A multi-data assessment of land use and land cover emissions from Brazil during 2000–2019

Brazilian ecosystems and especially forests play a fundamental role in regional and global carbon stocks and natural land C sinks. The Amazon forest is estimated to contain around 229–280 PgC in living biomass and soils (Malhi et al 2006, Gloor et al 2012), representing ∼10% of global land C stocks (Ciais et al 2013), and approximately 60% of its area is in Brazil. Carbon stocks in Brazilian ecosystems have been negatively impacted by significant land use and land cover change (LULCC) associated with demographic and agricultural expansion, resulting in large land-use emissions to the atmosphere (Houghton 2012, Aide et al 2013). Globally, in the last decade (2009–2018), a total of 1.5 ± 0.7 PgC yr⁻¹ was released to the atmosphere due to LULCC (Friedlingstein et al 2019). Despite a significant slowdown in deforestation in Brazil after a peak in 2004, mainly due to policy introduced to curb deforestation (Arima et al 2014, Godar et al 2014, West and Fearnside 2021), Brazil is still contributing with between 17% and 29% of global LULCC emissions (ELUC) (Friedlingstein et al 2019). However, divergent ELUC estimates for Brazil in the global carbon cycle budget have contributed to a large fraction of the corresponding overall global uncertainty (Bastos et al 2020).

In the global carbon budget (GCB 2020), ELUC is defined as the net anthropogenic LULCC flux and includes removals (e.g. from forest regrowth after harvest and agricultural abandonment) and emissions e.g. from clearing natural vegetation and transitions (Friedlingstein et al 2020). These definitions are different from those used by countries and FAO to estimate and report emissions from LULCC within the IPCC LULUCF category of the national greenhouse gas inventory (NGHGI). Within the IPCC guidelines, LULCC is limited to emissions associated only to anthropogenic-related processes driven changes in land use and land cover (see SI table 3 (available online at stacks.iop.org/ERL/16/074004/mmedia)). In the GCB, differently, the ELUC is estimated by two different bottom-up approaches, namely process-based and bookkeeping models (Friedlingstein et al 2019). An ensemble of process-based, Dynamic Global Vegetation Models (DGVMs) from the Trends in Net Carbon Exchange Project (TRENDY) (Sitch et al 2015), are applied using observed historical CO₂ in atmosphere, climate and LULCC fields.

An additional method is the carbon stock change approach of the IPCC (2003), followed by FAO for its ELUC estimates (Tubiello et al 2020). The advantage of this method is the possibility to perform complex calculations using a very small set of input data, while on the other hand it cannot distinguish necessarily between natural and anthropogenic fluxes. Bookkeeping models include Houghton and Nassikas (H&N, 2017) and bookkeeping of land use emissions (BLUE) (Hansis et al 2015). Both TRENDY and BLUE use the same LULCC dataset as spatially explicit input based on the History Database of the Global Environment (HYDE 3.2) (Goldewijk et al 2017) for annual change in pasture, rangeland and cropland area. There are many uncertainties related to these global LULCC datasets, since they use FAO statistics as input, which are provided as national aggregates (FAO 2020a) and then rely on a suite of methods to disaggregate that information spatially. With the increasing availability of land cover datasets based on Earth observation (EO) covering the last 30 years, it has been possible to integrate time varying remote sensing data with the FAO national statistics to generate new and improved spatially explicit global LULCC datasets. Several EO-based products of LULCC and deforestation have been developed in Brazil, such as the Amazon Deforestation Monitoring Project (PRODES) from the National Institute for Space Research (INPE), and the MapBiomas dataset, which was developed specifically for Brazilian biomes and provide annual LULCC maps for the whole of Brazil from 1985 up to present. However, these datasets have hitherto not been used to assess the impact of LULCC in global C-cycle assessments. Therefore there is an urgent need to improve estimates of ELUC for Brazil to better represent the spatio-temporal trends in future GCB annual assessments. Furthermore, accurate estimates of ELUC contribute to the quantification of emissions and removals from LULUCF processes that are needed to guide global and national policies to achieve the overarching goal of the Paris Agreement (UNFCCC 2015).

