Deforestation and agricultural fires in South-West Pará, Brazil, under political changes from 2014 to 2020

ABSTRACT The increasing deforestation and fires since 2019 raises concerns about the irreversible destruction of the Brazilian Amazon. Our goal was to better understand these changes in south-west Pará across different land-tenure and farm systems and between the terms of President Rousseff, Temer, and Bolsonaro. We reconstructed deforestation and fire history using all Landsat and Sentinel-2 observations from 2014 to 2020 and assessed, using quasi-experimental methods, the average treatment effects of each presidency on deforestation and fires across land-tenure and farm types. Deforestation nearly quadrupled to 1,201 km2, particularly during Bolsonaro in undesignated areas and conservation units and on medium-sized farms (p < 0.001). Burning increased to 4,805 km2 and in all tenure types (p < 0.001). The increase was strongest in agrarian settlements and conservation units and on medium and large farms. Our observations show the importance of clarifying land-tenure and re-strengthening disincentives of environmental infractions, which have been weakened specifically under President Bolsonaro.


Introduction
Brazil is the world's leading exporter of soy and beef (FAO, 2021;Rajao et al., 2020;Zalles et al., 2019).These commodities are often produced at the expense of forests in the Amazon and the Cerrado biomes (Barretto et al., 2013;Seymour & Harris, 2019).Between 2001 and 2013, around 75% of deforestation in the Brazilian Legal Amazon (BLA) was to establish pastures (Feltran-Barbieri & Feres, 2021;Tyukavina et al., 2017).Fires for deforestation and pasture maintenance in these land-use systems are a significant source of greenhouse-gas emissions and air pollution and increase the risk of soil erosion and soil nutrient loss (Aragão et al., 2018;Durrer et al., 2021;Figueiredo et al., 2020).Burning activities for management purposes are the most frequent ignition source for uncontrolled forest fires and lead to severe ecosystem degradation and losses in biodiversity (Aragao & Shimabukuro, 2010;Cano-Crespo et al., 2015).
CONTACT Benjamin Jakimow benjamin.jakimow@geo.hu-berlin.deGeography Department, Unter den Linden 6, Humboldt-Universität zu Berlin 10099, Germany This article has been republished with minor changes.These changes do not impact the academic content of the article.
Since 2019, the deforestation rates in the BLA surpassed all yearly rates since 2009 (INPE, 2021a).Similarly, fire use accelerated raising international concerns about the irreversible destruction of the Amazon forest (Barlow et al., 2020).These trends are largely attributed to agricultural lobbying in Brazilian politics that gained momentum with the impeachment of President Rousseff in 2016 (Fearnside, 2016) and more so after the inauguration of President Bolsonaro in 2019 (Barbosa et al., 2021;Pelicice & Castello, 2021).
What remains unclear, however, is how different types of farmers responded to related changes in environmental legislation and law enforcement (as outlined in section 2 in detail) in their management of farmland.Small farmers often operate under economic limitations to comply with environmental regulations or to access credits for technological investments.Yet, although small farms show high deforestation rates per property, the majority (54%) of potentially illegal deforestation occurs on larger farms (Fitz, 2018;Garrett & Rausch, 2015;P. Pacheco, 2009;Rajao et al., 2020;Sobreiro Filho et al., 2015).Indeed, only 2% of all properties in the Amazon and Cerrado are responsible for 62% of all potentially illegal deforestation.Moreover, at least 17% of beef exports from both biomes to the EU may be contaminated with illegal deforestation (Rajao et al., 2020).This being said, many studies report an increase in fires in the Amazon, but it is largely unknown how these differ between landtenure and farm sizes.
Our goal here was to better understand how political changes in Brazil affected land-use change and the use of agricultural fires in a typical agricultural frontier: the region of Novo Progresso in south-west Pará.Pará is the third largest meat producer in Brazil by cattle heads (IBGE, 2019) and has a total area of more than 20 million ha of pasture (C.M. Souza et al., 2020)-in other words, harboring extensive livestock systems with a stocking rate of approximately 1 AU/ha.South-west Pará features one of the Brazil's highest deforestation rates and became a priority region for increased monitoring and control to stop illegal deforestation and other violations of environmental regulations (Assunção & Rocha, 2019;Klingler et al., 2017;Nepstad et al., 2014).
Specifically, we asked the following questions: (1) How did trends of deforestation and the fire use change during the past three Brazilian presidencies?(2) How did these trends differ between different land-tenure zones?
(3) How did these trends differ between different farm sizes?

