Land-use change emissions based on high-resolution activity data substantially lower than previously estimated

The net CO₂ flux from land-use and land-cover change (E_LUC) is a key component of the global carbon cycle (Friedlingstein et al 2020). E_LUC includes the carbon transfer from soil and biomass to the atmosphere through e.g. deforestation, harvest activities, and pasture to cropland conversions as well as the uptake and storage of carbon from the atmosphere in biomass and soil through e.g. afforestation and regrowth of vegetation after abandonment of agricultural land or harvest (Pongratz et al 2014). These land-use and land-cover change (LULCC) activities can be targeted as means to reduce emissions or to re-sequester carbon (often called carbon dioxide removal or negative emissions technologies in the latter case) and will be essential for meeting the 1.5 ^∘C climate target (Griscom et al 2017, Harper et al 2018, Goldstein et al 2020, Crippa et al 2021). Especially, halting deforestation and forest degradation on the one side (Maxwell et al 2019, Roe et al 2019, Gatti et al 2021) and supporting afforestation and regeneration of natural forests on the other side are widely discussed, available, and effective measures for climate mitigation (Hoegh-Guldberg et al 2019, Lewis et al 2019, Roe et al 2019). The implementation of these also greatly influences national abilities to reach net-zero emissions (van Soest et al 2021).

Compared to fossil CO₂ emissions, estimates of E_LUC are subject to high relative uncertainties (Arneth et al 2017). In the Global Carbon Budget 2020 (GCB2020), the uncertainty in E_LUC estimates was specified to be, with a likelihood of at least 68% (±1σ), in the range of ±0.7 GtC yr⁻¹ based on a best-value judgement (Friedlingstein et al 2020). In relative terms, this translates to an uncertainty of 43.8% (in comparison, fossil CO₂ emissions: 5.2%). The high uncertainty of E_LUC estimates has various reasons as summarized by Pongratz et al (2021): different terminologies and definitions (Pongratz et al 2014, Grassi et al 2018, 2021, Malins et al 2020, Obermeier et al 2021), different model assumptions and parameters (Bastos et al 2020, Gasser et al 2020, Hartung et al 2021), and different considerations of management processes (Stocker et al 2014, Arneth et al 2017, Hartung et al 2021). Furthermore, several studies have attributed major parts of this uncertainty to underlying LULCC datasets. From a set of sensitivity experiments based on the high, low, and baseline LULCC scenarios, Hartung et al (2021) estimate that about 22% of the sensitivity in cumulative E_LUC stems from LULCC inputs. Similarly, Gasser et al (2020) find substantial differences between E_LUC estimates based on different versions of LUH2, LUH1 and Global forest resources assessments (FRAs). Houghton and Nassikas (2017) use different versions of FRA to highlight differences in E_LUC estimates after 1950, while Peng et al (2017) compile multiple historical plant functional type (PFT) maps and conclude that different transition rules result in large differences in E_LUC estimates. Moreover, different regional studies (Yu et al 2019: USA, Kondo et al 2021: Southeast Asia, Rosan et al 2021: Brazil) discuss the influence of underlying LULCC forcing data on E_LUC estimates.

For this study, we implemented the new LULCC dataset HIstoric Land Dynamics Assessment + (Winkler et al 2021, hereafter HILDA+) in the bookkeeping of land use emissions model (Hansis et al 2015, hereafter BLUE). HILDA+ is a global high-resolution data product with a spatial resolution of $0.01^{\circ} \times 0.01^{\circ}$, covering common LULCC classes and a decent time period (1900/1960–2019), which makes it suitable as LULCC forcing for carbon cycle models. BLUE is one of three bookkeeping models in the yearly global carbon budgets (GCBs) (Friedlingstein et al 2020, 2021). Within the high uncertainties associated with E_LUC, BLUE is generally in line with other bookkeeping model and dynamic global vegetation model (DGVM) estimates, such that we use it here as a representative state-of-the-art model to quantify E_LUC and expect our qualitative conclusions to be robust against the choice of model. Detailed comparisons of BLUE to other models can be found in Bastos et al (2021), Friedlingstein et al (2021), and Obermeier et al (2021). The implementation of HILDA+ in BLUE opens up the novel possibility to compare and evaluate E_LUC based on two spatially explicit and independently derived LULCC datasets. Given the high uncertainty arising from LULCC inputs, the verification of E_LUC estimates based on HILDA+ with estimates based on other LULCC forcings is an important step to identify causes of the E_LUC uncertainty. We take this opportunity to investigate mechanisms beyond the specific LULCC data and E_LUC model used and investigate the relevance of initialization time and, for the first time, the sensitivity of results to spatial resolution, highlighting a previously under-appreciated role of successive transitions in global carbon cycle modeling.

