The consolidated European synthesis of CO2 emissions and removals for EU27 and UK: 1990-2018

Ana Maria Roxana Petrescu, Matthew J. McGrath, Robbie M. Andrew, Philippe Peylin, Glen P. Peters, Philippe Ciais, Gregoire Broquet, Francesco N. Tubiello, Christoph Gerbig, Julia Pongratz, Greet Janssens-Maenhout, 5 Giacomo Grassi, Gert-Jan Nabuurs, Pierre Regnier, Ronny Lauerwald, Matthias Kuhnert, Juraj Balkovič, Mart-Jan Schelhaas, Hugo A. C. Denier van der Gon, Efisio Solazzo, Chunjing Qiu, Roberto Pilli, Igor B. Konovalov, Richard A. Houghton, Dirk Günther, Lucia Perugini, Monica Crippa, Raphael Ganzenmüller, Ingrid T. Luijkx, Pete Smith, Saqr Munassar, Rona L. Thompson, Giulia Conchedda, Guillaume Monteil, Marko Scholze, Ute Karstens, Patrick Brokmann and Han Dolman 10

fluxes from managed ecosystems. The work integrates recent emission inventory data, process-based ecosystem model results, data-driven sector model results, and inverse modelling estimates, over the period 1990-2018. BU and TD 50 products are compared with European national GHG inventories (NGHGI) reported under the UNFCCC in 2019, aiming to assess and understand the differences between approaches. For the uncertainties in NGHGI, we used the standard deviation obtained by varying parameters of inventory calculations, reported by the Member States following the IPCC guidelines. Variation in estimates produced with other methods, like atmospheric inversion models (TD) or spatially disaggregated inventory datasets (BU), arise from diverse sources including within-model uncertainty related 55 to parameterization as well as structural differences between models. In comparing NGHGI with other approaches, a key source of uncertainty is that related to different system boundaries and emission categories (CO2 fossil) and the use of different land use definitions for reporting emissions from Land Use, Land Use Change and Forestry (LULUCF) activities (CO2 land). At the EU27+UK level, the NGHGI (2019) fossil CO2 emissions (including cement production) account for 2624 Tg CO2 in 2014 while all the other seven bottom-up sources are consistent with the NGHGI and 60 report a mean of 2588 (± 463 Tg CO2). The inversion reports 2700 Tg CO2 (± 480 Tg CO2), well in line with the national inventories. Over 2011-2015, the CO2 land sources/sinks from NGHGI estimates report -90 Tg C yr -1 ± 30 Tg C while all other BU approaches report a mean sink of -98 Tg yr -1 (± 362 Tg C from DGVMs only). For the TD model ensemble results, we observe a much larger spread for regional inversions (i.e., mean of 253 Tg C yr -1 ± 400 Tg C yr -1 ). This concludes that a) current independent approaches are consistent with NGHGI b) their uncertainty is too large to allow a "verification" because of model differences and probably also because of the definition of "CO2 flux" obtained from different approaches. The referenced datasets related to figures are visualized at https://doi.org/10.5281/zenodo.4626578 (Petrescu et al., 2020a).

70
Global atmospheric concentrations of CO2 have increased 46% since pre-industrial times (pre-1750) (WMO, 2019). The rise of CO2 concentrations in recent decades is caused primarily by CO2 emissions from fossil sources.
Globally, fossil emissions grew at a rate of 1.3% yr − 1 for the decade 2009-2018 and accounted for 87% of the anthropogenic sources in the total carbon budget (Friedlingstein et al., 2019). In contrast, global CO2 emissions from land use and land use change estimated from bookkeeping models and dynamic global vegetation models (DGVMs) 75 were approximately stable during the same period, albeit with large uncertainties (Friedlingstein et al., 2019).
National GHG inventories (NGHGI) are prepared and reported under the UNFCCC on an annual basis by Annex I countries 1 , based on IPCC Guidelines using national activity data and different levels of sophistication (tiers) for well-defined sectors. These inventories contain time series of annual GHG emissions from the 1990 base year 2 1 Annex I Parties include the industrialized countries that were members of the OECD (Organization for Economic Co-operation and Development) in 1992 plus countries with economies in transition (the EIT Parties), including the Russian Federation, the Baltic States, and several central and eastern European states (UNFCCC, https://unfccc.int/parties-observers, last access: February 2020). 2 For most Annex I Parties, the historical base year is 1990. However, parties included in Annex I with an economy in transition during the early 1990s (EIT Parties) were allowed to choose one year up to a few years before 1990 as reference because of a non-representative collapse during the breakup of the Soviet Union (e.g., Bulgaria, 1988, Hungary, 1985-1987, Poland, 1988, Romania, 1989, and Slovenia, 1986). guidelines allow the use of atmospheric data for external checks within the data quality control, quality assurance and verification process (IPCC 2006 Guidelines, Chapter 6 QA/QC procedures). Only a few countries (e.g. Switzerland, UK, New Zealand and Australia) use atmospheric observations on a voluntary basis to complement their national inventory data with top down estimates annexed to their NGHGI (Bergamaschi et al., 2018).
For the post-2020 reporting (which will start in 2023 for the inventory of year 2021), the Paris Agreement 100 follows on the Kyoto Protocol and, at the EU level, the GHG Monitoring Mechanism Regulation 525 (2013) is replaced by Regulation 1999Regulation (2018 while Regulation 824 (2018) embeds the LULUCF sector with estimates based on spatial information in the EU Climate Targets of 2030. A key element in the current policy process is to facilitate the global stocktake exercise of the UNFCCC foreseen in 2023, which will assess collective progress towards achieving the near-and long-term objectives of the Paris Agreement, also considering mitigation, adaptation and paper, Petrescu et al., 2021,in press). However, CO2 emissions dominate the GHG fluxes and there is need for Monitoring and Verification Support capacity (Janssen-Meanhout et al., 2020) as the reduction of anthropogenic CO2 fluxes become increasingly important for the climate negotiations of the Paris Agreement, and where observationbased data can provide information on the actual situation. In addition, while fossil CO2 emissions are known to relatively high precision, LULUCF activities are generally much more uncertain (RECCAP, CarboEurope) and as 120 described below in sections 2.2. and 3.2.
The current study presents consistently derived estimates of CO2 fluxes from BU and TD approaches for the EU27 and UK, building partly on Petrescu et al. (2020b) for the LULUCF sector and on Andrew (2020) for fossil sectors, while laying the foundation for future annual updates. Every year (time "t") the Global Carbon Project (GCP) in its Global Carbon Budget (GCB) quantifies large-scale CO2 budgets up to year "t-1", bringing in information from 125 global to large latitude bands, including various observation-based flux estimates from BU and TD approaches (Friedlingstein et al., 2020 in review). Except for two sector-specific BU models based on national statistics (EFISCEN and CBM), we note that the BU observation-based approaches used in the GCB and in this paper are based on the NGHG estimates provided by national inventory agencies to the UNFCCC with differences coming from allocation.
They rely heavily on statistical data combined with Tier1 and Tier2 approaches. In our case, focusing on a region that 130 is well covered with data and models (Europe), BU also refers to Tier 3 process-based models or complex bookkeeping models (see section 2). At regional and country scales, no systematic and regular comparison of these observationbased CO2 flux estimates with reported fluxes at UNFCCC is yet feasible. As a first step in this direction, within the European project VERIFY (http://verify.lsce.ipsl.fr/), the current study compares observation-based flux estimates of BU versus TD approaches and compares them with NGHGIs for the EU27+UK and five sub-regions ( Figure 4). The 135 methodological and scientific challenges to compare these different estimates have been partly investigated before (Grassi et al., 2018 for LULUCF;Peters et al., 2009 for fossil sectors) but not in a systematic and comprehensive way including both fossil and land-based CO2 fluxes.
The work presented here represents many distinct datasets and use of models in addition to the individual country submissions to the UNFCCC for all European countries, which while following the general guidance laid out 140 in IPCC (2006) still differ in specific approaches, models, and parameters, in addition to differences in underlying activity datasets. A comprehensive investigation of detailed differences between all datasets is beyond the scope of this paper, though attempts have been previously made for specific subsectors (Petrescu et al., 2020b for AFOLU 3 ; Federici et al., 2015 for FAOSTAT versus NGHGIs). As this is the most comprehensive comparison of NGHGIs and research datasets (including both bottom-up (BU) and top-down (TD) approaches) for Europe to date, we focus here 145 on a set of questions that such a comparison raises: How can one fairly compare the detailed sectoral NGHGI to observation-based estimates? What new information do the observation-based estimates provide, for instance on the mean fluxes, spatial disaggregation, trends and inter-annual variation? What can one expect from such complex studies, where are the key knowledge gaps, what is the added value to policy makers and what are the next steps to take? 150 3 In the IPCC AR5 AFOLU stands for Agriculture, Forestry and Other Land Use and represent a new sector replacing the two AR4 sectors Agriculture and LULUCF We compare official anthropogenic NGHGI emissions with research datasets correcting wherever needed research data on total emissions/sinks to separate out anthropogenic emissions. We analyze differences and inconsistencies between emissions and sinks, and make recommendations towards future actions to evaluate NGHGI data. While NGHGI include uncertainty estimates, special disaggregated research datasets of emissions often lack quantification of uncertainty. While this is also a call to those developers to associate more detailed uncertainty 155 estimates with their products, here we use the median and minimum/maximum (min/max) range of different research products of the same type to get a first estimate of overall uncertainty. Table AA in Appendix A presents the methodological differences of current study with respect to Petrescu et al., 2020b. 2. CO2 data sources and estimation approaches 160 We use data of total CO2 emissions and removals from EU27 + UK from TD inversions and BU estimates, in addition to BU estimates from sector-specific models. We collected data of CO2 fossil and CO2 land 4 Table 1) described in detail by Andrew (2020), while for CO2 land estimates we used BU research-level biogeochemical models (e.g. DGVMs TRENDY-GCP, bookkeeping models, see Table 2). For TD we used global 170 inversions (GCP 2019, Friedlingstein et al., 2019) as well as regional inversions at higher spatial resolution (CarboScopeReg, EUROCOM (Monteil et al., 2019 andKonovalov et al. 2016).
The values are defined from an atmospheric perspective: positive values represent a source to the atmosphere and negative ones a removal from the atmosphere. As an overview of potential uncertainty sources, Appendix B presents the use of emission factor data (EF), activity data (AD), and, whenever available, uncertainty methods used 175 for all CO2 land data sources used in this study. The referenced data used for the figures' replicability purposes are available for download at https:// 10.5281/zenodo.4626578 (Petrescu et al., 2020a). We focus herein on EU27 and the UK. Within the VERIFY project, we have in addition constructed a web tool which allows for the selection and display of all plots shown in this paper (as well as the companion paper on CH4 and N2O), not only for the regions shown here but for a total of 79 countries and groups of countries in Europe. The website, located on the VERIFY project website: 180 http://webportals.ipsl.jussieu.fr/VERIFY/FactSheets/, is accessible with a username and password distributed by the 4 The IPCC Good Practice Guidance (GPG) for Land Use, Land Use Change and Forestry (IPCC 2003) describes a uniform structure for reporting emissions and removals of greenhouse gases. This format for reporting can be seen as "land based"; all land in the country must be identified as having remained in one of six classes since a previous survey, or as having changed to a different (identified) class in that period. According to IPCC SRCCL: Land covers the terrestrial portion of the biosphere that comprises the natural resources (soil, near surface air, vegetation and other biota, and water) the ecological processes, topography, and human settlements and infrastructure that operate within that system". Some communities prefer "biogenic" to describe these fluxes, while others found this confusing as fluxes from unmanaged forests, for example, are "biogenic" but not included in inventories reported to the UNFCCC. As this comparison is central to our work, we decided that "land" as defined by the IPCC was a good compromise.
project. Figure 4 includes also data from countries outside the EU but located within geographical Europe (Switzerland, Norway, Belarus, Ukraine and Rep. of Moldova).

