Commodity frontiers expand more slowly into tropical forests where forest smallholders are present

Significance

Tropical deforestation remains one of the greatest challenges for global sustainability. In deforestation frontier regions, diverse actors, including smallholders, Indigenous communities, or agribusinesses, make decisions that determine whether forests persist or disappear. Yet, how the interactions and competition among these actors influence the fate of forests is unclear. Studying a nearly 1 million km² region across two global deforestation hotspots, we show that forests are lost more slowly where forest-dwelling smallholders are present, even amid expanding commodity agriculture. This is an important finding, as we also show that agribusinesses predominantly expand over lands already used by smallholders rather than into unused areas. Supporting and securing smallholders’ land rights can thus strengthen local livelihoods and contribute to forest and biodiversity protection.

Abstract

The expansion of commodity agriculture into tropical forests leads to major social–ecological impacts. Yet little is known about where and how the capitalized actors driving this expansion interact with smallholders, particularly where smallholders live inside the forest. Two alternative outcomes of their interactions on deforestation are plausible: accelerated deforestation as all actors compete for land (i.e., land rush hypothesis) or, alternatively, slower deforestation as actors hinder each other’s expansion (i.e., competition hypothesis). We test these hypotheses for two global deforestation hotspots: the Dry Chaco and the Chiquitano Forests (together nearly 1 million km²), using a land-systems representation that captures spatially overlapping land-use actors. We integrate satellite-based forest-loss time series with an independent reconstruction of land-system change from 2000 to 2023 in a Bayesian regression framework. This reveals three key findings: First, capitalized agriculture has expanded mainly over areas used by forest-dependent smallholders and Indigenous communities, indicating strongly intensifying land competition. Second, where forest-dwelling smallholders persist, their presence can mitigate deforestation pressure from expanding commodity frontiers. Third, the presence of either forest-dwelling smallholders or conservation areas maintains forest cover, but their combined presence does not amplify the protection effect. Overall, our findings show that interactions between agribusinesses and smallholders are key factors that shape the social–ecological outcomes of frontier expansion in the tropics. Accounting for land-system overlaps in research and policymaking on frontiers can therefore improve efforts to safeguard forests while supporting traditional livelihoods, especially in contested, rapidly changing landscapes.

Tropical forest loss erodes nature’s contributions to local livelihoods and is a major driver of climate change and the accelerating biodiversity crisis (1, 2). Halting tropical deforestation is thus a global sustainability priority (3), enshrined in a wide range of international agreements and policy initiatives (4). These include ambitious goals to expand protected area networks (5), regulate supply chains (6), and implement land-use plans (7). Yet, achieving these aims is challenging because forest loss arises from complex processes involving diverse, interacting, and often competing land-use actors, each operating within distinct social–ecological contexts (8). This complexity makes it hard to clearly attribute forest loss and to target interventions effectively in space and along supply chains.

The need to understand deforestation dynamics is most urgent in frontier regions, where rapid resource exploitation—commonly for the expansion of commodity agriculture—can lead to rampant forest loss (9). Many of today’s deforestation frontiers are concentrated in the world’s tropical and subtropical dry woodlands, including seasonally dry forests, shrublands, and savannas (10). Largely unnoticed by policy or the public at large, many of these woodlands have recently undergone rapid conversion to commodity agriculture (11, 12), driven by investors and capitalized corporate actors (13), and facilitated by technological advancements (14), rising international demand, and the incorporation of remote regions into global trade networks (15). This contrasts with many historical deforestation frontiers, where settlers or colonists were the primary actors driving forest loss through engagement in small-scale, semisubsistence agriculture (9, 16). Where commodity frontiers expand today, capitalized actors increasingly interact and compete with those smallholders and Indigenous communities who have traditionally settled dry woodland regions and rely on woodlands for livestock herding, hunting, and collecting forest-based products, placing high pressure on these communities (17, 18). The smallholders and Indigenous communities often face insecure land-tenure arrangements and are disadvantaged in terms of financial capital and political power compared to capitalized actors (19). As a result, competition between such profoundly unequal actors often produces uneven distribution of benefits and burdens of frontier expansion (20, 21), marginalizes or displaces traditional smallholder and Indigenous communities (22), and drives the destruction of natural habitats—changes that are largely irreversible (23, 24). Understanding how different land uses interact in frontier regions is therefore crucial.

