Uncovering major types of deforestation frontiers across the world’s tropical dry woodlands

Tropical dry woodlands are rapidly being lost to agricultural expansion, but how deforestation dynamics play out in these woodlands remains poorly understood. We have developed an approach to detect and map high-level patterns of deforestation frontiers, that is, the expansion of woodland loss across continents in unprecedented spatio-temporal detail. Deforestation in tropical dry woodlands is pervasive, with over 71 Mha lost since 2000 and one-third of wooded areas located in deforestation frontiers. Over 24.3 Mha of deforestation frontiers fall into what we term ‘rampant frontiers’. These are characterized by drastic woodland loss and conditions favourable for capital-intensive agriculture, as seen in the South American Chaco and Southeast Asia. We have found many active and emerging frontiers (~59% of all frontiers), mostly in the understudied dry woodlands of Africa and Asia, where greater frontier monitoring is needed. Our approach enables consistent, repeatable frontier monitoring, and our global frontier typology fosters comparative research and context-specific policymaking. Agricultural expansion is responsible for tree loss in tropical dry woodlands, but the dynamics of such loss are not well understood. This study presents a global, high-resolution assessment of deforestation dynamics in dry woodlands and provides a tool for consistent monitoring.

T ropical deforestation is the principal driver of biodiversity loss as well as contributing to climate change, the spread of zoonotic diseases and the widespread degradation of ecosystem services, which disproportionately affects marginalized communities 1,2 . A major cause of deforestation is agricultural expansion 3,4 by different types of smallholders and agribusinesses 3,4 . The diverse forms of agricultural expansion, embedded in different social-ecological contexts, produce deforestation frontiers with heterogeneous severity, speed and spatial patterns of forest loss [5][6][7] . Navigating and structuring this complexity remains a major challenge for sustainability science 8 . In particular, identifying high-level, recurring frontier types is essential for the more context-specific policy responses often called for to govern frontier regions 9 . Likewise, identifying high-level frontier types would enable learning and comparative research across geographies 10 and help to build theory in sustainability science 8 . However, whether such generalizable, high-level deforestation frontier types exist remains very unclear.
Three factors have contributed to this knowledge gap. First, there are major geographical research biases towards a few specific regions 11 , while many others remain entirely overlooked 12 . Second, research on deforestation frontiers has typically focused on investigating dynamics within certain places or regions, such as in Amazonia 13 , Indonesia 14 or the Congo Basin 15 , but there is a clear lack of cross-regional and comparative assessments 16 . Third, global satellite-based time-series products describing forest and woodland loss at high spatial resolution have only recently become available 17 . These data could provide deep insights into how deforestation frontiers advance and help to identify typical, recurrent patterns of deforestation frontiers. Recently, detecting and assessing such archetypical human-environment interactions has become a major focus in sustainability science 18,19 , including for identifying static land systems 20 or driver/outcome constellations 8,21,22 . In this study we have pioneered the use of archetype analyses to identify the major, recurring patterns in global deforestation frontiers.
Deforestation frontiers, in particular, are poorly understood in the world's tropical woodlands (defined in our manuscript as including tropical dry forests, shrublands and wooded savannas with more than 10% tree cover). When compared with humid forests, these woodlands are overlooked by researchers, policymakers and the general public [23][24][25] . Yet these ecosystems harbour unique biodiversity, including many endemic species, and are some of the last refuges for iconic megafauna [24][25][26] . Tropical dry woodlands provide important ecosystem services that sustain the livelihoods of hundreds of millions of people and store globally relevant amounts of carbon 24,26,27 . Despite this, many woodland regions today are global deforestation hotspots 17 , including regions in South America (for example, Gran Chaco 28 and Cerrado 29 ) and Southeast Asia (for example, Cambodia 30,31 ). There is also increasing evidence that the same pervasive dynamics are emerging in African dry woodlands, such as in the miombo woodlands 17,25,32 . This translates into an urgent need to identify and monitor how deforestation frontiers expand into woodlands to allow for timely policy responses.
We have developed a methodology to identify and map deforestation frontiers, and applied it to carry out a global, high-resolution assessment of frontier dynamics in tropical dry woodlands. Specifically, we used a 20-year time series of tree loss to develop a consistent, high-resolution set of frontier metrics describing frontier severity, spatio-temporal patterns and development stage. Based on these data, we have established a nested typology of deforestation frontiers, allowing us to (1) map archetypical frontiers across all tropical dry woodlands, (2) describe the social-ecological characteristics of major frontier types and (3) assess the research effort across these types. then used in step 3 to develop three thematic typologies describing the severity, spatio-temporal patterns and development stage of the frontiers. Finally, in step 4, we grouped these thematic types into frontier archetypes.
Applying this framework across the world's tropical dry woodlands, according to our definition ( Supplementary Fig. 1), revealed that woodland grid cells occupied 1,884.97 Mha and contained about 1,044.13 Mha of woodland in the year 2000. We identified 577.36 Mha of woodland cells undergoing substantial deforestation processes from 2001 to 2020 (~31% of woodland cells in 2000), accounting for around 71.7 Mha of woodlands loss (92% of total loss). Of the frontier area, three-quarters (76.5%) is in the tropical and subtropical grasslands, savannas and shrublands biome, highlighting the importance for considering this biome when assessing woodland loss, with the remaining quarter (23.5%) in the tropical and subtropical dry broad-leaved forests biome. Deforestation frontiers occurred across the globe, with the highest prevalence in South America (46.2% of total frontier area in tropical dry woodlands) followed by Africa (43.8%), Asia (5.0%), North America (3.9%) and Australia and Oceania (here, Australia, 1.1%).
