Main

The Brazilian Legal Amazon (BLA) has the largest tropical rainforest area, the highest biodiversity and the largest amount of aboveground biomass in the world1,2,3. Since 2000, the Indigenous territories (ITs) and the protected areas (PAs) in the BLA have increased substantially and by 2013 they accounted for 43% of the total land area and covered about half of the total forest area in the region1. ITs and PAs have important roles in forest and biodiversity conservation and climate mitigation in the BLA4,5. Many studies have investigated the effects of the ITs/PAs in reducing deforestation and strengthening forest conservation in the BLA1,6,7,8,9,10, but their results differ substantially as different approaches and datasets were used. A comparison of deforestation rates between the PAs and the non-protected areas (non-PAs) showed that the deforestation rates in the PAs were 1.6–2.2 times6 or ten times1,7,8 lower than those in the non-PAs, respectively. PAs are usually located in remote areas with small deforestation pressure (Extended Data Fig. 1), whereas the non-PAs often have high deforestation pressure. The strict-protected PAs, located in more remote areas, have more forest cover and less deforestation pressure in comparison to sustainable-use PAs and ITs (Extended Data Fig. 2). The comparison of deforestation rate between PAs and non-PAs may not fully show the effects of the PAs in reducing deforestation. Spatial matching methods were applied to assess the effects of PAs on deforestation in the BLA9,10; however, due to extensive deforestation and forest degradation11, it is challenging to find locations in the non-PAs with conditions similar to those in each PA to serve as the controls for the spatial matching methods. Various studies have assessed the effects of ITs/PAs using the official Brazilian deforestation dataset (PRODES)6,7,8, which reports deforestation of primary forests based on the reference forest map in the 1980s and does not include secondary forests that are important for the carbon cycle and biodiversity conservation12.

The conflicts between forest conservation and socio-economic development in the BLA persist but vary over the years13,14,15,16. Forest conservation in ITs/PAs has encountered increasing threats from loosened environmental laws and regulations, changing governmental policies and massive economic development17,18, especially after 201219. About 100 × 106 ha, including at least 20% of the area in strict protection PAs and ITs, have applications pending for mineral prospecting or mining operations20,21. The satellite data show that illegal mining in the BLA substantially increased from 2000 to 2012 (an average area of 2,500 ha yr−1) and hit a record high area in 2020 (10,000 ha yr−1) amid widespread protests from Indigenous people22. Primary forest loss reached a decade high in 202023. Between March and September 2020, Brazil passed 27 legislative acts that weakened environmental protection24. Fines for violation of environmental and conservation laws dropped by 72% from March to August 2020, despite an increase in deforestation24. Besides, COVID-19 started to spread widely in the Amazon in early 2020, causing high mortality rates among Indigenous groups25, creating favourable conditions for encroachment of Indigenous lands by illegal loggers and miners.

PAs in the BLA have two types of governance and institution (national versus state PAs) and two types of management objective (strict protection versus sustainable use; Extended Data Figs. 1 and 2). To what degree were deforestation dynamics in the BLA and ITs/PAs under different governance and management affected by the COVID-19 pandemic and changing policies in recent years? Here, we investigate and report on (1) interannual changes in forest areas and deforestation in the BLA, ITs, PAs and non-PAs (that is, non-designated areas) and (2) the effects of ITs, national PAs (nPAs) and state PAs (sPAs) in reducing deforestation from 2000 to 2021 under the changing environmental laws and governance, natural disturbances and COVID-19. We generated annual evergreen forest maps by using daily images acquired by moderate resolution imaging spectroradiometer (MODIS) onboard the Terra satellite from 2000 to 2021 and the forest mapping algorithm1,2.

