There are four sites in the Brazilian Amazon where the top-down flight atmospheric measurements were collected (Gatti et al., 2014, 2021). The sites are ALF (8.8°S, 56.7°W) at the southeast, SAN (2.8°S, 54.9°W) at the northeast, RBA (9.3°S, 67.6°W) at southwest-central, and TAB (5.9°S, 70.0°W) until 2012, which has been replaced by TEF in 2013 (3.7°S, 66.5°W), at northwest central region (Figure 1). The influence region of each flight site is estimated as the envelope of backward air mass trajectories reaching the sites. In this study, we have considered the mean annual influence region of all years (2010–2018) of each flight collection site (Figures 1B–F) to estimate the bottom-up emissions and removals. Our study area follows the Brazilian Amazon forest biome delimitation of Olson et al. (2001) with an area of 4,215,763 km² (Figure 1A; from now on, we will refer to it as the Brazilian Amazon) and the influence regions of TAB, TEF, RBA, SAN, and ALF inside the Brazilian Amazon (Figures 1B–F).

At the air-sampling sites, samples were collected in a descending vertical profile from 4,420 m above sea level (a.s.l.) to ~300 m a.s.l. The sampling system was installed on board in a small aircraft. All vertical profiles were made during the period when the planetary boundary layer tends to be well-mixed (between 12:00 and 13:00 local time), such that the profiles integrate gas fluxes from large regions under its region of influence. Air samples were analyzed by a non-dispersive infrared analyzer for CO₂ and by gas chromatography with HgO reduction detection for CO.

For each vertical profile, the flux from the influence region of the site was estimated as the difference of ΔCO₂ of the observed vertical profile CO₂ mole fraction at the site, a so-called background mole representing the amount of CO₂ of the atmospheric column entering the Amazon from the tropical Atlantic via the trade winds. Then, ΔCO₂ is divided by the travel time of air parcels from the Atlantic, to the vertical profiling site, obtaining an average CO₂ flux estimate for the influence region of the site, following the equation:

where CO₂ is carbon dioxide dry air mole fraction in units of (mol m⁻³), z is the height above sea level (m), and t(z) is air mass travel time (s) from the coast to the site at level z. A detailed description of the vertical profile sampling and flux calculation is available in Gatti et al. (2014, 2021) and Basso et al. (2016, 2021)'s studies.

The method to calculate the influence regions uses individual air parcel back trajectories of each vertical profile calculated by the Hybrid Single-Particle Lagrangian Integrated Trajectory Model (HYSPLIT; Stein et al., 2015). This model uses 1-h resolution to get the trajectory density (number of back trajectories passed over a 1 × 1° grid cell), between ~300 and 3,500 m a.s.l. A total of 590 CO₂ profiles (flights) from 2010 to 2018, considering a minimum density of trajectories of 2.5% were used. Detailed methods of the calculation of influence regions and weighted trajectory density are given by Cassol et al. (2020) and Gatti et al. (2021).

The top-down carbon flux estimate integrates the effect of all the surface sources and sinks between the coast and the sampling site, i.e., it is the total carbon flux due to all natural and anthropogenic surface sources and sinks including fire flux (biomass burning emission), fossil fuel emissions, and NBE flux (Net Biome Exchange: total flux minus the fire flux; Gatti et al., 2021).

We combine multiple Landsat-based datasets to represent annual LUCC related to forest carbon change (Table 1). Here, we describe the bottom-up datasets and methods to create the annual LUCC map from 2010 to 2018, including deforestation, secondary forest, and forest degradation processes. We calculated the LUCC areas for the Brazilian Amazon (Figure 1A) and the influence regions of each site (Figures 1B–F). For the forest class, we used the MapBiomas collection 4.1 dataset (MapBiomas, 2020), hereafter referred to MapBiomas.

