gas and time coverage (Deng et al., 2021;Gütschow et al., 2016). As a result, many still lack a well-developed statistical infrastructure to provide detailed reports (Janssens-Maenhout et al., 2019). 80 Here we describe a new, comprehensive and synthetic dataset for global, regional and national GHG emissions by sector for 1970-2018 with a fast-track extension to 2019. Our focus is on GHG emissions from anthropogenic activities only. We build the dataset from recent releases of the "Emissions Database for Global Atmospheric Research" version 6 (EDGARv6) for CO2 emissions from fossil fuel combustion and industry (FFI), CH4 emissions, N2O emissions, and fluorinated gases (F-gases: 85 HFCs, PFCs, SF6 and NF3) . For completeness we add net CO2 emissions from land use, land-use change and forestry (CO2-LULUCF) from three bookkeeping models (Gasser et al., 2020;Hansis et al., 2015;Houghton and Nassikas, 2017). We provide an assessment of the uncertainties in each GHG at the 90% confidence interval (5 th -95 th percentile) by combining statistical analysis and comparisons of global emissions inventories with an expert judgement informed by the relevant scientific literature. 90  Darmenov and da Silva (2015) 2 Methods and Data

Overview 95
Our dataset provides a comprehensive, synthetic set of estimates for global GHG emissions disaggregated by 29 economic sectors and 228 countries. Our focus is on anthropogenic GHG emissions: natural sources and sinks are not included. 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) fluorinated gases (F-gases) comprising hydrofluorocarbons (HFCs), perfluorocarbons (PFCs), sulphur hexafluoride (SF6) 100 as well as nitrogen trifluoride (NF3). F-gases that are internationally regulated as ozone depleting substances under the Montreal Protocol such as chlorofluorocarbons (CFCs) and hydrochlorofluorocarbons (HCFCs) are not included. We provide and analyse the GHG emissions data both in native units as well as in CO2-equivalents (CO2eq) (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 Sixth Assessment Report (AR6) (Forster et al., 2021). We briefly discuss the impact of alternative metric 105 choices in tracking aggregated GHG emissions over the past few decades and juxtapose this estimates of 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 110 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).

115
We compile our dataset from four sources: (1) the full EDGARv6 release for CO2-FFI as well as non-CO2 GHGs covering the time period 1970-2018 ; (2) EDGARv6 fast-track data for CO2-FFI providing preliminary estimates for 2019 (and 2020) ; (3) CO2-LULUCF as the average of three bookkeeping models, consistent with the approach of the global carbon project (Friedlingstein et al., 2020). (4) 2019 non-CO2 emissions based on Olivier and Peters (2020). 120 As shown in Table 2, sectoral detail is organised along five major economic sectors as commonly used in IPCC reports on climate change mitigation (IPCC, 2014): energy supply, buildings, transport, industry as well as Agriculture, Forestry and Other Land-Use Changes (AFOLU). We devise a classification for assigning our 228 countries to regions, combining the standard Annex I/non-8 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.5053055. Table 2 -Overview of the two-level sector aggregation with reference to assigned source/sink categories conforming to the IPCC   130   reporting guidelines (IPCC, 2006, 2019) as well as relevant GHGs. Note that EDGAR v6 distinguishes biogenic CO2 and CH4 sources with a "bio" label, with all other sectors "fossil" by default, even if that source is not related to fossil fuel activities. The fossil/bio label is hence not descriptive in nature. Two HCFC gases (denoted with *) are included in the dataset, despite being neither PFCs nor HFCs (and hence regulated under Montreal). This is to preserve consistency with current and previous versions of EDGAR, which include these gases. Their total warming effect is low (~10 MtCO2eq in 2019) and the major HCFC sources are not included.  CO2, CH4, N2O, c-C4F8, C2F6, C3F8,  C4F10, C5F12, C6F14, CF4, HFC-125, HFC-134a, HFC-143a, HFC-152a, HFC-227ea, HFC-32, HFC-365mfc, NF3, SF6, HFC- (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 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) 145 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 150 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 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, 155 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 .

160
To compute emissions up to most recent years, a Fast-Track methodology is applied, as described in detailed in Oreggioni et al. (2021). The underlying idea is to extrapolate trends based on observed activity patterns in representative sectors. For CO2-FFI emissions, the fast track estimates were based on the latest BP coal, oil and natural gas consumption data (BP, 2021).
Emission updates for cement, lime, ammonia and ferroalloys production beyond 2018 are still based on stable statistics and in particular on US Geological Survey statistics, urea production and consumption on statistics from the International Fertilizer 165 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. Fast-track extensions for non-CO2 GHG emissions are developed from Olivier and Peters (2020). For CH4 and N2O these are based on agricultural statistics from Food and Agricultural Organization (FAO) (CH4 and N2O) of the United Nations, 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 170 of chemicals (N2O abatement). Finally, for F-gases the fast-track extension was based on the most recent national emission inventories, submitted under the UNFCCC (up to 2018). For all remaining countries and years, a simple extrapolation was used given the absence of international statistics. We apply this fast-track data by Olivier and Peters (2020)  We consider all fluxes of CO2 from land use, land-use change and forestry. This includes CO2 fluxes from the clearing of forests and other natural vegetation (by anthropogenic fire and/or clear-cut), afforestation, logging and forest degradation (including harvest activity), shifting cultivation (cycles of forest clearing for agriculture, then abandonment), and regrowth of 180 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.
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 185 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 mainly characterized by CH4 and N2O fluxes.
Since in reality anthropogenic CO2-LULUCF emissions co-occur with natural CO2 fluxes in the terrestrial biosphere, models 190 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 the literature and observations to describe the 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 195 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 carbon stocks in secondary forests, and include forest management practices such as wood harvesting.