Results

Increased areas of ITs/PAs for forest conservation

The number and area of ITs/PAs in the BLA increased from 1980 to 2018 with noticeable characteristics (Fig. 1a). The ITs area rose slowly from 0.2 × 106 ha in 1980 to 4 × 106 ha in 1988 but started to increase rapidly after 1988 and reached 74 × 106 ha in 2000 and 115 × 106 ha in 2016. This large and rapid increase in numbers and areas of ITs was driven by the 1988 Constitution’s requirement that the government demarcate all Indigenous lands within 5 years. Although the Constitution’s requirement has yet to be fulfilled, it stimulated substantial expansion of ITs, including 40 × 106 ha demarcated with financial support from the G7 Pilot Program to Conserve the Brazilian Rainforest. In 2000, the Brazil’s National Protected Areas System was officially established to expand PAs and better manage the PAs for forest conservation. The nPAs area increased from 9 × 106 ha in 1980 to 27 × 106 ha in 2000, 62 × 106 ha in 2008 and 66 × 106 ha in 2018. The sPAs area increased from less than 1 × 106 ha in 1980–1989 to 25 × 106 ha in 2000, 56 × 106 ha in 2007 and 60 × 106 ha in 2018. When lumped together, the total area of ITs/PAs substantially expanded from 10 × 106 ha in 1980 to 126 × 106 ha in 2000 and 241 × 106 ha in 2018. The numbers of ITs, nPAs and sPAs also had similar changes over time and substantially increased from 2, 11 and 2 in 1980 to 227, 70 and 95 in 2000 and 387, 146 and 191 in 2018, respectively.

Fig. 1: Cumulative areas and numbers of ITs and PAs and cumulative forest areas in ITs/PAs in the BLA.
figure 1

a, Cumulative areas and numbers of ITs, nPAs and sPAs from 1980 to 2018. ITs data in 2017 and 2018 are not available. b, Cumulative forest areas in ITs, nPAs and sPAs from 2000 to 2018.

As most ITs/PAs are covered by forests, the large increase of ITs/PAs strengthens forest conservation substantially in the BLA. Figure 1b shows the interannual change in forest area in ITs, nPAs and sPAs during 2000–2018. Forest area in ITs increased from 67 × 106 ha in 2000 to 105 × 106 ha in 2016, an increase of 55%. Forest area in nPAs increased from 25 × 106 ha in 2000 to 62 × 106 ha in 2018, an increase of 146%. Forest area in sPAs increased from 18 × 106 ha in 2000 to 50 × 106 ha in 2018, an increase of 169%. We overlaid the 2000 annual forest map (a total of 394 × 106 ha of forest) with the 2018 boundary maps of ITs/PAs and these ITs/PAs in 2018 covered 206 × 106 ha forest, accounting for 52% of the total forest area in 2000, which clearly shows the importance of the ITs/PAs for forest conservation.

Interannual change of forest area during 2000–2021

We used our annual forest maps to quantify the interannual changes in forest area during 2000–2021 (Fig. 2a). The total forest area in the BLA decreased substantially from 394 × 106 ha in 2000 to 366 × 106 ha in 2021, a loss of 28 × 106 ha (~7% of the forest area in 2000 or an area loss larger than Brazil’s state of Rondônia). The 2018 ITs/PAs boundary maps were overlaid on the annual forest maps to calculate the interannual change of forest area in ITs/PAs during 2000–2021. Total forest area in ITs, nPAs and sPAs decreased from 105.7 × 106 ha, 62.6 × 106 ha and 50.5 × 106 ha in 2000 to 104.9 × 106 ha, 62.2 × 106 ha and 50.0 × 106 ha in 2021, respectively, with the average loss rates of 0.04% yr−1, 0.03% yr−1 and 0.10% yr−1 over their forest areas in 2000 (Fig. 2b–d). As some of the ITs/PAs overlap each other, we combined the ITs/PAs together; the total forest area in the combined ITs/PAs decreased slightly from 206.4 × 106 ha in 2000 to 204.9 × 106 ha in 2021, an average annual loss rate of 0.08 × 106 ha yr−1 (0.04% yr−1). To our surprise, ITs/PAs together had a slightly larger forest area in 2018–2021 (204.6 × 106 ha) than in 2014–2017 (204.1 × 106 ha), which may be related to the recovery of forests after severe damage in 2015/2016 and, to a much lesser extent, from tree-planting projects. In comparison, the forest area in the non-PAs (Fig. 2e) decreased from 187.8 × 106 ha in 2000 to 161.3 × 106 ha in 2021, with an average annual loss rate of 1.3 × 106 ha yr−1 (0.7% yr−1), about 14 times more than the ITs/PAs. The loss of 27 × 106 ha forest area during 2000–2021 in the non-PAs accounted for ~95% of the total loss of forest area in the BLA over the same period.

