Tropical forest-transition landscapes: a portfolio for studying people, tree crops and agro-ecological change in context

ABSTRACT Nudging the development trajectory of tropical landscapes towards sustainability requires a global commitment and policies that take diverse contexts and forest transitions into account. Out-scaling and upscaling landscape-level actions to achieve sustainable development goals globally need to be based on understanding of extrapolation domains and interconnectivity of products and services. We evaluated three portfolios of tropical landscape observatories and quantified extrapolation domains across ecological zones, stages of forest transition, human development index (HDI), population density and potential prominence of four dominant tropical tree crops (arabica coffee, cacao, rubber and oil palm). The ASB Partnership for Tropical Forest Margins portfolio was focussed on active humid forest margins and the Poverty and Environment Network on early stages of forest transition. The portfolio of sentinel landscapes of the Forests, Trees and Agroforestry (FTA) research programme provides a 5% sample of pantropical area, 8% of people, 9% of tree cover and 10–12% of potential tree crop presence, with quantified biases across zones, transition stages and HDI. In the ‘water tower’ configuration, relatively high population density coincides with biodiversity, coffee expansion and contested ecosystem services. The extrapolation domain of the FTA portfolio includes trade-off (tree loss) and synergy (restoration) phases of tropical forest transition. EDITED BY Paolo Omar Cerutti


Introduction
Humans have a shared history with forests and trees (Williams 2003). This multifaceted relationship impacts several social, economic and environmental dimensions of our planetary and human systems. Humans converted and cleared forests to make space for agriculture, specific tree crops and development, often before they realized the value of what was being lost and the costs of restoring the ecosystem functions that society needs (Costanza et al. 2014).
Globally, there is no statistical relationship ( Figure 1) between an aggregated Sustainable Development Goals (SDG) performance metric (Sachs et al. 2016) and forest cover, once the latter is adjusted for human population density on the basis of a global relationship (Köthke et al. 2013). On one hand, this indicates that forest conservation is not essential for achieving the SDG; on the other, that it isn't a hindrance either. Much depends, of course, on the operational definition of 'forest' in this context (see below). The tropics as a whole are about 13% lower in aggregate SDG performance than the non-tropics (as can be seen from the regression equations), although there is a substantial range in SDG score as well as forest cover in both climatic zones. How can the benefits of forests be used while the downsides are avoided? Rural poverty and food, water and energy security interact with income opportunities, climate change, biodiversity, governance and partnership agendas within the comprehensive SDG framework (Van Noordwijk et al. 2016b).
Interactions between people, ecosystems and tree crops at the local level are well covered in many case studies (Wiersum 1997;Meyfroidt et al. 2014;Wunder et al. 2014). However, the variability of these interactions across the tropics is much less addressed. Geographically targeted, evidence-based policies for managing sustainable development trade-offs are scarce (Minang et al. 2014b). To support the development of such policies, the characterization of variability and extrapolation domains is important for deepening our understanding of sustainability in the tele-connected context of ruralurban and local-global relationships. This can be seen in attempts to reduce (or avoid further) deforestation through interventions in the value chainsfrom producers to consumersof tropical tree-crop commodities, which try to better couple the good intentions of companies and governments to the realities of producers, the environment and markets (Mithöfer et al. 2017): such efforts are best observed simultaneously at all levelslocal, national and globalin order to understand the full range of influences and effects.
The path to sustainability is likely to include increased agricultural productivity with more efficient use of resources (a local issue), improved value chains for commodities produced across landscapes (regional), functioning mechanisms that reward provision of ecosystem services (global), inclusive landuse planning and policies (regional and national), strengthened capacities and partnerships with better enabling conditions (local), and comprehensive, local-to-global, 'green growth' strategies. To achieve sustainability, a thorough understanding is needed of what is most apt for particular agro-ecological zones and which interventions (or light-touch 'nudges'; Thaler and Sunstein 2008) can be effectively applied. Timely and effective responses to recognized and newly emerging sustainability issues are fundamental in achieving the SDG.