200
The scientific definition of CO2-LULUCF emissions used here differs from the one applied in national greenhouse gas inventories (GHGI) or the inventory data provided by FAOSTAT . Concretely, this means that inventory data include natural terrestrial fluxes caused by changes in environmental conditions, e.g., effects of rising atmospheric CO2 ("CO2-fertilization"), climate change, and nitrogen deposition --sometimes called "indirect effects" as opposed to the direct anthropogenic effects of land-use change and management ) -when they occur on area that countries 205 declare as managed. Since environmental changes turned the terrestrial biosphere into a massive sink, removing about one third of annual anthropogenic emissions in the last decade (Friedlingstein et al., 2020), it is unsurprising that global emission estimates are smaller for inventory data than for the scientific definition (see Figure 1). About 3.2 GtCO2 yr −1 (for the period 2005-2014) was found to be explicable by these conceptual differences in anthropogenic forest sink estimation related to the representation of environmental change impacts and the areas considered as managed (Grassi et al., 2018). 210 These two conceptually different approaches have different aims: The scientific approach separates natural from anthropogenic drivers, i.e., effects of changes in environmental conditions from effects of land-use change and land management. By contrast, the inventory approach separates fluxes based on areas, with all those occurring on managed land being declared anthropogenic. Given that observational data of carbon stocks or fluxes cannot distinguish the co-occurring effects of 215 environmental changes and land-use activities, an area-based approach that does not require this distinction can more consistently be implemented across countries. These conceptual differences between scientific and inventory approaches have been acknowledged (Canadell et al., 2021;Petrescu et al., 2020a) and approaches have been developed to map the scientific and inventory definitions to each other (Grassi et al., 2018. For non-CO2 GHGs, drivers and areas coincide, such that FAOSTAT data for CH4 and N2O is complementary to bookkeeping CO2-LULUCF emissions. 220 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 summarised in Table 4. Since bookkeeping models do not include emissions from organic soils, emissions from peat fires and 225 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 land-use 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 from the global FAO data on organic soils emissions from croplands and grasslands (Conchedda and Tubiello, 2020). 230

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 uncertainties can be of a scientific nature, such as when a process is not sufficiently understood. They also arise from 235 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., 2020a(Petrescu et al., , 2021a(Petrescu et al., , 2021bSaunois et al., 2020a;Tian et al., 2020) (Petrescu et al., 2020b(Petrescu et al., , 2020aSaunois et al., 2020b;Tian et al., 2020); 2) by comparing estimates from multiple sources 240 and understanding sources of variation (Andres et al., 2012;Andrew, 2020a;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;, or to spatially distributed emissions .

245
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, 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 250 constraints are, therefore, particularly useful to bound total emission estimates (Petrescu et al., 2021b(Petrescu et al., , 2021a.  evaluated the uncertainty of the EDGAR's source categories and totals for the main GHGs (CO2-FFI, CH4, N2O). This 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 255 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.  assume full covariance between same source categories where similar assumptions are being used, and 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 260 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 dependent inventories (Saunois et al., 2016(Saunois et al., , 2020b. 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 265 comparable with aggregate country or regional uncertainty, estimated with the methods discussed above.
This section provides an assessment of uncertainties in greenhouse gas emissions data at the global level. The uncertainties reported here combine statistical analysis, comparisons of global emissions inventories and expert judgement of the likelihood of results lying outside a defined confidence interval, rooted in an understanding gained from the relevant literature. At times, 270 we also use a qualitative assessment of confidence levels to characterize the annual estimates from each term based on the type, amount, quality, and consistency of the evidence as defined by the IPCC (2014).
Such a comprehensive uncertainty assessment covering all major groups of greenhouse gases and considering multiple lines of evidence has been missing in the literature. The absence has provided a serious challenge for a transparent, scientific 275 reporting of GHG emissions in climate change assessments like those by IPCC's Working Group III or the UN Emissions Gap Report that have only more recently started to even deal with the issue UNEP, 2020). Most of the available studies in the peer-reviewed literature using multiple lines of evidence for their assessment have focused on individual gases like in the Global Carbon Budget (Friedlingstein et al., 2020), the Global Methane Budget (Saunois et al., 2020b) or the Global Nitrous Oxide Budget (Tian et al., 2020) or covered multiple gases, but mainly considered individual lines of evidence 280 (Janssens-Maenhout et al., 2019;. We adopt a 90% confidence interval (5 th -95 th percentile) to report the uncertainties in our GHG emissions estimates, i.e., there 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 Ciais et al., 2014). We note that national 285 emissions inventory submissions reported to the UNFCCC are requested to report uncertainty using a 95% or 2σ confidence interval. The use of this broader uncertainty interval implies, however, a relatively high degree of knowledge about the uncertainty structure of the associated data, particularly regarding the distribution of uncertainty in the tails of the probability distributions. Such a high degree of knowledge is not present over all the emission sectors and species considered here. Note that in some cases below we convert 1σ uncertainty results from the literature to a 90% confidence interval by 290 implicitly assuming a normal distribution. While we do this as a necessary assumption to obtain a consistent estimate across all GHGs, we note that this itself is an assumption that may not be valid. We have made use of the best available information in the literature, but note that much more work on uncertainty quantification remains to be done. Using IPCC uncertainty language, we cannot assign high confidence to the robustness of most existing uncertainty estimates.

CO2 emissions from fossil fuels and industrial processes 295
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., 2020b). However, 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). Some divergence between 300 these estimates (see Figure 1) are related to differences in the estimation methodology, conversion factors, emission coefficients, assumptions about combustion efficiency, and calculation errors (Andrew, 2020a;Marland et al., 2009). Key differences for nine global datasets are highlighted in Table 3 (see also Table 1 for further information on the inventories).
Another important 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 305 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 system boundary differences (Andrew, 2020a;Macknick, 2011).