By using BLUE, we make use of the computationally efficient design of the model that enables us to estimate E_LUC at the original resolution of HILDA+ at 0.01^∘. In the past, E_LUC has been estimated globally at 0.25^∘ resolution (Le Quéré et al 2018a, 2018b, Friedlingstein et al 2019, 2020, Bastos et al 2021, Hartung et al 2021), at 0.5^∘ resolution (Hansis et al 2015), at country level (Houghton and Nassikas 2017, Le Quéré et al 2018a, 2018b, Friedlingstein et al 2019, 2020, Bastos et al 2021), and at regional and biome level (Friedlingstein et al 2020, Gasser et al 2020). Thus, E_LUC estimates based on HILDA+ have an at least 25 times higher information content than any previous studies. The high resolution of HILDA+ allows us a spatially more precise detection of LULCC events and consequently a better location of E_LUC sinks and sources. Nevertheless, subgrid-scale omissions of transitions can still not be completely avoided, for which a field-scale resolution of roughly 1 ha would be needed (Wilkenskjeld et al 2014). An example of such subgrid-scale transitions are transitions from shifting cultivation (also called swidden agriculture/cultivation or slash-and-burn), which are small-scale land use systems with rotational cycles of shorter cultivation phases of annual crops and longer natural fallow phases of woody regrowth, separated by fire clearances (Mertz et al 2009). Using LULCC data of less than 100 m resolution, studies such as Spawn et al (2019) and Feng et al (2022) might be able to account for subgrid-scale transitions. However, these studies are restricted in their spatial extent (Tropics, USA), do not cover legacy fluxes due to their temporal limitation, and provide only specific component fluxes of E_LUC. The latter is a general problem of E_LUC estimates based on satellite-derived data of vegetation dynamics, such as forest cover changes (Hansen et al 2013): since land-use dynamics coincide with natural disturbances (e.g. natural wildfires or insect outbreaks), satellite-derived data of vegetation cover changes, although increasingly available at high resolution, cannot be used directly as input to carbon cycle models (Pongratz et al 2021). Typically, only component fluxes such as from cropland expansion of specific types of land-use-induced forest cover losses can be derived directly from satellite data. Due to the increasing availability of time series from satellite products, there is a clear tendency towards spatially higher resolutions of LULCC datasets and E_LUC estimates, but research on the influence of the resolution of underlying LULCC reconstructions on E_LUC estimates is limited.

HILDA+ provides annual data for the time period 1960–2019 and based on that data interpolated trends for the time period 1900–1960 (Winkler et al 2021). In comparison, LUH2 (Hurtt et al 2020, Chini et al 2021) covers a LULCC history dating back to AD 850 with data provided every 100 years until 1700, every 10 years between 1700 and 2000, and annually afterwards. To create annual LULCC maps, the data before 2000 is linearly interpolated between the above-mentioned time steps (Hurtt et al 2020). The importance of the starting year of a model simulation is analyzed by Hartung et al (2021) for cumulative LULCC fluxes. Accordingly, based on simulations starting in AD 850, 1700, and 1850, the uncertainty introduced by the initialization year amounts to 15% for estimates of cumulative E_LUC in the time period 1850 to 2014. However, it remains unclear to what degree the starting year influences estimates of the more recent years, which are most important, e.g. as reference years or for the global stocktake, and if an initialization in 1900 is sufficient for estimating emissions from 1960 onwards.

The goal of this study is to highlight spatial and temporal uncertainties in E_LUC estimates related to (a) LULCC reconstructions, (b) the resolution of the LULCC forcing, and (c) the initialization year.

For this study, HILDA+ is implemented in a bookkeeping model (BLUE) and results are compared to estimates of simulations based on LUH2, which is the LULCC dataset most commonly used in global E_LUC models. In simulations of BLUE, E_LUC fluxes from transitions between natural vegetation, cropland, and pasture, as well as from wood harvesting are considered (Hansis et al 2015). Vegetation and soil carbon densities for each combination of LULCC states and eleven PFTs are based on literature values and provided in Hansis et al (2015). Response curves derived from literature represent the carbon dynamics of different carbon pools following land-use changes and describe the decay and accumulation of vegetation and soil carbon. This includes the transfer of carbon to product pools of different lifetimes or the increase of carbon in different vegetation and soil pools due to regrowth of natural vegetation (Hansis et al 2015).

