BKs are explicitly designed to estimate LULUCF fluxes following land management and land-use changes by tracking the carbon content in soil, vegetation and product pools, i.e. stocks and fluxes of carbon in the land biosphere and between land and atmosphere. They combine spatial information on land-use activities with observation-based carbon stock densities and specific response curves of soil and vegetation carbon for each land-use conversion type (for more details refer to Sect.  and ). Using separate response curves for carbon release (emissions) and carbon uptake (removals), BKs model both gross fluxes explicitly, and net fLULUCF is derived as the sum of the two gross fluxes. BKs do not model the fluxes from peatland fire and drainage, but these emissions are added on top of the BK estimates in this study according to the approaches used in the GCBs; compare Sect. . We use simulations by three BKs that were conducted for the GCB2022 , namely (1) the “bookkeeping of land use emissions” model (hereafter BLUE22; ), (2) the “Houghton and Nassikas” model (hereafter H&N22; ) and (3) the compact Earth system model “OSCAR” (hereafter OSCAR22; ).

DGVMs are process-based models used to simulate the interaction between land surface and vegetation processes with the atmosphere. They approximate net fLULUCF as the difference in net biome productivity (NBP) of a simulation including LULUCF (similarly to the BKs, using spatial information on land-use activities) and a simulation excluding LULUCF (the latter using a time-invariant pre-industrial land-use map; for more details refer to Sect.  and ). DGVM simulations using transient (observed) environmental forcing are operationally available and can be used to estimate the impact of climate variability or long-term environmental changes on fLULUCF. In addition, DGVM simulations can be forced with constant environmental data prescribing, for instance, pre-industrial or present-day environmental conditions. Simulations using the latter setup most closely resemble conditions that occurred during the time when observed carbon densities (as used by BKs or inventories) were measured and are therefore the best DGVM setup to compare with BKs or inventories. Here we use nine DGVMs that provided simulations with present-day as well as transient environmental forcing within the project “Trends and drivers of the regional-scale emissions and removals of carbon dioxide” (TRENDY; ).

NGHGIs are the official country reports to UNFCCC that include estimates of greenhouse gas emissions and removals from LULUCF. They widely rely on empirical emission factors in combination with country-level data on land-use activities to estimate fLULUCF (for more details refer to Sect.  and ). The methods and report details strongly vary between and among non-Annex I and Annex I countries. Non-Annex I countries frequently use default IPCC emission factors and often only report fluxes from deforestation, while estimates from Annex I countries are often based on country-specific statistical or process-based models for all IPCC land-use categories (forest land, cropland, grassland, wetlands, settlements and other land). For the analysis presented here, we use the country database (DB) compiled by , as updated in (hereafter referred to as NGHGI DB), which includes GHG data from all available country reports submitted to UNFCCC, gap-filled when necessary to allow a complete time series from 2000 until 2020. According to UNFCCC guidelines, country reports should encompass all LULUCF fluxes from any areas considered managed and for which IPCC provide methodological guidance. In most cases, NGHGIs include natural and indirect anthropogenic fluxes (e.g. from CO2 fertilization; ). To improve comparability of the NGHGI DB data with modelled estimates, we adjust the NGHGI DB data to better match the processes and definitions of the modelled estimates (the basis in this study) by correcting for this so-called managed land issue (hereafter adjusted NGHGI DB). Following the approach described in , we therefore subtract from the NGHGI DB estimates those fluxes resulting from natural and indirect effects (i.e. human-induced environmental change) from managed land. These effects are estimated by the ensemble mean of 16 transient DGVM simulations without land-use changes from TRENDYv11 (corresponding to the “natural terrestrial sink” in recent GCBs of the GCP), except for Brazil and Canada filtered with an intact/non-intact forest mask . For Brazil and Canada, this approach uses the national gridded maps on managed and unmanaged forests used in the respective NGHGIs .

FAOSTAT fLULUCF estimates resemble the bottom-up approach of the UNFCCC data in the way that they are based on consistent underlying activity data – at grid cell or at country level – in combination with emission factors. FAOSTAT fLULUCF data are estimated by applying IPCC Guidelines to activity data generated either through official country reporting processes or through geospatial data analysed by FAO (for more details refer to Sect.  and ). FAOSTAT fLULUCF cover carbon emissions and carbon removals in forests and from deforestation, derived from national carbon stock change statistics and IPCC emission factors, as well as emissions from peatland fires and peat drainage, the latter two obtained from satellite imagery in combination with IPCC emission factors . The recognizes explicitly that both FAO activity data and FAO emissions estimates provide a valuable set of reference data that can be used for validation, quality control and quality assurance of the data submitted through NGHGIs.