Table 3 -System boundaries and other key features of global FFI-CO2 emissions datasets as published.
Comparison of some important general characteristics of nine emissions datasets, with green indicating a characteristic that might be considered a strength. Columns four 310 to six refer to CO2 emission estimates for industrial processes and product use. Since all datasets are under development, these details are subject to change. Further information on the individual inventories can be found in Table 1 Figure 1). However, this variability is almost halved when system boundaries are harmonised (Andrew, 2020a). 315 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 Annex I countries to the UNFCCCeven though some variation can occur for individual countries such as Kazakhstan, 320 Ukraine or Estonia, in general, or for certain years (see Figure  Uncertainties in CO2-FFI emissions arise from the combination of uncertainty in activity data and uncertainties in emission factors including assumptions for combustion completeness and non-combustion uses. CO2-FFI emissions estimates are largely 325 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 (Friedlingstein et al., 2020). Uncertainties are lower for fuels with relatively uniform properties such as natural gas, oil or 330 gasoline and higher for fuels with more diverse properties, such as coal (IPCC 2006;. 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).
However, recent versions of the global carbon budget include specific estimates for the cement carbonation sink and estimate 335 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 an average of 1.3% of a country's total fossil fuel use in the subsequent year with further more modest revisions later on 340 (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% for the U.S., ±15 -±20% for China and ±50% or more for countries with poorly developed or maintained statistical 345 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 statistics and more uncertain fuel properties (Ballantyne et al., 2015;Friedlingstein et al., 2020;Marland et al., 2009). 350 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 EDGARv4.3.2 and v5 (Janssens-Maenhout et al., 2019; at 2σ. It remains at the higher end of the ±5% -355 ±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 UNEP, 2020).  (Friedlingstein et al., 2020;Hansis et al., 2015); DGVM-mean -Multi-model mean of CO2-LULUCF emissions from dynamic global vegetation models (Friedlingstein et al., 2020); OSCARan earth system compact model (Friedlingstein et al., 2020;Gasser et al., 2020); HN -Houghton and Nassikas Bookkeeping Model (Friedlingstein et al., 2020;Houghton 370 and Nassikas, 2017); for comparison, the net CO2 flux from FAOSTAT is plotted, which comprises emissions and removals by forest land (FAOSTAT, 2021;Tubiello et al., 2021) (in contrast to the scientific definition this includes the natural terrestrial sink if occurring on managed land) and emissions from drained histosols under cropland/grassland (Conchedda and Tubiello, 2020  3.2 Anthropogenic CO2 emissions from land use, land use change and forestry (CO2-LULUCF) 380 CO2-LULUCF emissions are drawn from three global bookkeeping models. For 1990-2019, average net CO2-LULUCF emissions are estimated at 6.1, 4.3, and 5.6 GtCO2 yr -1 for BLUE, H&N, and OSCAR (Friedlingstein et al., 2020). Gross emissions 1990-2019 for BLUE, H&N, OSCAR are 17, 9.6 and 19 GtCO2 yr -1 , while gross removals are 11, 5.3, 13 GtCO2 yr -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 385 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 landuse 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 390 allocation of cleared and harvested material to fast turnover pools in BLUE compared to H&N; and (3) to the inclusion subgrid scale transitions .
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 CO2-395 LULUCF mean estimate in our data because the typical DGVM setup includes the loss of additional sink capacity, i.e. the additional sink capacity forests could have provided in response to environmental changes, in particular the rise in CO2, due to their long-lived biomass, but that is lost because large areas of forest were historically cleared for agriculture. The loss of additional sink capacity 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 400 the average estimate from 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 405 densities of bookkeeping models) can alter CO2 flux estimates substantially, but are included to varying extents in DGVMs, thus increasing model spread . 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 410 distinguishing natural from 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 the carbon embodied in trade products is not accounted for .
We base our uncertainty assessment on Friedlingstein et al. (2020) and take ±2.6 GtCO2 yr -1 as a best-value judgement for the 415 ±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 420 uncertainty of ±50 -±75% considered in AR5 . 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  as well as inventory data (Grassi et al., 2018). The constant absolute uncertainty estimate used in the global carbon budget (Friedlingstein et al., 2020) translates to a slightly lower relative uncertainty estimate, given that the mean of the CO2-LULUCF estimates has been increasing over the last few decades. Here we opt for a relative uncertainty 425 estimate of ±70% for a 90% confidence interval, which provides absolute uncertainty estimates for recent years that are consistent in magnitude with Friedlingstein et al. (2020).
Uncertainties can be much higher at a national level than at global level, since regional biases tend to cancel out. Land-use forcing has been identified as major driver of differences at regional and global level (Gasser et al., 2020;Hartung et al., 2021;430 Rosan et al., 2021), as have assumptions on carbon densities and the allocation of cleared or harvested material to slash or product pools of various lifetimes, for which accurate global data over long time periods is missing .
Although the bookkeeping models are conceptually similar, the bookkeeping estimates include country-specific information to different extents: for example, fire suppression (for the U.S.) is included in H&N (Houghton and Nassikas, 2017), but not the other estimates, and H&N includes peat drainage emissions only for Southeast Asia, while the FAO emissions estimates 435 for organic soil drainage added to BLUE and OSCAR cover all countries (Friedlingstein et al., 2020). The effect of smoothing the FAO cropland and pasture information, which can be very variable in some countries, with a 5-year running mean in H&N, while the annual data is used for the recent decades in HYDE underlying BLUE and OSCAR, must also be expected to contribute to the spread in estimates on a country level. Overall, great care has to be taken when comparing estimates of individual countries across models to not over-interpret differences. 440 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 CO2-LULUCF flux, since they usually only quantify vegetation biomass changes and exclude legacy emissions from the pre-satellite 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.  (Hansis et al., 2015); b (Houghton and Nassikas, 2017); c (Gasser et al., 2020); d (Hurtt et al., 2020); e (Chini et al.,455 2020); f (Nations, 2015); g based on rangeland-pasture distinction of the HYDE dataset (Klein Goldewijk et al., 2017) and forest cover map of Hurtt et al. (2020); see Friedlingstein et al. (2020) for details h (Conchedda and Tubiello, 2020); i (Hooijer et al., 2010) 460

Anthropogenic CH4 emissions
About 60% of total global CH4 emissions come from anthropogenic sources (Saunois et al., 2020b). 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 observation-based 465 upscaling for some specific sources such as geological sources (e.g. Etiope et al., 2019). Alternatively, top-down (TD) 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 470 carried out with various degrees of uncertainty depending of the methods and the degree of refinement of sectors, but often 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. ( , 2020b and Kirschke et al. (Kirschke et al., 2013).  (Darmenov and da Silva, 2013). Much like CO2 FFI, these inventories of anthropogenic emissions are not completely independent as they either follow the same IPCC methodology to derive emissions, rely on similar data sources (e.g., FAOSTAT activity data for agriculture, reported fossil fuel production), or draw on reported country inventory data (Petrescu et al., 2020a, e.g. Figure 4). 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 485 using a consistent approach for all countries, and country specific activity data, emission factor and technological abatement when available. FAOSTAT and EDGAR mostly apply a Tier 1 approach to estimate CH4 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.