Yet, the role of competition between land-use actors in shaping deforestation dynamics remains insufficiently understood. Two plausible alternative scenarios for how these interactions influence forest loss have been proposed. On the one hand, deforestation could be accelerated as agribusinesses and smallholders all seek to claim as much land as possible to expand various forms of agriculture (hereafter: land rush hypothesis), especially where frontier expansion by one actor group paves the way for other actors (9, 20) and where deforestation is a mechanism to manifest land claims (25). On the other hand, competition for land among actors, and the considerable frictions and sometimes open and fierce conflicts this brings, may slow down agricultural expansion and deforestation (hereafter: competition hypothesis) (26, 27).

Despite substantial efforts to attribute deforestation to land-use actors, major knowledge gaps persist in understanding the role of land competition for deforestation. In particular, four challenges have prevented a deeper understanding of how competition influences forest loss in the tropics. First, many studies have focused on generalized land-use “drivers” rather than specific actors, often correlating coarse spatial variables that–at best–are only very indirect proxies for actors. For example, advances in deep-learning allowed coarse attribution of deforestation to specific crops (28), but offer limited insight into which actors grow them (e.g., smallholders vs. agribusinesses). Second, studies using census data (29, 30) or tenure data (26, 31) have yielded important insights, particularly in the context of Indigenous land stewardship (32, 33), but fall short in areas where land titles are absent or contested—a widespread reality in frontier regions (34). Third, commodity frontiers often expand into land already used by smallholders or Indigenous communities rather than into “empty spaces” (35, 36), but there are limited data on the geographic distribution of these actor groups, especially of forest-dependent communities. This hinders assessments of how the interactions between these communities and new actors entering frontier regions influence deforestation dynamics. Finally, conservation is rarely treated as a land use, although the increasing expansion of conservation areas has made conservation itself a reason for land competition (37). While protected areas can limit the expansion of agribusinesses (38), they may also induce land scarcity that puts forest-dependent actors under pressure (39). The influence of these different interactions on deforestation outcomes remains poorly understood.

A land-systems perspective, which conceptualizes land use as a social–ecological phenomenon and explicitly represents land-use actors and their interactions, can potentially overcome these limitations and allow for more contextualized insights into deforestation dynamics. We build on recent efforts to define and map land systems in tropical dry woodlands (40, 41), by leveraging an approach that uniquely captures overlapping and co-occurring land systems. We focus on two dry woodland regions, the Dry Chaco and the Chiquitano Forest in Argentina, Bolivia, and Paraguay, both of which are global deforestation hotspots (42). An almost sixfold expansion of agriculture in this region has not only placed woodlands under immense pressure but has also generated fierce competition for land, leaving forest-dwelling smallholders particularly vulnerable (43). In these regions, two main categories of smallholders are commonly distinguished. The first comprises Indigenous communities who, historically nomadic, now practice mixed livelihoods combining hunting, craft-making, forest-product harvesting, and small-scale agriculture (17, 44). The second consists of small-scale, forest-dependent pastoralists of Spanish or mixed Indigenous-European descent, who settled the region in the 19th and early 20th century and practice forest-based grazing, farming, and hunting (22, 45). Both groups face high levels of poverty, tenure insecurity, and limited political power—conditions that systematically disadvantage them in competition with capitalized actors. Previous research has documented land-use change in the Dry Chaco and Chiquitano (42, 46), yet how land competition among different actors plays out, and how this shapes deforestation outcomes, remains unassessed. Focusing on the whole 1-million-km² region, we address two research questions:

RQ 1: How did land-systems change in the Chaco and Chiquitano Forest between 2000 and 2023, and how were these changes associated with deforestation?

RQ 2: How do land-system overlaps influence deforestation outcomes?

To address RQ2, we tested two hypotheses aligned with the competing narratives above: (H1): Overlaps between capitalized agriculture and smallholders accelerate deforestation (land rush hypothesis); versus (H2): Overlaps between capitalized agriculture and smallholders slow down deforestation (competition hypothesis). We also analyzed how conservation areas influenced land-system interactions and their outcomes.