BLUE simulations with three different LULCC inputs (HILDA+ at 0.25^∘ and at 0.01^∘, and LUH2 at 0.25^∘) were initialized. Four BLUE simulations were carried out based on HILDA+ at 0.25^∘ with different initialization years (1900, 1920, 1940, 1960), and six simulations with the HYDE 3.2 based LUH2 data that was used for the BLUE estimates in the GCB2020 (initialized in 1700, 1850, 1900, 1920, 1940, 1960). The runs with different years of initialization are important to identify the minimum required starting year for robust E_LUC estimates. The initialization year 1700 corresponds to pre-industrial times, the year 1850 marks the approximate beginning of the industrial era, and the years 1900, 1920, and 1940 relate to the time period of interpolated trends of HILDA+, while 1960 is the first data-driven year of HILDA+. The simulation with HILDA+ at 0.01^∘ was initialized in 1900.

Unlike LUH2, HILDA+ does not provide information on wood harvest and does not distinguish primary and secondary land, which is both required to capture important aspects of the carbon cycle. Thus, HILDA+ had to be processed and complemented before implementing it in BLUE. A detailed description of the processing of the data as well as a comparison of HILDA+ and LUH2 in terms of total area, spatial patterns, and annual change rates of LULCC states is provided in the supplementary materials (sections A and B) (available online at stacks.iop.org/ERL/17/064050/mmedia).

3.1. Differences in global estimates

Global E_LUC estimates based on HILDA+ and LUH2 differ in size and trends (figures 1 and 2). Total E_LUC estimates from the simulations with HILDA+ alternate around 1.0 PgC yr⁻¹ and decrease after 2012 from 1.3 to 0.8 PgC yr⁻¹ in 2019. Contrary, E_LUC estimates based on LUH2 decrease from 2.3 PgC yr⁻¹ in 1960 to about 0.9 PgC yr⁻¹ in 1999 and increase afterwards to 2.0 PgC yr⁻¹ in 2019. Gross source and sink fluxes are greater in estimates based on LUH2 compared to the one based on LUH2. Trends in the last two decades are dominated by emissions from cropland expansions, with increasing tendencies forLUH2-based estimates and decreasing tendencies for estimates based on HILDA+.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Overall cropland emission estimates are on average almost three times higher, and the sink from abandonment of agricultural land is more than twice as big in the simulation with LUH2 compared to the one based on HILDA+ (figure 2). The differences in cropland expansion and agricultural land abandonment estimates are connected to differences in the annual change rates of the LULCC input datasets. Due to the implementation of shifting cultivation in LUH2, gross gains and losses in cropland and secondary land area are higher in LUH2 compared to HILDA+, resulting in higher cropland emissions and a larger sink from agricultural land abandonment (figure 3). Compared to cropland expansion and agricultural land abandonment, emission estimates from pasture expansion and wood harvest are of similar magnitudes on a global level. However, larger regional differences exist for pasture emission estimates.

Zoom In Zoom Out Reset image size

3.2. Differences in regional estimates

3.3. Influence of spatial resolution

BLUE simulations, when forced with HILDA+ at 0.01^∘ and 0.25^∘ resolution (original HILDA+ resp. LUH2 res.) as LULCC input, reveal substantial differences mainly in component fluxes (figures 4, S6, S7, tables S1 and S3). Globally, the mean difference between the two simulations is 36 TgC yr⁻¹ for the time period 1960–2019. The highest differences in component fluxes are observed in Europe, South Asia, and the Mideast. In general, emission estimates from cropland and pasture expansion tend to be larger and the sink from abandonment of agricultural land tends to be greater at 0.25^∘ resolution.