490
Global anthropogenic CH4 emission estimates are compared in Figure 1. EDGARv5 has revised total global CH4 emissions 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 that are up to 50 MtCH4yr -1 lower before the 1990s compared to EDGARv5 and EDGARv4.3.2, but differences are smaller ranging from 1-13 MtCH4yr -1 since the 2000s (see Figure SM-495 1). The cause of these differences is a new procedure to separately estimate of the venting component for gas and oil, in the venting and flaring sector (1B2a/b2). Differences across different versions of the EDGAR dataset are shown in the Supplementary Material (Fig. SM-1). US-EPA show the lowest estimates probably due to missing estimates from a significant number of countries not reporting to UNFCCC (US-EPA2020 includes estimates from only 195 countries) and incomplete sectoral coverage. EDGARv6 estimates of anthropogenic CH4 emissions, as used here, are in the upper range of the different 500 inventories across most anthropogenic sources. However, none of these inventories cover CH4 emissions from forest and grassland burning, which amount to about 10-12 Mt yr -1 . Saunois et al (2020b) provide estimates of CH4 sources and sinks based on BU and TD approaches associated with an uncertainty range based on the minimum and maximum values of available studies (because for many individual source and 505 sink estimates the number of studies is often relatively small). Thus, they do not consider the uncertainty of the individual estimates. As shown in Table 5, uncertainties in total global CH4 emissions across all anthropogenic and natural sources are comparatively small from TD approaches at ±6% -a range larger than errors in transport models only (Locatelli et al., 2015).
However, this uncertainty on total emissions is probably underestimated as the 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 ±5%, much lower 510 than EDGARv4.3.2 and v5 uncertainties derived from the detailed analysis of (Janssens-Maenhout et al., 2019) and , reaching around ±45% at 2σ. Saunois et al. (2020) reported uncertainty of 10-15%, which translates to an uncertainty of about ±10% to ±30% depending on the category, with larger uncertainty in the fossil fuel sectors than in the agriculture and waste sector (Saunois et al., 2020b). However, these uncertainties are also underestimated as they do not consider the uncertainty in each individual estimate, which includes potential uncertainties in activity data, emission factors, 515 and equations used to estimate emissions.
Uncertainties in EDGAR CH4 emissions using a Tier 1 approach are estimated at -33% to +46% at 2σ, but there is great 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 520 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 uncertainties in national anthropogenic emissions of about ± 32% for the 24 member countries of OECD, and up to ±57% for other countries, whose activity data are more uncertain (Janssens-Maenhout et al., 2019).  Figure  530 1. This does not consider uncertainty on each individual estimate. c Uncertainty calculated as ((min-max)/2)/mean*100 from individual estimates for the 2008-2017 decade. This does not consider uncertainty on each individual estimate, which is probably larger than the range presented here. * Mainly due to difficulties in attributing emissions to small specific emission sector.

535
The most recent UN emissions gap report (UNEP, 2020) gives an uncertainty range for global anthropogenic CH4 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 CH4 emissions for a 90% confidence interval. This is justified by the larger uncertainties reported in uncertainties studies on the EDGAR dataset (Janssens-Maenhout et al., 2019;Solazzo et al., 540 2021) as well as for FAO activity statistics by Tubiello et al. (2015).

Anthropogenic N2O emissions
Anthropogenic N2O emissions occur in a number of sectors, namely agriculture, fossil fuel and industry, biomass burning, and waste. The emissions from the agriculture sector have four components: direct and indirect emissions from soil and water bodies (inland, coastal, and oceanic waters), manure left on pasture, manure management, and aquaculture. Besides these main 545 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 (Tian et al., 2020)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 (Tian et al., 2020). in EDGARv6 lead to an average increase of 1.5% yr -1 in total N2O emissions estimates between 1999 and 2018 compared to 560 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 ( Figure 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") (Tian et al., 2020) and FAOSTAT are on average 1.5 Mt N2O yr -1 higher than those from GAINS and EDGAR during 1990-2016, due to much higher estimates of direct emissions from fertilised soils and manure left on pasture. GCP provides the largest estimate, because 565 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 (Tian et al., 2020). EDGAR estimates of anthropogenic N2O emissions as used in this dataset should therefore be considered as lower bound estimates. Differences in N2O emissions across different versions of EDGAR are shown in Figure SM 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, storage and applications (Galloway et al., 2010;Steinfeld et al., 2010) as well as nutrient, crops and soils management (Ciais et al., 2014;Shcherbak et al., 2014). Emission factors are also uncertain (Crutzen et al., 2008;Hu et al., 2012;IPCC, 2019;575 Yuan et al., 2019) and there remains several sources that are not yet well understood (e.g. peatland degradation, permafrost) (Elberling et al., 2010;Wagner-Riddle et al., 2017;Winiwarter et al., 2018). Model-based estimates face uncertainties associated with the specific model configuration as well as parametrisation (Buitenhuis et al., 2018;Tian et al., 2018Tian et al., , 2019. Total uncertainty is also large because N2O emissions are dominated by emissions from soils, where our level of process understanding is rapidly changing. processes and product use (14%) and other (112%). 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 , but is comparable with the ranges for anthropogenic emissions in the global N2O budget (Tian et al., 2020). Overall, we assess the relative uncertainty for global anthropogenic N2O emissions at ±60% for a 90% confidence interval. 590 Table 6 -Comparison of four global N2O inventories: EDGAR ; GCP (Tian et al., 2020); GAINS (Winiwarter et al., 2018); FAOSTAT (FAOSTAT, 2021;Tubiello, 2018;Tubiello et al., 2013) Figure 2 for the years 1980 -2016 (or a subset thereof, depending on the availability of the top-down estimates). Where available, the various top-down estimates agree with each other within uncertainties. The magnitude of the difference between the WMO (2018) and EDGAR estimates varies markedly between species, years and versions of EDGAR; for several HFCs, the top-down and bottom-up estimates often agree within 635 uncertainties for EDGARv6 (but much less often in v5), whereas for c-C4F8, the top-down estimate is more than 100 times the EDGAR estimates. Some similarities and differences have been previously noted for earlier versions of EDGAR (Lunt et al., 2015;Mühle et al., 2010Mühle et al., , 2019Rigby et al., 2010). For SF6, the relatively close agreement between EDGAR v4.0 and a topdown estimate has been discussed in Rigby, et al. (2010). They estimated uncertainties in EDGAR v4.0 of ±10% to ±15%, depending on the year, and indeed, top-down values were consistent within these uncertainties. However, the agreement is 640 now poorer during the 1980s in EDGARv6. For some PFCs (e.g., CF4, C2F6), it was previously noted that some assumptions within EDGAR v4.0 had been validated against atmospheric observations, hence EDGAR might be considered a hybrid of top-down and bottom-up methodologies for these species (Mühle et al., 2010). However, it is unclear for which other species similar validation has taken place, or how these assumptions vary between versions of EDGAR.