Fig. 2: Interannual changes of forest areas in the BLA from 2000 to 2021.
figure 2

a–e, Forest areas in the BLA (a), ITs (b), nPAs (c), sPAs (d) and non-PAs (e). f, Forest area map in 2000. g–i, Forest area change rates in ITs (g), nPAs (h) and sPAs (i). Forest area change rates are calculated from annual forest areas between 2000 and 2021 on the basis of linear regression analysis at the 90% confidence level.

Geographically, the interannual change of forest area in ITs, nPAs and sPAs from 2000 to 2021 had noticeable spatiotemporal patterns. The trend analysis of forest areas (Fig. 2f–i) showed divergent dynamics among ITs/PAs. ITs/PAs in the southern and eastern portions of the BLA (the ‘arc of deforestation’) had the largest losses of forest area. The ITs/PAs in the northern portion of the BLA either had no significant change in forest area or had increased forest area. In total, 39.8% of nPAs (59 out of 146 nPAs), 40.4% of ITs (154 out of 387 ITs) and 44.0% of sPAs (84 out of 191 sPAs) had significant forest area loss from 2000 to 2021 (P < 0.1).

Spatiotemporal dynamics of primary forest area loss

We used the 2001 forest map as the reference map and identified the first year the forest pixels in 2001 were classified as non-forest pixels during 2002–2021 (Fig. 3a,b) and we counted the number of pixels with a change from forest to non-forest (forest loss) in a year as annual gross forest area loss over the BLA, ITs, PAs and non-PAs from 2002 to 2021 (Fig. 3c–g). The cumulative gross forest area losses during 2002–2021 were 49 × 106 ha for the BLA, including 2.1 × 106 ha for ITs, 1.2 × 106 ha for nPAs, 2.8 × 106 ha for sPAs and 43.1 × 106 ha for non-PAs. The combined ITs/PAs had a 5.9 × 106 ha gross forest area loss from 2002 to 2021, accounting for ~12% of total gross forest area loss in the BLA, which clearly indicates the critical role of the ITs/PAs in forest conservation.

Fig. 3: Annual gross forest area losses in the BLA from 2002 to 2021.
figure 3

a, Spatial distribution of the year a forest pixel in 2001 was classified from forest to non-forest (forest loss) in the ITs/PAs. b, Spatial distribution of the year a forest pixel in 2001 was classified from forest to non-forest (forest loss) outside ITs/PAs. c–g, Annual gross forest area loss in the BLA (c), ITs (d), nPAs (e), sPAs (f) and non-PAs (g). We used the 2001 forest map as the reference and identified the first year that forest pixels became non-forest pixels during 2002–2021. The dark grey pixels in a and b are forest in 2001.

The interannual change of gross forest area loss in the BLA (Fig. 3c) during 2002–2021 reveals three interesting results. First, annual gross forest area losses in 2005, 2007, 2010 and 2015 were substantially larger than those in previous and subsequent years. Years 2005, 2007, 2010 and 2015 were characterized by strong El Niño events (Extended Data Fig. 3), high air temperature or severe drought26. Second, annual gross forest area loss in 2013 was least during 2000–2021. Year 2013 was a year with high air temperature without an El Niño or Atlantic Multidecadal Oscillation event. Third, when those years (2005, 2007, 2010, 2013 and 2015) are taken out, the temporal dynamics of annual gross forest area loss show five noticeable phases and each of these phases lasted a few years: (1) increased forest area loss in 2002–2004, (2) reduced forest loss in 2006–2009, (3) increased forest area loss in 2011–2014, (4) decreased forest area loss in 2016–2018 and (5) increased forest area loss in 2019–2021. In 2018–2021, the annual gross forest area loss rates increased 3.6 times in ITs/PAs, larger than the increase in non-PAs (1.6 times), indicating increasing deforestation pressure and an alarming signal in ITs/PAs.