185
UNFCCC NGHGI (2019) emissions are country estimates covering the period 1990-2017. The Annex-I parties to the UNFCCC are required to report emissions inventories annually using the Common Reporting Format (CRF). This annual published dataset includes all CO2 emissions sources for those countries, and for most countries for the period 1990 to t-2. Some eastern European countries' submissions begin in the 1980s. Revisions are made on an irregular basis outside of the standard annual schedule.

CO2 fossil emissions
CO2 fossil emissions occur when fossil carbon compounds are broken down via combustion or other forms of oxidation or via non-metal processes such as for cement production. Most of these fossil compounds are in the form of fossil fuels, such as coal, oil, and natural gas. Another category are fossil carbonates, such as calcium carbonate 195 and magnesium carbonate, which are used as feed stocks in industrial processes, and whose decomposition also leads to emissions of CO2. Because CO2 fossil emissions are largely connected with energy, which is a closely tracked commodity group, there is a wealth of underlying data that can be used for estimating emissions. However, differences in collection, treatment, interpretation and inclusion of various factors such as carbon contents and fractions of oxidized carbon, lead to methodological differences (Appendix A, Table A1) resulting in differences of emissions 200 between datasets (Andrew 2020). In contrast to BU estimates, atmospheric inversions for emissions of fossil CO2 are not fully established (Brophy et al., 2019), though estimates exist. The main reason is that the types of atmospheric networks suitable for fossil CO2 atmospheric inversions have not been widely deployed yet .
In this analysis, the BU CO2 fossil estimates are presented and split per fuel type and reported for the last year when all data products are available (Andrew 2020). In addition to the BU CO2 fossil estimates, we report a fossil 205 fuel CO2 emission estimate for the year 2014 from a 4-year inversion assimilating satellite observations. In order to overcome the lack of CO2 observation networks suitable for the monitoring of fossil fuel CO2 emissions at national scale, this inversion is based on atmospheric concentrations of co-emitted species. It assimilates satellite CO and NO2 data. While the spatial and temporal coverage of these CO and NO2 observations is large, the conversion of the information on these co-emitted species into fossil fuel CO2 emission estimates is complex and carries large 210 uncertainties. Therefore, we focus here on the comparison between the uncertainties in the inversion versus the magnitude and variations of BU estimates without discussing system boundaries and constraints of each of these products (which are instead discussed in Andrew 2020). The detailed descriptions of each of the data products described in Table 1 are found in Appendix A1.  Konovalov-et-al.pdf and Grassland) from both land class remaining 6 (land class remains unchanged) and land class converted 7 (land class 225 changed in the last 20 years). The Wetlands, Settlements, Other Land categories are included in the discussion on total LULUCF activities (incl. Harvested Wood Products (HWP)) presented in sections 3.3.1, 3.3.3 and 3.3.4. Not all the classes reported to the UNFCCC are present in FAOSTAT or other models; in addition some models are sectorspecific. We use the notation of "FL-FL", "CL-CL", and "GL-GL" to indicate forest, cropland, and grassland which remain the same class from year to year. We present separate results from sector-specific models reporting carbon 230 fluxes for FL-FL, CL-CL and GL-GL (the models EPIC-IIASA, ECOSSE, EFISCEN, CBM), those including multiple land use sectors and simulating land use changes ( e.g. Dynamic Global Vegetation Models (DGVMs) ensemble TRENDY v7 (Sitch et al., 2008, Le Quéré et al., 2009), and those employing bookkeeping approaches (H&N (Houghton & Nassikas, 2017) and BLUE (Hansis et al., 2015)). The detailed description of each of the products described in Table 3 is found in Appendix A2.
This system is part of the EUROCOM ensemble, but new runs were carried out for the VERIFY project. The results