To test these hypotheses, we use a regional land-systems typology for the Dry Chaco and Chiquitano Forests, along with spatially detailed maps of 18 systems (40, 41). These systems encompass five capitalized farming systems (agribusiness cropping, Mennonite capitalized farming, cattle grazing, cattle fattening, speculative clearing), two small-scale farming systems (medium-scale mixed farming, Mennonite small-scale farming), three forest-dwelling systems (forest-dependent grazing, Indigenous forest use with secure land tenure, Indigenous forest use on nonformalized lands), a forestry system (commercial logging), four conservation systems (strict state area protection, less restrictive state area protection, private reserves, Indigenous reserves), a mining system, as well as areas not assigned to any of these land systems. We thus test our hypotheses for both broad land-system groups, as well as detailed systems. We first reconstructed land-system change from 2000 to 2023. Next, we used an independent, satellite-derived woodland change dataset for the same period (42) and a Bayesian regression framework to explore how specific land systems related to woodland loss, and how spatial overlap among systems led to different deforestation outcomes.

Results

Land-System Change 2000–2023.

Most land systems in the Chaco and Chiquitano Forests expanded, with the most pronounced growth being in capitalized land systems, which expanded by 137% (especially agribusiness cropping and cattle grazing), conservation land systems, which expanded by 97%, and medium-scale mixed farming, which expanded by 209%. The systems experiencing area losses were forest-dependent grazing (−12%) and speculative clearing (−43%) (Fig. 1). Moreover, there was a pronounced process of Indigenous land titling, converting almost 100,000 km² of nonformalized Indigenous land use into formalized tenure. Unassigned areas made up 18% of the region in 2000, more than half of which were assigned to one or more land systems by 2023. Consequently, the total area covered by any land system increased over this period.

Fig. 1.

Area changes in land systems in the Dry Chaco and Chiquitano Forests 2000 and 2023 (see SI Appendix, Fig. S2 for a breakdown into countries).

The overall pattern of land-system transitions between 2000 and 2023 was characterized by the expansion of capitalized and conservation land systems, which increasingly replaced or overlapped with forest-dwelling systems (Fig. 2A). Forest-dwelling systems, especially forest-dependent grazing, declined in area over time, largely due to transitions to cattle grazing and agribusiness cropping. Indigenous lands remained relatively stable, with the main change being tenure formalization. By contrast, capitalized land systems expanded notably, especially into areas previously used for forest-dependent grazing and Indigenous land use, and to a lesser degree, into areas not assigned to any of the land systems in 2000. Conservation land systems also expanded substantially. This was mainly driven by a growth in less restrictive protected areas and Indigenous reserves, particularly over former forest-dependent grazing, Indigenous land, and unassigned areas (Fig. 2).

Fig. 2.

Land-system transitions between 2000 and 2023. Columns represent transitions from or to a focal land-system group: Left, from any land system in 2000 to capitalized land systems in 2023; Middle, from forest-dwelling systems in 2000 to any other land systems in 2023; Right, from any land system in 2000 to conservation systems in 2023. The first row shows how focal-group cells transitioned between land systems, with the striped segment indicating the share of area that overlapped with another land system in both years for (A) capitalized land systems, (B) forest-dwelling systems, and (C) conservation land systems. The second row zooms in on those overlapping areas and details which land systems were involved, plotted in absolute area for both years for land systems overlapping with (D) capitalized land systems, (E) forest-dwelling systems, and (F) conservation land systems.

Overlaps between land systems more than doubled over time, driven largely by growing overlaps between forest-dwelling and capitalized land systems (Fig. 2B). Forest-dwelling land systems showed the largest increase in overlaps, particularly with cattle grazing and agribusiness cropping (+192%). By 2023, overlaps between forest-dwelling and capitalized systems covered almost 100,000 km². Overlaps involving conservation land systems remained constant, at around 115,000 km² from 2000 to 2023, although overlaps between less restrictive protected areas and forest-dwelling systems grew by 78%. Overall, the doubling of overlapping areas between 2000 and 2023, despite only a 12% increase in the total area mapped with at least one land system, indicates a substantial rise in the interactions among actors involved in these systems.

Relationship of Land Systems and Woodland Loss.