Varying effects of ITs/PAs on annual forest area loss

To investigate the effect of ITs/PAs on reducing forest area loss after they were established, we selected those ITs/PAs that were established after 2002 and analysed the average annual gross forest area loss rates before and after their establishment years (Fig. 4). The results show that 72 ITs, 32 nPAs and 38 sPAs had substantially reduced annual forest area loss rates but 49 ITs, 21 nPAs and 27 sPAs still had small to moderate increases in annual forest area loss rates; the other 24 ITs, 2 nPAs and 7 sPAs had no change in forest area loss rates as they had little or no forest area loss (Fig. 4a–c). There was no clear geographical cluster among those ITs/PAs with reduced or increased forest loss rates (Fig. 4d–f). Most of the sPAs in the northern and western BLA had small increases in annual forest area loss rates. In terms of institution and governance, the average annual gross forest area loss rates were reduced substantially for the nPAs (36%, 13.7 × 103 ha) and the ITs (30%, 10.7 × 103 ha) but only slightly for the sPAs (5%, 2.7 × 103 ha; Fig. 4g and Extended Data Fig. 4a). In terms of management objectives, the average annual gross forest area loss rates were reduced substantially for the strict-protection PAs (48%, 8.6 × 103 ha) and the ITs (30%, 10.7 × 103 ha) but only moderately for the sustainable-use PAs (11%, 7.7 × 103 ha; Fig. 4h and Extended Data Fig. 4b).

Fig. 4: The effects of ITs/PAs on annual forest area loss rates in the BLA (2001–2021).
figure 4

a–c, Comparison of average annual gross forest area loss rates before and after ITs/PAs establishment (after the year 2002) by ITs (a), nPAs (b) and sPAs (c). d–f, Spatial distribution of the changes in annual gross forest area loss rates before and after ITs/PAs establishment by ITs (d), nPAs (e) and sPAs (f). g,h, The changes in average annual gross forest loss rates before and after the ITs/PAs establishment by governance (nPAs, ITs and sPAs) (g) and management (strict protection and sustainable use) (h). All ITs/PAs established before and during 2002 are white polygons.

Different deforestation dynamics inside different PA types

Forest area in nPAs with the strict-protection objective (Fig. 5a) varied slightly during 2000–2013 but decreased moderately in 2014–2016, with a net loss of 0.14 × 106 ha (0.45%) from 2000 to 2021. Forest area in the nPAs with the sustainable-use objective (Fig. 5a) decreased continuously from 2000 (31.7 × 106 ha) to 2021 (31.5 × 106 ha), a net loss of 0.22 × 106 ha (0.7%). Forest area in sPAs with the strict-protection objective (Fig. 5b) decreased from 5.91 × 106 ha in 2000 to 5.88 × 106 ha in 2021, a net loss of 0.03 × 106 ha (0.51%). Forest area in sPAs with the sustainable-use objective (Fig. 5b) decreased from 44.65 × 106 ha in 2000 to 43.53 × 106 ha in 2016, a net loss of 1.12 × 106 ha (2.5%) but it had a modest recovery by 2021 (44.19 × 106 ha), thus a net loss of 0.46 × 106 ha (1.0%) from 2000 to 2021.

Fig. 5: Annual forest areas and gross forest area loss rates in PAs with different governance and management regimes from 2000 to 2021.
figure 5

a, Annual forest areas in nPAs (strict protection, left Y-axis, and sustainable use, right Y-axis). b, Annual forest areas in sPAs (strict protection, left Y-axis, and sustainable use, right Y-axis). c, Area proportion of annual gross forest area loss to forest area in 2001 in nPAs (strict protection and sustainable use). d, Area proportion of annual gross forest area loss to forest area in 2001 in sPAs (strict protection and sustainable use).