Here we present a critical analysis and evaluation of LULCC in Brazil and associated C emissions over the 21st century (2000–2019). We use two versions of the HYDE land use dataset, HYDE 3.2 (Goldwijk et al 2017; used in GCB 2019) and the new HYDE 3.3 based on updated FAOSTAT statistics (FAO 2020a, 2020b) and time-varying land cover data from the European Space Agency Climate Change Initiative (ESACCI-LC 2017). First, we evaluate the two versions of HYDE using the national LULCC MapBiomas product. Then, we compare simulated ELUC for Brazil based on the two HYDE versions and MapBiomas using the process-based JULES-ES and BLUE bookkeeping models. Finally, we discuss simulated ELUC dynamics from BLUE and JULES-ES using HYDE 3.3 and MapBiomas with the other published estimates.

2.1. Land use datasets

2.2. Estimates of ELUC

2.3. Analyses

2.3.1. Land use change analysis

To assess the differences between the LULCC products for Brazil and to understand LULCC dynamics, HYDE was evaluated against the national MapBiomas dataset. MapBiomas was processed as follows: first each category is reclassified with the proportion of crop and pasture at 30 m spatial resolution. The following class aggregation applies to cropland for MapBiomas: the categories, Annual and Perennial Crop and Semi-Perennial Crop were defined to contain 100% cropland in each pixel; the category Mosaic of Agriculture and Pasture as 40% cropland. For pastures we consider Pasture category as 100% and Mosaic of Agriculture and Pasture to contain 60% of pasture as most of these mosaic categories are used in extensive cattle ranching and small-scale agriculture. Then we re-grid to 0.5 degree to generate gridded maps of cropland and pasture cover fraction.

The resulting maps were compared against ESA CCI cropland categories 10, 11, 12, 20 which contain 100% cropland (table 1 SI description). Further categories were included, such as mosaic categories 30 (cropland >50%), 40 (cropland <50%), assuming a cropland area proportion of 60% and 40%, respectively, based on (Liu et al 2018). Note, ESA CCI does not have an explicit 'pasture' category, rather it is included in the 'cropland' category. For representation, we report the sum of cropland and pasture categories. These maps between land cover and land use categories introduce a level of uncertainty that is currently poorly quantified. Furthermore, the choice of the underlying land cover map introduces uncertainties, for instance results discussed herein are likely different from those that would be obtained using MODIS land cover maps rather the ESA CCI (e.g. see FAO 2020b).

To test the spatial similarity of the LULCC maps from HYDE against MapBiomas we apply the fuzzy numerical method implemented in the Map Comparison Kit 3 application (Visser and De Nijs 2006). It is a cell-by-cell comparison method for numerical maps that also considers the neighbourhood to show the similarity of each pair of grid-cells in a range between 0 (distinct) and 1 (identical). Here, we adopt the default settings provided by the algorithm, with an exponential decay function called Halving distance equal to 2 and neighbourhood distance equal to 4 grid-cells. This provides a spatial assessment showing the location and severity of the disagreement between two maps.

Since agricultural area change is currently used to infer tree cover loss in DGVMs, we calculated the pairwise Pearson correlation between increase of net land cover (cropland + pasture) with the vegetation cover loss (deforestation) from MapBiomas c5 including both primary and secondary vegetation and the following natural vegetation categories in MapBiomas: forest, savanna and natural grassland. We performed this correlation analysis for the changes between 2000 and 2019 comparing the grid-cells with increase in LULCC from HYDE 3.2, HYDE 3.3 and MapBiomas to grid-cells with vegetation cover loss for the same period.

2.3.2. Carbon emissions from LULCC

HYDE 3.3 and MapBiomas fraction maps were re-gridded for using in JULES-ES at N96 resolution (1.25° latitude × 1.875° longitude) with Climate Data Operators (CDOs) using first-order conservative remapping method. For this analysis, we also considered converted rangeland fraction from HYDE 3.3 in order to better represent the LULCC changes across the Amazon arc of deforestation (low human population density but transitions to extensive pastures), which was not considered in JULES-ES, TRENDY-v9. We run JULES-ES using the LULCC fields (cropland, pasture and converted rangeland) from HYDE 3.3 and MapBiomas (cropland and pasture) with the same configuration as in TRENDY-v9 (GCB2020), including time varying climate, CO₂, and nitrogen deposition following the TRENDY protocol. The simulations using HYDE 3.3 extend over the period 1700–2019 whereas the MapBiomas simulations are over the more recent period 1986–2019. JULES-ES is spun-up to steady state conditions and then we perform the following experiments: S2 (CO₂ and climate forcing varying, land-use constant at 1700 for HYDE 3.3 and 1986 for MapBiomas) and S3 (CO₂, climate and land-use time variant) over 1700–2019 for HYDE 3.3 and 1986–2019 for MapBiomas. ELUC is diagnosed as the difference in net biome productivity at a grid-cell level between the two runs, i.e. S2–S3:

$$\begin{equation*}\left( {{\text{S}}2} \right){\text{ NBP}} = {\text{NPP}} - {\text{RH }} - {\text{HARVEST}}\end{equation*}$$

$$\begin{align*}\left( {{\text{S}}3} \right){\text{ NBP}}& = {\text{NPP}} - {\text{RH}} - {\text{HARVEST}}\\ & \quad - {\text{PRODUCT}}\_{\text{DECAY}}\end{align*}$$

where NPP is net primary productivity, RH is the heterotrophic respiration, and the human disturbances in this study are represented by Harvest which is the crop harvest and Product decay which is the decay flux from wood product pools (Jones et al 2011). Further explanation on the MapBiomas legacy flux estimation can be found in the supplementary information and in SI table 5.

BLUE used the same configuration as in GCB2019, however land-use forcing (LUH2) was replaced by the two HYDE versions and wood harvest as well as sub-grid scale transitions (i.e. shifting cultivation) were consequently not considered because HYDE only provides net area changes per land cover category, and does not consider gross transitions within our large grid-cell areas (>27 km). Both BLUE simulations were initialized in 1961 for HYDE and 1986 for MapBiomas. In order to minimize differences to the GCB2019 setup to allow a cleaner comparison, HYDE 3.2 and HYDE 3.3 were re-gridded to a spatial resolution of 0.25° × 0.25° with CDO using first-order conservative remapping method. The re-gridded HYDE data was processed to match the pre-processing of the land-use forcing data used in other BLUE simulations. In particular rangeland areas were considered to imply clearing of natural vegetation only when the forest/non-forest map of LUH2-original land cover data set in BLUE-indicates forest. Potential vegetated grid cell fractions not in cropland or pasture were split into primary and secondary land according to the proportion in LUH2. The set-up of BLUE simulations can be found at SI table 5.

Finally, we discuss our new ELUC results using HYDE 3.3, MapBiomas and other published estimates. We compared ELUC estimates from JULES-ES and BLUE (2000–2019) for the whole of Brazil with the H&N global bookkeeping model (Houghton and Nassikas 2017), the FAOSTAT (Tubiello et al 2020), and national inventory datasets: the 4th National Communication of GHG to the United Nations Framework Convention on Climate Change (UNFCCC) (Brazil MCTI 2020) and provides estimates up to 2016, and the SEEG LULUCF estimates based on MapBiomas dataset (De Azevedo et al 2018). Comparison of ELUC for Amazonia and Cerrado is done using estimates from INPE-EM for the biome boundaries. The INPE-EM ELUC (Aguiar et al 2012) estimate is based on the official deforestation data from the National Institute for Space Research (INPE) for both biomes using the 2nd order estimates (i.e. instantaneous and legacy C emissions). A table with the model components considered to estimate ELUC is in SI table 2.

3.1. Land use and land cover changes in Brazil

We find temporal agreement between both remote sensing-based products (ESA CCI and MapBiomas) and HYDE 3.3 at country level. The three datasets agreed on the peak LULCC in Brazil between 2003 and 2005 with a negative trend thereafter opposite to the positive trend shown by HYDE 3.2 (figure 1(a)). This period corresponds to the peak of deforestation in the Brazilian Amazon between 2003 and 2004 and is followed by a slowdown in deforestation rates due to the implementation of governmental regulations to reduce deforestation thereafter. Additional economic factors (West and Fearnside 2021), and improvement in the use of verification and detection tools based on remote sensing, e.g. PRODES and the DETER program for near-real time deforestation detection also contributed to the deforestation slowdown in Brazil. However, both ESA CCI and consequently HYDE 3.3 showed lower LULCC after 2005 compared to MapBiomas. This may partially be due to the pervasive increase in small-scale deforestation (<1 ha) (Kalamandeen et al 2018), which may remain undetected using the 1 km change detection implemented in ESA CCI LC methodology (ESACCI-LC 2017). In addition, the total LULCC in HYDE 3.3 reflected the updated cropland and land under permanent meadow and pasture area in FAOSTAT (FAO 2020b) based on the decadal Brazilian Agriculture census. To balance the total land area, FAO increased the residual area in another category called 'other land' which is not used by the HYDE dataset and may include a proportion of the LULCCs associated with deforestation. Further explanation of the other land category can be found in the supplementary information. All these changes contribute to lower net LULCC observed using HYDE 3.3 data when compared to MapBiomas (figure 1(a)).