645
When species are aggregated into F-gas total emissions, weighted by their current 100-year GWPs based on IPCC AR6 (Forster et al., 2021), we note that in the left panel of Figure 3 the Olivier & Peters (2020) [EDGARv5FT] estimates are around 10% lower than the WMO 2018 values in the 1980s. Subsequently, EDGARv5FT estimates grow more rapidly than the top-down values and are almost 30% higher than WMO 2018 by the 2010s. EDGARv6 emissions are around 10% lower than the WMO 2018 values throughout. Given that detailed uncertainty estimates are not available for all EDGAR F-gas species, we base our 650 uncertainty estimate solely on this comparison with the top-down values (see Figure 3, left panel), and therefore suggest a conservative 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 EDGAR do not include chlorofluorocarbons (CFCs) and hydrochlorofluorocarbons (HCFCs), which are groups 655 of substances regulated under the Montreal Protocol. Historically, total CO2eq 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 2018 emissions of the gases included in EDGARv5 and v6 (1.3 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 CO2eq emissions of the HFCs, HCFCs and CFCs were 660 approximately equal by 2016, with a smaller contribution from PFCs, SF6, NF3 and some more minor F-gases. Therefore, the GWP-weighted F-gas emissions in EDGAR, which are dominated by the HFCs, represent less than half of the overall CO2eq F-gas emissions in 2016.

Aggregated GHG emissions
Based on our assessment of relevant uncertainties above, we apply constant, relative uncertainty estimates for GHGs at a 90% 665 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 GHGs in terms of CO2eq emissions, we are taking the square root of the squared sums of absolute uncertainties for individual (groups of) gases, using 100-year Global Warming Potentials (GWP-100) with values from IPCC AR6 (Forster et al., 2021, Section 7.6 and Supplementary Material 7.SM.6) to weight emissions of non-CO2 gases but excluding uncertainties in the metric itself 670 (see Section 3.7). Overall, this is broadly in line with IPCC AR5 , but provides important adjustments in the evaluation of uncertainties of individual gases (CH4, F-gases, CO2-LULUCF) as well as the approach in reporting total uncertainties across GHGs. The contribution of SLCF emissions to total GHG emissions expressed in CO2eq thus depends critically on the choice of GHG 690 metric and its time horizon. However, even for a given choice, the metric value for each gas is also subject to uncertainties.

GHG emission metrics
For example, the GWP-100 for biogenic CH4 has changed from 21 based on the IPCC Second Assessment Report (SAR) in 1995 to 28 or 34 based on the IPCC AR5 (excluding or including climate-carbon cycle feedbacks) , and to 27 based on IPCC AR6. These changes and remaining uncertainties arise from parametric uncertainties, differences in methodological choices, and changes in metric values over time due to changing background conditions. 695  Parametric uncertainties arise from uncertainties in climate sensitivity, radiative efficacy and atmospheric lifetimes of CO2 and non-CO2 gases, etc. The IPCC AR6 assessed the parametric uncertainty of GWP for CH4 as ±32% and ±40% for time horizons of 20 and 100 years, ±43% and ±47% for N2O, and ±26-31 and ±33-38% for various F-gases (Forster et al., 2021). The uncertainty of GTP-100 for CH4 was estimated at ±83%, which is larger than the uncertainty 700 in a forcing-based metric due to due to uncertainties in climate responses to forcing (e.g., transient climate sensitivity).
 Methodological choices introduce a different type of uncertainty, namely which indirect effects are included in the calculation of metric values and the strength of those feedbacks. For CH4, indirect forcing caused by photochemical decay products (mainly tropospheric ozone and stratospheric water vapour) contributes almost 40% of the total forcing from CH4 emissions. More than half of the changes in GWP-100 values for CH4 in successive IPCC 705 assessments from 1995 to 2013 are due to re-evaluations of these indirect forcings. These 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 CH4 without climate-carbon cycle feedbacks to be 28, 710 whereas including this feedback would raise the value to between 31 and 34 (Gasser et al., 2016;Myhre et al., 2013;Sterner and Johansson, 2017). The IPCC AR6 decided to include clinate-carbon cycle feedbacks by default and no longer reports values without climate-carbon cycle feedbacks (Forster et al., 2021). Overall, we estimate the uncertainty in GWP-100 metric values, if applied to an extended historical emission time series, as 725 ±50% for CH4 and other SLCFs, and ±40% for non-CO2 gases with longer atmospheric lifetimes (specifically, those with 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).