Differences in annual gross forest area loss rates were small and not significant between the PAs with different governance and management (Fig. 5c,d). The sPAs with the sustainable-use objective had the highest average gross forest area loss rate (0.31 ± 0.14% yr−1 of the forest area in 2001), followed by nPAs with the strict-protection objective (0.11 ± 0.05% yr−1) and nPAs with the sustainable-use objective (0.09 ± 0.04% yr−1). The sPAs with the strict-protection objective had the lowest average gross forest area loss rate (0.04 ± 0.04% yr−1) due to these PAs being far away from the ‘arc of deforestation’ and therefore were under little deforestation pressure (Fig. 4f). The gross forest area loss rates increased from 2018 to 2021; these losses were probably related to the loosened forest conservation policies during the Bolsonaro presidential administration that began in January 201917,18,27. Gross forest area loss from 2018 to 2021 increased 1.5 times in nPAs with the strict-protection objective and 5 times in nPAs with the sustainable-use objective. Gross forest area loss from 2018 to 2021 increased 12.4 times in the sPAs with the strict-protection objective and 4.3 times in sPAs with the sustainable-use objective.

Discussion

Interannual change of forest area and deforestation

Annual forest maps from our mapping tools, PRODES28 and Global Forest Watch (GFW)29 were evaluated for South America30 and the BLA1, showing higher accuracy in our annual forest maps. The interannual changes of forest area and deforestation during 2000–2017 from these datasets were reported1. Here, we extended the data record from 2017 to 2021 (Fig. 6). We compared our forest area with the newly developed MapBiomas31 forest area and they had similar interannual change trends in forest area in the BLA (Extended Data Figs. 5 and 6). During 2000–2020, the average annual forest area from our dataset is ~4% lower than the MapBiomas and our forest area declined by 8% (32.7 × 106 ha), close to the MapBiomas (7%, 27.9 × 106 ha). As the MapBiomas dataset has not provided annual gross forest area loss data to the public yet, it was not used for detailed analysis in this study. Annual gross forest area loss from our dataset was 1.0 × 106 ha in 2018 and increased to 2.9 × 106 ha in 2021 (Fig. 6a). Annual gross forest area loss from the PRODES deforestation dataset was 0.8 × 106 ha in 2018 and increased to 1.3 × 106 ha in 2021 (Fig. 6b). Compared to the PRODES dataset, annual gross forest area loss from our dataset was higher but had a similar temporal trend. Our previous study explained the differences in forest area loss estimates between the PRODES dataset and our forest map, using annual forest maps in the BLA from 2002 to 20161. Annual gross forest area losses from the GFW dataset in 2018–2020 were slightly higher than those from our dataset (Fig. 6c). Annual forest area loss from the GFW dataset has been modified in numerous ways and these forest area loss data may not be comparable in those years32. Despite their differences in forest definition, in the satellite image data used and, in the forest mapping algorithms, all three datasets report that annual forest area loss in the BLA increased from 2018 to 2021. Together these results provide strong satellite-based evidence of increased forest area loss in 2020/2021.

Fig. 6: Comparison of annual gross forest area loss rates from this study, the official Brazilian deforestation dataset PRODES and GFW in the BLA from 2002 to 2021.
figure 6

a, Annual gross forest area loss rates from this study. b, Annual gross forest area loss rates from PRODES. c, Annual gross forest area loss rates from GFW.

Effects of ITs/PAs on forest conservation

Several studies analysed the PRODES deforestation datasets and assessed the effects of the ITs/PAs on forest conservation under changing laws, policies and climate change9,33,34. One study found that during 2000–2008 ITs/PAs in the states outside of the ‘arc of deforestation’ had little impact on deforestation within their boundaries but ITs and nPAs in the ‘arc of deforestation’ were more effective in reducing deforestation than were sPAs9. Another study reported that 91 sPAs established between 2005 and 2016 reduced deforestation both within their boundaries and in their adjacent surroundings during 2005–201733. The third study also reported that ITs/PAs reduced deforestation during 2000–2010 in the BLA34. The PAs with the strict-protection objective reduced deforestation more than PAs with the sustainable-use objective and the ITs were particularly effective at reducing deforestation in locations with high deforestation pressure34.