Three typologies of deforestation frontiers. In step 3, we combined the frontier metrics from step 2 that described how deforestation progressed in space and time. This resulted in our thematic frontier typologies (see Methods, Supplementary Text 1 and Supplementary Figs. 2 and 3). We combined metrics pairwise to yield three typologies describing the severity, spatio-temporal patterns and development stage of frontiers. We classified each metric into three classes, combining metrics in pairs resulted in nine frontier types for each of our three themed typologies. Our first typology was geared towards assessing deforestation severity, combining the metrics of baseline woodland and the percentage of total woodland loss ( Fig. 2 and Supplementary Fig. 4). Frontiers with high woodland cover and low woodland loss were most extensive (260.00 Mha, 45.0% of total frontier area), followed by medium woodland cover and low woodland loss (125.75 Mha, 21.8%). These frontiers were most common in central and southern Africa (for example, Northern Congolian forest-savanna and Central Zambezian wet miombo woodlands) and North America (for example, Puerto Rican dry forests). Over 40.02 Mha were classified as high severity frontiers, having both a high percentage of total woodland loss and high and medium baseline woodland (4.6 and 2.4%, respectively). High severity frontiers occurred mostly in South America (for example, Dry Chaco), and were more prevalent within South America and Asia (for example, Southeast Indochina dry forests and Southern Vietnam lowland dry forests in Asia).
In our second typology, we explored spatio-temporal patterns of woodland loss, combining metrics of speed of woodland loss and fragmentation ( Fig. 3 and Supplementary Fig. 4). The most common type of frontier had medium speed of woodland loss with high levels of fragmentation (141.20 Mha, 24.5% of total frontier area). This type of frontier was more prevalent within Africa (for example, Guinean forest-savanna and Western Congolian dry forests) and parts of South American (for example, Ecuadorian and Lara-Falcón dry forests). A further 85.00 Mha were classified as slow frontiers with high fragmentation (14.7%), and were more prevalent within North America (for example, the Chiapas Depression and Central American dry forests) and parts of South America (for example, Cauca Valley and Patía Valley dry forests). In addition, 51.54 Mha were classified as fast frontiers with medium fragmentation (8.9%), mainly within Australia (for example, Cape York Peninsula), South America (for example, Dry Chaco) and Asia (for example, Southeast Indochina dry forests).
In our third typology, we explored the development stage of frontiers, combining metrics of activeness and remaining woodland ( Fig. 4 and Supplementary Fig. 4)   classified as active or emerging frontiers (58.9%), most commonly in Africa and Asia.
Archetypical deforestation frontiers. Our final step combined the frontier types identified by our three themed typologies (severity, spatio-temporal patterns and development stage) to identify frontier archetypes, defined here as high-level frontier patterns across geographical contexts (step 4, Fig. 1). By qualitatively grouping combinations of types from our typologies, we derived five such frontier archetypes: inactive frontiers, consolidated frontiers, fragmented frontiers, looming frontiers and rampant frontiers ( Fig. 5 and Supplementary Fig. 5). Importantly, the goal here was not to characterize and classify all frontier cells, but to group those frontier cells that clearly undergo recurrent and comparable patterns. In other words, our typologies (characterizing frontier processes in all cells) and archetype analyses (identifying high-level and recurring frontier patterns) should be seen as complementary.
Inactive frontiers correspond to old frontiers with high baseline woodland, yet still have high levels of remaining woodland. This archetype covered about 75.33 Mha (13.0% of all areas identified as frontiers), mostly in Africa (for example, Northern Congolian forest-savanna and Zambezian-Limpopo woodlands).
Consolidated frontiers have little remaining woodland, are old and deforestation is relatively slow (Fig. 5 and Supplementary Fig. 4). This archetype corresponded to 68.56 Mha (11.9% of all areas identified as frontiers), and was most prevalent within North America (for example, Sierra de la Laguna dry forests), South America (for example, Tumbes-Piura dry forests) and Australia (for example, Kimberly tropical savanna).
Fragmented frontiers are active frontiers producing highly fragmented woodland loss patterns, mainly in Africa (for example, Southern and Western Congolian forest-savanna) and South America (for example, Maranhão Babaçu forests and Lara-Falcón dry forests). This archetype covered about 114.43 Mha (19.8% of all areas identified as frontiers).
Looming frontiers are frontiers emerging at slow-to-medium speed ( Fig. 5 and Supplementary Fig. 4). This archetype covered about 91.82 Mha (15.9% of all areas identified as frontiers), and was most prevalent within Africa (for example, Guinean and Victoria Basin forest-savanna) and Australia (for example, New Caledonia dry forests).
Finally, rampant frontiers refer to rapidly advancing frontiers in situations with high baseline woodland and high woodland loss. This archetype expanded over about 24.3 Mha (4.2% of all areas identified as frontiers), and such frontiers were most prevalent within South America (for example, Dry Chaco and Chiquitano dry forests), followed by Asia (for example, Southeast Indochina and Southern Vietnam lowland dry forests). The remaining areas classified as frontiers in tropical dry woodlands regions, but not falling into one of our five archetypes, corresponded to about 202.92 Mha (35.1% of all areas identified as frontiers).