Although the different approaches summarized above (and described in detail in Appendix A) all aim at quantifying CO2 fluxes from LULUCF, they differ substantially. Some of the key differences between the approaches are briefly described below; further details are provided in the results in Sect. 4.2 and 4.3 and in the Appendix.

Uncertainties in fLULUCF estimations from modelling approaches mainly arise at high spatial resolutions , from differences in underlying land-use and land-cover information , missing observational constraints , differences in process complexity and in the degree of implementation of LULUCF practices , inconsistencies in common terminology and definitions , and different model assumptions and setups .

Most of the investigated modelling approaches use the spatially explicit LUH2-GCB2022 dataset as LULUCF forcing . The BK model H&N22 uses FAO activity data at country level, and OSCAR22 uses both LUH2-GCB2022 and FAO activity data. To analyse the impact of different land-use forcing data, we further use BLUE data from the GCB2019 (hereafter BLUE19). As the BLUE model code was not changed between the GCB2019 and GCB2022, this allows the direct impact of the LULUCF forcing data to be isolated, which for these GCBs was based on HYDE3.2 and HYDE3.3, respectively. The main innovations in HYDE3.3 were the provision of yearly output from 1950 onwards, the update of the onset of agriculture based on new radiocarbon data and archaeological expertise indicating more spatial heterogeneity, the use of the latest satellite data with increased spatial resolution on an annual basis from 1992–2018 from the European Space Agency (ESA) and MapBiomas data for Brazil for the period 1985–2020, and the inclusion of more sub-national cropland and pasture data.

BKs do not explicitly represent biogeochemical processes and do not directly use environmental forcing data and thus usually exclude effects from climate change and meteorological or climate variability. DGVMs, in contrast, incorporate biogeochemical processes implemented in their modelling scheme, and – additionally to land-use change data – they use environmental forcing data as input. Consequently, DGVM estimates under transient environmental conditions include the long-term response to changing environmental conditions (e.g. atmospheric CO2 increase, nitrogen deposition and fertilizer applications) and effects of climate variability. Due to the setup to calculate LULUCF emissions from DGVMs, transient DGVM fLULUCF estimates include the loss of additional sink capacity (LASC), representing carbon fluxes in response to environmental changes on managed land (typically croplands with low carbon sink capacity and fast turnover rates) as compared to potential natural vegetation (typically forests with large carbon sinks and low turnover rates), which leads to higher flux estimates compared to BKs towards the end of the simulated periods .

In contrast to modelling approaches that provide globally consistent historical fLULUCF analysis over the entire simulated period (e.g. from 1700 onward), country reports only cover the most recent decade for which observations and statistics exist (starting around 1990–2000, depending on the dataset and country) and are restricted to the territories and components that countries report. In addition, fLULUCF estimates from country reports rely on different definitions and assumptions than global models and therefore do not allow, for example, a realistic derivation of countries' climate change mitigation contributions that are consistent with the pathways to achieve the climate targets of the Paris Agreement as derived by integrated assessment models. One major difference is the definition of managed land, with NGHGIs having comparatively larger areas of managed land than models . Further, NGHGI estimates that rely on direct observations (e.g. national forest inventories) include both direct and most indirect anthropogenic effects on managed lands, which can only be separated based on modelling approaches . Consequently, most NGHGI fLULUCF estimates include large parts of the natural response to recent environmental change in managed lands (e.g. larger vegetation carbon sink due to so-called CO2 fertilization) and, thus, estimate larger anthropogenic CO2 removals (and widely lower net fLULUCF) than global models . In order to make the NGHGI and BK land-use flux estimates comparable, here we translate NGHGI estimates by removing the fluxes that models attribute to the natural land sink (compare the adjusted NGHGI DB data described in Sect. ).

Reported fLULUCF estimates remain highly uncertain, particularly in many developing countries, as country reports to UNFCCC do not explicitly separate managed from unmanaged forest land or only report forest-related fluxes . Other countries report fluxes from additional LULUCF practices, such as from natural peatland that is converted to agriculture or from land that is converted to human settlements (not included in modelled estimates). While human-induced degradation from logging and fires is often included in national reports, forest degradation fluxes are hardly existent in BKs.