Our study uses a longer (2000–2021) and updated dataset to assess the effects of the ITs/PAs for forest conservation in the BLA. Our results over the 2000–2013 period agree with the findings from these previous publications and showed that ITs/PAs reduced deforestation within their boundaries. PAs with the strict-protection objective had small forest area losses in 2000–2013 but large forest area losses in 2013–2021 (Fig. 5). ITs/PAs in the states outside of the ‘arc of deforestation’ reduced deforestation within their borders (Fig. 4), which differs from that of ref. 9, possibly due to the different study periods or to limited data availability in PRODES caused by cloud cover1. Deforestation fronts also reached or encroached into some ITs/PAs (Fig. 3a and Extended Data Fig. 7). Those PAs with large deforestation areas were more likely to experience downgrading, downsizing and degazettement35. To avoid deforestation and reconcile the conflicts between forest conservation and socio-economic development, both intensive agriculture production in the deforested areas and forest restoration projects need to be fully explored. While these measures are important for environmental quality and for generating employment, they cannot be expected to have a ‘land sparing’ effect in reducing gross deforestation rates19. PAs with the sustainable-use objective, where small resident populations have low-impact uses of natural resources36, are designed to promote conservation. They have been demonstrated to offer a win–win solution for biodiversity conservation and socio-economic development37; however, substantial deviations from the intended low-impact uses have sometimes occurred38.

Investigation of various driving factors and development of models that predict the effects of ITs/PAs on reducing deforestation is important for stake-holders and decision-makers. Several studies identified a few driving factors (for example, Indigenous land rights, institutional context and property rights of ITs) and different methods (spatial matching and regression discontinuity model)39,40,41,42. Our exploratory data analyses (Supplementary Information) reveal the importance and limitation of the governance (ITs, nPAs and sPAs), management (strict protection and sustainable use), sizes and locations of the ITs/PAs and forest areas within the ITs/PAs for reducing deforestation. Our preliminary study also highlights the need for casual analysis and more efforts by the research community, stake-holders and decision-makers, more data collection (for example, politics43, social-economic conditions and management practices in individual ITs/PAs), integrated analytics and models across multiple spatial and temporal scales.

Challenges for forest conservation in ITs/PAs

The changes in laws, policies, agriculture (soybean and beef cattle) and climate strongly affect forest conservation44,45 and recent changes in those factors suggest that forest conservation in the ITs/PAs and non-PAs could face increasing challenges in the coming years. As shown in Extended Data Fig. 8, the Action Plan for the Prevention and Control of Deforestation in the Legal Amazon (PPCDAm)23,46, which reduced deforestation, had been implemented since 2004 but was interrupted in 2019 by the Bolsonaro’s administration. The Forest Code and the credit restrictions to owners of deforested lands implemented by the Brazil’s Central Bank as of 200819, which also reduced deforestation, had also been weaken since 201247. The Bolsonaro’s administration is well known for its proximity to agribusiness18,48 and since 2019 has implemented various measures that weaken forest policies and laws and impede their enforcement. In 2021, there was an even stronger pushback against the Brazilian legal framework governing the PAs: five draft bills (PL 490/2007, PL 191/2020, PL 2633/2020, PL 2159/2021 and PLS 510/2021) would further loosen constraints to the economic activities in the ITs, reduce government authority over PAs and provide incentives for agricultural expansion in these areas. The Brazilian Supreme Court is currently trying a case (RE 1017365) that could restrict the constitutional provisions that favoured the demarcation of Indigenous lands and could result in only areas effectively possessed by Indigenous peoples in 1988 being granted demarcation as Indigenous lands. In addition, as the prices of soybeans and beef increased substantially in 2021 (Extended Data Fig. 9), more deforestation for agriculture expansion is likely to continue.

Our study also reveals that annual gross forest area loss rates were high in the El Niño and tropical Atlantic dipole drought years, such as 2005, 2007, 2010 and 2015 (Fig. 3), which may be related to the extensive tree mortality caused by the drought–fire interactions45. Over the period 2002–2021, a total of 10.4 × 106 ha forest was burned in ITs/PAs and 55.1 × 106 ha forest in non-PAs and out of these burned forest areas deforestation occurred in 31% for ITs/PAs and 48% for non-PAs (Extended Data Fig. 10). As the Amazon is projected to have more frequent and severe droughts in the future49, how to manage and protect the remaining intact forests in the ITs/PAs continues to be a major concern.