Social-ecological characteristics of frontiers.
To characterize the social-ecological context of our identified frontiers, we described them according to a set of spatial determinants, selected based on prevalent frontier theories (see Methods). Specifically, we determined the key characteristics of our frontiers based on proximate drivers of woodland loss, capital inputs to farming (captured by field size), market integration (captured by accessibility) and agroecological suitability (Figs. 2-4 and Supplementary Fig. 6). Inactive frontiers mostly showed patterns of smallholder agriculture, very small field size, very low accessibility and marginal suitability (Fig. 6). Consolidated frontiers mostly showed patterns of smallholder agriculture, large field size, very high accessibility and marginal suitability. Fragmented frontiers mostly showed patterns of smallholder agriculture, very small field size, very high accessibility and marginal suitability. Looming frontiers mostly showed patterns of smallholder agriculture, diverse field size, varied accessibility and marginal suitability. Rampant frontiers mostly showed patterns of commodity-driven woodland loss, large field size, very low accessibility and moderate-to-marginal suitability. Other frontiers that did not fall into the five archetypes shown in Fig. 6 showed patterns of smallholder agriculture, large field size, very high accessibility and marginal suitability.
Research effort across frontier archetypes. To assess research effort across our frontier archetypes in tropical dry woodlands, we carried out a systematic literature search and identified 155 relevant studies (see Methods). Of these, 35 and 49 studies only analysed deforestation before and after 2000, respectively. Moreover, 126 focused on national to subnational scales, and 29 studied woodland deforestation across several countries or continents. The greatest number of studies was recoded in South America (66 records), followed by Africa (52), North America (19), Asia (7) and Australia (1). The most researched ecoregions were Cerrado (36), Dry Chaco (35) and the West Sudanian savanna (23), while the least researched ecoregions were in Australia and Asia, notably Southeast Indochina dry forests (4) and Central Indochina dry forests (5; Supplementary  Fig. 8). Assessing the relative research effort across our types and archetypes showed that rampant frontiers received the most research attention, while the least researched archetypes were inactive frontiers (Supplementary Table 3). Research in South America showed a positive bias towards rampant frontiers, while in the other continents this bias was negative (Supplementary Table 3).