The main differences between FAOSTAT data and the NGHGI database are explained by a more complete coverage in the NGHGI database than FAOSTAT, in particular the inclusion of non-biomass carbon pools and non-forest land uses . Additional differences stem from (1) differences in activity data, for instance, different time series of the forest area which may be communicated independently to UNFCCC and FAO by different national agencies; (2) differences in emission factors, including carbon stock data and their sub-national resolution as well as their changes over time; and (3) differences in scale, especially considering that forest fluxes may be computed at grid cell or sub-national scale in many countries using remote sensing or information from national forest inventories. Additionally, the FAO area data do not distinguish between managed and unmanaged areas and, thus, do not in principle separate anthropogenic and non-anthropogenic drivers . The nature of data reported to FAO indicates that in most cases the area considered in FAOSTAT estimates is similar to the one in the NGHGIs, but often the effects of environmental changes (i.e. natural and indirect anthropogenic effects) are not included as in the NGHGIs. For this reason, we did not correct FAOSTAT estimate data here.

Country-level net fLULUCF estimates from bookkeeping models strongly vary across countries. Net LULUCF emissions are highest (in descending order) in Brazil, Canada, China, DR Congo, India, Indonesia, Nigeria and Russia based on cumulative estimates between 1950 and 2021 (Fig. a). The eight countries with the highest cumulative net emissions from LULUCF are either located in carbon-rich, forested tropical regions (Brazil, DR Congo, Indonesia and Nigeria) and/or encompass vast territories (Brazil, Canada, China, India and Russia). These top eight emitters emitted more than ∼ 53 % of the total net LULUCF emissions in the period 1950–2021 and are thus of outstanding importance for climate change mitigation via LULUCF emission reductions. In particular in Indonesia, the estimated emissions per area are higher than in all other 185 countries studied, which illustrates an enormous pressure on the terrestrial carbon stocks in the tropics (Fig. ). However, it should be noted that the resulting call for action is not limited to these top emitters, as large parts of national LULUCF emissions, particularly in the tropics, are caused by consumption elsewhere .

During the second half of the 20th century, high net LULUCF hotspots became increasingly concentrated in the countries of the Global South. As stated in the GCB2022 , more than 50 % of most recent net emissions from LULUCF occur in only three countries, namely Brazil, DR Congo and Indonesia – all located in the tropics (Fig. b). The pronounced land-use change impact in the tropics is also evident from the fLULUCF estimates calculated per area and per capita, with the 10 largest net emitters all located in the tropics (Figs.  and ). However, most of these emissions are embodied in trade and are caused by consumption in industrialized regions such as Europe, the United States and China . A trend towards fewer countries comprising larger shares of global net LULUCF emissions is found when comparing cumulative and most recent annual fLULUCF estimates. In the period 1950–2021, the top 22 net emitters comprised more than 75 % (top 43 emitters more than 90 %) of cumulative net emissions from LULUCF, while in 2011–2021, only 15 countries make up 75 % (top 33 emitters more than 90 %) of the global net emissions. In China, the fLULUCF estimates from the BKs show an remarkable turnaround, turning China from the third-highest cumulative emitter in 1950–2021 to the country with the third-highest net removals in 2011–2021.

The large number of net emitting countries and the fact that BKs estimate a global net source of carbon from LULUCF to the atmosphere are in stark contrast to the pledges to achieve the goals associated with the Paris Agreement and the reported global carbon removal from LULUCF when summed across all NGHGIs . Of note, despite widely included net negative emissions from LULUCF in the NDCs, including the need for carbon dioxide removal (CDR) techniques, BKs estimate net removals from LULUCF only for very few countries (both cumulatively between 1950–2021 as well as more recently in 2011–2021). Substantial country-level net removals from LULUCF are only found in the United States; some European countries; and, in the most recent decade, in China. However, large relative uncertainties in BK estimates remain, within the cumulative net fLULUCF estimates (globally ∼ 38 %, and much higher in specific countries, for example, in China ∼ 150 %, the United States ∼ 680 %; see Table ) and recent annual fLULUCF estimates (globally ∼ 31 %, in China ∼ 700 % and Russia ∼ 1000 %). In order to explain the uncertainties related to fLULUCF estimates from BKs, the latter need to be compared to estimates from other approaches, and the underlying drivers for the differences in fLULUCF estimates need to be investigated.