An example of an intervention rooted at the global level is 'certification', which can be seen as a response to a loss of trust between producers and consumers. With global commodity trade driving agricultural expansion into tropical forests, efforts to influence the drivers of land-use change require globally applicable standards, regulations and incentive mechanisms (Mithöfer et al. 2017). Major pathways of land-use change for tree crops have been described by Meyfroidt et al. (2014), where the land was either directly derived from forests or converted from pasture or crop land (that had historically been derived from forests). The production of commodities, such as tropical timber, cacao, coffee, palm oil and rubber, is often associated with negative impacts on biodiversity, ecosystem services and social wellbeing, especially when production stems from a direct conversion pathway ). There have been abundant efforts across the tropics to improve local land governance, including using a 'landscape approach', that promotes change towards sustainable production and protection systems Minang et al. 2015). Whilst these efforts have been generating invaluable lessons that encourage replication in other areas with similar characterizations of human, ecosystem and tree-cropping patternswhich constitute extrapolation domainssystematic management of knowledge is not yet available.
The multiplicity of pathways towards solving the challenges facing sustainable development needs to be better understood if we are to secure a joint future for humanity and other life on the planet (Steffen et al. 2015). However, much current data and research are context-, system-and time-specific, highlighting the need for more global evidence on feedback and trends to inform the decisions of policymakers and others about land use and related issues. Such evidence also requires fine-tuning to the dynamics of any given system, where decisionmakers operate often without the time or budget for detailed studies and data collection. For example, solutions that are presented to decision-makers as opposites from which to choose (such as 'land sparing versus land sharing': the segregation or integration of multiple land uses within a landscape) may share more commonalities once examined in a particular landscape (Lusiana et al. 2012), suggesting that the empirical basis for generic theory is weak. Within this context, we argue that trustworthy extrapolation domains for data need to be established, maintained and communicated. 'Case studies' are not able to provide wide-ranging conclusions, regardless of data analysis and metastudy compilation methods if the bias in their site selection cannot be quantified.
Case researchers do not aspire to select cases that are directly representative of diverse populations and they usually do not and should not make claims that their findings are applicable to such populations except in contingent ways. (George and Bennett 2005) Decision-makers aiming for immediate results may hastily try to replicate success stories from case studies without sufficient contextualization within local, regional and global domains. Extrapolation errors from 'case study' data can affect the three dimensions of usable knowledge (Clark, van Kerkhoff, et al. 2016;Clark, Tomich, et al. 2016): (1) salience of the underlying research (not being relevant to decisions in the actual context); (2) credibility (underlying data or system descriptions might appear to be unrelated to  (Sachs et al. 2016) and the difference in forest cover between what is reported and expected on the basis of human population density (Köthke et al. 2013); SL indicates countries included in the sentinel landscape portfolio.
The prevailing extrapolation mode is to describe individual case studies and identify general patterns through 'meta-studies' (Van Vliet et al. 2016). This method has several weaknesses. The geographic distribution of case studies quantifying ecosystem services in Africa, for example, probably tells more about the areas to which researchers have easy access rather than those where underlying issues are the most prominent and important (Kuyah et al. 2016). As methods tend to vary between studies as well as context, it is hard to judge whether differences in results reflect the bias and sampling and measurement errors of the researchers' methods or real variations in conditions and trends. In addition, further bias can be introduced through selective publication of 'significant' results (Forstmeier 2015;Kicinski et al. 2015).
A number of efforts have been made to harmonize methods in global comparative studies on the patterns and drivers of deforestation, and accompanying loss of biodiversity and ecosystem services, in relation to economic development and poverty alleviation. By far, the most extensive networked tropical studies have been carried out by the ASB Partnership for Tropical Forest Margins (ASB), Poverty Environment Network (PEN) and, more recently, the CGIAR Research Program on Forests, Trees and Agroforestry (FTA).
Since 1993, ASB has engaged with landscapes in Peru, Brazil, Cameroon, Indonesia and Thailand (Murdiyarso et al. 2002;Tomich et al. 2005Tomich et al. , 2007Clark, Tomich, et al. 2016;Gillison et al. 2013;Minang et al. 2014a). PEN had study sites in 24 countries Jagger et al. 2014;Wunder et al. 2014). Other research efforts, such as the Landscape Mosaics project ) and the Diversitas Agrobiodiversity network (Jackson et al. 2012), partially overlapped with ASB and PEN in their choice of study sites. The breadth and value of the data gathered from such extended networks can be seen, for example, in statements from PEN, such as '28% of rural income is derived from forest' , which has associated confidence intervals based on variation within the data set (3 continents, 24 countries, 58 sites, 333 villages and 7978 households). This provides a first 'post hoc' perspective on extrapolation possibilities within the domain of the data set.