245
are plotted separately to illustrate two points: 1) that the CSR runs for VERIFY are not identical to those submitted to EUROCOM (VERIFY runs from CSR included several sites that started shortly before the end of the EURCOM inversion period), and 2) the CSR model was used in four distinct runs in VERIFY, that differ in the spatial correlation of prior uncertainties and in the number of atmospheric stations whose observations are assimilated. By presenting CSR separate from the EUROCOM results, one can get an idea of the uncertainty due to various model parameters in 250 one inversion system, with one single transport model.  Rh, NEE and NBP 1990-2018Bradbury et al., 1993Jenkinson., 1977, 1987Smith et al., 1996, 2010a Gt CO2eq from all sectors (incl. LULUCF) and 4.21 Gt CO2eq (excl. LULUCF) (Appendix B1, Figure B1a). LULUCF only contributed 0.28 Gt CO2 in 2017. This number is consistent with a variety of independent emission inventories (Andrew 2020 andPetrescu et al., 2020b CO2 fossil emissions are dominated by the energy sector, which includes emissions from energy use in energy 290 industries (heat and electricity, industry, transport, and buildings). Out of the remaining three sectors (IPPU, agriculture and waste), IPPU contributes the most to the CO2 emissions, in the EU27+UK these emissions contributed 7.1 %, 7.5 %, 5.6 % and 6.4 % from the total NGHGI, EDGAR v5.0 (2017), CEDS (2014) and PRIMAP (2015) respectively. For agriculture and waste, overall, emissions are very small, accounting in the EU27+UK in 2017 for 0.3% (NGHGI) and 0.4 % (EDGAR v5.0) respectively, therefore this difference is negligible for the total C budget.

Bottom-up estimates by source category
While Figure 1 was made to assist explanation of differences between datasets disaggregated by sector (e.g., energy industry, transport etc.), in Figure 2 we present CO2 fossil emissions results from EU27+UK split by major source categories (solid, liquid, gas). As in Andrew (2020), we observe good agreement between all data sources and 300 UNFCCC NGHGI (2019) data at this level of regional aggregation. The figure presents estimates for the year 2014, as that was the most recent year when all sources reported estimates. BP 9 (2019), CEDS (v_2019_12_23), and EDGAR 10 v5.0 (2020) do not publish emissions split by fuel type at the country level and the latter two are shown as dark grey while the former is shown separating gas from liquid/solid.
While the datasets agree well, there are some differences. The EIA (2020) estimate is higher than others, 305 largely because it includes international bunker fuels in liquid-fuel emissions. The IEA (2019) excludes a number of sources from non-energy use of fuels as well as all carbonates. GCP's total matches the NGHGIs exactly by design but remaps some of the fossil fuels used in non-energy processes from 'Others' to the fuel types used. BP, CEDS, and EDGAR v5.0 all report total emissions very similar to the UNFCCC NGHGI (2019).

9
For BP, the method description allows for emissions from natural gas to be calculated from BP's energy data, but the data for solid and liquid fuels are insufficiently disaggregated to allow replication of BP's emissions calculation method for those fuels.   Figure 3 represents the first attempt to evaluate our single inversion of CO2 fossil emissions, based on satellite CO and NO2 measurements, against BU estimates. The particular inversion reported here provides emission totals for EU11 11 + Switzerland and these exclude non-fossil fuel emissions (Konovalov et al. 2016, Konovalov & Lvova, 2018 330 Figure 3: A first attempt in comparing BU CO2 fossil estimates from eight data sets with a TD fast-track inversion (Konovalov and Lvova, 2018). The data represents EU11 + Switzerland for the year 2014. The uncertainty bar on the inversions represents the 2σ confidence interval.

CO2 land fluxes
This section presents an update to the benchmark data collection by Petrescu et al., 2020b on CO2 emissions and removals from the LULUCF sector (excluding energy-related emissions, but including emissions from land use change, emissions from disturbances on managed land, and the natural sink on managed land), expanding the scope of that work by adding TD estimates from inverse model ensembles and additional BU models run with higher-340 resolution meteorological forcing data over the EU27+UK.
Land CO2 fluxes result from CO2 emissions/removals from one land type converted to another (e.g., forests cleared for croplands), as well as emissions/removals from land occupied by terrestrial ecosystems (depending on the dataset, this may be from managed or unmanaged land, which complicates comparisons with NGHGIs). Such fluxes typically include emissions and sinks in soils and carbon shifts due to harvests, including emissions from the decay of 345 harvested wood products (HWP). Some estimates are specific to a given vegetation/sector type (i.e., only cropland or grassland). As discussed by Petrescu et al., 2020b, the analyzed fluxes therefore relate to emissions and removals from direct LULUCF activities (clearing of vegetation for agricultural purposes, regrowth after agricultural abandonment, wood harvesting and recovery after harvest and management) but also indirect LULUCF for CO2 fluxes due to processes such as responses to environmental drivers (i.e., climate change and CO2 fertilization) on managed land 12 .
Additional CO2 fluxes may occur on unmanaged land, but these fluxes are very small. According NIRs, all land in the EU27+UK is considered managed, except for 5% of France territory.
The indirect CO2 fluxes on managed and unmanaged land, are part of the land sink in the definition used in IPCC Assessment Reports or the Global Carbon Project's annual global carbon budget (Friedlingstein et al., 2019), while the direct LULUCF fluxes are termed "net land-use change flux". Grassi et al. (2018) have shown that the inclusion or exclusion of the indirect sink on managed land in LULUCF is a key reason for discrepancy between reporting and scientific definitions.
Several studies have already analyzed the European land carbon budget from different perspectives and over several time periods using GHG budgets from fluxes, inventories and inversions (Lyussaert et al,, 2012), flux towers (Valentini et al., 2000), forest inventories (Liski et al., 2000, Nabuurs et al., 2018 and IPCC Guidelines (Federici et al., 2015, Tubiello et al., 2020, in addition to the first benchmark data collection of BU estimates (Petrescu et al., 2020b).
Achieving the well-below 2 o C temperature goal of the PA requires, among other things, low-carbon energy technologies, forest-based mitigation approaches, and engineered carbon dioxide removal (Grassi et al., 2018. Currently, the EU27 + UK reports a sink for LULUCF and forest management will continue to be the 365 main driver affecting the productivity of European forests for the next decades (Koehl et al., 2010). For the EU to meet its ambitious climate targets, it is necessary to maintain and even strengthen the LULUCF sink (COM(2020) 562). Forest management, however, can enhance (Schlamadinger et al., 1996) or weaken (Searchinger et al., 2018) this sink. Furthermore, forest management not only influences the sink strength, it also changes forest composition and structure, which affects the exchange of energy with the atmosphere (Naudts et al., 2016), and therefore the 370 potential of mitigating climate change (Luyssaert et al., 2018;Grassi et al., 2019). Meteorological extremes (made more likely through climate change) can also affect the efficiency of the sink (Thompson et al., 2020). Therefore, understanding the evolution of the CO2 land fluxes is critical to meet the goals set out in the Paris Agreement.