To test the robustness and explain the large spread in modelled country-level fLULUCF estimates, we compare and discuss the differences of the net estimates from the BK ensemble and the DGVM ensemble with respect to the characteristics of the individual modelling approaches, particularly regarding the underlying land-use forcing data and, for the DGVM ensemble, the environmental forcing data.

Country-level net fLULUCF estimates of BKs and DGVMs from 1950 onward agree generally well in the investigated countries (Fig. ; refer to Fig.  for RECCAP2 regions and Figs.  and  for all 186 investigated countries).

In most countries, modelled estimates from BLUE19 and (present-day) DGVM simulations show consistent temporal evolutions and peaks in emissions, when using the same land-use forcing data (compare the “difference of the modelling approach”, derived as the present-day TRENDYv8 mean minus BLUE19 which share the same (LUHv2) land-use forcing data; purple line in Fig. ). These consistent trends and peaks lead to comparably high correlation coefficients between the BLUE19 and present-day DGVM estimates, for example, in Brazil, China and Nigeria. However, the correlation coefficients for the estimates from the different modelling approaches are generally rather low, which is due to the interannual variability, which is captured by the DGVMs but usually not by the BKs. In line with this, the particularly low correlation coefficients between DGVMs and BKs in the United States and Russia result from high interannual variability in combination with a low signal from land-use changes. Yet, the fLULUCF estimates also substantially differ in some of the countries with pronounced land-use changes, despite identical land-use forcing data. The most striking difference occurs in 1997 in Indonesia. The year 1997 was an El Niño year, causing high carbon emissions from organic soils in Indonesia, which are included in the BK estimates but not in the DGVM estimates used (in line with the GCP assessments).

BLUE19 generally yields higher net emission estimates compared to the present-day TRENDYv8 ensemble (except for Nigeria), which indicates a tendency towards higher estimates when using the BK approach. To investigate this further, we compare the estimates of all three BKs from 2022 (BKs 2022; note that BLUE model code was not changed between 2019 and 2022 versions). BLUE22 tends to estimate the highest emissions among the BKs in most countries and during most of the time, indicating that the BK mean might agree better with the mean of the present-day DGVM simulations compared to BLUE19. This can mainly be explained by higher differences in the carbon densities between natural and managed areas assumed in the BLUE model in conjunction with the fact the BLUE captures the full extent of (LUH2-based) gross transitions (compare Sect.  and description of the BKs in Sect.  and ).

The strong influence of the land-use forcing data is highlighted by the often differing trends and peaks in fLULUCF from BLUE19 based on HYDE3.2 and BLUE22 based on HYDE3.3 (compare the “LUHv2 forcing difference”, derived as BLUE22 minus BLUE19, in Fig. ) and the huge ranges in the BK estimates in many regions. Large LUHv2 forcing differences, as indicated by particularly low correlation coefficients, are found in Brazil, DR Congo and Nigeria, with substantially larger estimates using the more recent HYDE3.3 data. The particularly big differences for Brazil mainly result from an improved representation of deforestation patterns through the inclusion of MapBiomas land cover data in the newer HYDE version, and, for DR Congo, they result from the inclusion of revised data from the FAO . In contrast, BLUE22 has lower fLULUCF estimates than BLUE19 in Canada (mainly in the 1950s) and, in the most recent decade, in China, India, Indonesia and Nigeria. In particular for tropical countries, lower emissions may be related to decreased cropland expansion in the updated HYDE data .

Emission peaks around 1960 occur in several regions (e.g. Canada, DR Congo and Russia), mainly in the 2022 estimates (BLUE22) but not in the 2019 estimates (BLUE19). Those peaks can be explained by an artefact in HYDE3.3, resulting from the merge of historical HYDE data (used up to 1960) with FAOSTAT data (used from 1961 onward) . The 1960 peaks thus do not represent any real-world land-use change fluxes and should be corrected in future updates of HYDE (as has already been achieved for Brazil; compare Sect. A1 and ). For Brazil, an additional peak in 2004 corresponds to increased deforestation for cropland and pasture, followed by a slowdown in deforestation rates due to governmental regulations, which is only captured in HYDE3.3 through the inclusion of ESA CCI Land Cover data .