Changing politics
Since 1965, the Brazilian Forest Code (FC) is the central law to mitigate the unruled expansion of agriculture into native vegetation in Brazil.Its two essential requirements are that landowners must maintain a Legal Reserve (LR) of natural vegetation on, in the Amazon biome, 80% of their property and protect sensitive Areas of Permanent Protection (APP), e.g.around hilltops and rivers to avoid erosion and water pollution (Soares-Filho et al., 2014).These requirements were weakly enforced, and deforestation rates continuously raised in the BLA, showing the highest rates in the federal states of Mato Grosso and Pará.Acknowledging the need to protect the Amazon rainforest to counteract dangerous global change, a set of important programs, combatting deforestation, was initiated during the first presidency of Lula de Silva (2003Silva ( -2010)), e.g. the Action Plan for Prevention and Control of Deforestation in the Amazon (PPCDAm) and the Low Carbon Agriculture Plan (ABC Plan).Deforestation rates in the BLA were reduced by 83% from 27,772 km 2 in 2004 to 4,571 km 2 in 2012, which were the lowest rates observed since the start of the Brazilian deforestation monitoring program PRODES in 1988(INPE, 2021b;Nepstad et al., 2014).
Legal insecurities, unclear rules of vegetation restoration on properties -not conformant with the FC -and demands for social justice to support small farms led to a set of new instruments for environmental regulation (Table 1) and the revision of the FC under President Dilma Rousseff in 2012 (Brancalion et al., 2016;Soares-Filho et al., 2014).With the impeachment of Dilma Rousseff in Table 1.Important policies that affect or regulate land use in the Brazilian Amazon during the presidencies Rousseff, Temer, and Bolsonaro.D = Decree, PEC = constitutional amendment proposal, R = resolution, MP = executive order (Azevedo-Ramos et al., 2020;Brancalion et al., 2016;Caetano, 2021;Pelicice & Castello, 2021;Soares-Filho et al., 2014) May 2016, Michel Temer became the acting President.His presidency was characterized by a growing influence of the agribusiness ('ruralistas') and mining sector and came alongside with different law proposals trying to facilitate large infrastructure projects, i.e. highways and hydropower projects, by weakening environmental protection standards (Fearnside, 2016).Furthermore, the constitutional amendment 241/2016 drastically capped public expenditures for environmental protection and law enforcement.
The election campaign of Jair Bolsonaro in 2018 and his inauguration as President in January 2019 increased the polarization of the Brazilian society and was followed by a 'Tsunami' of policies and decisions that weakened environmental protection (Barbosa et al., 2021;Pelicice & Castello, 2021;Vale et al., 2021).As a result, in August 2019, satellites observed increasing numbers of fires in the Amazon and raised international concerns up to high-level diplomatic disputes around the G12 summit (Andrade, 2019).In 2020, the number of land conflicts reached its highest value (1,576 cases) since the beginning of case recordings in 1985 (Ennes, 2021), and 10,851 km 2 of forest was lost -the highest deforestation rate since 2008 (INPE, 2021a).The forced impotence of the environmental protection system became remarkably visible in our study area, where local producers announced 10 August 2019 as 'Day of fire' to clear pastures and forests 'as much as they can' and show the president their 'will to work' (Caetano, 2021;Piran, 2019).