Zoom In Zoom Out Reset image size

Spatially, there were large differences between the global LULCC products and MapBiomas (figure 2). HYDE 3.2 had the largest LULCC located in SE Brazil associated with cropland expansion and lower changes in pasture area. These changes are not consistent with MapBiomas (figure 2), were large losses of natural vegetation are found in Amazonia and the Cerrado biomes, and an intensification of cropland areas mostly concentrated in repurposed pastures in the SE Brazil (Zalles et al 2019). HYDE 3.2 used ESA CCI LC baseline for 2010 but to estimate year-to-year changes it employs an algorithm to allocate transitions within a country giving preference for conversion of lands near existing agriculture and with high NPP. This may not be sufficient to capture deforestation in more remote regions in a large country like Brazil, hence the LULCC allocation in HYDE 3.2 was centred in SE Brazil in consolidated areas.

Zoom In Zoom Out Reset image size

Although HYDE 3.3 showed lower magnitude change compared to MapBiomas, the main gain with the updated version was the spatial allocation of LULCC. The similarity analysis showed that 20.2% of the grid-cells from HYDE 3.2 are dissimilar (similarity = 0) when compared to MapBiomas in grid-cells mainly located in the SE and NW Brazil (SI figure 3) indicating a larger spatial inconsistency in the LULCC for this region. HYDE 3.3 on the other hand, had only 2.2% of the grid-cells with similarity index equal to 0 when compared with MapBiomas and those grid-cells tend to be spatially spread, indicating an improvement on the spatial allocation. The spatial distribution of the similarity index and the frequency histogram is shown in SI figure 3. Additionally, the net changes from MapBiomas and natural vegetation cover loss showed a strong correlation (R = 0.94; SI figure 4), with net changes mainly due at expense of forest loss. This indicates that net observed LULCC change can be used as a proxy of the deforestation process. A pixelwise correlation comparison between both HYDE net LULCC versions and MapBiomas vegetation loss, indicated a superior performance in HYDE 3.3 (R = 0.55, SI figure 4) compared with HYDE 3.2 (R = 0.094; SI figure 4). This result demonstrates that although HYDE 3.3 still underestimates the changes, it is able to better reproduce and allocate spatially the deforestation pattern than HYDE 3.2.

3.2. Land use and land cover change emissions (ELUC) in Brazil

Simulated average ELUC for Brazil after the peak in 2004 was 0.34 PgC yr⁻¹, 0.18 PgC yr⁻¹ and 0.32 PgC yr⁻¹ in JULES-ES simulations with HYDE 3.2 (2005–2017), HYDE 3.3 and MapBiomas (2005–2019), respectively. ELUC with the BLUE model was 0.19 PgC yr⁻¹, 0.11 PgC yr⁻¹ and 0.39 PgC yr⁻¹ for HYDE 3.2. HYDE 3.3 and MapBiomas, respectively (SI table 4). Although simulations based on HYDE 3.2 showed higher emissions than HYDE. 3.3-based, both models when forced with HYDE 3.3 showed a negative trend after 2004. This is in agreement with the trend from our reference run based on MapBiomas LULCC forcing. Therefore, differences between MapBiomas and HYDE 3.3 simulations can reach up to 0.3 PgC yr⁻¹ in some years due the higher LULCC in MapBiomas than HYDE datasets (figure 3(a)).

Zoom In Zoom Out Reset image size

At the biome level, simulations with HYDE 3.3 and MapBiomas agreed on a downward trend after 2004 (p < 0.05, figure 3(b)) for the Amazon biome which is also consistent with INPE-EM bookkeeping model based on official deforestation data. However, for the same period simulations with HYDE 3.2 had an opposite direction and showed an increase of ELUC in both the Amazon and Cerrado biomes with the greatest increase in the latter (figure 3(c)). The update in HYDE also reflects an improvement in the contribution of each biome to the country total ELUC. Our results indicated that with the HYDE 3.2 version the Cerrado biome (Brazilian savannas) was the highest contributor (40.7%–61%) in contrast to Amazon (14.6%–22.5%) (SI figure 5). This is opposite to data from INPE-EM, which confirms that Amazon emissions are on average 59% higher than Cerrado emissions. The spatial improvement in the allocation method of HYDE 3.3 increased the contribution of Amazon biome to the total ELUC, now responsible for about 40% and the Cerrado 24%–32% of the country's total (SI figure 5). These improvements are important to better represent spatially the human disturbances across the Brazilian biomes.