However, the bias in extrapolating data beyond a research domain, area or geographic location to the level of the whole tropics is often difficult to assess because it depends on the 'external validity' of the portfolio of sites. As a consequence, the adoption of research findings and quantitative estimates in policies risks overstating or understating the evidence. But such problems can be avoided if the selection of research sites is part of a stratified sampling design within the desired extrapolation domain, using insights into (or hypotheses of) major dimensions of variability. The results from such a design can be used jointly with recommendations based more explicitly on theory to inform policies and interventions.
Against this background, we will here discuss the FTA research programme initiated in 2011. At the beginning of the research programme, it was deemed important to develop a portfolio of so-called sentinel landscapes that could be used to monitor patterns of change across the tropics and subtropics. The ambition was to integrate research across multiple scalestrees, farms, landscapes and governance domainsand provide international public goods, including 'change of theory' if existing 'theories of change' were found inadequate. The 'sentinel' terminology was derived from public health, where research to recognize generic patterns can lead to more effective prevention rather than only curing individuals (Shepherd et al. 2015). We will here focus on two questions: (1) Can the 'forest transition' theory, in combination with climate and watershed data, be used for a rigorous typology of interactions between humans, ecosystems and tree cropping in which any set of sampling sites can be compared to the global frequency of similar conditions? (2) What is the external validity of ASB, PEN and the FTA sentinel landscape portfolio for the pantropical extrapolation domain, according to such a typology? What are the consequences of the representativeness for the conclusions drawn from the three global comparative studies?
In the discussion, we will further focus on how useful the derived typology is for empirical data collection aimed at advising extrapolation domains from lessons learned locally and at drawing global, scientifically rigorous conclusions that can inform development and increase impact in FTA or similar research-indevelopment programmes. Before explaining the methods used in answering these questions, we provide further background on the forest-transition theory and biases in Theories of Induced Change (ToIC).

Forest-transition theory
Forest-transition theory ( Figure 2) documents, describes and tries to explain non-linear changes in tree cover (loss of natural forests followed at some point by an increase in planted and managed trees) across the development trajectory of a country (Mather and Needle 1998). A spatial pattern of trees close to and far away from settlements was interpreted by Von Thünen (1842) on the basis of distance-dependent costs of maintenance and protection versus value of the resource as a Theory of Place (ToP). Common versions use time, population density or distance to cities as the X-axis (Van Noordwijk et al. 1997;Angelsen and Rudel 2013;Köthke et al. 2013;Matthews et al. 2014;Minang et al. 2014a;Jouvet and Wolfersberger 2015). The theory provides a unified scheme for the multiplicity of changes (including in direction and rate of change) of tree cover that can take place at local and national scales. It offers a solid basis for a stratified sampling scheme at the forestry-agriculture interface because it logically connects direct, observable patterns (changes in forest and tree cover) with the drivers of change, the consequences and the response options (Van Noordwijk et al. 2011). As detailed elsewhere (Van Noordwijk and Minang 2009), however, discrepancies exist between commonly used national and international definitions of 'forest', leading to confused time-series data sources (Grainger 2008). Common institutional concepts of forests are a complement of and incompatible with agriculture (Van Noordwijk et al. 2008). Meanwhile, tree-based concepts recognize the dichotomy as a continuum (Zomer et al. 2016). 'Agroforestry' and 'Community-Managed Forests' emerged institutionally as 'boundary object' (Van Noordwijk 2014). The lack of distinction between natural and planted forests (Sasaki and Putz 2009) has been a major obstacle to creating policies for sustainable forest, plantation and agroforest management. The confusion is substantial when the term forest is combined with the concept of 'restoration' (Chazdon et al. 2016), as either a return to the left or progression to the right side of the forest transition curve. In the midst of these terminological confusions, the term 'tree-cover transitions' has emerged as empirically observable relative to forest transitions (Ordonez et al. 2014;. However, given the common use of 'forest' transition, we will use this term throughout to refer to 'tree-cover' transitions. Ultimately, most human land use falls outside of the various categories of 'forest' even though it begins with semi-permanent or permanent cultivation within a forest that leads to loss of tree cover and, finally, reclassification of land as 'non-forest'. This biophysical change interacts with institutional claims of forests as a domain controlled by the state. The 1217 Charter of the Forest in England, detailing royal concessions made in the Magna Carta, included 'disafforestation' as a return of royally claimed land to local control. At the earliest stage of the forest-transition curve, when population density is low, much of the forest is 'old growth' (but not necessarily 'pristine' because humanity has had a long and complex history of engagement with forests) and human occupancy is largely determined by biophysical suitability. At low population densities, shifting cultivation (also known as swidden or fallow rotation) has often been an effective forestmanagement practice . Extraction of forest products, followed by conversion to other land uses, leads to further modification, degradation and deforestation, providing financial resources for more economic development. It is worth noting that large-scale logging of forests has, despite claims of sustainability and participatory forest management, rarely led to sustainable use beyond a first or second logging cycle, despite opportunities to do so (Putz et al. 2012;Bulkan and Palmer 2016). Usually before the last forest patch has disappeared, however, a transition point is reached where tree cover starts to increase (Köthke et al. 2014), either by a 'push' (people move to cities and abandon land which then regrows) or 'pull' (scarcity increases the value of the products and services of the planted or remnant natural forests).