Discussion
Tropical dry woodlands are under high and rising pressure globally, yet have been neglected by sustainability research, policymaking and planning 12   dry woodlands globally yielded four major insights. First, the loss of woodlands is widespread, with more than 71 Mha of tropical dry woodlands lost since 2000 and one-third of all woodland areas within deforestation frontiers. However, existing research efforts have focused mostly on deforestation in the humid tropics with only a few studies in tropical dry woodlands, highlighting the urgent need to ramp up research there. Second, rampant frontiers are widespread and chiefly associated with expanding commodity agriculture, which, given the rising demand for agricultural commodities, suggests that pressure on these woodlands will continue to stay high. Third, much of the world's tropical dry woodlands fall into early frontier stages, with many emerging frontiers in Africa, most of which remain overlooked and understudied. Together, this emphasizes the need for monitoring and forward-looking sustainability planning in those regions as frontier dynamics unfold. Finally, despite the high diversity in deforestation speed, patterns and drivers in frontiers, we identified five common, high-level archetypes of frontier dynamics that occur globally. These archetypes provide a basis for comparative research and cross-regional learning in sustainability science. More generally, our frontier metrics and typologies provide a flexible, repeatable and scalable approach to foster more context specificity in research, policymaking and land-use planning in frontier regions.
Deforestation frontiers occurred in a third of our study area, leading to the loss of over 71 Mha of tropical dry woodlands over the period 2000 to 2020. Over the same period, about 147 Mha of tropical moist forests were lost 17,33 . Yet, despite major woodland losses and the huge social-ecological impacts these losses entail 1,24,27 , research efforts have been lagging behind those in other threatened regions, such as tropical moist forests 12,24 . Moreover, among understudied active frontiers in tropical dry woodlands, those in Africa and Asia are particularly under-researched compared with old frontiers in South America. This geographic bias can be explained by structural inequalities in research institutions, access and agendas across continents 11 , the preferences of funding organizations and the fact that general deforestation studies often do not separate dry and moist ecosystems 34 . Fostering tropical dry woodland research, particularly on Asian and African frontiers, would allow for timely policy and management responses in these threatened ecosystems.
Our second main finding was that rampant frontiers were mostly associated with conditions typical for capitalized commodity agriculture (commodity-driven woodland loss representing most of the deforestation, large field sizes, higher agroecological suitability and low accessibility areas). Commodity production drives rampant frontiers in several of those regions, such as the Dry Chaco 28 , where export-oriented soybean production and cattle ranching have created one of the world's most severe deforestation hotspots 35,36 . Commodity production is likely a key frontier driver in less well researched regions, such as the Chiquitano forests, Venezuelan Llanos or Indochina dry forests 16,30,37 . It is well recognized that commodity production, particularly for international markets, is increasingly driving tropical deforestation 36,38 . Actors operating in commodity frontiers typically have the capital to overcome accessibility constraints, operate over large areas to leverage internal economies of scale, are active across international borders and target remote regions where land prices are still low 36,39 . Policy responses, such as land-use planning, often lag behind in rampant frontier regions 30 , and when frontiers slow down due to local resistance, regulation or spatial constraints 40,41 , land-use pressure might be displaced to other regions 29,30 . Rapid regulatory action should therefore be deployed in rampant frontiers, combining area-based interventions (for example, zoning and native habitat protection) as well as actor-focused ones (for example, supply chain agreements) 16,30 .
Many African frontiers were classified as fragmented or inactive frontiers, associated with conditions typical for smallholder regions (high prevalence of smallholder agriculture and small field sizes) 4,6 .
Smallholders use land for subsistence, employment and income through the exploitation of forest resources or the expansion of small-scale agriculture that feeds both local and global markets 32,42 . Fragmented frontiers thus likely represent active frontiers driven by smallholders and occur mostly in more accessible areas. In contrast, inactive frontiers are no longer active and might reflect pioneer efforts to advance into poorly accessible areas 36,43 . Such inactive frontiers could also indicate areas where initial waves of resource exploration have built a technical, social or institutional basis for future, rapid reactivation 44 . The final archetype identified was consolidated frontiers, where frontier dynamics slow down due to the scarcity of remaining forest. This archetype was common in South America, where some regions (for example, Caatinga) have a long history of deforestation 45 .
A key advantage of our methodology is that it can uncover emerging frontiers, of which we found many, particularly in African woodlands. These frontiers emerge because of the availability of unexploited land suitable for agriculture, growing population densities, high dependency on woodland resources for energy and possible investments from actors leaving consolidated frontier regions 1,4 . Frontiers in Africa currently develop under varying trends, with large-scale operations increasing in some areas such as the miombo forests 16,32 . Within emerging frontiers, we identified an archetype of looming frontiers that develop slowly, yet are ubiquitous across Africa. Improved monitoring of woodland loss in these looming frontiers would provide early warning signs of frontiers that have the potential to become rampant. Given what is at stake, including some of the last wild places with intact communities of unique megafauna 24 as well as vast areas where local communities depend on forests and woodlands for their livelihoods 1 , forward-looking sustainability planning of African emerging frontiers is of the utmost importance.
Typologies and archetype approaches are a powerful tool for identifying and mapping high-level human-nature interactions. The archetypes approach that we have developed can be useful for at least three purposes. First, our approach allows us to develop policy responses appropriate for specific, yet recurring contexts. For example, while in emerging frontiers there is an urgent need to focus on increasing protection and ensuring the rights of indigenous peoples and local communities, in consolidated frontiers it may be more important to plan restoration 8,16 . Similarly, while supply chain measures are likely key in rampant frontiers driven by agribusiness actors, in fragmented frontiers dominated by smallholders, policies fostering local participation by recognizing local needs, knowledge and constraints could help protect remaining forests 46,47 . Second, archetypes can enable learning and comparative research across geographies 10,18 . For example, we can apply lessons from successful and failed interventions in South American frontiers to socially and ecologically similar conditions in Africa, or more generally to prevent leakage outcomes of policies against rampant frontiers 29,30 . Third, archetypes can be useful to build theory in sustainability science 8 as our types and archetypes are boundary objects that can intersect with additional information, can be used to compile and synthesize case studies across frontier types or can be used to foster understanding of causal effects and mechanisms underlying deforestation frontiers 10,18 . This could reveal, for instance, when and why frontiers emerge, under which conditions they become rampant or when inactive frontiers are reactivated.
We have developed a typology of deforestation frontiers for tropical dry woodlands using the best available data. Our approach is open and flexible, and while we derived our typologies for the purpose of our study, our methodology can easily be adjusted or refined to generate other typologies and archetypes for different applications. Likewise, our approach can easily absorb updated datasets to consistently trace frontier dynamics and archetypes over time. Our work adds to recent advances in global studies of deforestation frontiers 16 , by considering frontier dynamics, and by extending beyond only mapping deforestation hotspots (~92% of forest loss in tropical dry woodlands falls under our definition of frontiers). Nevertheless, a few limitations need to be mentioned. First, recent work has shown that tree cover in dryland areas can be under-represented 48,49 . The Global Forest Change (GFC) dataset we used here is no exception and potentially underestimates tree cover at low tree densities 50 . The result would be a subsequent underestimation of frontier dynamics in very sparsely wooded ecosystems. Second, the GFC dataset does not distinguish between native woodland and tree plantations. Retaining these areas is important as frontier processes in some tropical dry woodlands include the expansion of tree plantations 31,51 . As a result, we cannot rule out the overestimation of frontier activeness in the few regions where tree plantations are widespread (for example, Uruguayan savanna, Vietnam lowland dry forests and southern Cerrado 52 ). Third, some forest loss might be due to fires, which are often an indication of the frontier-making process 8,37 , but might also occur naturally. Fourth, as with any global analyses, uncertainty remains, and our results are therefore likely most insightful when interpreting the broad-scale dynamics we have uncovered across the world's tropical dry woodlands. We explicitly caution against interpreting and analysing our results at the level of individual cells or localities. Lastly, to generate our archetypes, we qualitatively grouped frontier types, combining both deductive and inductive reasoning. This allowed us to highlight common patterns across geographies while building on prior knowledge and theories.
Our analysis is open for expansion and updating as tree-loss time series grow. Currently, about 35% of frontier cells are not grouped into any frontier archetype, illustrating the diversity of frontier processes in tropical dry woodlands. This diversity is better represented in our three detailed thematic typologies.
Deforestation frontiers are rapidly advancing in tropical dry woodlands, driving biodiversity loss, degrading global ecosystem services and impacting local livelihoods. To better understand these dynamics, we have developed an analytical and operational framework to identify recurrent frontier patterns, which we applied here to provide the first typology and map of global frontier dynamics in tropical dry woodlands. This confirmed that those woodlands are being lost rapidly, that these losses are weakly understood and poorly studied, and that particularly rampant woodland losses are associated with commodity agriculture, particularly in South America and Asia. However, our approach also uncovered the importance of emerging African frontiers, many of which remain virtually unstudied, highlighting the urgent need for more targeted research and sustainability planning. More generally, the frontier types and archetypes we identified can facilitate comparative studies of frontier development, contextualized policymaking and planning, and cross-regional learning. Our approach is generalizable and repeatable across scales, time and geographies, and its value will increase over time as forest-loss time series grow, providing an innovative tool for monitoring deforestation frontier dynamics. Together, this constitutes a major step towards governing deforestation frontiers and promoting more sustainable land use in the world's tropical dry forests.