Estimates of European and regional total CO2 land fluxes
We present results of total CO2 land fluxes from EU27 + UK and five main regions in Europe: North, West,

375
Central, East (non-EU) and South. The countries included in these regions are listed in Appendix A, table A. Figure 4 shows the total CO2 fluxes from NGHGI for both 1990 base year and mean of 2011-2015 period.
We aim with this period to bring together all information over a five year period for which values are known in 2018.
In fact this can be seen as a reference for what we can achieve in 2023, the year of the first global stocktake, where for most UN Parties the reported inventories will be compiled only up to the year 2021. Given that the GST is only 380 repeated every 5 years, a five-year average is clearly of interest.
The CO2 fluxes in Figure 4 include direct and indirect LULUCF on managed land. The total UNFCCC estimates include the total LULUCF emissions and sinks (by the UNFCCC definition) belonging to all six IPCC land classes and HWP (see section 2.3, Appendix B1, Figure B1b). We plot these and compare them with fluxes simulated with statistical global datasets, bookkeeping and biosphere models, sector-specific models and inversion model 385 ensembles. The error bar represents the variability in models estimates as the min and max values in the ensemble.

395
For all regions and the EU27+UK, we note considerable disagreement between the BU and TD results. We mostly see that BU (observation-based and process-based models) agree well with the NGHGIs, while inversions, in particular EUROCOM, report very strong sinks and high variability of the results compared to the BU estimates. We believe that, in general, the differences we see between regions' TD and BU results are linked to model-specific set-400 ups and definition issues explained in detail in sections 3.3.2 (process-based models and NGHGIs), 3.3.3 (DGVMs, bookkeeping models and NGHGIs) and 3.3.4 (all BU, TD and NGHGIs). As the current analysis is a first attempt to quantify EU27+UK estimates as a whole, we aim in the future to deepen the analysis for regional/country results.

405
In Figure 5 we show the CO2 LULUCF flux decadal change from UNFCCC NGHGI (2019). The contribution of each category ("remaining" and "converted") to the overall reduction of CO2 emissions in percentages between the three mean periods (grey columns are the mean values over 1990-1999, 2000-2009 and 2010-2017). The "+" and the "-" signs represent a source and a sink to the atmosphere. LUC(-) are the land use conversion changes that increase the strength of the LULUCF sink between two averages; LUC(+) are the land use conversion changes that decrease 410 the strength of the overall LULUCF sink. Note that the sectors inside LUC(-) may be sources or may be sinks, but between the two average periods, they become more negative. For the period between 1990-1999 mean and 2000-2009 mean the overall reduction is -9.5 % (i.e., increased land sink) with positive contribution from FL-FL and LUC(+) (wetlands, settlements and other land conversions) contributing to weakening the overall sink (+3.5 %) 13   We see that HWP emissions are by far the major contributor but in different directions across the two periods, from strengthening the sink between 1990-1999 and 2000-2009 to reducing the sink in the second period. This is 430 mostly due to the specific accounting approach where a reduction on the amount of harvest, such as the one occurred after the economic crisis in 2008, progressively reduced the inflow of raw material and, taking into account the decay rate applied to each commodity, this further reduced the C stock within the same pool. Therefore, Figure 5 suggests that carbon emissions from HWP decay became greater than the amount of carbon entering HWP in recent decades.

435
In this section we present annual total net CO2 land emissions between 1990-2018 i.e., induced by both LULUCF and other (environmental changes) processes from class specific models as well as from models that simulate some or all classes.

440
we compare modelled net biome production (NBP) estimates (including soil plus living and dead biomass C stock change) simulated with class-specific ecosystem models to UNFCCC and FAOSTAT data consisting of net carbon stock change in the living biomass pool (aboveground and belowground biomass) associated with forests and net forest conversion including deforestation.
The Forest Land estimates, which remain in this class (FL-FL) in Figure 6, were simulated with ecosystem 445 models (CBM, ORCHIDEE, EFISCEN) (described in Appendix A2 and Appendix For ORCHIDEE, the model shows a high inter-annual variability in carbon fluxes because ORCHIDEE operates on a sub-daily time step for most biogeochemical and biophysical processes except for a daily time step for "slow" processes like carbon allocation in the vegetation reservoirs, while all other models involved in this comparison use forest inventory data which is reported every few years (i.e., five years for FRA). ORCHIDEE results indicate that

15
The Global Forest Resource Assessment (FRA) is the supplementary source of Forest land data disseminated in FAOSTAT, http://www.FAO.org/forestry/fra/en/ climatic perturbations and extreme events (multi-month droughts, in particular) can have significant impacts on the net carbon fluxes depending on when they occur. This is to some extent supported by dendrometer data although highly varying per site and tree species obscuring a significant net effect (Scharnweber et al., 2020). It should also be noted that dendrometer data measures carbon stored in individual trees, while the NBP reported in figures in this paper include fluxes from litter and soil respiration. The variability of the weather data affects all components of the carbon 465 dynamics in the ecosystems (hence NBP), with for instance impacts on C assimilation rates, length of the growing season, dynamics of respiration rates and allocation of the carbon in the plant (cf. Figure 1 and 2 in Reichstein et al. (2013)).
The UNFCCC NGHGI uncertainty of CO2 estimates for FL-FL across the EU27+UK, computed with the error propagation method (95% confidence interval) (IPCC, 2006), ranges between 23 % -30 % when analyzed at the  Some of the reasons for differences between estimates we see in Figure 6 are linked to different activity data 485 (e.g. forest area) the models use, for example the stronger sink reported by FAOSTAT compared to UNFCCC NGHGI.
By analyzing three of the forest area products (ESA-CCI LUH2v2 (Hurtt et al., 2020) used in ORCHIDEE, FAOSTAT and UNFCCC) we found the following:  For this study, the ORCHIDEE model used a so-called ESA-CCI LUH2v2 PFT distribution (a combination of the ESA-CCI land cover map for 2015 with the historical land cover reconstruction from LUH2 (Lurton 490 et al., 2020)), and assumes that the shrub land cover classes are equivalent to forest. In terms of area, the original ESA-CCI product corresponding to our domain of the EU-27+UK shows shrub land equal to about 50 % of the tree area in 2015. A similar analysis using the FAOSTAT domain Land Cover, which maps and disseminates the areas of MODIS and ESA-CCI land cover classes to the SEEA land cover categories (http://www.fao.org/faostat/en/#data/LC), shows that shrub-covered areas are around 495 20 % of that of forested areas for the EU-27+UK. The impact of classifying shrubs as "forests" on the total carbon fluxes could therefore account for a significant percentage of the differences between ORCHIDEE and other results in Figure 6. ESACCI LUH2v2 does not include the 20-year transition period, as included in the IPCC reporting guidelines. This could be 1% of the forests in Europe, but there is a considerable uncertainty in that based on the transition data seen between the maps.  The three process-based models report sinks in most years (means of -12, -49 and -23 Tg C respectively), contrary to the NGHGIs, which report a small but constant source over the whole period (mean of 5.6 ± 3.5 Tg C) with almost no inter-annual variability by construction. The source reported by NGHGIs, at EU level, is mostly attributed to emissions from cropland on organic soils 16 in the northern part of Europe which emit CO2 due to C 525 oxidation from tillage activities. As an example, Finland and Sweden report together more than half of the total area of organic soil in Europe. Organic soils are an important source of emissions when they are under management practices that disturb the organic matter stored in the soil. In general, emissions from these soils are reported using country-specific values when they represent an important source within the total budget of GHG emissions. In the 16  southern part of Europe, the two categories (CL-CL and GL-GL) are a sink, due to a lack of organic soils in those 530 regions and due to an abandonment trend of land converting arable land to grassland (EU NIR, 2019). In addition, NGHGIs assume that all aboveground biomass of non-woody crops re-enters the atmosphere at harvest. In models like ORCHIDEE and EPIC-IIASA, only part of the aboveground biomass is harvested and enters the atmosphere, and the rest (approximately 50% of the aboveground carbon) enters the soil and decays. Given more favorable growing conditions due to climatic changes and CO2 fertilization, this can lead to more carbon entering the soil in ORCHIDEE 535 in recent decades, which is driving the CL-CL sink observed in the model. The strongest sink reported by ECOSSE model is linked to the soil C model (RothC) used, which simulates a large 'inert pool' which thus leads to a slower C turnover time in the soil (compared to ORCHIDEE or EPIC-IIASA) and thus to significantly larger sink. This 'respiration' aspect of RothC will be addressed in the next synthesis.
According to Ciais et al., 2010, a small carbon source would be a realistic assumption for croplands and in line with 540 the NGHGI report. Thus, while the NGHGI and the three process-based models show a different sign of the CO2 flux, the difference is a result of the processes included/definitions used in each approach, as explained above.
For the inter-annual variability all three models follow the same dynamic, but the impacts of climate extremes are different with significantly larger impacts in ORCHIDEE. While ORCHIDEE shows a strong reaction to drought impacts changing from a sink to a source (e.g. for 2003, which is reported as a very dry year ), the 545 other two models follow ORCHIDEE's variation, but show less extremes. As ECOSSE directly simulates the annual net primary production (NPP) (i.e., internal component model (MIAMI) implemented in ECOSSE) and not the intraannual gross primary production (as in ORCHIDEE), the impact of season specific climate anomalies is smaller than in ORCHIDEE.  Figure 8 shows the CO2 flux of the grassland remaining grassland category, GL-GL. Grassland definition in the IPCC includes rangelands and pasture land that is not considered as cropland, as well as systems with vegetation that fall below the threshold used in the forest land category. This category also includes all grassland from wild lands 560 to recreational areas as well as agricultural and silvo-pastural systems, subdivided into managed and unmanaged, consistent with national definitions (Petrescu et al., 2020b). The NGHGIs of countries in the EU-27+UK report emissions from managed pastures only, which, in 2010, represented a minimum of 58 % (Chang et al., 2016) of the total managed grassland area in the EU. Since almost all European grasslands are somehow modified by human activity and have to a major extent been created and maintained by agricultural activities, they could be defined as 565 "semi-natural grasslands", even if their plant communities are natural (EU LIFE, 2008). Therefore, NGHGIs report a small mean source over 1990-2017 (9 Tg C) primarily due to the use of EFs from national statistics which are linked to intensive management practices applied to grasslands in the EU.
Out of all the models used in this study, only ORCHIDEE and ECOSSE report fluxes from this category.
Grasslands in ORCHIDEE do not undergo any specific management and are not separated from pasturelands. Therefore, discrepancies between ORCHIDEE and the NGHGI data result in the first reporting a mean sink over 1990-2017 of -12 Tg C while official inventories report a small source, as explained above. The sink in ORCHIDEE is due to the fact that the CO2 fertilization effect increases the NPP over time and also increases input of C to the soil, which then leads to increased soil C stocks. The strong sink simulated by ECOSSE (-94 Tg C in mean) is the result of using a limiting scenario where intensively managed grasslands, i.e., high grazing intensity and high yield removal, are not 575 included, thus favoring high soil carbon storage. These effects are similar to that seen in croplands (see above), resulting from the CO2 fertilization effect.