Study region
Our study region comprises a corridor of 150,000 km 2 (Figure 1a+b) in southern Pará, along the federal highway BR-163 and around the municipality of Novo Progresso.The region features the following land-tenure types: undesignated lands, public lands like conservation units, indigenous territories, and military areas, agrarian settlements that are registered by the Institute for Colonization and Agrarian Reform (INCRA), and private land.
INCRA created nine agrarian settlements, which are either 'traditional' settlements for settlers with a focus on agriculture, specifically cattle ranching, or 'environmentally distinctive' projects, providing land to traditional populations while aiming at low-deforestation rates and sustainable forest management (S1).According to land-tenure data (IPAM, Stabile et al., 2022), most farms (Figure 1c) exist in undesignated areas (63%), followed by agrarian settlements (19%) and conservation units (10%) with low-protection status.The size of a single fiscal unit (FU) ranges in our study region from 75 to 100 ha (Law 8.629/1993).Categorized by their FU into 'small' (≤4 FU), 'medium' (>4 to 15 FU), and 'large' (>15 FU) properties (INCRA 2021, and S2), most are considered as 'small' and eligible to register as family farms or rural family enterprises (Decree 9.064/2017).

Mapping deforestation and burned areas
We based our mapping procedure on all available Landsat 7 + 8 (Collection 1, L1T, EROS, 2020) and Sentinel 2 A+B observations for the study region, with a cloud coverage less than 80%, between 1 January 2014 and 31 October 2020 (S3).All imagery was downloaded and preprocessed to analyze ready data with FORCE (Frantz, 2019).Based on this workflow (Figure 2), we obtained a spatially and spectrally harmonized time series of bottom of atmosphere reflectance images with 30 m spatial resolution and six spectral bands.The FORCE quality assessment images were used to derive a water mask for each year, by masking pixels that had been flagged as water in at least five observations of the same year.Depending on the year and region, our time series consisted of zero to 38 cloud-free observations per pixel in 2014 (median 14, Landsat only) and up to 98 in 2018 (median 35, Landsat + Sentinel-2, S4).
From our time series, we calculated six spectral indices which are known to be sensitive to changes in vegetation cover, burned areas, soil, and soil moisture.Once available, we used the Clear Observation Sequence (COS, Jakimow et al., 2018) approach to classify the time series of spectral indices: First, we calculated for each observation t and pixel a Clear Observation Sequence (COS).This is non-clouded and non-shadowed, i.e. the 'clear' sequence t a ; t; t b f g, with t a and t b being the closest 'clear' predecessor and successor of t.For each COS, we extracted the spectral indices and stacked them into a COS feature stack (S4).
We then visually identified reference polygons for 'natural vegetation' (primary and secondary forests and dense shrublands), 'agriculture' (pastures and cropland), 'tilled land', 'burned areas', 'deforestation', and 'others' (e.g.infrastructure and open water).Additionally, each polygon was labeled with a range of dates, in which the polygon class became visible in the Landsat and Sentinel-2 time series.Based on the reference polygons, we extracted a reference database of >300,000 pixels from the COS feature stacks.For each class, we clustered the available reference pixels using K-Means clustering and prepared a stratified random sampling of 2,500 training pixels per class (15,000 labeled pixels in total), which we used to train a Random Forest (RF).The RF was applied to the entire time series of COS feature stacks used to derive the class probabilities for each observation.We aggregated these class probabilities into annual scores for a total of 7 years and six classes (Jakimow et al., 2018, S4).
The annual scores were input to a rule-based mapping that, for example, intersects the burned area scores with other land-cover scores, to create a burned area and land cover map (BLCM) with six final classes for each year (2014-2020): natural vegetation; not-burned deforested and burned deforested land; not-burned agricultural land and burned agricultural land; and other land covers (e.g.mining, settlements, infrastructure, and water).We then applied a minimum mapping unit of 1 ha for the two agriculture classes and others.Deforested areas were only allowed to occur in primary forests, i.e. forests never mapped as deforested in Griffiths et al. (2018) or a previous BLCM.Removal of secondary natural vegetation was labeled as agriculture or others.We applied a minimum mapping unit of 3 ha to reduce commission errors.Finally, we assessed for each deforestation pixel the year of deforestation according based on our BLCMs and the land-cover maps from Griffiths et al. (2018).
We validated our maps following best practices (Olofsson et al., 2014) by generating for each year a random sample of ~350 validation points, considering the sampling strata of our six map classes.We additionally (a) accounted for the effects of omission errors caused by the disproportional high area weight of large classes (e.g.natural vegetation and agriculture) and introduced an additional 500 m buffer (Olofsson et al., 2020) around deforested (burned and not burned) areas and (b) minimized the effect of spatial autocorrelation by ensuring a minimum distance of 200 m between validation points.We then visually examined each individual point based on Landsat and Sentinel-2 and, where available, high-resolution imagery (i.e.Google Earth, selected RapidEye, and Pléiades images were available for the vicinity of Novo Progresso).Using this sample, we generated an error matrix and calculated the adjusted area estimates for each map class and confidence intervals (Olofsson et al. (2014) and S7).