Methods
Definition of tropical dry woodlands. Many definitions of tropical dry forests or woodlands have been proposed 23,24 . Here, we followed previous work on these systems globally 23,53,54 and used an inclusive definition. Specifically, we focused on all forests, shrublands and savannas falling into two biomes according to the updated biome classification of Dinerstein et al. 55 : (1) tropical and subtropical dry broad-leaved forests and (2) tropical and subtropical grasslands, savannas and shrublands. Within these biomes we defined woodlands as all areas with a minimum tree cover threshold of 10% in the year 2000, based on the GFC dataset. Tree cover in this dataset refers to vegetation taller than 5 m (refs. 17,53 ). Thus, all forests, shrublands and savannas exceeding this threshold are collectively referred to as tropical dry woodlands for the purpose of our manuscript.
Tropical dry woodlands are generally characterized by a marked dry season of 3 months or more, with average annual rainfall from 250 to 2,000 mm, as well as often mesotrophic soils 23,53,56 . This results in typical, diverse vegetation structure within these woodlands, with semideciduous and deciduous trees, drought-resistant shrubs or succulents, as well as grasses. There are strong social-ecological differences both within and between woodland regions 23,53,56 . Tropical dry woodlands have long been a preferred zone for human settlement due to favourable agroclimatic conditions, resulting in various degrees of transformation. Tropical dry woodland regions therefore differ in terms of land-use history, dominant land-use practices or different use of fire as a management tool 57 . Importantly, some tropical dry woodland regions have retained megafauna, such as elephants or giraffes, which can substantially impact on vegetation patterns, while megafauna has been lost from other regions historically or recently 58 .
A major activity in tropical dry woodlands is agriculture, including cropping and rearing livestock. Agriculture has caused major transformations of these regions, both historically and recently 23 . Historically, these transformations were mainly driven by expanding subsistence agriculture, including shifting agriculture 24 . While subsistence agriculture still dominates over large swaths of tropical dry woodlands 27,59 , land-use change in many of these regions today is driven by market-oriented actors expanding industrialized agriculture to produce commodities for domestic and international markets 28,60 . This has caused major tropical dry woodland loss in many regions, including the South American Chaco and Cerrado (for example, soy, maize and beef), Indochina dry forests (for example, rubber and coffee) or Zambia (for example, tobacco and cotton), which have recently experienced some of the highest deforestation rates worldwide 16,17 . Tropical dry woodlands are additionally exploited for their forest resources, including firewood and timber extraction, as well as charcoal production 56,61 . Fuelwood remains an important energy source in many countries of Africa, translating into a major driver of woodland loss there 56,61 .