Zoom In Zoom Out Reset image size

Additional idealized BLUE simulations with artificial LULCC input data (section G in supplementary materials) revealed that these differences are related to the occurrence of successive transitions in grid cells, i.e. these grid cells experience at least two, but mostly more transitions in the covered time period. In the prepared HILDA+ dataset at 0.01^∘ resolution, 84% of the global land grid cells do not undergo any transition between 1900 and 2019, 10% experience one transition and 5% have more than one transition (table S4). In comparison, in Europe 21%, in South Asia 15%, and in the Mideast 7% of the grid cells have two or more transitions. Oceania (34%) and USA (9%) have high numbers of grid cells with successive transitions as well. However, the differences in component fluxes in these two regions are rather small, balancing out spatial differences. Also other regions have substantial amounts of successive transitions, but relative to the total transitions less than Europe, South Asia or the Mideast. Figure 5 illustrates the effect of successive transitions at different resolutions.

Zoom In Zoom Out Reset image size

3.4. Influence of initialization

E_LUC estimates of simulations with different initialization years show relatively small differences when the initialization year is at least 60 years prior to the analyzed time period (figure 6). The difference in cumulative E_LUC estimates of the time period 1960–2019 for the simulation based on HILDA+ (res. 0.25^∘) and initialized in 1900 versus the simulation based on HILDA+ (res. 0.25^∘) and initialized in 1920 is less than 2%. For simulations based on LUH2 and initialized in 1700 and 1900, the difference in cumulative E_LUC estimates (1960–2019) is less than 0.1% (figure S9). The difference of cumulative E_LUC emissions of later time periods such as 1990–2019 or 2010–2019 is even smaller, since E_LUC estimates with different years of initialization converge with increasing time.

Zoom In Zoom Out Reset image size

The alignment of E_LUC emission estimates based on different underlying LULCC forcing data differs globally, between regions and in certain regions depending on the time periods. Comparing E_LUC estimates from the BLUE model based on HILDA+ with estimates based on LUH2, the highest agreement for the total E_LUC is reached for Europe and Central America, while estimates for Brazil, China, and Oceania disagree substantially (figure S5). For other regions, the level of agreement varies over time. For Europe, a high consensus among estimates based on different LULCC forcing data and models is confirmed by several studies (Gasser et al 2020, Bastos et al 2021, Petrescu et al 2021). For Brazil, similar to our analysis, Rosan et al (2021) find little agreement of E_LUC emission estimates based on HYDE 3.2, the newer HYDE 3.3 version, and a national LULCC forcing. However, the suggested decline of total E_LUC emission estimates based on HYDE 3.3 by Rosan et al (2021) in the last two decades cannot be reproduced by our estimates based on HILDA+ due to increasing emissions from pasture expansion. The change in trend in the global E_LUC estimates that occurred in the GCB2020 as compared to the GCB2021 (Friedlingstein et al 2021) and that resulted from the change from a HYDE3.2 to a HYDE3.3 based LULCC forcing as described by Rosan et al (2021) is thus not confirmed by our simulations based on HILDA+ for Brazil. The decreasing trend in global emissions described in section 3.1 (figure 1(C)) for the last two decades in E_LUC based on HILDA+ in contrast to LUH2 is instead strongly attributable to Southeast Asia, where cropland emissions are revised down in our simulations using HILDA+. For Southeast Asia, a regional study (Kondo et al 2021) that uses DGVMs and bookkeeping models with different LULCC forcing data concludes a higher reliability of estimates based on LUH1 (Chini et al 2014) compared to the ones based on LUH2 for the region. The estimates based on HILDA+ confirm decreasing E_LUC emissions in Southeast Asia since the 2000s, although they suggest a later peak than Kondo et al (2021). For the USA, Yu et al (2019) reason that E_LUC emission estimates based on LUH2 overestimate the carbon sink, when comparing it to estimates based on a national land cover dataset. Contrary, our estimates based on HILDA+ do not suggest such a substantial overestimation compared to estimates based on LUH2 for the USA. These regional examples highlight a lack of agreement between different LULCC datasets and the implementation of LULCC dynamics in different models, in particular on regional level. Newer estimates do not necessarily converge. Given the fact that the most recent years are most important for tracking mitigation efforts such as policies to halt deforestation or reforestation programs, the disagreement of LULCC datasets since 2000 urgently needs to be resolved.