The best-performing Bayesian regression model clearly showed that some land systems present in 2023 were associated with woodland loss between 2000 and 2023 (capitalized and small-scale farming systems), whereas others maintained woodland cover over time (forest-dependent grazing, Indigenous, and conservation systems) (Fig. 3A and SI Appendix, Fig. S3 and Table S8). Pronounced woodland losses (i.e., negative effects) were estimated for the groups of capitalized (−0.23, SE 0.005) and small-scale farming land systems (−0.23, SE 0.004). In other words, for each percentage point increase in the area covered by these land systems, woodland cover was predicted to decline by 0.23 percentage points between 2000 and 2023 (on average, holding other variables constant). These estimates are based on 3 × 3 km grid cells, aggregating finer-scale land-cover and land-system data, so the same cell can contain both woodland and agricultural areas. In contrast, the effects of the other, noncapitalized land systems were close to zero and either slightly negatively associated with woodland cover–as forest-dependent grazing (−0.009, SE 0.0007) and nonformalized Indigenous land use (−0.003, SE 0.001)–or slightly positive–as formalized Indigenous lands (0.002, SE 0.0001) and conservation land systems (0.005, SE 0.0005). Given negligible woodland regrowth in the area from 2000 to 2023, effect estimates close to zero can be interpreted as indicative of avoided deforestation. Sensitivity analyses using spatially thinned datasets produced consistent coefficient estimates, indicating that spatial structure in our data is unlikely to bias our results (SI Appendix, Fig. S4 and Table S9).

Fig. 3.

Conditional effects of different land systems on woodland cover change. Panel (A) shows the average effects of land systems, depending on the percentage area coverage within 9 km² grid cells. Panels (B–D) show interaction effects of land systems in areas where they overlap with other land systems’ presence. Panels (E–G) show interaction effects of land systems in areas where they overlap with typical land systems transitions. The comparison of the second and third rows is insightful: Panel (B) shows the effect of capitalized agriculture land systems being present in 2000–2023 vs. panel (E) of these systems appearing or expanding in 2000–2023. Similarly, (C) shows the effect of forest-dependent grazing present throughout 2000–2023 vs. (F) these systems disappearing in this period. Finally, panel (D) shows the group of conservation land systems present throughout the period, vs. (G) these systems being absent throughout. Lines show the estimated mean effect and the 95% credible interval (shaded area), based on the model’s output (posterior distribution. These estimates are calculated while holding all other predictors at their average value or reference category (see SI Appendix, Fig. S3 for plotted posterior regression coefficients with SE).

More generally, model comparison showed that all land system groups contributed to explaining changes in woodland cover, but at different levels (SI Appendix, Tables S5 and S6). For the groups of capitalized, medium-scale farming, and conservation land systems, the Bayesian regression models did not improve when replacing the summed areas of systems in these groups with variables capturing the individual land systems. For example, including five separate variables for agribusiness cropping, Mennonite farming, cattle grazing, cattle fattening, and speculative clearing (rather than a single variable representing total area covered by these capitalized land systems) did not increase model performance, suggesting similar effects on woodland cover change for these systems. In contrast, the forest-dwelling group had more heterogeneous effects: splitting them into variables capturing forest-dependent grazing, formalized Indigenous land use, and nonformalized Indigenous land systems separately did improve predictive accuracy, highlighting that these systems had distinct effects on woodland cover.

Land-System Interactions and Woodland Loss.

We found robust evidence for strong interactions among land systems. In other words, the impact of a given land system on woodland loss was significantly influenced by whether it overlapped with other land systems (Fig. 3 B–7G and (SI Appendix, Fig. S3 and Table S8). Most notably, the interaction effects of forest-dependent grazing and formalized Indigenous land use with capitalized land systems suggested a mitigating effect on woodland cover loss, providing strong support for the competition hypothesis over the land rush hypothesis. Although the average effect of these two systems was close to zero (Fig. 3A), their interaction when present in areas where capitalized land systems expanded had positive effects on forest cover (0.16, SE 0.006 and 0.16, SE 0.009 for forest-dependent grazing and formalized Indigenous land use, respectively) (Fig. 3E). This indicates that when forest-dwelling systems are present in the landscapes where capitalized land systems expand, it lowers the deforestation that occurs in these landscapes. In line with this, the negative impact of capitalized land systems was substantially more pronounced in areas where the forest-dependent grazing system was lost (−0.81, SE 0.007) (Fig. 3F) than in average (Fig. 3A) or where forest-dependent grazing was maintained (Fig. 3C). Formalized Indigenous land use also had a pronounced positive effect on woodland cover, when forest-dependent grazing was lost (0.61, SE 0.026) (Fig. 3F). In contrast, the effect of nonformalized Indigenous land use remained close to zero and mostly negative across contexts (Fig. 3 A–G), indicating that land titles are critical for the positive effect of Indigenous land use. The overall negative effect of small-scale farming on woodland cover (Fig. 3 A–D and F–G) also turned into a positive effect on forest cover in the context of expanding capitalized land systems (Fig. 3E). Our results thus did not provide any evidence for a land rush scenario where overlaps among different actor groups accelerate deforestation.