Additionally, the environmental forcing plays an important role in many regions, as can be assessed by comparing transient and present-day TRENDYv8 simulations (Fig.  – the “Present-day versus transient difference”; refer to Fig.  for RECCAP2 regions and Figs.  and  for all 186 investigated countries). Estimates of fLULUCF under present-day environmental forcing tend to be higher compared to transient estimates in the earliest simulated periods, particularly in tropical regions with high LULUCF activity (e.g. in Brazil, China, DR Congo and India), which is also reflected in the cumulative emissions estimates (see numbers displayed in Fig. ). This can be explained by higher carbon stocks under present-day environmental forcing compared to transient forcing (the multi-DGVM mean global vegetation carbon stock increased by ∼ 23 % from 664 to 815 PgC from 1800 until today), particularly in early simulation periods (when atmospheric CO2 concentration was substantially lower than today; ). Differences between present-day and transient fLULUCF estimates become smaller as the simulations progress, as transient environmental conditions approach present-day conditions and additionally accumulate the loss of additional sink capacity . It is noteworthy that towards the end of the simulated period, transient fLULUCF estimates even tend to exceed present-day estimates, due to the steadily accumulating loss of additional sink capacity, which globally comprises ∼ 0.8 ± 0.3 GtC yr-1 (∼ 40 %) of transient fLULUCF in the period from 2009–2018 . In the temperate zone (e.g. in Russia and the United States), variations in fLULUCF estimates due to environmental conditions rather depend on the inter-annual meteorological and climate variability. Here, carbon stocks have not been enhanced by higher CO2 concentrations as homogeneously as in the tropics, and climate change has even led to decreased carbon stocks in some regions . What is striking here is the particularly low correlation coefficient in the United States. Similar to the difference between the modelling approaches, this results primarily from the combination of a low land-use signal with very high interannual variability in environmental conditions.

As described above (and in the Appendix), the multiple modelling approaches differ in their underlying assumptions, their implemented process complexity and parametrizations, and the input forcing data used. These differences partly lead to highly differing estimates, in particular at finer spatial scales, which decreases the accuracy of some country-level estimates from models. The differences in country-level net fLULUCF estimates that are due to the modelling approach versus changes in land-use forcing (approximated by the LUHv2 forcing difference for the BK model BLUE) or environmental forcing (difference between present-day and transient DGVM simulations) are of a similar order of magnitude, yet with very different spatial patterns. Whether the modelling approach, land use or environmental forcing is the dominating factor depends on the specific country and partly also on the specific time period. The modelling approach had the highest influence on the cumulative estimates, for example, in Brazil, Canada, India, Indonesia and Russia, whereas the LUHv2 forcing difference in the BK model BLUE was higher in China, DR Congo and Nigeria. For the DGVMs, the land-use forcing data impacted cumulative fLULUCF estimates widely more strongly compared to the environmental forcing (except in the United States), although of similar magnitude. It should be emphasized that the strong influence of environmental factors, which is reflected in the up- and downswings of the DGVM estimates, is likely to become more important under future climate conditions with more frequent and intense extreme environmental conditions and potentially decreasing LULUCF activities as set out by the Glasgow Leaders' Declaration on Forests and Land Use.

Despite broad agreement in most country-level net fLULUCF estimates when averaged across multiple models, individual BK model output differs strongly in some countries (compare BK range in Figs.  and ), and, not surprisingly, country-report-based estimates mostly diverge even more (Fig. ; refer to Fig.  for RECCAP2 regions and Figs.  and  for all 186 investigated countries).

Differences in annual net fLULUCF estimates across the three BKs (based on simulations with the most recent land-use forcing) are particularly high in Canada and the United States throughout the 20th century and before and after emission peaks (e.g. 1960 and 2011 in DR Congo, 1980s in China, and in the 1950s in India). As stated above, the high differences related to emission peaks are predominantly due to the use of different land-use forcing data, which is reflected by the fact that H&N22 (based on FAO data) does not capture any of these peaks, while BLUE22 and OSCAR22 models (both of which use LUH2 data) show very similar trends and peaks. In China, for example, the H&N22 land-use forcing data assume a steady increase in forest areas from 1950, while the LUH2 data show decreasing forest areas until 1990 and relatively stable forest areas thereafter . The estimates of OSCAR22 often lie close to the multi-model mean of the BKs, which can be explained by the fact that OSCAR22 uses both FAO and LUH2 forcing data to derive a best-guess estimate for fLULUCF . The huge uncertainties in Canada, China and the United States cannot fully be explained by the net fLULUCF but are discussed in Sect.  in relation to the gross fluxes.