As the secondary forest class is missing in MapBiomas, we obtained this class from Silva Junior et al. (2020), which is based on the same version of MapBiomas. A secondary forest is defined as a pixel with an anthropogenic class (e.g., agriculture or pasture) that turns into a forest again. The deforestation data were also derived from MapBiomas using an algorithm similar to that of Silva Junior et al. (2020). Deforestation is defined as a forest pixel that was replaced with any other class. For forest degradation, we used the data from Bullock et al. (2020), in which any pixel containing a forest disturbance is considered as forest degradation (e.g., selective logging, fire, and wind damage that did not result in a change in land cover or deforestation at a 0.09-ha scale). Bullock et al. (2020) methods used Landsat spectral endmember fractions to calculate the Normalized Degradation Fraction Index (NDFI). Other LUCC classes, planted forest, pasture, agriculture, and other non-forest natural formation, were obtained from MapBiomas as well. We merged some MapBiomas classes (Table 2) according to the Brazilian Third and Fourth National Communications on greenhouse gasses (3NC and 4NC; MCTI, 2016, 2020a) to the United Nations Framework Convention on Climate Change (UNFCCC) and the Brazilian System for Estimating Greenhouse Gas Emissions 8th (SEEG8) (Albuquerque et al., 2020; Alencar et al., 2020).

Considering that the LUCC dataset employed different objectives and methods for mapping secondary forest (also secondary forest loss), degradation and deforestation classes, there may be some overlap between these classes. For example, one pixel classified as degradation by Bullock et al. (2020) can be classified as a secondary forest by Silva Junior et al. (2020). That is why we made an overlaying analysis between these datasets containing deforestation, degradation, and secondary forest gain and loss for 3 years (2010, 2014, and 2018). The overlaps between these classes are equal to or < 1.8% of the total pixels (see Supplementary Table 1). Thus, we made our annual LUCC maps merging the datasets according to the order shown in Figure 2.

To quantify the CO₂ emissions and removals, different spatial analyses for each annual LUCC class in each influence region and within the Brazilian Amazon were made (Figure 3). In the case of TAB, we used the emission/removal estimates from 2010 to 2012 and for TEF from 2013 to 2018 because the TAB flight site was replaced by TEF in 2013. According to the IPCC (2006), to calculate GHG emissions, you multiply the activity data (our annual areas of each LUCC) by an emission/removal factor (representative value to quantify the emissions or removals). To designate emission/removal factors, we made a separate spatial analysis for each LUCC (Figure 3). The detailed emission/removal factors ratios are described in Table 3.

The carbon stored in the forest and other non-forest natural formation classes also released into the atmosphere by deforestation, degradation, and secondary forest loss is determined by the biomass, assuming a carbon fraction of 0.5 in forest biomass (IPCC, 2006). To quantify the amount of carbon in the forest and the CO₂ emissions by deforestation and degradation, we used the carbon map of the 4NC (Fourth National Communications on greenhouse gasses) of Brazil to the UNFCCC (MCTI, 2020a). The 4NC map is based on field biomass (from the Brazilian Forest National Inventory and other projects) and airborne LiDAR data from the EBA project (EBA, 2016), which was extrapolated from 2016 to 2010 (the base year of the 4NC; MCTI, 2020b). The 4NC carbon map represents the carbon stocks in 2010, including aboveground biomass, belowground biomass, dead wood, and litter, derived from carbon ratios of the aboveground biomass (MCTI, 2020b).

Intact forest carbon stocks are not accounted for in national communications (NCs) on greenhouse gases, which only consider anthropogenic emissions/removals. As carbon stock changes in intact forests are relevant for our study, we used the annual mean carbon change of ground forest biomass plots of RAINFOR (Phillips et al., 2017) from 2000 to 2009 (last decade with data). We used the forest class of the 4NC map only to account for the C stocks. The mean carbon change values for each Amazon biogeographic region [Brazilian Shield (BrSh), southwest Amazon (SW), the central-east Amazon (EC), the Guyana Shield (GuSh), and the central-west (W)] of Feldpausch et al. (2011, 2012) were multiplied by our total intact forest area of each biogeographic region (Table 3).

In the case of deforestation, we assumed that all the CO₂ from the forest carbon pools (aboveground biomass, belowground biomass, dead wood, and litter) is released to the atmosphere at the moment of the clear cut, the forest carbon data are from the 4NC map. To calculate the carbon removal and loss of secondary forest, we used the secondary forest extent and loss data of Silva Junior et al. (2020) and the secondary forest growth regions of Heinrich et al. (2021; Table 3). Furthermore, we assumed that the secondary forest carbon content of the prior year is released into the atmosphere the following year (secondary forest loss class).