Conservation land systems had a positive effect on woodland cover throughout contexts, but with different magnitudes: The effect was highest when capitalized land systems expanded (0.35, SE 0.008) (Fig. 3E) and lowest when forest-dependent grazing was maintained (0.02, SE 0.002) (Fig. 3F).

Discussion

Given the stark social–ecological consequences of tropical deforestation, halting further loss of these forests has become a global sustainability priority. This challenge is especially acute in the world’s tropical dry forests, many of which have recently become deforestation hotspots due to expanding commodity agriculture, which in turn triggers the expansion of conservation areas. As capitalized producers and conservation actors enter and expand in these regions, they interact—and increasingly complete—with those who have traditionally used land and forests there, including forest-dependent pastoralists, and Indigenous communities. Yet, how these interactions shape deforestation outcomes has remained unclear, mainly due to a lack of spatially explicit data on where these actors and their land claims overlap.

For the Chaco and Chiquitano Forests in South America, we find a massive expansion of capitalized land systems over land used by forest-dwelling smallholders. Yet, where forest-dwelling smallholders persist, their presence can mitigate deforestation from expanding commodity frontiers and even complement or substitute the protective effect of conservation land systems. These results underscore the potential importance of forest-dwelling actors as landscape stewards and highlight opportunities for creating cobenefits between protecting nature and protecting traditional livelihoods.

Capitalized Agriculture Does not Expand Mainly into Empty Lands.

Our analysis shows increasing overlaps among land systems over time, signaling intensifying land competition. Most notably, capitalized agriculture expanded over many areas previously associated with forest-dwelling land systems, particularly Indigenous lands (with or without formal titles) and forest-dependent grazing (+192%). In many cases, this process resulted in the replacement of forest-dwelling systems by large-scale, capitalized producers (Fig. 2), suggesting that expanding and consolidating commodity frontiers often lead to the exclusion or displacement of forest-dwellers. While factors such as smallholder conflicts (25), rural outmigration (47, 48), or voluntary shifts to intensified pastures may also contribute to smallholder decline, these processes frequently stem from eviction and pressures on land (45, 49).

Our findings resonate with other studies documenting the marginalization of forest-dwellers by agribusinesses, both in the study region and globally (22, 50, 51), through compensation, rent demands, or coercive tactics, including threats or violence (19). Forest-dependent people in the study region and elsewhere typically comprise communities with low tenure security and weak, or no, political representation. These people face stark power asymmetries compared to capitalized producers who have access to financial and technological capital. State support for agribusinesses, for instance, through the acquisition of state land at low prices, policies facilitating large-scale land investments or favoring export-oriented commodity production (25, 52), or the deliberate lack of control and law enforcement in frontier regions (53, 54), can further exacerbate these disparities. Although large-scale, commercial agriculture or land investments can yield short-term benefits for smallholders, such as through employment or the availability of improved technologies, there is little evidence for long-term livelihood gains without strong policy safeguards (55, 56).

Where Forest-Dwellers Persist, Their Presence Slows Down Deforestation.

A central finding from our work was the considerable support for the competition hypothesis, postulating that competition for land can slow down deforestation, specifically among capitalized and forest-dwelling land-use actors. Indeed, we found that where forest-dwelling smallholders and Indigenous communities persist in the landscapes in expanding commodity frontiers, woodland loss is reduced and mitigated by 16% in average. Our work thus contributes to a growing body of studies highlighting the positive effects of Indigenous land stewardship on maintaining tropical forests (33, 57, 58), often contingent on secure tenure, as also reflected in our findings (44, 59) and, broadly, more equitable governance arrangements (60). As a result, it is increasingly acknowledged that Indigenous stewardship, coevolved with local ecologies, is often compatible with forest conservation through values based on collective ownership, traditional agroecological practices, and governance principles inclusive to more-than-human beings (58, 61). Thus, the widespread replacement of Indigenous communities without secure land tenure by agribusinesses in our study region (Fig. 2) represents a missed opportunity to protect both forests and Indigenous livelihoods.