Beyond the effects of the land-use forcing data, BK model differences in net fLULUCF can be explained by several individual model specifics. The upper limit of the BK range is predominately defined by BLUE22, particularly during phases of high BK uncertainty, while the lower limit is often defined by H&N22. This is mainly due to different assumed carbon densities, which, depending on the ecosystem, are particularly high in the BLUE model and low in the H&N model (compare Appendix and ). Moreover, the inclusion of sub-grid-scale transitions in BLUE22 (and OSCAR22) increases emission estimates compared to H&N22. Additionally, H&N22 assumes conversion of natural grasslands to pasture, while BLUE22 and OSCAR22 allocate pasture proportionally on all natural vegetation that exists in a grid cell , yielding lower fLULUCF for H&N22. In addition, different turnover periods for the HWPs in the different BKs lead to varying BK estimates after significant changes in LULUCF practices. The exception of slightly higher emission estimates from H&N22 in Indonesia (mainly in the 1990s) and DR Congo (from 1970s until 2007) is due to much higher carbon removals due to afforestation and reforestation in BLUE22 and OSCAR22 (not shown) and faster increasing emissions from deforestation in the H&N22 model.

Country-report-based net fLULUCF estimates are substantially lower than BK estimates in most of the investigated countries; in particular NGHGI DB has the lowest emission estimates in almost all investigated countries. Much of this discrepancy (globally adding up to ∼ 1.6 GtC yr-1 for the period 2001–2020) can be explained by different definitions, particularly regarding two points (compare Appendix and ). (1) Natural and indirect human-induced fluxes are included, such as those resulting from increased forest regrowth due to higher atmospheric CO2 concentration and N deposition, in many country reports. (2) The area assumed to be managed land is larger and is therefore included in the fLULUCF assessment in the country reports compared to the BKs. This is confirmed by the fact that the adjusted NGHGI DB data, where the natural land sink on managed land is subtracted, agree better with the BK estimates than the NGHGI DB data for most countries.

However, for individual countries, other reasons, such as omitted fluxes due to incomplete reporting of land uses and carbon pools, also (partly) explain the differences between modelled and reported fLULUCF estimates . In the United States, lower country-report-based estimates may additionally result from the inclusion of CO2 removals from urban vegetation , which is not considered in the BK estimates. In addition, despite the increased methodological complexity of the approaches being used by many developed countries, it was shown that reported fLULUCF estimates for the United States still remain uncertain, mainly due to uncertainty in inputs, model parameters and plot-based sampling (e.g. ). The Canadian NGHGI report discounts fluxes (emissions and removals) on areas affected by wildfires and severe insect disturbances and reports them in a separate category , which can explain the strongly increased difference for the adjusted NGHGI DB estimate . In China, the large gap between country-report-based and modelled estimates might be explained by high carbon removals from afforestation and ecological restoration projects considered in the country reports but not fully included in the land-use data used by the models (compare Sect.  and ).

The generally good agreement in the fLULUCF estimates from BKs and inventories in Indonesia is due to the dominance of the added peat data, with a pronounced interannual variability controlled mostly by El Niño–Southern Oscillation patterns on top of land use .

LULUCF flux estimates from FAOSTAT are largely in agreement with BK estimates, with some important differences, notably lower estimates in China and Russia and higher estimates in DR Congo and Canada. FAOSTAT estimates are generally higher than the NGHGI DB estimates except for the period from 2000–2019 in Russia and Nigeria and the emission peak in Brazil in 2004. As for the NGHGI DB estimates, lower FAOSTAT estimates compared to BK estimates are likely due to the inclusion of all fluxes on managed land (compare adjusted NGHGI DB data). Differences between FAOSTAT and NGHGI DB estimates can be explained by the generally more complete coverage of carbon fluxes in the latter and differing approaches to estimate forest fluxes, where FAO applies a carbon stock change approach based on observed forest data from FRA, while NGHGI reports are based on the use of a simple carbon stock change approach or a gain loss approach by the scaling up of forest growth rates based on IPCC default factors to forest land estimates (refer to Appendix and , for a detailed description of the differences between UNFCCC country, NGHGI DB and FAOSTAT estimates).