Developing a typology of deforestation frontiers.
To typify deforestation frontiers in tropical dry woodlands, we developed a four-step analytical framework (Fig. 1).
In step 1, to map deforestation frontiers in tropical dry woodlands, we relied on baseline tree cover and annual tree-loss time-series data from 2000 to 2020 from the GFC dataset 17 . We aggregated the data from 30 × 30 m 2 resolution to a resolution of 3 × 3 km 2 . The GFC dataset has an overall accuracy of >99% (ref. 17 ), and aggregating these data to a lower resolution increases accuracy further 62,63 . The aggregation resulted in an annual time series of woodland loss for the period 2000 to 2020 at an unprecedented spatial and temporal resolution of analysis of deforestation frontiers. Our subsequent analyses were based on this annual tree-loss time series. Our tree-loss analysis included both woodland conversion and degradation processes, which we jointly refer to as deforestation. We did not include tree-gain data from GFC because (1) these are not available annually, (2) are not available for our full analysis period, (3) are less accurate than tree loss and (4) lump together gains in tree plantations, tree crops and natural woodland regrowth 64 . Although assessing tree cover gain in frontiers would be interesting, there is currently no robust dataset allowing us to do this at the spatial and temporal resolution of our analysis. To identify areas potentially qualifying as frontiers, we first selected all cells that had more than 5% of forest cover in 2000, to exclude areas with very low initial forest cover and with little potential to display frontier processes. Next, we selected cells that had an average annual percentage of forest loss of at least 0.5% within a consecutive 5-year period 13 . This ensures potential frontier areas are in line with the common conceptualization of frontiers as progressively expanding 43 and through this causing considerable forest loss.
In step 2, to characterize deforestation frontiers, we created a set of frontier metrics based on our aggregated time series of woodland loss and cover. These frontier metrics go substantially beyond conventional change analyses as they represent multiple facets of frontier expansion, allowing for a deeper characterization of frontiers in space and time. Specifically, we derived metrics capturing (1) the severity of the frontier (that is, the percentage of total woodland loss and baseline woodland), (2) the spatio-temporal pattern of frontier development (that is, the speed and fragmentation of woodland loss) and (3) the development stage of frontiers (that is, frontier activeness and woodland left). We calculated the percentage of total woodland loss by dividing total woodland loss by baseline woodland, where the latter is woodland cover in the year 2000. We calculated speed as the maximum rate of change of woodland loss through the time period, and fragmentation as the maximum value of edge density of the spatial pattern of woodland loss through the time period. We categorized the activeness of the frontier based on when in the time period the frontier was detected. Finally, we considered the extent of woodland left after the period of woodland loss analysed (Supplementary Text 1 and Supplementary Fig. 2).
In step 3, we combined the frontier metrics into three thematic typologies. To do so, we first classified each metric into three classes, informed by data or literature when available (Supplementary Text 1). Our first thematic typology focuses on severity, and we combined baseline woodland with the total percentage of total woodland loss to show the accumulated impact of the frontier over the period analysed. The rationale for this typology is that the accumulated impact of deforestation frontiers can result from the varied ability of frontier actors to capture rents 13,36 . The second thematic typology focuses on spatial and temporal patterns, where we paired speed and fragmentation to show how frontiers progress. The rationale is that the spatio-temporal patterns in frontiers result from different contexts underlying frontier expansion. For instance, commodity-driven frontiers might result in faster and less fragmented frontier progression, given their higher access to technologies and capital 36,65,66 . Our third thematic typology represents the frontier development stage, where we paired the activeness of the frontier with the amount of forest left to understand when a frontier is active and the potential of further frontier activity. This typology further allows us to indicate the status of a frontier in a pre-to-post frontier gradient 13,36,67 . To identify frontier types for each of our three thematic typologies, we classified each frontier metric into three subclasses. This classification follows former frontier studies 13,36 and a set of heuristic rules that best represent the variability in our metrics. The combination of metrics and their subclasses in pairs yielded nine frontier types per thematic typology (further information on each metric is provided in Supplementary Text 1).
In step 4, we qualitatively grouped frontier types into a second level of high-level frontier archetypes by combining specific types, based on the most prevalent types and insightful combinations thereof, potentially reflecting distinct mechanisms of frontier development, and configurations of commodity frontiers, post frontiers, smallholder frontiers and pioneer frontiers 13,36,43 . These archetypes were built to provide a coherent understanding of the most common frontier patterns, complementing the full description of frontiers captured by the typologies. Thus, not all frontiers are included in our archetypes, and future studies could adjust our archetypes to highlight different, or more, archetypes based on the specific goals of the analyses. In other words, our archetype exercise is an approach that can be purposefully adapted, and our implementation serves as an example. Finally, we quantified the extent of frontier types and archetypes, as well as the total area occupied by frontiers. We further quantified the proportions of frontier types and archetypes within continents and ecoregions. For visualization purposes, in the frontier types and archetypes maps (Figs. 2-5), we applied a moving window calculation based on the modal value of neighbouring 3 × 3 grid cells. The combination of our thematic typologies (step 3) and our archetypes (step 4) allows us to structure complexity in frontier processes and explore emerging patterns at high thematic resolution while identifying high-level, recurring frontier archetypes with high generalization potential.