For the forest degradation class, we used the 4NC biomass map as a base of forest biomass before degradation, assuming that 45.8% of aboveground carbon remains (54.2% of carbon is lost), according to Rappaport et al. (2018) and Assis et al. (2020). For belowground carbon, dead wood, and litter, see the ratios used in Table 3.

The removal factors (C stocks and annual increment) of other non-forest natural formation classes (which includes the grassland class) and planted forests class were obtained from the SEEG8 (based on the 3NC), which used the same MapBiomas LUCC classes. As we are focusing on the land use, land use change, and forestry (LULUCF) classes, in the carbon emissions or removal analyses, we excluded the classes: others, pasture, agriculture, water bodies, and non-observed. The SEEG8 and the 3NC (MCTI, 2016; Alencar et al., 2020) also do not include these classes. In the bottom-up approach, we are not including specifically the fire fluxes, even though part of fire fluxes is already included in the deforestation and forest degradation processes. Also, energy (fossil fuels), industrial processes, and agriculture sectors are not considered.

Each influence region has a different trajectory density; warmer colors represent the highest densities (Figures 1B–F). The bottom-up CO₂ emissions and removals were calculated using the weight of the trajectories' density. After calculating the CO₂ emissions and removals of the bottom-up approach, we compared them with the top-down CO₂ fluxes.

To compare bottom-up with top-down estimates, we projected bottom-up analyses on the influence regions of each of the atmospheric CO₂ observation sites. The largest influence region is from RBA, covering 95% of the Brazilian Amazon area. TAB is the second largest area (82%), followed by TEF with 56%. RBA and TAB had the largest forest area in 2010 with 82 and 85%, respectively. TEF forest area was 82% of the total area in 2013. ALF (in the southeast) and SAN (in the northeast) have a smaller influence region with 35 and 16%, respectively, and the forest area in 2010 is also smaller (59 and 79% of the total area). All the annual LUCC areas are shown in Figure 4 and Supplementary Table 2, and their percentages in Figure 5. As observed in Figure 1, the influence regions may overlap and, therefore, are not cumulative.

In 2010, the Brazilian Amazon had 82% (3,265,968 km²) of its area covered by forest. Between 2010 and 2018, ~97,066 km² of its forests were lost, an area larger than Portugal (Figure 4), even though it was only 3% of the Brazilian Amazon forested area. In the same period, the forest loss in the TEF region of influence was 43,239 km² (2%), in RBA 94,921 km² (3%), in SAN 29,036 km² (5%), and in ALF 50,031 km² (4%; Figure 5).

The second largest land use class in the Brazilian Amazon is pasture (452,039 km² in 2018; Figures 4, 5). It is also the class for which the total area varied the least. During the period 2010–2018, in the whole Brazilian Amazon, the pasture had increased by 10,455 km²; in RBA, the pasture increased by 9,361 km² (Figure 5); in TAB, pasture remained constant; in TEF, the increase was 11,774 km² reaching 226,998 km² in 2018. In percentage, SAN (23%) and ALF (19%) have almost double the pasture area in 2018 than RBA (11%), TAB (9%), and TEF (11%). Agriculture was very small in TAB (1,123 km²) and SAN (731 km²) in 2010 and TEF (1,659 km²) in 2013. In the Brazilian Amazon and RBA, agriculture areas almost doubled in 2018, reaching 1.4% (25,855 km²). In ALF and SAN, the agricultural area tripled in 2018 compared with 2010, increasing 3,345 and 10,290 km², respectively. Secondary forest is the third largest LUCC class in all the influence regions. In ALF and SAN, the secondary forest percentage is larger than the other sites; these were 7.6% (48,879 km²) in SAN and 4.7% (67,221 km²) in ALF in 2010, reaching 9.3% (59,215 km²) and 5.9% (84,242 km²) in 2018, respectively (Figure 5).

The water class covers an area of 139,878 km², and the other non-forest natural formation class covers an area of 100,477 km². Other non-forest natural formation classes covered 2 and 3% by area in all influence regions except for TEF with 3.9% and SAN with 5.7%. Water class is also ~3%, except for SAN, which is the influence region with the most water at 5.2% (Supplementary Table 2).