580
In this section we attempt to present a comprehensive analysis of CO2 emissions and sinks for the LULUCF sectors. Here we try to compare the sum of all categories and sectors of the NGHGIs discussed in Figure 5   The bookkeeping models, BLUE and H&N, calculate net emissions from land use change including immediate emissions during land conversion, legacy emissions from slash and soil carbon after land-use change, regrowth of secondary forest after abandonment, and emissions from harvested wood products when they decay. They thus do not account for the net fluxes occurring in the "remaining" land categories due for 610 instance the CO2 fertilization effects or climate changes. One exception to this are fluxes from wood harvested, which is a primary source of emission on managed forest land and also included in bookkeeping models. As seen before in Figure 5, this component can present a significant flux.
Given all these differences in terms of activities, the comparison in this section should be considered as a 615 first step that raises both important aspects of the C cycle and questions that need to be addressed in the future. Going toward a more specific comparison of only net land-use change fluxes would require additional considerations. In GCP's annual global carbon budget, this term is estimated by global DGVMs as the difference between a run with and a run without land-use change and by bookkeeping models. Such an estimate is given in Figure 13 in Petrescu et al., 2020b for forest land. While attractive, such a plot does not fully resolve the differences mentioned above. In 620 particular, questions remain about net vs. gross land use change, managed vs. unmanaged land, and emissions from wood harvest. In addition, UNFCCC "convert" emissions (i.e., emissions resulting from land that has been converted from one type to another) are calculated for 20 years following conversion. FAOSTAT, DGVMs, and bookkeeping models typically only include "convert" fluxes from the year following conversion, although bookkeeping models can more easily include this transition period.

625
17 According IPCC 2006 guidelines the reporting is done for the five LULUCF carbon pools: above-ground biomass, belowground biomass, dead wood, litter, and soil organic matter   Grassi et al., 2018a, Petrescu et al., 2020b. Unmanaged area in the EU27+UK is negligible and sums up only 4 Mha. The similarities between bookkeeping models and UNFCCC can be explained by the fact that, despite a smaller forest sink in H&N, they both report a small sink in non-forest land 645 uses while for these land uses UNFCCC reports a source (Figures 7 and 8).
The UNFCCC LULUCF estimates contain CO2 emissions from all six land use classes and HWP, including remaining classes and conversion to and from a class to another. ORCHIDEE (-93.9 Tg C) shows large variabilities (black diamonds), mostly following the temporal patterns of the mean from TRENDY v7 DGVMs (-103 Tg C) (grey bars) as detailed above. Note again that ORCHIDEE is also part of the TRENDY ensemble, but with a different 650 meteorological forcing (coarser resolution, 0.5°) than the one used within the VERIFY project (around 0. 1° resolution).
The differences between bookkeeping models and UNFCCC and FAOSTAT are discussed in detail in Petrescu et al., 2020b cf. Fig. 12, who concludes that the key difference between bookkeeping models, on the one hand, and FAOSTAT and UNFCCC methodologies, on the other, is that the latter are based on the managed land 655 proxy (Grassi et al., 2018a). ORCHIDEE model and the TRENDY v7 ensemble means show much higher inter-annual variability due to the sensitivity of the model fluxes to highly variable meteorological forcing and the models' subdaily time steps which allow for much more rapid responses to changing conditions (i.e. 2003 extreme drought year), as already discussed in the previous sections. The incorporation of variable climate data and the fact that DGVM models simulate explicitly climate impacts on CO2 fluxes, which inventories and bookkeeping do not, explain these 660 differences.
DGVMs estimate net land-use emissions as the difference between a run with and a run without land-use change, and their estimate includes the loss/gain of additional sink capacity, that is, the sink that favors the environmental changes (e.g. CO2 fertilization). This sink created over forest land in the simulation without land use change is "lost" in the simulation with land use change (i.e., deforestation) because agricultural land lacks the woody 665 material and thus has a higher carbon turnover (Gasser et al., 2013, Pongratz et al., 2014and cf. Figure 12 in Petrescu et al., 2020b. This different definition from bookkeeping models historically implies on average higher carbon 'land use' emissions from DGVMs when an ecosystem is converted to another with a lower carbon density, even if all postconversion carbon stocks changes were the same in DGVMs and bookkeeping models. Tg C), with a difference of 48 Tg C yr -1 that is well within the mean uncertainty of the regional inversion ensemble (about 250 Tg C yr -1 ). It also matches well with the TRENDY v7 DGVMs trend which is smaller (+7.3 Tg C yr -2 ) than 685 that of the global GCP inversions (-16 Tg C yr -1 ). On the other hand, the large range of variability in the EUROCOM ensemble estimates (+ 335 Tg C in 2015 to -615 Tg C in 2013) demonstrates that there is still a very significant uncertainty in the TD estimates. This variability seen from the TD estimates is primarily due to uncertainties in atmospheric transport modeling, boundary conditions and uncertainty inherent to the limitation of the observation network.