Analysis of land-tenure types
We assessed the differences in deforestation and burned areas across the study area and the three different presidencies (i.e.Rousseff 2014-2015, Temer 2016-2018, and Bolsonaro 2019-2020) by statistically comparing deforestation and fire rates while accounting for confounding factors.To do so, we created for each year a random sample of 50,000 points from the annual area of potential treatment, i.e. deforested areas and primary vegetation areas that could have been deforested as well as deforested areas and farmland that both could have been burned.We then extracted for each point the corresponding land cover and confounding variables that influence farmland suitability, such as terrain roughness and distances to settlements and slaughterhouses (Table 2).Additionally, we added a dummy variable representing the presidency during which deforestation or burning occurred.We use the annual mean and minimum values of the 3-monthly Standardized Precipitation-Evapotranspiration Index (SPEI, Vicente-Serrano et al., 2010) as further explanatory variables.They characterize temperature and precipitation conditions that favor fires, i.e. generally dry years (low mean SPEI values) and years with extraordinary dry seasons (low min SPEI values, Fig. S8-7).
We then estimated two models in the general form of: where y represented the binary outcome (i.e.deforestation/not deforestation and burned/unburned, respectively), tenure the tenure category, and presidency the political administration which formed our treatment.In addition, the spei03mean represents as proxy for generally dry years and spei03min for years with a highlighted dry season.
Prior to model estimation, we calculated inverse probability weights from covariate balancing propensity scores (CBPS) for each data point to ensure an optimal balance of the covariates (Austin, 2011;Imai & Ratkovic, 2014;Schleicher et al., 2020).For that, we used the WeightIt and CBPS packages in R (Greifer, 2021): (2) Once we ensured optimal covariate balance across land-tenure types (S8), we used the weights from (2) in our general model (1) to estimate the average treatment effect (ATE) of the zonation type on deforestation and burning.Because they differed too much in covariates from the other land-tenure classes and showed little deforestation and fires only, indigenous territories and military areas were excluded from the ATE assessment.

Analysis of farm size effects
Finally, we assessed how deforestation and burning varied with farm size between the three presidencies.We used a property dataset from the Amazon Environmental Research Institute (IPAM) that is compiled from federal and state land-tenure registers and solves property overlaps by prioritizing a higher ranked registration system (Stabile et al., 2022).The farms were grouped into 'small', 'medium', and 'large' properties by size in fiscal units (FU, Section 3).
We then estimated the ATE of the respective presidency on the relative deforestation and burning rates by fitting linear models in the form: Prior to the estimation, we again accounted for other confounding factors affecting deforestation and burning by applying covariate weighting.Contrary to the point-level analysis, however, we summarized our covariates at the property level by calculating the mean of these variables (Table 3).