730
For the purpose of this paper, we use GWP-100 metric values from the IPCC AR5 (Myhre et al., 2013) without climate-carbon 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. 735 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 additional warming relative to a given reference level.
Sustained emissions of SLCFs such as CH4 do not cause the same temperature response as sustained emissions of CO2. 740 Showing superimposed emission trends of different gases over multiple decades using GWP-100 as equivalence metric therefore does not necessarily represent the overall 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 simple climate model emulator FaIRv1.6.2 and calibrated to reproduce the pulse-response functions for each gas consistent with the IPCC AR6 (see Forster et al., 2021, Supplementary Material 7.SM.3). There are some differences compared to the contribution of 745 each gas, based on GHG emissions expressed in CO2eq using GWP-100 (see Figure 8), in particular a greater contribution from CH4 emissions to historical warming. This is consistent with warming from CH4 being short-lived and hence having a more pronounced effect in the near-term during a period of rising emissions. Nonetheless, Figure 4 highlights that weighting emissions based on GWP-100 does not provide a vastly different overall story than modelled warming over the historical period when emissions of all gases have been rising, with CO2 being the dominant and CH4 being the second most important 750 contributor to GHG-induced warming. Other metrics such as GWP* (Cain et al., 2019) offer an even closer resemblance between cumulative CO2eq emissions and temperature change, relative to a specified starting point, especially if SLCF emissions are no longer rising but potentially falling, as in mitigation scenarios. the FaIR reduced-complexity climate model. Major GHGs and aggregates of minor gases as a timeseries in a) and as a total warming bar chart with 90% confidence interval added in b). Bottom row: contribution from short-lived climate forcers as a timeseries in c) and as a total warming bar chart with 90% confidence interval added in d). The dotted line in c) gives the net temperature change from short-lived climate forcers. F-Kyoto/Paris includes the gases covered by the Kyoto Protocol and Paris Agreement as well as the HFCs, while F-other includes the gases covered by the Montreal Protocol but excluding the HFCs.
Global anthropogenic GHG emissions grew continuously slower than world GDP across all (groups of) individual gases 800 resulting in a sustained decline in the GHG intensity of GDP as shown in Figure 5. The only exception is the group of F-gases for which the GHG intensity of GDP has significantly increased since 1970, with a marked acceleration during the 1990s and  (Forster et al., 2021). Coloured shadings show the associated uncertainties at a 90 % confidence interval without considering uncertainties in GDP 815 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 Bank, 2021).
The continuous growth in global anthropogenic GHG emissions since the 1970s was mainly driven by activity growth in three 820 major sectors: energy supply, industry and transportation (see Table SM-2; Fig. SM-4). In energy supply and transportation, average annual emissions were about 2.3 and 2.2 times larger for 2009-2018 than for 1970-1979, respectively, growing from 8.4 to 19 GtCO2eqyr -1 and 3.6 to 8.0 GtCO2eqyr -1 , respectively. In industry, average annual GHG emissions were 1.8 times larger, growing from 7.3 GtCO2eqyr -1 in 1970-1979to 13 GtCO2eqyr -1 in 2009. At the sub-sector level, electricity & heat and road transport are the largest segments, growing 2.9 and 2.6 times between 1970-1979 and 2009-2018, respectively, 825 from an average 4.6 to 13 GtCO2eqyr -1 , and 2.2 to 5.7 GtCO2eqyr -1 . The fastest growing sub-sector has been process emissions from cement, which is 4.1 times larger in 2009-2018 compared to 1970-1979, and currently accounts for an average 1.4 GtCO2eqyr -1 . Other rapidly expanding sectors are international aviation (2.8 times larger on 1970-1979 levels), chemicals (1.9 times larger), metals (1.7 times larger), and waste (1.7 times larger). Growth in GHG emissions in AFOLU and buildings has been much more moderate with average annual GHG emissions only about 25% and 10% higher for 2009-2018 than for 1970-830 1979.
Most GHG emissions growth occurred in Asia and Developing Pacific as well as the Middle East, where emissions more than tripled from 6.3 GtCO2eqyr -1 and 0.8 GtCO2eqyr -1 in 1970-1979 to 23 GtCO2eqyr -1 and 2.8 GtCO2eqyr -1 in 2009-2018, respectively (see Table SM-1). Over the same time period GHG emissions grew 2.2 times in Africa and 1.7 times in Latin America and the Caribbean, while average annual anthropogenic GHG emissions levels in developed countries and Eastern 835 Europe and West-Central Asia remained stable. However, Figure 6 highlights important variability at the country level. Note that these country-level estimates exclude CO2-LULUCF emissions, because we assign low confidence to them. First, GHG emissions growth is taking place against the background of large differences in per capita GHG emissions between and within regions. For example, GHG emissions in developed countries have stabilized at high levels of per capita emissions compared to most other regions. Similarly, some countries in the Middle East are among the largest GHG emitters in per capita terms, 840 while other countries of the region such as Yemen have seen comparatively little economic development showing low levels of per capita emissions. Second, the growth in GHG emissions has also been highly varied. For example, several developed countries in Europe such as UK, Germany or France have lower GHG emissions levels today than in the 1970s. In other countries like the US GHG emission levels are still considerably higher today even though they have recently started reducing GHG emissionsunlike Australia or Canada, which have until now only begun stabilizing their GHG emission levels. A 845 comprehensive assessment of country progress in reducing GHG emissions can be found in Lamb et al. (2021b). In Figure 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 Socioeconomic Pathways (SSPs) for different levels of radiative forcing in 2100 using six different IAMs (Riahi et al., 855 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) (Eyring et al., 2016) 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. Peters, 2020b, 2020a;Pedersen et al., 2020;Schwalm et al., 2020). 860 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 SSP3-7.0 and SSP5-8.5 markers, in terms of both levels and growth rates. CO2-FFI emissions still tend towards the higher end of the 865 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. 870 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 Wm -2 forcing at the end of the century, while they are about halved in others. Reductions in CH4 emissions are a bit more 875 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 landuse change emissions (Hausfather and Peters, 2020b). 880  (Forster et al., 2021). The Shared Socioeconomic Pathways (SSPs) are from the SSP database version 2 (Riahi et al., 885 2017;Rogelj et al., 2018b). See also: https://tntcat.iiasa.ac.at/SspDb/). Highlighted scenarios are the markers used in CMIP6 (O'Neill et al., 2016) after harmonisation (Gidden et al., 2019).