690
Additional analyses are still ongoing with the different inversion ensembles to analyze the factors controlling the large difference obtained when compared to BU approaches (for instance, the effect of the a priori fluxes, observation sites, a priori flux and observation uncertainties, and boundary conditions). This paper should be taken cautiously as a first comparison at a spatial scale not investigated so far (i.e., EU27+UK).  Also, noteworthy is that the global inversions provide reliable results at a global scale (following the atmospheric global CO2 growth rate) but the ranges of estimates when considering continental to regional scales 715 increase significantly due to the difficulties of the inversion systems to separate regional fluxes (e.g. Friedlingstein et al 2020). Note also that these systems are still primarily designed for large scale flux estimates (for instance the CarboScope global system uses a transport model at coarse spatial resolution (4° x 5°) and an error correlation length of 1000 km over land). The regional inversions (EUROCOM and CarboScopeReg) are still systems in development with additional complexity due to the treatment of the boundary fluxes (compared to the global systems).

720
For the models, differences result from choices in the simulation setup and depend on the type of model used bookkeeping models, DGVMs, or inventory-basedand whether fluxes are attributed to LULUCF emissions due to the cause or place of occurrence (indirect fluxes on managed land included in NGHGIs and FAOSTAT e.g., changes due to human-induced climate change, including CO2 fertilization and nitrogen deposition changes) (Petrescu et al., 2020b). Table 3 below highlights these differences by presenting an overview of processes included in the models, 725 seen for the moment as the main cause of discrepancies between estimates shown in Figure 10.

Summary and concluding remarks
The overview and variety of data products described in this study is the first of a series of European CO2 synthesis papers presenting and investigating differences between UNFCCC NGHGIs, bottom-up data-based 750 inventories, high resolution observation-based BU models, and TD approaches represented by both global and regional inversions.
The CO2 fossil emissions dominate the anthropogenic CO2 flux in the EU27+UK. Fossil CO2 emissions are more straightforward to estimate than ecosystem fluxes. Different BU methods have only minor differences with respect to the NGHGI. These differences can often be attributed to different definitions or assumptions about activity 755 data or emission factors or by the allocation of fuel types to different sectors (see Fig. 2, section 3.2). Currently, TD methods, albeit only a single inversion using CO/NOx proxies to determine CO₂ fossil emissions, show broad agreement with the BU estimates. The TD inversion is not yet capable of verifying the minor differences between the BU estimates. However, a substantial decrease in the level of uncertainty is expected in the near-term with the largescale deployment of observation networks dedicated to detecting fossil fuel emissions (e.g., with launch of the CO2M 18 760 constellation in 2025, Maenhout al., 2020). In the short-term, methodological improvements and the potential coassimilation of existing CO2 satellite data are also expected to lead to significant decreases in the uncertainty.
The CO2 land fluxes belong to the LULUCF sector, which is one of the most uncertain sectors in UNFCCC reporting due in part to the fact that these fluxes can be either sinks or sources. The IPCC guidelines prescribe methodologies that are used to estimate the CO2 fluxes in the NGHGI, but differences between countries continue to 765 exist due to the use of specific national circumstances (as permitted under the 2006 IPCC guidelines). When we analyzed the estimates from multiple BU sources (inventories and models) we observe similar sources of uncertainties: (a) differences due to input data and structural/parametric uncertainty of models (Houghton et al., 2012) and ( (Pongratz et al., 2014;Grassi et al., 2018b, Petrescu et al., 2020b. More accurate estimates for LULUCF data will be needed in the post-2020 reporting for the EU27 and UK since the LULUCF sector will now 770 contribute to the EU's 2030 targets. To better assess natural variability and trends we believe a reconciliation of BU and TD estimates should focus on clearly defined activities over a given period (e.g. 5 years) and regions as presented in Figure 4. The considerable differences in the agreement between BU and TD estimates from regional split are related to areas and for some regions (e.g., Eastern Europe) sparseness of observation data. Regarding the detailed sector-specific and inversion results (Figures 6,7,8,9 and 10) often differences come from choices in the simulation 775 setup and depend on the type of model usedbookkeeping models, DGVMs, inventory-based or inversion ensembles.
Results also differ based on whether fluxes are attributed to LULUCF emissions due to the cause (e.g., direct or indirect) or place of occurrence. For example, indirect fluxes on managed land are included in NGHGIs and FAOSTAT, while additional sink capacity (e.g. Petrescu et al, 2020b) is included in estimates from process-based models (e.g., ORCHIDEE or TRENDY DGVMs). A more in depth analysis of regional/country level is foreseen as 780 part of the overall long-term VERIFY's objectives.
All observation-based BU estimates for LULUCF presented in this study show similar magnitudes and trends compared to the NGHGIs but generally differ in the specific values. We notice stronger similarities between NGHGIs and models using national forest inventory data (e,g. CBM, EFISCEN). For cropland and grassland sector-specific models (ECOSSE, EPIC-IIASA) the differences between their results and the NGHGIs are due to differences in input data, process representation (in particular those linked to soil organic matter decomposition) and management representation. In general, management is one of the main drivers for the carbon balance of croplands and grasslands.
However, spatial data on management is scarce and can have high uncertainty. For EPIC-IIASA specifically, the regional carbon simulation results for managed cropland are almost evenly impacted by model parameterization, soil input accuracy, and crop management regionalization (Balkovič et al. 2020). For the overall estimation of emissions 790 from LULUCF activities on all land types (Figure 9), the comparison is made more challenging as results from both land use and land use changes are presented. Comparing only the "effect of land use change" (conversion) is nontrivial and presents an area for improvement to be handled in next synthesis.
Observation-based BU estimates of LULUCF provide large year-to-year flux variability (Figures 6,7,8,9), contrary to the NGHGIs, primarily due to the effect of varying meteorology especially through the duration and The next steps needed to improve and facilitate the reconciliation between BU and TD estimates will include 1) as already discussed in Petrescu et al., 2020b, BU process-based models incorporating unified protocols and guidelines for uniform definitions should be able to disaggregate their estimates to facilitate comparison to NGHGI and 2006 IPCC practices (i.e., managed vs. unmanaged land, 20-year legacy for classes remaining in the same class, distinction of fluxes arising solely from land use change); 2) for sector-specific models, especially for cropland and 810 grassland, improving treatment of the contribution of soil organic carbon dynamic to the budget; 3) for TD estimates, using the Community Inversion Framework currently under development (Berchet et al., 2020) to better assess the different sources of uncertainties from the inversion set-ups (model transport, prior fluxes, observation networks); 4) for the overall comparison of BU and TD fluxes, incorporating the contribution of lateral fluxes of carbon by human activities and rivers that connect CO2 uptake in one area with its release in another .

815
From this analysis we demonstrate that a complete, ready-for-purpose monitoring system providing annual carbon fluxes across Europe does not yet exist. Therefore, for consistent future estimates to be used in the global stock take exercise to reach the Paris Agreement targets, significant effort must still be undertaken to reduce the uncertainty across all potential methods used in such a system (e.g. Maenhout et al., 2020).

VERIFY project
VERIFY's primary aim is to develop scientifically robust methods to assess the accuracy and potential biases in national inventories reported by the parties through an independent pre-operational framework. The main concept is to provide observation-based estimates of anthropogenic and natural GHG emissions and sinks as well as associated uncertainties. The proposed approach is based on the integration of atmospheric measurements, improved emission 830 inventories, ecosystem data, and satellite observations, and on an understanding of processes controlling GHG fluxes (ecosystem models, GHG emission models).
Two complementary approaches relying on observational data-streams will be combined in VERIFY to quantify GHG fluxes:

1) atmospheric GHG concentrations from satellites and ground-based networks (top-down atmospheric inversion
835 models) and 2) bottom-up activity data (e.g. fuel use and emission factors) and ecosystem measurements (bottom-up models).
For CO2, a specific effort will be made to separate fossil fuel emissions from ecosystems fluxes. For CH4 and N2O, we will separate agricultural from fossil fuel and industrial emissions. Finally, trends in the budget of the three GHGs will be analysed in the context of NDC targets.