Additionally, the underlying data on forest land differ, with the NGHGI DB database reporting much greater forest areas and forest carbon removals (in particular for non-Annex I countries). Moreover, NGHGIs of Annex I and the largest non-Annex I countries also include non-biomass carbon pools and non-forest land uses, while, except for organic soils, FAOSTAT only includes above- and belowground biomass pools . The higher estimates of NGHGI DB compared to FAOSTAT in Brazil are caused by larger deforestation and afforestation areas in the Brazilian report to UNFCCC compared to FRA and the fact that it considers gross deforestation and afforestation, while FAOSTAT reports net deforestation and afforestation directly (which might increase emissions in the NGHGI due to the asymmetry in instantaneously occurring gross emissions versus slowly increasing gross removals over the long term).

To get more insights into the underlying drivers of country-level net LULUCF estimates, we split them into gross fluxes, namely gross emissions (or “sources”) and gross removals (or “sinks”). As stated in the Introduction, here we define these gross fluxes as the sum of all fluxes related to those LULUCF practices that typically lead to emissions or removals, respectively. Gross emissions are mainly caused by deforestation, peatland degradation, biomass burning, the decay of HWPs and biomass left on site after harvest . Gross removals are mainly associated with afforestation and reforestation including forest regrowth after agricultural abandonment, as well as forestry cycles and the restoration of other (non-forest) ecosystems. In the following, we present and discuss gross fLULUCF derived from the three BKs for the nine selected countries (Fig. ; refer to Fig.  for RECCAP2 regions and Figs.  and  for all 186 investigated countries).

Similarly to the net fluxes, gross fluxes modelled by the three BKs show widely similar trends and agree on the timing of emission peaks. Peaks in net emissions are predominantly due to peaks in gross emissions, while the time series of gross removals is much smoother, consistent with the slower pace of vegetation regrowth. Consequently, the short-term evolution of net fLULUCF is much more influenced by the dynamics of highly fluctuating gross emissions than by the dynamics of rather slowly changing gross removals. However, the decreasing trends in net fLULUCF estimates across many regions (globally from 1.6 ± 0.7 GtC yr-1 in the 1960s to 1.1 ± 0.7 GtC yr-1 in the period from 2011–2020) mainly relate to a steady increase in the gross removals (globally from -1.9 ± 0.4 to -2.7 ± 0.4 GtC yr-1 in the same period), which exceeded the increase in gross emissions (globally from 3.4 ± 0.9 to 3.8 ± 0.6 GtC yr-1 in the same period; ). This smooth evolution towards increased removals results from increased carbon sequestration on previously managed land, mainly due to forest regrowth and soil recovery. Environmental changes, such as those resulting from CO2 fertilization, are not modelled by BKs (except the changing carbon densities in OSCAR).

In most regions, uncertainties in net fLULUCF are due to uncertainties in gross emissions rather than uncertainties in gross removals. The largest differences, and thus pronounced uncertainties, in gross emissions from BKs are found for India and Canada as well as in the years before and after most emission peaks in many regions. Uncertainties related to the peaks in the 1960s mainly stem from the merging of two different datasets. In China, the pronounced peak in the 1980s is caused by spurious signals in the LUHv2 data, inherited from an abrupt cropland increment in the FAO data . Because cropland area is quantified relative to forest proportions, an increasing cropland area causes decreasing forest area (and vice versa), while China's afforestation projects were largely implemented in drier and previously unmanaged and unforested lands, increasing the total forest area without replacing croplands . Similar to net fLULUCF, the highest emission estimates are generally derived from BLUE22 and the lowest emissions predominantly from H&N22. As stated above, this can mainly be explained by different process representation and parametrization in the models (compare Sect. ). The exception of higher OSCAR22 estimates in Brazil in recent decades likely results from higher deforestation rates since 2004 and shorter turnover times for HWPs in OSCAR compared to the other BKs. In addition, OSCAR uses the averaged biome-specific carbon densities taken uniformly over the country which may overestimate emissions in particular in large countries covering differing types of the same biome (e.g. different types of forest), if land-use transitions predominantly happen in regions with lower carbon densities.