The main LUCC classes related to carbon emissions including deforestation, degradation, and secondary forest loss cover a smaller area when compared to other LUCC classes, but they are by far the most important for carbon emissions. In general, secondary forest loss is larger than deforestation and degradation. Nonetheless, in dry years like 2010, 2015, and 2016, degradation emissions are larger than deforestation and secondary forest loss emissions. The area of these three classes increased at all sites, in terms of the percentage of each influence region, and is largest in ALF and SAN (Figure 5).

Forest carbon stocks are smaller near deforested areas. The mean forest carbon stocks for each influence region are shown in Figure 6 and Table 4. The highest carbon density areas are in the Guyana Shield biogeographic region, mainly northeast of TAB, TEF, and RBA, with a mean carbon stock of 316 MgCha⁻¹. The central-east Amazon biogeographic region has a mean carbon stock of 260 MgCha⁻¹, with high C stocks in the west and less in the east where part of ALF and SAN are, due to the high historical deforestation. The west biogeographic region covers the southwest of all influence regions, with lower carbon stocks than the central and north Amazon, with a mean of 202 MgCha⁻¹. The lowest carbon stocks are found in the southwest Amazon biogeographic region, with 172 MgCha⁻¹ in part of Acre State, which also represents the smallest area.

The mean forest carbon stocks per influence region follow the same pattern, higher in the north, and northeast than in the central and eastern Amazon. TAB, TEF, and RBA have mean carbon stock close to 145 MgCha⁻¹. SAN has a small area of high carbon stocks in the northeast, but the rest are forest patches with less carbon density than the rest of the influence regions. ALF has more carbon stocks in the east, but in the southeast, there are forest patches with less carbon density. For both the SAN and ALF influence regions, the mean carbon stock is ~125 MgCha⁻¹ (Table 4).

There are conspicuous differences between the bottom-up and top-down CO₂ fluxes in our study period (2010–2018), in both the Brazilian Amazon and each influence region (Figure 7 and Supplementary Table 3). First, we show the results of the bottom-up emissions and removal estimates considering the total (sum) fluxes from 2010 to 2018, also looking at each year and then the net bottom-up estimates (emissions less removals). In the case of top-down estimates, we show the total flux, fire flux, and NBE flux (total CO₂ estimates minus fire emissions; Figure 7). The top-down fire emissions are not explicitly considered in the bottom-up estimates, but deforestation, degradation, and secondary forest carbon loss emissions partially result from fire. Thus, we indicate the differences between the top-down total CO₂ flux with the net bottom-up CO₂ flux.

Constructing an annual LUCC map from 2010 to 2018 using several datasets was essential to show the relevance of the different LUCC dynamics, including deforestation, forest degradation, and secondary forest growth and loss. The forest is the largest class within the Brazilian Amazon, but anthropogenic activities related to pasture, secondary vegetation, deforestation, and forest degradation areas have a large extent. There are large and constant pasture areas that could be used to expand agriculture in the Brazilian Amazon and are also potential areas for secondary forest growth. Thus, some studies suggest that there is no need to open new deforestation frontiers (Souza et al., 2020).

The influence regions used in this study represent the annual mean air mass back-trajectory densities from all flights from 2010 to 2018, so we showed the annual LUCC in these areas. However, if additional years of atmospheric measurements (e.g., 2019 and 2020) are included, the number of trajectories would increase, and the boundary of each influence region may change. These modifications could impact the calculation of CO₂ fluxes.

Forest degradation, unlike deforestation, has challenges in terms of mapping and monitoring using remote sensing techniques (Costa et al., 2019). As a result, only a few datasets are available for this variable. Consequently, forest degradation is not currently included in NCs or SEEG GHG estimates (Wiltshire et al., 2022). However, our study along with the findings of Assis et al. (2020) indicates that forest degradation areas exceed deforestation and secondary forest loss during dry years (2010, 2015, and 2016). Therefore, we agree with the recommendation of Silva Junior et al. (2021), for the inclusion of forest degradation emissions in NCs.

As expected, the utilization of LUCC datasets employing different methodologies and approaches leads to variations in LUCC extent, consequently resulting in differences in CO₂ bottom-up estimates. Fortunately, LUCC datasets continue to progress. For example, the new MapBiomas collections now include secondary forests, deforestation, and fire. Additionally, TerraClass and IBGE are preparing updated versions including new years. As a result, the accuracy and representation of LUCC will continue to improve.