Our findings reveal an additional pattern: Non-Indigenous, forest-dwelling smallholders mitigate woodland loss in frontier regions where capitalized actors expand. In the Chaco and Chiquitano Forests, these forest-dependent pastoralists have lived in the area for generations and rely on the forest for grazing, the collection of firewood, the production of fence posts and timber, subsistence hunting, and the collection of nonwood forest products (22, 51). Yet, they are often overlooked in policy debates and receive less outside support from NGOs and allies from public, policy, and academia than Indigenous communities (45). Their land-use practices have often been raised as a cause of environmental degradation, including loss of natural grasslands, forest degradation, and overhunting (62, 63). Although our analysis focuses on deforestation rather than forest degradation, our results challenge simple narratives portraying such smallholders as agents of forest destruction. While the immediate surroundings of their homesteads are often negatively affected by livestock grazing, the collection of firewood, nontimber forest products, and hunting (64), vegetation typically stabilized at greater distances (65) and similar patterns have been observed across drylands globally (66, 67). These observations point to the broader stewardship that forest-dwellers can play, particularly amid frontier pressure. Given the elevated risk of displacement, securing land tenure, offering legal protection, and recognizing their rights are essential steps toward reducing inequalities and safeguarding forest-dependent livelihoods.

Conservation Mitigates Woodland Loss But Does not Amplify the Positive Effect of Forest-Dwellers.

Conservation land systems independently reduced woodland loss (Fig. 3). Yet, in line with our findings about the decisive impact of land-system interactions, the magnitude of this effect was context dependent. While the conservation effect was strongest where commodity agriculture expanded, it was weakest where forest-dwelling smallholders were present and persisted in the landscape. This suggests that expanding conservation areas substitutes, rather than amplifies and reinforces, the mitigating effects of forest-dwelling smallholders on woodland protection. While several studies have demonstrated the role of Indigenous stewardship in maintaining tropical forests (33, 44), fewer have examined how formal conservation areas interact with forest-dweller presence (58, 68, 69). This gap is remarkable, as protected area expansion frequently occurs on lands historically used by Indigenous people or other traditional forest-dwellers, as in our case (Fig. 2) (70). In such contexts, overlooking overlaps risks undermining the additionality of conservation interventions–particularly in light of scarce conservation funding (71). Moreover, the institutional reorganization that accompanies strict protected areas expansion risks reshaping actor constellations, contributing to marginalizing local people, or leading to unjust and unwanted conservation outcomes (72, 73). More inclusive approaches that recognize forests as social–ecological systems, such as other effective area-based conservation measures (OECMs) offer a promising way forward (74).

Some progress in this direction has recently happened in the Dry Chaco and Chiquitano Forests where new conservation areas are mainly less restrictive, multiuse reserves, creating opportunities for conservation-livelihood cobenefits. Although alliances between Indigenous groups and conservation actors have occasionally advanced joint goals of land rights and forest protection in our study region and elsewhere (75, 76), such collaborations remain uncommon, especially for non-Indigenous smallholders. Furthermore, recent advances in Indigenous land titling and zoning laws recognizing sustainable smallholder use, are now again under pressure from political pushes toward environmental deregulation and agro-industrial lobbying (25, 77).

As agricultural commodity frontiers continue to expand into remote tropical and subtropical forests, the need to halt deforestation is urgent. Viewing frontiers as places of interaction among vastly different actors—most notably capitalized agriculture on the one hand and forest-dwelling smallholders or Indigenous communities on the other—provides a more accurate understanding of the social–ecological dynamics at play. Our findings highlight that land systems frequently compete and overlap, and these overlaps can markedly shape deforestation outcomes. Yet, such complexities are rarely accounted for in land-use and conservation planning, often rendering forest-dwelling smallholders invisible. Our land-systems approach offers a way forward, through a concept and toolkit to identify, map, and thus to more adequately consider diverse land-use actors, their overlaps and interactions, in sustainability research, planning, and policymaking.