Trends of deforestation and burned areas
Between 2014 and 2020, 3.8% of the region was deforested, which is an area equivalent three times the size of São Paulo, the largest city in Brazil.Until the mid of Temer's presidency in 2017, the rates decreased to 331 km 2 , but more than doubled in the next year, and reached their highest value under Bolsonaro in 2020 (1,201 km 2 , Figure 3 and S5).Similarly, the burned area was lowest under Temer in 2018 (2,343 km 2 ) and highest under Bolsonaro in 2020 (4,805 km 2 ).In all years, the majority of burned areas had been deforested more than 5 years before (Figure 4).Only 3% (2017) to 15% (2020) of burned areas were linked to deforestation in the same year.
The overall accuracies of the BLCMs (Figure 5) reached on average 99% (minimum 97% (2018), maximum 100% ( 2016)).The class-specific user's (UA) and producer's (PA) accuracies were generally highest for natural vegetation (mean UA/PA = 100%/100%) and (unburned) agricultural land (97%/97%).All classes were mapped with a UA of ≥90%, except of unburned deforestation with a mean UA of 69%, often because being mapped as burned deforestation instead.The accuracy for the total deforested area was still high with a mean UA of 92% (85% in 2017 to 100% in 2014) and a mean PA of 89% (74% in 2018 to 100% in 2017).Similarly, the accuracy of total burned area showed a high mean UA of 92% (79% in 2018 to 96% in 2014 and 2016) and a mean PA of 90% (74% in 2018 to 100% in 2016).More details on the accuracy assessment are shown in S5.

Differences between land-tenure systems
Summarizing across our entire observation period, deforestation occurred mostly in undesignated lands, conservation units, and agrarian settlements (Figure 3).In terms of average treatment effect (ATE), conservation units were least and agrarian settlements most affected by deforestation.From Rousseff to Temer, the ATE for deforestation often showed small, insignificant reductions (Table 4, a +c).Strongest significant increases in deforestation occurred during Bolsonaro in conservation units (+0.6) and undesignated areas (+0.6).Moreover, the likelihood of burning increased significantly across all tenure types, again strongest under Bolsonaro (Table 4, b+d).Areas burned in conservation units were often younger in terms of years since deforestation.Compared to that, fires on private lands mostly occurred on areas which had already been deforested 10 or more years ago (Figure 4).The effect of the minimum SPEI value, i.e. years with especially dry months, was highly significant for burned areas, but mostly showed lower magnitudes than the presidency variable.

Differences between farm sizes
Although small, medium, and large farms contributed differently to the farming land (20%, 35%, and 45%, respectively), deforestation was most intensive on small-and medium-sized farms (32%, 34%, and 33%, respectively).In particular, large farms contributed to the steep deforestation increase in conservation units starting with the Temer presidency (Figure 6a).Overall, this presidency did not show significantly lower deforestation on farmlands but increased thereafter, particularly on large and medium farms (Table 5).The removal of natural vegetation (i.e.deforestation of primary vegetation and re-clearing of secondary vegetation) always exceeded rates of regrowing natural vegetation.The increase of forest clearing was reciprocal to the farm size, with+13% to 54%, +7% to 39%, and +4% to 22% on small, medium, and large farms, respectively.

Discussion
Brazil successfully reduced its deforestation rates until 2014, but rates dramatically increased again since then.The observed surges in deforestation and fires in the Amazon are largely attributed to policies that came into power after the impeachment of President Rousseff (Table 1, Reydon et al., 2020).With respect to the deforestation frontier in South-West Pará, our work provides the following key insights: (1) Deforestation significantly increased under Bolsonaro, reaching the highest rates since 2009.(2) While being still lower than for other tenure types, ATE on deforestation increased strongest in conservation units, i.e. public lands associated with the highest environmental protection levels, and undesignated lands.(3) The percentage of deforested farmland increased strongest on medium and large farms and more so during the Bolsonaro presidency.(4) Fires significantly increased in undesignated areas, conservation units, agrarian settlements, and on private land during the Temer government and even more under Bolsonaro.The majority of burned areas (≥75%) were mapped in already-deforested areas.