840
The objectives of VERIFY are: Objective 1. Integrate the efforts between the research community, national inventory compilers, operational centres in Europe, and international organisations towards the definition of future international standards for the verification of GHG emissions and sinks based on independent observation.
Objective 2. Enhance the current observation and modelling ability to accurately and transparently quantify the sinks and sources of GHGs in the land-use sector for the tracking of land-based mitigation activities.
Objective 3. Develop new research approaches to monitor anthropogenic GHG emissions in support of the EU commitment to reduce its GHG emissions by 40% by 2030 compared to the year 1990.
Objective 4. Produce periodic scientific syntheses of observation-based GHG balance of EU countries and practical policy-oriented assessments of GHG emission trends, and apply these methodologies to other countries.
For more information on project team and products/results check https://verify.lsce.ipsl.fr/.

Bottom-up emission estimates
For further details, see Andrew (2020) 865

UNFCCC NGHGI (2019)
The Annex-I parties to the UNFCCC are required to report emissions inventories annually using the Common Reporting Format (CRF). This annual published dataset includes all CO2 emissions sources for those countries, and for most countries for the period 1990 to t-2. Some eastern European countries' submissions begin in the 1980s.
Revisions are made on an irregular basis outside of the standard annual schedule. For complete description see for each category and greenhouse gas, and then aggregates these uncertainties, for all categories and greenhouse gas components, to obtain the total uncertainty for the inventory. The Tier 2 method for uncertainties determination is the same, in principle, but it also considers the distribution function for uncertainties and carries out aggregation using Monte Carlo simulation. In the Tier 2 method, the process also necessarily includes the determination of the probability 880 density function for both parameters.

EDGAR v5.0
The first edition of the Emissions Database for Global Atmospheric Research was published in 1995. The dataset now includes almost all sources of fossil CO2 emissions, is updated annually, and reports data for 1970 to n-

BP
BP releases its Statistical Review of World Energy annually in June, the first report being published in 1952.
Primarily an energy dataset, BP also includes estimates of fossil-fuel CO2 emissions derived from its energy data. The emissions estimates are totals for each country starting in 1965 to year n-1. For complete description see Andrew, 900 2020.

CDIAC
The original Carbon Dioxide Information Analysis Center included a fossil CO2 emissions dataset that was long known as CDIAC. This dataset is now produced at Appalachian State University, and includes emissions from fossil fuels and cement production from 1751 to n-3. Fossil-fuel emissions are derived from UN energy statistics, and 905 cement emissions from USGS production data. For complete description see Andrew, 2020.

EIA
The US Energy Information Administration publishes international energy statistics and from these derives estimates of energy combustion CO2 emissions. Data are currently available for the period 1980-2016. For complete description see Andrew, 2020.

IEA
The International Energy Agency publishes international energy statistics and from these derives estimates of energy combustion CO2 emissions including from the use of coal in the iron and steel industry. Emissions estimates start in 1960 for OECD members and 1971 for non-members, and run through n-1 for OECD members' totals, and n-2 for members' details and non-members. Estimates are available by sector for a fee. For complete description see

GCP
The Global Carbon Project includes estimates of fossil CO2 emissions in its annual Global Carbon Budget publication. These includes emissions from fossil fuels and cement production for the period 1750 to n-1. For complete 920 description see Andrew, 2020.

CEDS
The Community Emissions Data System has included estimates of fossil CO2 emissions since 2018, with an irregular update cycle. Energy data are directly from IEA, but emissions are scaled to higher-priority sources, including national inventories. Almost all emissions sources are included and estimates are published for the period 1750-2014.

925
Estimates are provided by sector. For complete description see Andrew R. M., ESSD, 2020.

Fast-track fossil CO2 emission inversion
The so called KL18 inversion product (Konovalov and Lvova 2018)   and from an assessment of the uncertainties in the CO2/CO and CO2/NOx emission ratios (based on their spatial variability). The preliminary results indicate that the uncertainty in the information from the CO inversion is too high to provide reliable estimates of the CO2 fossil emissions when using CO satellite data only, or to provide weight to 950 this information when using CO2 fossil estimates from both the CO and NOx inversions. The estimates based on NO2 data are close to EDGAR v4.3.2 in 2012. These estimates are quite constant over the 4-year period while we assume that the CO2 fossil emissions followed a significant negative trend during this period. The analysis shows that the uncertainties in these estimates can explain the difficulty to detect such a trend.

Bottom-up CO2 estimates UNFCCC NGHGI 2019 -LULUCF
Under the convention and its Kyoto Protocol national greenhouse gas (GHG) inventories are the most important source of information to track progress and assess climate protection measures by countries. In order to

ORCHIDEE
ORCHIDEE is a general ecosystem model designed to be coupled to an atmospheric model in the context of 990 modeling the entire Earth system. As such, ORCHIDEE calculates its prognostic variables (i.e., a multitude of C, H2O and energy fluxes) from the following environmental drivers: air temperature, wind speed, solar radiation, air humidity, precipitation and atmospheric CO2 concentration. As the run progresses, vegetation grows on each pixel, divided into thirteen generic types (e.g., broadleaf temperate forests, C3 crops), which cycle carbon between the soil, land surface, and atmosphere, through such processes such as photosynthesis, litter fall, and decay. Uncertainty in the ORCHIDEE model arises from three primary sources: parameters, forcing data (including spatial 1015 and temporal resolution), and model structure. Some researchers argue that the initial state of the model (i.e., the values of the various carbon and water pools at the beginning of the production run, following model spinup) represents a fourth area. However, the initial state of the model is defined by its equilibrium state, and therefore a strong function of the parameters, forcing data, and model structure, with the only independent choice being the target year of the initial state. Out of the three primarily areas of uncertainty, the climate forcing data is dictated by the 1020 VERIFY project itself, thus removing that source from explaining observed differences among the models, although it can still contribute to uncertainty between the ORCHIDEE results and the national inventories. The land use/land cover maps, another major source of uncertainty for ORCHIDEE carbon fluxes, have also been harmonized to a large extent between the bottom-up carbon budget models in the project. Parameter uncertainty and model structure thus densities at the much coarser spatial resolution of the so-called COSCAT regions. All estimates were then downscaled to 0.1° using the spatial distribution of European inland water bodies. Note that in contrast to Hastie et al. (2019), the areal distribution of lakes was extracted from the HYDROLAKES database (Messager et al., 2016), to be consistent with the estimates of inland water N2O and CH4 presented by Petrescu et al., 2021 in press at ESSD. the European forests. It is currently applied to 26 EU Member States, both at country and NUTS2 level (Pilli et al., 2016b).