Spatially, the main differences between the HYDE-based simulations are in SE Brazil and the arc of deforestation in Amazonia (figures 3(d)–(i)). Both models using HYDE 3.2 simulate higher emissions concentrated in the SE Brazil (i.e. Sao Paulo, Mato Grosso do Sul and Parana states) (figures 3(b) and (d)) when compared to the Amazon region, which is unrealistic given the dynamic of recent trends in deforestation hotspots in Brazil (i.e. Amazon and north of Cerrado regions as shown in SI figure 4). Therefore, both HYDE-based simulations still spatially underestimate the emissions in the Amazon 'Arc of deforestation' as shown by the spatial simulated emissions based on MapBiomas (figures 3(f) and (i)).

Zoom In Zoom Out Reset image size

The goal of this study was to evaluate global land use products used to estimate Brazil ELUC emissions comparing with additional country-specific data (MapBiomas) and estimate the impact of the new HYDE 3.3 dataset on the simulated ELUC. One of the factors of uncertainty and disagreement in the mean and trends in ELUC for Brazil is the driving LULCC dataset (Bastos et al 2020, Gasser et al 2020). Our results show that the HYDE 3.3 including updated FAO statistics and an improvement in the allocation method using multi-annual remote sensing-based ESA CCI land cover) product better distributed the LULCC and consequently ELUC across Brazil compared to the previous version. Unlike ELUC based on HYDE 3.2, our new HYDE 3.3-based estimates now agree on an overall negative trend after 2004 as shown in our reference MapBiomas simulation and other national datasets (NGHGI and SEEG LULUCF), FAOSTAT and H&N global bookkeeping models (figure 4 and SI table 4). Although there are still differences in the scale of emissions between the estimates in figure 4, these can be attributed to different methodological approaches, processes considered, and input data used to estimate ELUC (SI table 3).

Challenges remain in order to further improve representation of LULCC and ELUC in regional and global assessments. In general, DGVMs and some bookkeeping models use change in agricultural areas based on statistical data reported by countries to FAO as a forcing for tree-cover loss. By using only the agricultural areas based on FAOSTAT such as in HYDE dataset, our results demonstrated that we underestimate the ELUC compared to remote sensing-based data. This occurs because of the limitation of relying on the country reports and the assumption that all changes in natural vegetation loss can be inferred from changes in agricultural land without consideration of part of the 'other land' category which may be associated to the deforestation process. A recent study showed that about 13% of the tree cover loss in Brazil goes into a long transitional land category which is land not converted to be used by agricultural activities and may be associated to land grabbing (Zalles et al 2021). Thus, these areas are potentially included as a residual in the 'other land' category from FAO and not being considered in the ELUC estimates based on HYDE dataset. When using a remote sensing-based product to extract the observed changes in LULCC areas, such as the MapBiomas, the models were able to reproduce the overall pattern of vegetation loss and consequently ELUC trends over Brazil. Another caveat is that as we do not consider gross land cover changes and expect lower simulated ELUC in the early 2000s using the HYDE 3.3 dataset, also the regrowth sink in the late 2000s will likely be underestimated (Arneth et al 2017). A future challenge to the DGVMs and bookkeeping models such as BLUE will be to incorporate tree-loss directly from remote-sensing products. Nonetheless, this presents other challenges, such as the representation of legacy fluxes which occur on a timescale longer than the current availability of remote sensing datasets (Pongratz et al 2014). We believe this shortcoming will be overcome on the medium-term as longer time-series become available from satellite data. Moreover, global LULCC products with medium to coarse spatial resolution may not capture the increasing small-scale deforestation in Amazon (Kalamandeen et al 2018) which can contribute to an underestimate in the LULCC and its interannual variability. Variations in vegetation cover loss are associated with economic changes and environmental policy (Macedo et al 2012, West and Fearnside 2021), climatic events such as ENSO (El Niño–Southern Oscillation), which can contribute to spread of fires in intact forests (Alencar et al 2006, Aragão et al 2018), facilitating land cover conversion through fire, and also government decisions (Barlow et al 2020, Cardil et al 2020), which can lead to a high interannual variability. Using datasets based on semi- and decadal scale values such as in the updated version of FAO land-use statistical data for Brazil will result in lower year-to-year variation in vegetation loss and ELUC variability from DGVMs and Bookkeeping models. As shown in SI table 3 approaches vary in terms of LU processes included, e.g. shifting cultivation/sub-grid transitions, wood harvest, each of which could lead to an increase in ELUC estimates (Arneth et al 2017). In addition, forest degradation (selective logging, forest fire, edge-effects and fragmentation) is a growing threat, and may surpass deforestation in terms of both area and C emissions in several recent years (Aragão et al 2018, Assis et al 2020, Bullock et al 2020, Matricardi et al 2020, Silva Junior et al 2020), but is still not included in global DGVMs, bookkeeping models nor in national inventories. Hence improvements in DGVMs and global Bookkeeping models to explicitly use remote sensing derived tree-cover loss dataset and represent degradation processes are needed to further improve the representation of human disturbances on ecosystems and deliver better estimates of ELUC. Further efforts are still needed to better align (or map) the concepts between ELUC and LULUCF (Grassi et al 2018).