Materials and Methods

To understand how land competition among different land systems shapes deforestation outcomes in the Dry Chaco and Chiquitano Forests, we developed a reconstruction of land-system change for the period 2000–2023 and integrated it with a satellite-based time series of forest loss in a Bayesian regression framework (Fig. 4). Using our land-system maps, we examined how land systems changed over time and statistically quantified how these changes were associated with woodland loss.

Fig. 4.

(A) Conceptual and methodological framework for mapping and analyzing land-system dynamics and overlaps, as well as their (interaction) effects on deforestation in the study region between 2000 and 2023. (B) Study region.

Mapping Land System Change.

We mapped and validated land systems for the years 2000 and 2023 (Fig. 4A), building on an existing conceptualization of land systems for our study region (40, 41) (SI Appendix, Supporting Information Text). Specifically, we developed an approach to map land systems in an expert-based, bottom–up manner that allows land systems to overlap (41). To do so, we organized expert workshops to identify spatial indicators signaling the presence of a specific land system. Ethical approval of the research was granted by the Ethical Review Board of the European Research Council and by the Ethics Committee of the Faculty of Natural Sciences, Humboldt-Universität zu Berlin (Ref. Ares(2023)123438). All participants provided informed consent prior to participation, following an approved informed consent protocol (Ref. Ares(2024)3908626). Based on iterative evaluations of mapping outcomes using literature and high-resolution satellite imagery, we derived a final set of conditions that meaningfully represents the individual systems. We validated our system maps for 2023 by overlaying them with a large dataset of geolocated interview surveys of land-use actors conducted in the region (41).

Here, we used the same approach to reconstruct land systems for the year 2000. We applied the same conditions and indicators representing land-cover, institutional, and production-related features (41). Our mapping approach was system specific. For some systems, we combined segmentation of relevant land-cover classes with characteristic land-use features, such as cattle confinements or large-scale homogeneous croplands. For others, we intersected geospatial data on homesteads with land cover and applied resource use footprints as reported in regional literature (e.g., for forest-dependent grazing). Other system maps also relied on spatially explicit cadastral or jurisdictional delineations, such as conservation areas. For a detailed description, see SI Appendix, Table S2. We carried out all processing using Python and QGIS (78, 79). The final land system maps were generated at 300 m resolution.

To rule out circular reasoning when analyzing the relationship between land systems and deforestation, we conducted sensitivity analyses. First, we assessed the sensitivity of natural vegetation cover in the forest-dependent grazing system to the minimum woodland cover threshold used for mapping this system. Results showed that the proportion of natural vegetation in this system was insensitive to the threshold applied (SI Appendix, Supporting Information Text). Second, we tested the sensitivity of the buffer applied around forest homesteads (5 km) to signal their area of influence on woodlands, initially set based on literature and expert knowledge (44, 80). Our sensitivity analysis revealed that unassigned areas decreased substantially when increasing the buffer size, suggesting that our mapping is a conservative estimate of the spatial influence of forest-dependent people. Nonetheless, the share of natural vegetation in unassigned areas remained largely unaffected when changing this buffer (SI Appendix, Supporting Information Text).

To investigate changes and overlaps in land systems, we combined the 2000 and 2023 maps into a single, pixel-based dataset capturing the presence or absence of each land system in both years. This allowed us to quantify overlaps and analyze land-system transitions over time. We summarized changes in total area and system overlaps and plotted transitions among three major land-system groups: capitalized, forest-dwelling, and conservation systems.

Statistical Modeling.

We used Bayesian generalized linear regression models to analyze how different land systems—and particularly their overlaps—relate to deforestation (Fig. 4A) (81). To do so, we overlaid our land-system maps with annual land-cover data from a 30 m resolution time-series dating back to 1985 (42), distinguishing eleven classes. We aggregated the land-cover maps (30 m) and land-system maps (300 m) into a common grid with 3 × 3 km cells, based on the area percentage of each land system and cover per grid cell (see SI Appendix, Table S7 for summary statistics of the resulting dataset). Hence, we considered land systems as overlapping and interacting when more than one land system was present within the same 3-km cell. As a measure of deforestation, we used the change in woodland cover (i.e., forest, shrubland, and savannas) between 2000 and 2023, expressed in percentage points per 3-km cell.