Policies and land-use changes
This work highlights land-use differences between farmers that operate in different land-tenure systems and on different farm sizes.Dubious (or unclear) legal status of ownership is the most important constraint for economic development and compliance with environmental regulation (Araujo et al., 2009;Azevedo-Ramos & Moutinho, 2018;Reydon et al., 2020), i.e. to avoid deforestation or use fire-free management alternatives (P.Pacheco, 2012;R. Pacheco et al., 2021).Our quantitative approach builds a clear link between changes in presidencies and landuse change.Our study region has been a priority region for operations against illegal deforestation since 2008.Due to the high deforestation rates and 'blacklisting' of municipalities like Novo Progresso, farmers hardly obtained credits, while the level of law enforcement by IBAMA field inspections and fining of environmental violations was high (Assunção & Rocha, dos Santos Massoca & Brondízio, 2022;Klingler & Mack, 2020).The defunding of the MMA under president Temer, but especially under the Bolsonaro government, which changed key leader positions with politically opportune persons (Abessa et al., 2019), weakened IBAMA's capability to combat deforestation (Ferrante & Fearnside, 2019).IBAMA reduced its ground operations by 60%, stopped its practice of destroying confiscated equipment, and lost protection by local police (Hecht, 2020)-and the number of environmental fines became the lowest in the decade (Vale et al., 2021).This weakening of previously effective command-and-control mechanisms to disincentive environmental infractions (Börner et al., 2014) and the signals of the Bolsonaro government to tolerate infractions are plausible explanations for the increases in deforestation, which both our maps and the statistical analysis suggest.
In 2017, the Brazilian congress discussed partially downgrading two conservation units in our study area from National Forests to Areas of Environmental Protection (MP 756/2016 andMP 758/ 2016).This would have allowed private ownership, agriculture, forest use, and legalize land occupations (Branford, 2017).During the Temer-Bolsonaro transition, decrees 9.309/2018 and 10.165/2010 attempted to change the dates to recognize land regularization (Salomão et al., 2021).The initiatives were stopped but are examples of repeated attempts to regularize illegal occupations in our study region, encouraging deforestation and irregular land use.
Most agricultural fires are likely illegal, as they are, by law, permitted for subsistence farming only and require an official license (Silveira et al., 2022).Reduced fining of environmental crimes may have contributed to the increase of agricultural fires too.In addition, the defunding of ICMBio is likely to have reduced the effectiveness of prevention measures, like integrated fire management programs (Oliveira et al., 2021) in the four conservation units of our study region.

Deforestation, farmers, and fires
Similar to our observations, PRODES has shown the increase of deforestation, especially under Bolsonaro (Silva Junior al., 2021), which was accompanied by an increase in the deforestation patch size (Trancoso, 2021) and often occurred in undesignated areas and public lands (A.Alencar et al., 2022;Salomão et al., 2021).
In accordance with S.S.D. Silva et al. (2021), we also observed most burnings in undesignated areas, conservation units, and agrarian settlements and an increase of fires in 2019.Likewise, our conservation units showed higher fractions of burning for deforestation compared to agricultural fires.We mapped higher fractions of burned area in already-deforested areas: 89% in conservation units to 98% on private land (Figure 4), compared with a range of 49% in conservation units to 69% in agrarian settlements of Acre (S.S.D. Silva et al., 2021).These differences may be caused not only by differences in the study regions but also by our differing mapping approach.Our study further shows that most fire is used for maintenance of agricultural land, i.e. to sanitize pastures or to re-clear fallow pastures (Barlow et al., 2020).This is consistent with field interviews in the region, where fire is part of an often extensive management strategy, i.e. to control pest species and prevent the spread of pasture sudden death syndrome that affects locally foraged species (Barlow et al., 2020;Eri et al., 2020;M.A. Silva et al., 2018).The dependence of small farmers on fire might explain why observations of burned areas were more likely in agrarian settlements, which are associated with communities where access to and experience with firefree practices are limited and economically constrained (Carmenta et al., 2018;Fonseca-Morello et al., 2017).The lack of budget allocated to INCRA has worsened this situation, as can be seen by the decreased number of contracts for technical assistants (INCRA, 2022; M.L. Souza et al., 2022).