Another major difference in E_LUC estimates, mainly in tropical regions, are much higher emissions from cropland expansion and a larger sink from abandonment of agricultural land (cropland and pasture) in estimates based on LUH2. As explained in section 3.1, this is connected to the implementation of shifting cultivation in LUH2 and the omission of it in HILDA+. According to Heinimann et al (2017), the area influenced by shifting cultivation is spatially limited to roughly 280 Mha in the tropics between 30^∘ S and 30^∘ N. The inclusion of shifting cultivation in models, usually treated as a net vs. gross transition issue, is reported to lead to higher E_LUC estimates (Stocker et al 2014, Wilkenskjeld et al 2014, Hartung et al 2021). Arneth et al (2017) estimate an increase by 20%–30% when considering processes such as shifting cultivation. Furthermore, Bastos et al (2020, 2021), Gasser et al (2020) highlight substantial differences due to the implementation of gross transitions in estimates based on LUH2 compared to estimates based on other LULCC datasets. We do not find considerably higher E_LUC estimates based on LUH2 and HILDA+ that can be attributed to shifting cultivation as long as we consider the total E_LUC. Despite much higher annual area gross changes of cropland and secondary land in certain tropical regions in LUH2 compared to HILDA+, which we ascribe to the implementation of shifting cultivation in LUH2, the component fluxes of cropland expansion and agricultural land abandonment mostly compensate for each other, and as a consequence total E_LUC estimates match fairly well in most of the affected regions (at least before the increase in the last two decades, which is not connected to shifting cultivation). Similarly, Gasser et al (2020) note that shifting cultivation has a long-term effect of zero net emissions in the OSCAR model. Based on our findings, we argue that (a) gross transitions and shifting cultivation should be treated differently and (b) the implementation of shifting cultivation in LULCC reconstructions and carbon cycle models needs to be reconsidered. As described in section 3.3, in LULCC reconstructions with low resolution more area is assumed to be under transition compared to the same data at high resolution ('effect of successive transitions'), which shows that the rotational cycles of shifting cultivation cannot accurately be represented at 0.25^∘ resolution, neither can they at 0.01^∘ resolution, since patches of shifting cultivation are usually maximum a few hectares in size (Villa et al 2020, Bruun et al 2021). Moreover, several case studies (Bruun et al 2009, 2021, McNicol et al 2015, Terefe and Kim 2020) point out substantial differences in the carbon fluxes of the expansion and abandonment cycles of shifting cultivation compared to other expansion or abandonment transitions (e.g. clearing of former shifting cultivation areas for palm oil plantations), due to different regrowth rates and soil carbon dynamics. It remains unclear, if these drawbacks of current implementations in models can fully explain the large influence that shifting cultivation has on global and regional E_LUC component fluxes according to simulations based on LUH2 or if the implementation of shifting cultivation in LUH2 leads to an additional overestimation.

The spatial resolution of the LULCC input data has a significant influence on E_LUC component fluxes. Our estimates based on gross transitions of HILDA+ at 0.01^∘ and 0.25^∘ resolution and the BLUE experiments with artificial LULCC input revealed that component fluxes are smaller at higher resolutions, which can lead to overall higher or lower total E_LUC estimates. As described above, these differences are caused by successive transitions. According to Winkler et al (2021), successive transitions were prevailing in the Global North (USA, Europe, Australia) and rapidly growing economies such as India, Nigeria, and Turkey. Most of the transitions in these regions were changes between managed and unmanaged land (crop/pasture to secondary land or reverse) (Winkler et al 2021). However, potential explanations are needed for these diverse and region-specific high land-use dynamics: in the USA cropland abandonment was driven over time by federal policies and changes in commodity prices among others (Chen and Khanna 2018, Hendricks and Er 2018, Lark et al 2022), in Mediterranean Europe and Australia certain pasture-shrubland dynamics were influenced by climatic and socioeconomic changes (Eldridge and Soliveres 2014, Rolo and Moreno 2019), in Eastern Europe the agricultural sector experienced massive changes following the breakdown of the former Soviet Union (Prishchepov et al 2013, Schierhorn et al 2019), in Turkey a mix of industrialization, urbanization, and migration led to rapid changes in land use practices (Tanrivermis 2003), in India the heavy usage of irrigation and fertilizer enabled agricultural intensification (Ambika et al 2016, Chen et al 2019), and in Nigeria conversions to cultivated land dominated LULCC dynamics (Arowolo and Deng 2018). Moreover, crop rotation or mixed crop-livestock systems may also be linked to the observed successive transitions in Australia, the USA, and Europe (Peyraud et al 2014, Rosenzweig et al 2018, Ghahramani et al 2020).