The highest differences in gross removal estimates among the BKs are found in India, Russia and the United States. In India, this may result from greater removals due to the inclusion of sub-grid-scale transitions in BLUE22 and OSCAR22, while H&N22 estimates rather negligible removals. It is noteworthy that in India, the large uncertainties in gross emissions and removals from BKs translate into a huge uncertainty in net fLULUCF in the 1950s, but subsequently uncertainties in gross fluxes cancel out, yielding only small uncertainties for the net flux. In Russia, the models agree in a decreasing removal trend despite a considerable spread, with large removal estimates by H&N22 and small estimates by OSCAR22. In recent decades, these decreasing removals in Russia can partly be explained by the decreasing trend in the abandonment sink as was shown for the BLUE model by in addition to intensified logging and wood harvest activities that cause ongoing deforestation . In the United States, the large BK range for net fluxes is predominantly due to large uncertainties in the removal estimates, while the gross emission estimates agree well among BKs. This removal-driven net fLULUCF uncertainty can be explained by the inclusion of fire management in the United States in H&N22, leading to large removal estimates, while BLUE22 and OSCAR22 show much lower removal estimates.

To further investigate the importance of gross fluxes, we calculate the ratio of net fLULUCF to the sum of gross fLULUCF (net-to-gross ratio, with the sum of gross fluxes defined as the range between gross emissions and gross removals) for the three BKs (Fig. ; refer to Fig.  for RECCAP2 regions and Figs.  and  for all 186 investigated countries). Ratios close to 1 (close to -1) indicate that the net flux reflects mostly gross emissions (gross removals) and only very small gross removals (gross emissions). Ratios between 0 and 0.5 (between 0 and -0.5) indicate that the net flux represents only a small fraction of the occurring gross sources (sinks), which is the case in most countries during most of the time investigated. Ratios close to zero indicate that gross fluxes are largely compensating each other, which might indicate a sustainable land management that causes gross removals to largely offset gross emissions.

Seven out of the nine investigated countries (Brazil, China, India, Canada, Russia, Nigeria and the United States) show decreasing country-level ratios of net-to-gross fLULUCF from 1950 onward, indicating increasingly compensating gross emissions and gross removals. This is mostly due to increasing gross removals from LULUCF, while at the same time the gross emissions did not increase in such a pronounced way (particularly in Brazil, Canada, China, India and Nigeria). In some of the countries, the ratios became even negative over time (China, India, Russia, and the United States), indicating that gross removals became larger than gross emissions. Large negative net-to-gross ratios indicate that gross removals are much larger than gross emissions, and thus the (negative) net flux is mostly controlled by carbon removals from the atmosphere (e.g. in the most recent decade in European countries, Japan and Türkiye; compare Fig. a and Tables A1–A3).

In contrast, increasing net-to-gross ratios over time are found in Indonesia and DR Congo (particularly, in the most recent decade), mainly due to strongly increasing gross emissions, which are not compensated by equally large gross removals, despite increasing removals also observable in these countries (see Fig. ). High positive net-to-gross ratios in the most recent decade reveal large gross emissions that are not compensated by gross removals and are mainly found in the tropics and the Southern Hemisphere, in particular in Argentina, Angola, Paraguay, Bolivia, Papua New Guinea and Tanzania (Fig. a).

Large uncertainties in the net-to-gross ratio in Canada and China in the second half of 21st century are caused by strongly varying emission estimates and in the same period in the United States, from strongly varying removal estimates.

Near-zero net-to-gross ratios indicate that gross fluxes counterbalance each other; i.e. gross removals compensate gross emissions. Country-level near-zero net-to-gross ratios can be found in China, India, Russia and the United States, particularly in the most recent decade (compare Figs.  and a). Grid-cell-wise analysis revealed, however, that pronounced gross fluxes occurred also in these countries. This highlights that considering only the net land-use CO2 fluxes might miss the importance of potentially large gross fluxes. This is especially true when net fluxes are estimated on a larger scale, such as at the country level and particularly for very large countries, since here the opposing gross fluxes, which often occur spatially separately, are more likely to be offset (compare, for example, United States and China in Fig. c and d). This, furthermore, highlights the need for spatial explicit analysis of the net as well as gross fLULUCF and that country commitments based on net LULUCF fluxes can still be associated with large emission fluxes. Similarly, the rather vague commitment of the Glasgow Leaders' Declaration on Forests and Land Use to halt deforestation, without stating whether this accounts for gross or net transitions, can lead to strongly varying forest flux trajectories . In line with , we therefore argue that climate mitigation measures should focus on gross fluxes from LULUCF rather than net fluxes.