The northeast Amazon region exhibits areas of high forest carbon stocks. Interestingly, these high-carbon areas also show significant divergences in biomass datasets, including both field measurements and biomass maps (Ometto et al., 2014; Tejada et al., 2020). On the other hand, areas with higher agreement between datasets, such as the southeast Amazon (Brazilian arc of deforestation), are located in regions with lower carbon stocks, closer to previously deforested or degraded forest areas.

The spatial and temporal scales of biomass maps present limitations for estimating CO₂ emissions/removals related to deforestation, degradation, and secondary forest dynamics. Biomass maps represent carbon stocks for specific years; for instance, the 4NC map represents the carbon stocks in 2010 and the Baccini et al. (2012) map represents the carbon stocks in 2000. Unfortunately, the availability of multitemporal (yearly) biomass data is insufficient to assess the carbon changes in secondary forests and degradation. However, with the advent of BIOMASS (Carreiras et al., 2017) and GEDi (Dubayah et al., 2022) new remote sensing missions, designed to have high-resolution, and improved temporal and spatial data, assessing secondary forest and forest degradation annual carbon change will become more feasible.

For the bottom-up estimates, we aggregated several methodologies and approaches to designate the appropriate emission/removals factors for each LUCC class dynamics. Environmental factors influencing the growth of primary and secondary forests have been considered for carbon removal estimations. While primary forest carbon stocks are not considered in NCs (or SEEG), except for protected areas, they play a significant role in carbon sequestration. Therefore, we used the mean carbon change (growth of primary forest) derived from the RAINFOR field data (Phillips et al., 2017), considering the biogeographic regions determined by Feldpausch et al. (2011, 2012). The mean carbon change forest growth, as indicated by Phillips et al. (2017), shows a decline in carbon sink over the decades. We used the last decade of the study (2000–2010), which exhibits a 40% decrease in mean forest carbon change compared with the 1980–1990 period. If this trend continues, we can expect a further decline in mean forest carbon change during the 2010–2020 decade. Regarding secondary forest growth, we considered forest age and the location in the Amazon, based on the regions identified by Heinrich et al. (2021). Considering the great extent of secondary forests, it is crucial to map the dynamics of this LUCC class.

In estimations of bottom-up emissions, the main emission classes are deforestation, forest degradation, and secondary forest loss. These classes have contributed to the Amazon region becoming a carbon source since 2015. It is crucial to include these LUCC classes in estimates of emissions. Forest degradation, despite not being considered in NCs (or SEEG), is the largest emission source during dry years (i.e., 2010, 2015, and 2016). Dry years (mainly 2015 and 2016) have high emissions in the estimates of bottom-up and top-down fluxes. Furthermore, removals of primary forests, except those in protected areas, are also not included in the NCs and SEEG despite being a large carbon sink, having absorbed −5,901 TgCO₂ in the Brazilian Amazon from 2010 to 2018.

Although the bottom-up and the top-down approaches are not completely comparable, this study reveals interesting insights when comparing both approaches. The bottom-up carbon fluxes depend on LUCC input data (activity data) and the emission/removal factors designation; we assumed constant emission factors without including climate variability. The difference in CO₂ flux estimates, according to the LUCC data, is clear when comparing the SEEG and NC estimates (Wiltshire et al., 2022). On the other hand, top-down estimates capture the forest response to climatic conditions during drier or wetter years, as well as the seasonality of atmospheric fluxes. The top-down atmospheric measurements implicitly include sectors of more emissions and removals (e.g., agriculture, industry) than the LULUCF sector fluxes considered in this study. This may explain why the total top-down estimates are higher than the net bottom-up estimates. To better assess the bottom-up net fluxes, it would be pertinent to include specific fire emissions and consider additional sectors such as agriculture. The continuity of the Amazonian atmospheric measurements is essential to understand the carbon balance in the Amazon, as the deforestation and forest degradation processes have been increasing sharply in recent years. Analyzing both bottom-up and top-down approaches will be necessary to reach the climate targets set by the Paris Agreement (Wiltshire et al., 2022).