Characteristics of deforestation frontiers.
To further characterize our frontiers, we identified a set of spatial determinants (Supplementary Text 2), building on theories of frontiers, land rent theory and location theory 8,36,43 . We used four global datasets. First, we used a classification of the dominant drivers of tree loss (for example, commodity-driven forest loss) 68 . Second, we used a dataset on field size, as large field sizes proxy capital inputs to farming 69 . Third, to characterize market accessibility, we used a dataset of travel time to cities 70 . Finally, we used a dataset that proxies agricultural suitability 71 via the summed agroclimatic potential for low-input, rain-fed agriculture of main crops and commodities. For the forest-loss driver's dataset, particularly in Africa, the class 'shifting agriculture' encompasses various forms of subsistence and market-oriented forms of agriculture practised by smallholders. In Africa specifically, the importance of shifting agriculture in a strict sense has been decreasing recently 72 . We thus refer to the shifting agriculture class as 'smallholder agriculture' in our manuscript, encompassing shifting agriculture as well as other forms of smallholder agriculture. Each characterizing variable had five categories (Supplementary Text 2). We overlayed the deforestation frontier types with all spatial determinants. We extracted the median value of each spatial determinant per pixel of analysed frontiers and then calculated the share of each spatial determinant category for each type and archetype. The dataset of dominant drivers of forest loss did not overlap with 5.4% of our frontiers, and the field size dataset with 54.70% of our frontiers. For each combination, frontiers with no overlap are not displayed in the results.

Research effort across frontier types.
To evaluate past research effort on deforestation in tropical dry woodlands, and how this research effort relates to our frontier types, we carried out a systematic literature review following the Preferred Reporting Items for Systematic Review and Meta-analysis (PRISMA) approach 73 . We selected a set of keywords characterizing the object of study (tropical dry woodlands) and process studied (woodland loss). We tested keywords interactively in the Institute for Scientific Information (ISI) Web of Knowledge. We finally used the following search string: "("Frontier*" OR "Defor*") AND ("*Dry Forest*" OR "*Savanna*" OR "Dry tropical forest*" OR "Woodland*")". We searched the ISI Web of Knowledge (https://www.webofknowledge.com/) and Scopus (https://www.scopus.com/) databases using these keywords in July 2019, the search yielding 2,146 records. We removed duplicates, leaving 1,439 records ( Supplementary Fig. 7).
We then reviewed the remaining records using an inclusion criterion by reading, depending on necessity, the title, abstract and then the full text. The inclusion criteria retained papers that quantitatively or qualitatively analysed forest loss in our study area. We included papers that directly or indirectly calculated or inferred forest loss to analyse other social-ecological components (for example, carbon storage) if the forest-loss data were disclosed. We finally selected 155 records for further analysis. Next, we reviewed the full text of each of these records to classify each record on geographical location, ecoregion studied, and spatial and temporal scale of the analysis. We georeferenced these records to ecoregion resolution.
To relate frontier archetypes with research effort, we compared the occurrence of each ecoregion's frontier archetypes with the number of studies per ecoregion. For that, we calculated the weighted occurrence of archetypes by the number of studies by multiplying the share of archetypes by the share of records found in each ecoregion. We then summed the ecoregion-level weighted occurrences by archetype and divided it by the total share of archetype and by continent, resulting in the research effort bias by archetype and by continent, respectively.
Reporting Summary. Further information on research design is available in the Nature Research Reporting Summary linked to this article.