This study used a new global LULCC dataset based on the integration of time varying remote sensing data with updated national statistics to generate a new global land use dataset. HYDE 3.3 is shown to be superior in the spatial allocation to an earlier version based on only a single year satellite baseline. In particular, it reproduces the general spatial pattern of LULCC across Brazil when compared with national datasets, however it still underestimates the LULCC changes due to limitations associated with the assumption of a one-to-one correspondence between natural vegetation loss and changes in agricultural land based on statistical reports. When HYDE 3.3 was applied as input to a processed-based DGVM and a global bookkeeping model, both simulated a negative trend in ELUC for Brazil, in agreement with national and other global estimates. The simulations with both HYDE 3.3 and MapBiomas also identify the Brazilian Amazon as the largest contributor to the total country ELUC. In summary, improvements in LULCC datasets have resulted in consistent estimates of ELUC trends across different methodologies for Brazil. These advances will likely improve GCB estimates, contributing to reduced uncertainty in the global estimates and improve our understanding of the global carbon cycle.

This research has been supported by the Newton Fund through the Met Office Climate Science for Service Partnership Brazil (CSSP Brazil), RECCAP2 project which is part of the ESA Climate Change Initiative (contract no. 4000123002/18/I-NB), and the H2020 European Institute of Innovation and Technology (4C; Grant No. 821003). V H was supported by a NERC GW4+ Doctoral Training Partnership studentship from the Natural Environment Research Council (NE/L002434/1). R G acknowledges support from the European Commission through Horizon 2020 Framework Programme (VERIFY, Grant No. 776810). FAOSTAT is produced with the support of its member states and FAO regular programme funding. L M M acknowledges the UK Natural Environment Research Council funding (UK Earth System Modelling Project, UKESM, Grant NE/N017951/1). L E O C A was supported by CNPq (processes 305054/2016-3 and 442371/2019-5). C vR was supported by CNPq grant 314780/2020-3.

T M R and S S conceived the ideas and designed the methodology, T M R, R G, J P, K K G and M O S compiled the data. K K G developed new HYDE3.3 datasets iterating with T M R on Brazil. R G applied BLUE, and T M R and M O S applied JULES-ES using HYDE3.3 and MapBiomas. T M R analyzed the data, and results discussed with all authors. T M R and S S led the writing of the manuscript to which all authors contributed. F N T provided support in terms of FAO data on land use and GHG emissions as well as in providing mapping between land use classes. The views expressed in this publication are the author's and do not necessarily reflect the views or policies of FAO. We thank Renan Milagres Lage Novaes from EMBRAPA and Giorgia De Santis from FAO to help clarify the nature of FAOSTAT land use data used in the manuscript. We also thank Clement Albergel from ESA for his helpful comments on the manuscript.

MapBiomas data is freely available at http://mapbiomas.org/ and ESA CCI LC at www.esa-landcover-cci.org. HYDE 3.3 will be available upon request to (kees.kleingoldewijk@pbl.nl). The simulations will be available upon reasonable request to the correspondent author.

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