Summary

The results of this study confirm the critical role of ITs/PAs in forest conservation and raise serious concerns about increased deforestation in the BLA in 2018–2021, especially in the ITs/PAs. The large infrastructure projects50 and the severe COVID-19 pandemic could further increase deforestation and forest degradation in the ITs/PAs. Both deforestation and forest degradation had severe potential impacts on biodiversity and carbon stock in the BLA2,51,52. More attention is urgently needed to strengthen the environmental policies and laws, uphold the existing legal protections and resist the changes being staged in Brazil’s National Congress. More investments from the Brazilian government, private companies and international organizations for the expansion and management of ITs/PAs are also critically needed53. Further reducing deforestation and forest degradation and supporting forest conservation and Indigenous people could prevent passing the tipping point for the Amazon forest ecosystems to flip into savanna ecosystems54.

Methods

Annual forest cover data from MODIS forest mapping tool

We published annual forest maps in the BLA during 2000–2017, which were generated by using the time-series MOD09A1 data product (8 d temporal resolution, 500 m spatial resolution) and the forest mapping algorithm1,2,55. The MOD09A1 8 d composite selects the best-quality observation within each 8 d period. We used cloud-free MOD09A1 observations based on the quality layer. The forest mapping algorithm is mainly based on the unique features of evergreen forest in terms of vegetation greenness, land surface water content and phenology1,2,55, specifically, all good-quality observations for an evergreen forest pixel in a year have the enhanced vegetation index (EVI) ≥ 0.2 and the land surface water index (LSWI) ≥ 0. Our annual MODIS forest maps in the BLA had high overall accuracy (>97%) when they were validated by three independent ground reference datasets1: (1) 18 5 × 5 km2 sample blocks from the Global Land Cover Validation Reference Dataset, which was produced from analyses of very high spatial resolution images at 2 m spatial resolution (a total of 1,268 pixels at the 500 m spatial resolution); (2) 416 10 × 10 km2 sample blocks from the TREES-3 forest and non-forest dataset at 30 m spatial resolution (a total of 262,514 pixels at the 500 m spatial resolution); and (3) 1,991 stratified random sample pixels at 500 m spatial resolution for forest changes generated by visual interpretation of time-series Landsat-5 TM, Landsat 7 ETM+ and Landsat 8 OLI images at 30 m spatial resolution. Our annual forest maps have area and spatial distribution similar to the forest maps derived from microwave images1, which are less affected by frequent clouds in the Amazon.

In this study, we made a minor improvement to the forest mapping algorithm, as a few pixels were contaminated by clouds or aerosols but may not be detected by the quality layer, which resulted in the EVI values dropping substantially although LSWI and the normalized difference vegetation index (NDVI) did not change much. Thus, we added another criterion to identify evergreen forest, that is, we classified a pixel as evergreen forest when it had >90% cloud-free observations that met the criteria of EVI ≥ 0.2 and LSWI ≥ 0 and had annual minimum LSWI ≥ 0. We analysed MOD091 data from February 2000 to December 2021 and we generated the annual evergreen forest maps from 2000 to 2021. We applied a 3 yr consistency check procedure to reduce the potential error in the annual maps of evergreen forest. We used the annual evergreen forest maps to generate two reports. One report is on forest area by year. In this report, we counted primary forest and secondary forest together as forest (no separation into these two categories). The second report is on ‘primary forest’ (used 2001 as the reference year) and primary forest area loss. Here, we tracked which year deforestation first occurred for individual pixels. It is possible that some of forest pixels in 2001 were ‘reforested or recovered forest’. As we have no data before 2000, we do not know how many pixels there were for this case, thus we kept this caveat during our data analysis and result interpretation.