To estimate the effect of different land systems on woodland loss, we first identified which land systems contributed to explaining changes in woodland loss. To achieve this, we modeled change in woodland cover by incrementally adding different individual land systems, or land-system groups, as independent variables and comparing the performance of the resulting models (SI Appendix, Supporting Information Text). To reduce potential omitted-variable bias arising from biophysical and accessibility gradients known to influence deforestation, we evaluated slope, elevation, length of growing season, and accessibility as travel time to population centers of at least 50,000 inhabitants (82) as candidate control variables. Using model comparison, we retained only variables that improved predictive performance, resulting in the inclusion of length of the growing season in the final model. Second, we examined how the effect of different land systems changes depending on the co-occurrence with other land systems. We created a variable that captured the transition among land system, categorized as i) never present, ii) present over time, iii) present in 2000 but disappeared until 2023, or iv) not present in 2000 but appeared until 2023, noting that certain transitions might not exist for specific land systems (e.g., capitalized agriculture or conservation areas present in 2000 but disappeared until 2023 is not plausible). This allowed us to assess the effect of land-system transitions.

To assess and minimize the potential effect of spatial autocorrelation, we applied a systematic spatial thinning procedure, retaining only every third grid cell in x- and y-direction. This resulted in a total of 9,729 cells and substantially reduced residual spatial autocorrelation (Global Moran’s I decreased from ~0.4 to ~0.1). We generated three independent thinned datasets, each created by shifting the 3 × 3 sampling grid by one cell in the x and y direction. We estimated the full set of regression models on each of the three thinned datasets and compared the results both across the thinned samples and against the full-data model to assess the robustness of model coefficients.

We used a Student’s t-distribution to model the likelihood of changes in natural vegetation cover, because our data were centered around zero but showed a higher-than-expected number of extreme values—especially large negative changes—which the t-distribution can handle better than a normal distribution. We standardized all numerical variables, determined weakly regularizing priors (SI Appendix, Supporting Information Text), and specified the models as follows:

$$\mathrm{\Delta }{W}_i\sim t\left(\nu ,{\mu }_i,\sigma \right),$$

$$\begin{array}{cc}{\mu }_i\hfill & ={\alpha }_i+{\beta }_{LS1}LS{1}_{2023i}+{\beta }_{LS2}LS{2}_{2023i}+\cdots +{\beta }_{LSn}LS{n}_{2023i}\\ & LS\_T{r}_i:\left({\beta }_{L{S}_{T1}}LS{1}_{2023i}+{\beta }_{L{S}_{T2}}LS{2}_{2023i}+\cdots +{\beta }_{L{S}_{Tn}}LS{n}_{2023i}\right)\\ & {\beta }_{slope}slpo{e}_i+{\beta }_{elev}Ele{v}_i+{\beta }_{LGS}LG{S}_i+{\beta }_{access}acces{s}_i\end{array},$$

where ∆W_i indicates change in woodland cover between 2000 and 2023 for every datapoint (3-km grid cell) i, predicted by the presence of different land systems in 2023 (LS_2023i) in interaction with the transitions of specific land systems of interest (LS_Tr_i), and the control variables slope (slope_i), elevation (elev_i), length of growing season (LGS_i), and accessibility (access_i). In simpler terms, our model explored how different land systems influence change in woodland cover, depending on whether other land systems are present, changing, or absent. For example, to test the impact of maintaining forest-dependent grazing, we included an interaction with the possible transitions of forest-dwelling land systems across the four types of transitions: presence, absence, decline, or expansion.

We sampled 4,000 realizations of the posterior distribution using Markov Chain Monte Carlo simulations with four sampling chains running for 2,000 iterations and a warm-up period of 1,000 iterations. We verified model convergence based on the R-hat statistic and the trace plots and evaluated model fit based on posterior predictive checks, that is, predicting new hypothetical data sampled from the posterior predictive distribution and comparing it to a random draw of observed data. We performed all modeling through the brms package in R, as an interface to the Bayesian inference engine Stan (83–85).

Supplementary Material

Appendix 01 (PDF)

Data, Materials, and Software Availability

Aggregated geospatial data, code data have been deposited in GitLab https://scm.cms.hu-berlin.de/pratzema/land-systems-and-deforestation (86).