Data availability
All datasets used here are publicly available and are referenced. Data outputs from this study are publicly available on Zenodo at https://doi.org/10.5281/ zenodo.6141799. The methodological steps are described in the Methods and Supplementary Information.

Code availability
The code used for the development of frontier metrics, typologies and archetypes in this study is permanently and publicly available on Zenodo at https://doi. org/10.5281/zenodo.6141799.

nature research | reporting summary
April 2020 Corresponding author(s): DBPR NATSUSTAIN-210610240 Last updated by author(s): Jul 1, 2021 Reporting Summary Nature Research wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency in reporting. For further information on Nature Research policies, see our Editorial Policies and the Editorial Policy Checklist.

Statistics
For all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.

n/a Confirmed
The exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement A statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly The statistical test(s) used AND whether they are one-or two-sided Only common tests should be described solely by name; describe more complex techniques in the Methods section.
A description of all covariates tested A description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons A full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient) AND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals) For null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted

Software and code
Policy information about availability of computer code Data collection No software was used to collect data.

Data analysis
Forest loss was summarized in Python 3.8, deforestation frontiers metrics were calculated and typologies developed in RStudio 1.3.1056. Types and archetypes analyses were plotted with ggplot2 and alluvial. Maps were made using ArcMap 10.5. Literature review was conducted in MAXQDA 2020 and results summarized in MS Excel. The code used for the development of frontier metrics, typologies and archetypes in this study is permanently and publicly available on Zenodo at https://doi.org/10.5281/zenodo.6141799.
For manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors and reviewers. We strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Research guidelines for submitting code & software for further information.

Data
Policy information about availability of data All manuscripts must include a data availability statement. This statement should provide the following information, where applicable: -Accession codes, unique identifiers, or web links for publicly available datasets -A list of figures that have associated raw data -A description of any restrictions on data availability Datasets used in this analysis are publicly available from the references provided within the paper. Forest cover and loss data are available at: https:// data.globalforestwatch.org/; Accessibility data are available at: https://www.edenextdata.com/?q=content/jrc-accessibility-map-estimated-travel-time-nearest-citypopulation-50000; Agricultural suitability data are available at: https://www.gaez.iiasa.ac.at/; Drivers of forest loss data are available at: https:// data.globalforestwatch.org/; Field size data is available here: https://pure.iiasa.ac.at/id/eprint/15526/. Ecoregion data is available at: https:// ecoregions.appspot.com/. Data outputs from this study are publicly available on Zenodo at https://doi.org/10.5281/zenodo.6141799.

nature research | reporting summary
April 2020 Field-specific reporting Please select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.

Life sciences Behavioural & social sciences Ecological, evolutionary & environmental sciences
For a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf

Ecological, evolutionary & environmental sciences study design
All studies must disclose on these points even when the disclosure is negative.

Study description
We developed and mapped deforestation frontier metrics at 3 km spatial resolution in tropical dry forest and woodlands for the period between 2000 and 2020. Deforestation frontiers metrics were combined to yield typologies and archetypes of deforestation frontiers. Types and archetypes were overlayed with spatial determinants of deforestation frontiers and results summarized. A literature review was conducted to evaluate the spatial distribution of past research effort on TDF deforestation.

Research sample
There was no sampling, as we have continuous coverage data (i.e., maps).

Sampling strategy
Sampling strategy is not relevant for this study.

Data collection
We collected and integrated data from a variety of sources, which are needed to reproduce results presented in this work.
Timing and spatial scale The forest cover and loss product cover the period between 2000 to 2020 and at 30m resolution. Dominant drivers of forest loss cover the period 2001 to 2015 at 10 km resolution. Global field size map was for the year of 2015 at 1km resolution. Agricultural suitability data was available for the period baseline 1961-1990 and at a five arc-minute grid-cell resolution. Global map of travel time to cities to assess inequalities in accessibility was available for the year of 2015 at 1 km resolution. Analyses were conducted at a 3km resolution, and cover tropical dry forests and woodlands worldwide.

Data exclusions
No data were excluded from the analysis.

Reproducibility
The study is fully reproducible by acquiring necessary datasets and applying the same methodology, or by re-running developed custom code.