Annual forest area data from the MapBiomas project

The MapBiomas project was launched in 2015 and generates annual land-cover and land-use maps in Brazil31. The algorithm theoretical basis document (ATBD) of MapBiomas presents the cross-reference of the MapBiomas land-cover and land-use classes with classes from other classification systems, including Food and Agriculture Organization of the United Nations (FAO), Brazilian Institute of Geography and Statistics (IBGE) and National Greenhouse Gas Emissions Inventory. The MapBiomas project (Collection 6) uses six steps to generate the annual land-cover and land-use maps by analyses of surface reflectance data from Landsat Thematic Mapper (TM), Enhanced Thematic Mapper Plus (ETM+) and the Operational Land Imager and Thermal Infrared Sensor (OLI-TIRS). The first step is to generate annual Landsat mosaics for specific temporal windows. The second step is to derive the spectral and temporal attributes from Landsat spectral bands to train the Random Forest classifier. The third step is to generate annual land-cover and land-use maps in each biome and cross-cutting theme using the random Forest algorithm (the classification of aquaculture, mining, irrigation, rice and citrus is based on the U-Net convolutional neural network classifier) and training samples. The fourth step is to apply spatiotemporal filters to reduce noise, including gap fill, spatial filter, a 3–5 yr temporal filter, frequency filter and incident filter. The fifth step is to merge the filtered land-cover and land-use maps of each biome and cross-cutting themes and apply the spatiotemporal filters again. The sixth step is the accuracy assessment based on the 75,000 independent samples per year from 1985 to 2018. At the level-1 (forest, non-forest natural formation, farming, non vegetation area and water) classes, the land-cover and land-use maps have 91% global accuracy. We used the level-1 class of forest from the MapBiomas land-cover and land-use maps in this study. For more information about the MapBiomas project see https://mapbiomas-br-site.s3.amazonaws.com/Metodologia/ATBD_Collection_6_v1_January_2022.pdf.

Annual forest cover loss data from the GFW

The 30 m GFW (v.1.8) annual forest cover loss in 2001–2020 was generated by using the decision-tree algorithms and time-series Landsat images acquired during the growing season29. In terms of the year in which a pixel experienced forest loss, the data producers reported that they are 75% confident that the forest loss occurred within the stated year and 97% confident that it occurred within the period from 1 yr before to 1 yr after the stated year.

Annual deforestation data from PRODES

The PRODES dataset28 is generated by Brazil’s National Institute for Space Research (INPE) and is the official Brazilian deforestation dataset. PRODES selected high spatial resolution images (tens of metres) from Landsat-5/7/8, China–Brazil Earth Resources Satellite (CBERS-2/2B), Indian Remote Sensing Satellites (IRS-1) and United Kingdom-Disaster Monitoring Constellation-2 (UK-DMC-2) with two criteria: (1) acquisition dates close to the reference date (1 August) and (2) cloud cover as low as possible. Both visual interpretation and digital image classification were used to identify forest and non-forest pixels.

Active fire data during 2002–2021

We used the MODIS active fire data product: MODIS/Terra Thermal Anomalies/Fire 8-Day L3 Global 1 km SIN Grid. We only used nominal and high-confidence observations.

The ITs and PAs data

We downloaded the data for ITs and PAs from the Amazonian Network for Socio-environmental Information (RAISG)56. The dataset covers the 1980–2018 period and its attributes include the boundary maps (shapefiles) of individual ITs/PAs, names, years of establishment, governance (by national or state government agency) and management objectives (strict protection or sustainable use; Extended Data Figs. 1 and 2). There are 387 ITs, 146 nPAs and 191 sPAs within the BLA in this study, including the establishment of 145 ITs, 55 nPAs and 72 sPAs after 2002.

The boundary map of BLA

BLA includes nine states: Amazonas, Pará, Mato Grosso, Amapá, Roraima, Acre, Rondônia, Tocantins and Maranhão.

Geospatial data analyses

We used the resultant annual evergreen forest maps, the boundary maps for the BLA, individual ITs/PAs in ArcGIS software to calculate annual forest area by the BLA, ITs and PAs. We also used the evergreen forest map for 2001 as the reference map and calculated the annual gross forest area loss from 2002 to 2021. We grouped individual PAs first by governance—(1) nPAs and (2) sPAs—and second by management objectives—(1) strict protection and (2) sustainable use. We then compared the forest area and deforestation dynamics among these four types of PA.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.