Methods, uncertainties, advantages
Our study provides a comprehensive reconstruction of deforestation and fire in south-west Pará.We mapped burned areas that were clearly visible in optical Landsat and Sentinel-2 data, hence underestimating low-intensity burns where top tree or shrub cover remained unaffected, or where burn scars were barely visible after the fire event.Deforestation fires were mapped if a significant share of primary vegetation was removed in the same year.This excludes wildfires and forest degradation, which were not in the focus of our research.Our approach also does not allow us to draw conclusions about whether the burns were intentional or unintentional.
Our approach allowed us to assess burned areas also within smallholder farming systems, with small property sizes.Visual comparisons with PRODES deforestation maps and (30 m) MapBiomas burned area (A.A.C.Alencar et al., 2022) showed that our maps are often more accurate, spatially and temporally (S6, S9), when the aforementioned fine-scale patterns prevail.A quantitative comparison of our maps with those based on different mapping workflows, and different datasets, will provide more detail on what was beyond the scope of this study.
To our best knowledge, we were the first to control confounding factors that may influence deforestation and burning, in the form of spatially varying topographic and climatic variables.Forthcoming studies could also consider further socio-economic factors like household income, education, cultural background, or agricultural prices, all which influence management decisions (Nascimento et al., 2019;Santos et al., 2021).Furthermore, we do not differentiate between federallevel and state-level governance, which could have had different, and even opposite, influences on land-use regulations.

Conclusion
Our study highlights the usefulness of quasi-experimental methods for comparing land use and land-use change under different political regimes.In combination with spatial and temporal metrics, they adequately resolved remote sensing-based maps.Using our methodology, we were able to clearly identify how deforestation and agricultural fires accelerated during the presidencies of Temer and Bolsonaro.The diversity in outcomes among legal farm sizes informs our understanding about how different farmers responded to recent shifts in environmental politics and policies in Brazil.Such a differentiated view allows judging policy impacts on land-use regimes.In areas of conservation and cultural importance, this can be useful for safeguarding biodiversity and carbon stocks and upholding local livelihoods.Overall, our observations show the importance of solving land-tenure insecurities and strengthening measures -those weakened under President Bolsonaro -to disincentive environmental infractions.More broadly, this research specifically reveals how fast changes in national governance regimes may impact local decision-making.That is, deforestation and fire use of small and large agricultural producers in Southern Amazonia.

Figure 1 .
Figure 1.Study region around Novo Progresso and the BR-163 highway in south-west Pará, Brazil.Farm sizes grouped by number of fiscal units they occupy.

Figure 3 .
Figure 3. Annual deforestation by land-use zone with and without burning.The error bars represent the uncertainty in the area estimates resulting from our accuracy assessment.Note the different scales of the y-axes for agrarian settlements, conservation, and undesignated areas (top) as compared to military, indigenous, and private areas (bottom).

Figure 4 .
Figure 4. Burned area by years since deforestation and land tenure.Absolute area (top) and by fraction of annual burned area (bottom).Error bars relate to the total area burned.Military areas and indigenous territories skipped because of low fires rates.

Figure 5 .
Figure 5. Year of deforestation map (left) and burned area and land cover map (BLCM) for 2019 with exemplary subset of intensive deforestation south of Novo Progresso.

Figure 6 .
Figure 6.Deforested (a) and burned agricultural (b) areas (exclusive that deforested) by farm size and land-tenure system.Relative values relate to the year with highest value per tenure class.

Table 2 .
Weighting variables and explanatory variables to analyze average treatment effects of land-tenure type and presidencies.

Table 3 .
Weighting covariates and explanatory variables to analyze average treatment effects of farm size and presidencies.