2 -Global GHG emissions for the last decade 2009-2018
There is high confidence that global anthropogenic GHG emission levels were higher in 2009-2018 than in any previous decade 890 and GHG emission levels have grown across the most recent decade. Average annual GHG emissions for 2009-2018 were 55±5.9 GtCO2eqyr -1 compared to 46±5.3 and 39±4.8 GtCO2eqyr -1 for 2000-2009 and 1990-1999, respectively (2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016)(2017)(2018) 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). and India (0.95 GtCO2eqyr -1 ) (Figure 8). Among the major emitters, fastest GHG emissions growth was observed for Turkey with average annual rates of 4.2% yr -1 between 2009 and 2018, followed by Indonesia (3.8% yr -1 ), Saudi Arabia (3.4% yr -1 ), India (3.2% yr -1 ), Pakistan (3.1% yr -1 ), and China (2.3% yr -1 ). GHG emission reductions achieved by countries over the last decade are comparatively small even though there is a growing number of countries on sustained emissions reductions 905 trajectories (Lamb et al., 2021b;Le Quéré et al., 2019b). The US showed the largest net anthropogenic GHG emissions reductions of 0.14 GtCO2eqyr -1 between 2009 and 2018, resulting from reductions of about the same size in CO2 emissions from a switch from coal to gas in the context of the shale gas expansion. Other countries with decreasing GHG emission levels were Australia (-0.01 GtCO2eqyr -1 ), Germany (-0.02 GtCO2eqyr -1 ), and the United Kingdom (-0.12 GtCO2eqyr -1 ), where the latter shows the fastest average annual reductions at a rate of 2.9% yr -1 in the sample (Figure 8)in line with some GHG 910 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 Strefler et al., 2018). Further information on country contributions to GHG emission changes since 1990san important reference for UN climate policyare shown in supplementary Figure   SM GtCO2eqyr -1 according to the national inventories 1.2% lower than the data presented here for the same countries. Both datasets, however, agree in terms of the average annual growth rates over the last decade (2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016)(2017)(2018), which stood at -0.4% for the total GHG emissions of the Annex I countries. For single countries there is still some divergence in growth rates observed between the national inventories and the dataset presented here (Figure 8,  Sectoral GHG emissions were either stable or increased between 2009 and 2018. There is high confidence that no substantive GHG emissions reductions were observable for entire sectors at the global level. The most substantial growth was observed in the metal industry with an average annual growth rate of 3.4% yr -1 between 2009 and 2018, followed by the chemical industry 930 (2.5% yr -1 ), road transport (2.0% yr -1 ), electricity and heat (1.9% yr -1 ), and the cement industry (1.7% yr -1 ) (Figure 8, panels d and e). International and domestic aviation, which are small in their contribution to global GHG emissions (and are therefore not shown in Figure 8 e-f), are exhibiting even larger growth rates of 3.8% yr -1 (0.69 GtCO2eq yr -1 ), and 3.7% yr -1 (0.39 GtCO2eq yr -1 ), respectively. indicates its preliminary nature and lower confidence associated with it. Panel b: Waterfall diagrams juxtaposes GHG emissions for 2018 in CO2eq units using GWP-100 values from the IPCC's AR6, AR5, AR4, and AR2, respectively. Error bars show the associated uncertainties at a 90 % confidence interval (see Section 3). Panels c and d show relative (in % yr -1 ) and absolute (in GtCO2eq yr -1 ) average annual changes in GHG emissions for a selection of the largest emitting countries (contributing 75% of global GHG emissions in 2018) excluding CO2-LULUCF emissions as uncertainties in our estimates are too high for country-level reporting. The yellow dots represent the emissions data 945 from UNFCCC-CRFs (2021) that were accessed through Gütschow et al. (2021a). Further comparisons with CRF data are provided in Figures SM-3 and SM-4. Panels e and f show relative (in % yr -1 ) and absolute (in GtCO2eq yr -1 ) changes in GHG emissions for a selection of the largest emitting sectors (see Table 2) (contributing 90% of global GHG emissions in 2018).
In 2018, emissions were 1.1±0.11 GtCO2eqyr -1 or 1.9% higher than the 57±6.0 GtCO2eq in 2017. Most of this growth 960 (0.78±0.062 Gt yr -1 , 2.1% yr -1 ) was related to increases in CO2-FFI emissions. Also F-gas emissions (0.067±0.020 GtCO2eqyr -1 , 5.2% yr -1 ) and CO2-LULUCF emissions (0.12±0.08 Gt yr -1 , 2.1% yr -1 ) increased significantly, but we assign less confidence in the magnitude of the growth particularly for CO2-LULUCF due to the high uncertainties attached. Emissions in CH4 and N2O were rather stable between 2017 and 2018, with growth rates of 0.8% yr -1 and 0.4% yr -1 , respectively. Given prevailing uncertainties there is low confidence that GHG emissions have never been higher than in 2018 as suggested by the data, but 965 high confidence that average annual GHGs emissions have never been higher for a decade than in 2009-2018 (see Friedlingstein et al., 2020;Hoesly et al., 2018).

Fast-track estimates for GHG emissions in 2019
GHG emissions in 2019 are estimated at 59±6.6 GtCO2eq yr -1 . This is 2.2% higher (1.26±0.64 GtCO2eq yr -1 ) than emissions in 2018, and an increase in the annual growth rate compared to 2017-2018 of 1.9% (1.05±0.11 GtCO2eq). These estimates are 970 in large parts derived from less complete information and there is less confidence in the exact magnitude. The magnitude of the recent emissions growth is particularly uncertain, because a large portion of emissions growth between 2018 and 2019 anthropogenic CO2-LULUCF emissions of 6.6±4.6 Gt yr -1 . This was due to a surge of fire emissions from peat burning, 975 deforestation and degradation, occurring in principally in equatorial Asia and the Amazon, and substantially exceeding average rates in the previous decade (Friedlingstein 2020;GFED4.1s;van der Werf et al., 2017). Anthropogenic fire processes are not captured well by the underlying land-use datasets. SFurther, the 2019 estimate was extrapolated for all three bookkeeping estimates by applying additional information on emissions from equatorial Asia peat fires and tropical deforestation and degradation fires (GFED4.1s; van der Werf et al., 2017) in a similar way (Friedlingstein 2020). This explains the consistent 980 upward trend for all three bookkeeping estimates for 2019.

Discussion
In this article we provide a comprehensive, detailed dataset for global, regional, national and sectoral GHG emissions from 990 anthropogenic activities covering the last five decades . This is based on the EDGARv6 GHG emissions inventory, but additionally includes a fast-track update to 2019 and data on CO2-LULUCF emissions from three 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). We note that national emissions inventory submissions reported to the UNFCCC are requested to report 995 uncertainty using a 95% (2σ) confidence interval. The use of this broader uncertainty interval implies, however, a relatively high degree of knowledge about the uncertainty structure of the associated data, which is not present over the emission sectors and species considered here.
Our uncertainty assessment is broadly consistent with previous assessments focussing on all GHGs ; 1000 UNEP, 2020), but we provide some important updates. Our evidence-informed uncertainty judgements are higher for CO2-LULUCF (±70% rather than ±50%) and CH4 (±30% rather than ±20%) drawing from the global carbon budget (Friedlingstein et al., 2020) and the available literature (Janssens-Maenhout et al., 2019;. We note the limited literature on the uncertainties in F-gas emissions in global emissions inventories and recognize the divergence between bottom-up inventory estimates and top-down atmospheric measurements for individual F-gases. Our revised uncertainty estimate for 1005 aggregate F-gas emissions of ±30% (rather than ±20%) reflects the smaller aggregate deviation observed for aggregate F-gas emissions across species. We further acknowledge that we apply the same uncertainty estimates to our fast-track extension to 2019 even though the 2019 estimates themselves will be more uncertain. However, our analysis almost exclusively focusses on the data up to 2018 that is based on full data releases, where our global uncertainty estimates are applied.