Uncertainty
Based on the model framework, each stand is described by area, age and land use classes and up to 10 classifiers based on administrative and ecological information and on silvicultural parameters (such as forest 1055 composition and management strategy). A set of yield tables define the merchantable volume production for each species while species-specific allometric equations convert merchantable volume production into aboveground biomass at stand-level. At the end of each year the model provides data on the net primary production (NPP), carbon stocks and fluxes, as the annual C transfers between pools and to the forest product sector.
The model can support policy anticipation, formulation and evaluation under the LULUCF sector, and it is 1060 used to estimate the current and future forest C dynamics, both as a verification tool (i.e., to compare the results with the estimates provided by other models) and to support the EU legislation on the LULUCF sector (Grassi et al., 2018a).
In the biomass sector, the CBM can be used in combination with other models, to estimate the maximum wood potential and the forest C dynamic under different assumptions of harvest and land use change (Jonsson et al., 2018).
Uncertainty: Quantifying the overall uncertainty of CBM estimates is challenging because of the complexity of each With the help of biomass expansion factors, stem wood volume is converted into whole-tree biomass and subsequently to whole tree carbon stocks. Information on litter fall rates, felling residues and natural mortality is used as input into the soil module YASSO (Liski et al. 2005), which is dynamically linked to EFISCEN and delivers information on Uncertainties: Sensitivity analysis on EFISCEN v3 is described in detail by Schelhaas et al. 2007 (the manual). Total sensitivity is caused by especially young forest growth, width of volume classes, age of felling and few more. Scenario uncertainty comes on top of this when projecting in future.

EPIC-IIASA (croplands)
The Environmental Policy Integrated Climate (EPIC) model is a field-scale process-based model (Izaurralde et al., 2006;Williams, 1990) which calculates, with a daily time step, crop growth and yield, hydrological, nutrient and carbon cycling, soil temperature and moisture, soil erosion, tillage, and plant environment control.
Potential crop biomass is calculated from photosynthetically active radiation using the radiation-use efficiency concept modified for vapor pressure deficit and atmospheric CO2 concentration effect. Potential biomass is adjusted to actual biomass through daily stress caused by extreme temperatures, water and nutrient deficiency or inadequate aeration.
The coupled organic C and N module in EPIC (Izaurralde et al., 2006) distributes organic C and N between three pools of soil organic matter (active, slow, and passive) and two litter compartments (metabolic and structural). EPIC calculates potential transformations of the five compartments as regulated by soil moisture, temperature, oxygen, large-scale data on land cover (cropland), soils, topography, field size, and crop management practices aggregated at a 1×1 km grid covering European countries (Balkovič et al., 2018(Balkovič et al., , 2013. In VERIFY, a total of ten major European crops including winter wheat, winter rye, spring barley, grain maize, winter rapeseed, sunflower, sugar beet, potatoes, soybean, and rice were used to represent agricultural production systems in Europe. Crop fertilization and irrigation were estimated for NUTS2 statistical regions between 1995 and 2010 (Balkovič et al., 2013). For VERIFY, the simulations were carried out assuming conventional tillage, consisting of two cultivation operations and mouldboard ploughing prior to sowing and an offset disking after harvesting of cereals. Two row cultivations during the growing season were simulated for maize and one ridging operation for potatoes. It was assumed that 20% of crop residues are removed in case of cereals (excluding maize), while no residues are harvested for other crops.
Uncertainties in EPIC arise from three primary sources which were in detail described by ORCHIDEE. A detailed 1115 sensitivity and uncertainty analysis of EPIC-IIASA regional carbon modelling is presented in Balkovič et al. (2020).

ECOSSE (grasslands)
ECOSSE is a biogeochemical model that is based on the carbon model ROTH-C (Jenkinson and Rayner, 1977;Jenkinson et al. 1987;Coleman and Jenkinson, 1996) and the nitrogen-model SUNDIAL (Bradbury et al. 1993; 1120 Smith et al. 1996). All processes of the carbon and nitrogen dynamics are considered (Smith et al., 2010a,b).
Additionally, in ECOSSE processes of minor relevance for mineral arable soils are implemented as well (e.g. methane emissions) to have a better representation of processes that are relevant for other soils (e.g. organic soils). ECOSSE can run in different modes and for different time steps. The two main modes are site specific and limited data. In the later version, basis assumptions/estimates for parameters can be provided by the model. This increases the uncertainty but makes ECOSSE a universal tool that can be applied for largescale simulations even if the data availability is limited. To increase the accuracy in the site-specific version of the model, detailed information about soil properties, plant input, nutrient application and management can be added as available.
During the decomposition process, material is exchanged between the SOM pools according to first order rate equations, characterised by a specific rate constant for each pool, and modified according to rate modifiers dependent on the temperature, moisture, crop cover and pH of the soil. The model includes five pools with one of them are inert. The N content of the soil follows the decomposition of the SOM, with a stable C:N ratio defined for each pool at a given pH, and N being either mineralised or immobilised to maintain that ratio. Nitrogen released from decomposing SOM as ammonium (NH4+) or added to the soil may be nitrified to nitrate (NO3-).
For spatial simulations the model is implemented in a spatial model platform. This allows to aggregate the 1135 input parameter for the needed resolution. ECOSSE is a one-dimensional model and the model platform provides the input data in a spatial distribution and aggregates the model outputs for further analysis. While climate data are interpolated, soil data are represented by the dominant soil type or by the proportional representation of the different soil types in the spatial simulation unit (this is in VERIFY a grid cell).
Uncertainties in ECOSSE arise from three primary sources: parameters, forcing data (including spatial and temporal 1140 resolution), and model structure. These uncertainties are not yet quantified.

1145
The BLUE model provides a data-driven estimate of the net land use change fluxes. BLUE stands for 'bookkeeping of land use emissions'. Bookkeeping models (Hansis 2015, Houghton 1983

TRENDY v7
The TRENDY (Trends in net land-atmosphere carbon exchange over the period 1980-2010) project represents a consortium of dynamic global vegetation models (DGVMs) following identical simulation protocols to investigate spatial trends in carbon fluxes across the globe over the past century. As DGVMs, the models require climate, carbon dioxide, and land use change input data to produce results. In TRENDY, all three of these are harmonized to make the results across the whole suite of models more comparable. In the case of VERIFY, we used

1205
While describing the details of all the models used here is clearly not possible, DGVMs calculate prognostic variables (i.e., a multitude of C, H2O and energy fluxes) from the following environmental drivers: air temperature, wind speed, solar radiation, air humidity, precipitation and atmospheric CO2 concentration. As the run progresses, vegetation grows on each pixel, divided into generic types which depend on the model (e.g., broadleaf temperate forests, C3 crops), which cycle carbon between the soil, land surface, and atmosphere, through such processes such Uncertainties in TRENDY v7 are model specific and described by Le Quéré et al., 2018. The spread of the 14 TRENDY models used by this study (Fig. 9) gives an idea of the uncertainty due to model structure in dynamic global vegetation models, as the forcing data was harmonized for all models.

CarboScope-Regional, GCP 2019 (CTE, CAMS, CarboScope) and EUROCOM
Top-down estimates of land biosphere fluxes are provided by a number of different inverse modeling systems that use atmospheric concentration data as input, as well as prior information on fossil emissions, ocean fluxes, and 1225 land biosphere fluxes. The land biosphere fluxes, and in some systems the ocean fluxes, are estimated using a statistical optimization involving atmospheric transport models. The inversion systems differ in the transport models used, optimization methods, spatiotemporal resolution, boundary conditions, and prior error structure (spatial and temporal correlation scales), thus using ensembles of such systems is expected to result in more robust top-down estimates. Top-down estimates at regional scales (up to 0.25°x0.25° resolution) for the period 2006 -2015 are taken from the six models used within EUROCOM (Monteil et al., 2019). These inversions make use of more than 30 atmospheric observing stations within Europe, including flask data and continuous observations. The CarboScope-

1235
Regional (CSR) inversion system (also included within the EUROCOM ensemble) was also run for the extended   Sensitivity and uncertainty analysis of EPIC-IIASA regional soil carbon modelling (Balkovič et al. 2020

Data availability
All raw data files reported in this work which were used for calculations and figures are available for public download at https:// 10.5281/zenodo.4626578 (Petrescu et al., 2020a). The data we submitted are reachable with one click (without the need for entering login and password), with a second click to download the data, consistent with the two click access principle for data published in ESSD (Carlson and Oda, 2018). The data and the DOI number are subject to future updates and only refers to this version of the paper.