The resolution-dependent 'effect of successive transitions' has not been described in the literature so far, although different studies discuss the importance of spatial resolution and transition types for E_LUC estimates in other respects. For example, Wilkenskjeld et al (2014) point out that a coarser resolution of net LULCC data leads in a reduction in area affected by LULCC and thus affects E_LUC estimates. Several studies highlight the importance of using gross over net LULCC transitions to account for the actual area changes (i.e. Hansis et al 2015, Arneth et al 2017, Bayer et al 2017, Bastos et al 2020, 2021). However, Yue et al (2018) conclude from simulations with sub-grid secondary forests of different age classes that the contribution from gross transitions to overall E_LUC estimates tend to be overestimated due to the non-consideration of age classes in most models. The findings from Yue et al (2018) go in a similar direction as our observation that successive transitions are not adequately represented in gross transitions at coarse resolution (nor with net transitions), and consequently different land areas are affected by successive transitions, when compared to the same LULCC data at high-resolution. It is likely that the 'effect of successive transitions' is also of greater importance for DGVMs and other bookkeeping models.

Our simulations starting at different years showcase the importance of a prudent choice for the year of initialization. E_LUC estimates of simulations based on HILDA+ for 2019 differ by more than 5% when initialized in 1960 compared to simulations with the same LULCC forcing but initialized in 1900. Further, the results indicate that the influence decreases over time and differences between simulations with earlier and later starting years become marginal after a few decades. The simulations highlight that (a) the initialization year needs to be well before the satellite area to capture present-day fluxes accurately (at least 95% similarity in cumulative emission estimates compared to simulations starting 20 years earlier), (b) a lead time of 60 years seems sufficient (95% similarity criterion, see above) and (c) the time period covered by HILDA+ starting in 1900 is suitable for the estimation of E_LUC after 1960 without introducing large uncertainties due to the initialization year.

E_LUC estimates have high uncertainties, which are partly caused by underlying LULCC datasets among other drivers and parameters. The implementation of a new LULCC reconstruction dataset (HILDA+) in a bookkeeping model (BLUE) enabled us to evaluate and compare E_LUC estimates based on HILDA+ to E_LUC estimates based on the widely-used default LULCC dataset LUH2. Results show that global E_LUC estimates based on HILDA+ are substantially lower than estimates based on LUH2. Regionally, a pattern of higher E_LUC emissions from cropland expansion and a larger sink from agricultural land abandonment in estimates based on LUH2 can be observed in most tropical regions. The larger sources and sinks can partly be explained by the inclusion of shifting cultivation in LUH2, which raises questions about the influence of shifting cultivation on the global carbon cycle and the implementation of shifting cultivation in LULCC datasets and carbon cycle models. Another significant difference are opposing trends of E_LUC estimates globally and in many regions in the last two decades. These substantial differences highlight the need for more reliable LULCC reconstructions for more accurate and robust E_LUC estimates. Independent estimates for the evaluation of LULCC dynamics, including knowledge of regionally specific LULCC activities, component-specific evaluations, and complementing default global runs, such as in the GCBs, by alternative LULCC data could increase the understanding of differences and provide better estimates of uncertainties. Furthermore, we run simulations based on LULCC data at different spatial resolutions (0.01^∘ vs. 0.25^∘) and find significant differences in E_LUC component fluxes. The reason for this phenomenon are successive transitions. These cannot adequately be represented at the coarse resolution, which has the effect that at the coarser resolution overall a larger area is affected by LULCC events. Moreover, a lead time of at least 60 years has been found crucial to account for legacy emissions and retrieve robust E_LUC estimates. This rather long lead time to capture legacy emissions, together with the need for ancillary data or methods to split anthropogenic from natural drivers of land use dynamics, challenges the application of purely satellite-based LULCC datasets, although their often high spatial resolution could provide an important step forward to capture successive transitions. Both the sensitivity to spatial resolution and initialization year are qualitatively independent of the concrete LULCC dataset, such that we recommend accounting for these issues in future studies with other LULCC activity data and carbon cycle models.

R G and K W acknowledge support from the European Commission through Horizon 2020 Framework Programme (VERIFY, Grant No. 776810). S B was supported by German Stifterverband für die Deutsche Wissenschaft e.V. in collaboration with Volkswagen AG. This work used resources of the Deutsches Klimarechenzentrum (DKRZ) granted by its Scientific Steering Committee (WLA) under Project ID bm0891.

The data that support the findings of this study are available upon reasonable request from the authors.