1010
Our analysis involves aggregating GHG emissions into a single unit using GWP-100 values from IPCC AR6, which include climate feedbacks. By doing so we follow the practice taken in UNFCCC inventory reporting 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 (Forster et al., 2021) (see Section 3.7). We therefore also use a simple climate model to assess the warming contribution by the individual groups of gases and find that for the historical period when 1015 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 a comprehensive uncertainty analysis that covers CO2-LULUCF as well as F-gas emissions, we estimate the overall uncertainty of aggregated GHG emissions by simply adding the individual uncertainties judgements by (groups of) gases in quadrature under the assumption of their independence. Over time, uncertainties fluctuate between 10% and 14% depending on the composition of gases within the aggregate. 1020 Comprehensive uncertainty analysis of EDGAR data covering all GHGs should be performed in the future, building on . We also provide an initial estimate of metric uncertainty arising from the aggregation of individual GHGs into a single unit (see Section 3.7).
We have used the scientific definition for CO2-LULUCF emissions, in line with our intention to identify GHG fluxes 1025 attributable to human activities and in line with the GCP's (Friedlingstein et al., 2020) and IPCC WG1's (Canadell et al., 2021) approach that split natural from anthropogenic drivers. We have acknowledged that this differs from national GHGI (Grassi et al., 2018) or inventory data provided by FAO , which should be used if consistency in definition with, e.g., the UNFCCC inventories is required. Net CO2-LULUCF emissions estimates are substantially smaller based on inventory data over managed land, because the environmental drivers (e.g. CO2 fertilisation) of terrestrial sink on managed land are 1030 attributed to anthropogenic emissions in UNFCCC inventories. This highlights the potential of land in emissions reduction efforts: on the one hand, net emissions from land-use activities should be minimized by reducing gross emissions (e.g., by stopping deforestation and degradation) and increasing gross removals (e.g., by reforestation) (Roe et al., 2019); on the other hand, vegetation acting as a natural sink to anthropogenic CO2 emissions should be retained, be it via managed land, as in the inventories, or via pristine vegetated lands. 1035 While the distinction between the scientific approach (drivers) using bookkeeping models and the inventory approach (areas) is clear, and methods to map between approaches have been suggested (Grassi et al., 2018, the attribution of environmental and anthropogenic changes differs between methods. Further, it should also be mentioned that system boundaries partly differ across datasets, and FAOSTAT data (Conchedda and Tubiello, 2020) is currently limited to CO2 fluxes 1040 related to forests and emissions from drainage of organic soils under cropland/grassland, excluding other managed land or agricultural conversions. In principle, bookkeeping and DGV models include all fluxes, but are often coarse in their description of management, which observation-based approaches capture . Several authors (Grassi et al., 2018;Obermeier et al., 2020;Pongratz et al., 2014) have shown the strong dependence of CO2-LULUCF emissions estimates on the time a certain land-use change event happened to occur, if environmental changes are represented transiently over time, as is 1045 the case for typical simulations with dynamic global vegetation models. This dependence is eliminated by using bookkeeping estimates, as done here.
Comparisons with other global emissions inventories highlights the comprehensive nature of our dataset covering anthropogenic sources of GHG emissions. However, there are still some important data issues. In particular, F-gas emissions 1050 estimates for some individual species in EDGARv5 and v6 do not align well with atmospheric measurements and the F-gas aggregate emissions over the last decade either overestimate top-down estimates by around 30% (EDGAR v5) or underestimate them by around 10% (EDGAR v6). Furthermore, EDGAR and official national emission reports under the UNFCCC do not comprehensively cover all relevant F-gases species. In particular, chlorofluorocarbons (CFCs) and hydrochlorofluorocarbons (HCFCs), which are regulated under the Montreal Protocol, have historically contributed more to CO2eq emissions as well as observed warming than the F-gases included in our study. 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 important role as a driver of future warming. We also find a need for more transparent methodological documentation of some of the available inventoriesparticularly for F-gas emissions. Moreover, recent work on the global methane budget (Saunois et al., 2020b) and the global nitrous oxide budget (Tian et al., 2020) further suggest discussions on whether global 1060 inventories should be further expanded in terms of their reporting scope.
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 GHG emissions have increased every decade. Emission growth has been varied, but persistent across different (groups of) gases. While CO2 has accounted 1065 for almost 75% of the emission growth since 1970 in terms of CO2eq as reported here, the combined F-gases have grown much faster than other GHGs, albeit starting from very low levels. Today, they make a non-negligible contribution to global warming (see Figure 4) to the 2010s has been the largest since at least the 1970s when the dataset startsand within all human history as suggested by available long-term data (e.g. Friedlingstein et al., 2020;Gütschow et al., 2016Gütschow et al., , 2021bHoesly et al., 2018). While there is 1080 a growing number of countries today on a sustained emission reduction trajectory (Lamb et al., 2021b;Le Quéré et al., 2019a), GHG emissions are growing over time for all global sectors and sub-sectors in our dataset mirroring global GHG emissions trends that are characterized by distinct patterns of development and industrialization. It is therefore important to study the drivers of these reductions as well as patterns of emission growth in more detail at the regional level (Lamb et al., 2021a) and systematically evaluate the impact of climate-relevant policies on regional drivers and trends. 1085 There is a growing availability of global datasets on anthropogenic emissions sources over the last 10-20 years (see Table 1).
However, such global emission inventories often heavily rely on relatively simple Tier-1 estimation methods and few use more complex Tier-2 or Tier-3 methods. Comparison of our estimates with UNFCCC-CRFs by Annex I countries shows considerable discrepancies for some gases and countries (see Figure 8, Figure SM-3, Figure SM-4). On aggregate, there is a 1090 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 differences, as has been done, for example, for CO2 emissions (Andrew, 2020a) and for Europe across all GHGs (Petrescu et al., 2020b(Petrescu et al., , 2021b(Petrescu et al., , 2021a. Figure 9 further highlights the lack of recent official national emissions inventories for many non-Annex I countries. The BURs are also associated with less stringent reporting requirements in terms of sector, gas and time coverage (Deng et al., 2021;Gütschow et al., 2016). This 1095 highlights the important role of global inventories such as EDGAR, CEDS, PRIMAP-hist, FAOSTAT or those from IEA or BP among others that are equally as comprehensive in scope to those from Annex I countries. Despite the importance of highquality 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. Additionally, there are significantly fewer independent estimates for full GHG accounting, in contrast to fossil CO2 emissions. In sectors where production 1100 efficiencies are changing rapidly, as is often the case in developing countries, using emission estimates based on Tier-1 methodologies may 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 1105 emissions factors or atmospheric inversions (IPCC, 2019). This would also support the formulation of more nuanced climate policy goals that reflect changes in emissions intensities as entry points for more comprehensive and ambitious targets to reduce absolute 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). In parallel, it might be a pragmatic way forward to continue and intensify work on comprehensive, up-to-date 1110 global emissions inventories such as EDGAR or CEDS as well as synthetic datasets as presented here or PRIMAP-hist. Future extensions of this work could update country and sector-specific data from UNFCCC inventories wherever possible and available. It could also make sense to add missing emissions componentsparticularly, in non-CO2 emissions from AFOLU and develop fast-track methods to extend the inventories from the last available inventory year to the most recent year.
Keeping global warming well below 2°C and pursuing efforts towards 1.5°C requires dedication and cooperation between 1115 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 1130 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