Historically, a wide range of nations have experienced this transition, following changes in population densities and percentages of tree cover that are often associated with the development of new, non-forest, non-agricultural sources of national income. The non-linear temporal pattern of forest and tree-cover change consists of trade-off and synergy stages between environmental and economic functionality (Tomich et al. 2005;. For example, replacement of natural forests by tree crops that feed global value chains may have a small net effect on tree cover but a large one on ecosystem services (Ordonez et al. 2014).
A number of hypotheses on patterns in tropical tree-cover transitions, as well as the consequences for local livelihoods of such changes in quantity and quality of tree cover, ecosystem services and biodiversity, were framed at the start of the FTA (Table 1), with the sentinel landscapes portfolio (see below) as the tool. However, the extrapolation domain of this portfolio has not yet been formally assessed.

Addressing biases in theory of induced changes
In the past, development agencies used 'log frames' as a planning tool, trying to manage activities towards achieving planned positive outcomes and impacts beyond a project's lifespan. The terminology of ToC has become common (Weiss 1997;Connell and Kubisch 1998), with emphasis on understanding how and why an initiative is expected to work. This may be more specifically described as a ToIC (Van Noordwijk et al. 2016a). Impacts are quantified as the difference in system performance between the trajectory of a system without intervention (but subject to external variability and change), described with a ToC, and a trajectory with planned interventions on top of external variability and change, described by a ToIC. The impact literature commonly recognizes the non-intervention counterfactuals as the weakest part of the logical chain, especially for natural resource management and policy research (Renkow and Byerlee 2010). As there is likely to be multiple feedback rather than a simple causal pathway, the use of information on context (ToP) can clarify our understanding of how systems work in a ToC . ToP starts from 'it matters where you are' or 'context matters' but continues to analyse how to describe and understand the 'where' beyond geographic coordinates or 'what is current context' and 'where else is it like this'. ToP potentially deals with a near-infinite dimensional space. A more pragmatic ToP seeks key dimensions of variation in social-ecological systems' structures, functions and dynamics and suggests that similarity along those dimensions can, to some extent, have predictive power in responses to new technologies, institutions or Table 1. Hypotheses of interest in the FTA to be tested in the sentinel landscape portfolio.
Six hypotheses on patterns in tree-cover transition Tree cover in landscapes changes in quality, quantity and pattern in non-linear fashion; depending on the operational definition of 'forest', tree-cover transitions at certain scales show a 'forest transition' graph of decline followed by recovery (basic forest-transition hypothesis) Temporal tree-cover transitions can be understood as the result of time-variant processes, with increases in human population density (or rather the logarithm of it) linked to decreases in natural forest cover and increases in human development index (HDI) (or other economic indicators) linked to increases in tree cover (population density and welfare hypothesis) The spatial pattern in quality and quantity of tree cover, from urban areas with (surrounding) trees to areas with few trees and open-field agriculture towards remaining natural forests, show more than coincidental resemblance with the temporal dynamics of hypothesis 2 because both patterns reflect benefits derived from tree cover relative to other land-cover types (spatial forest-transition hypothesis) Institutional change from a 'forest' to an 'agrarian' regime of tenure and control is essential for the transition from decline towards increase of tree cover (agroforestation or tenurial-reform hypothesis) What happens in one part of the tree-cover transition is linked at the level of drivers and/or actors to other parts of the landscape because (1) profitability of tree planting depends on access to tree and forest products elsewhere; (2) migration flows modify human population density in sink and source areas and (3) landscape-wide rules instigated to address specific issues in parts of the curve (e.g. 'illegal logging' control) affect actors elsewhere (landscape-linkage hypothesis; the 'sparing' hypothesis that agricultural intensification saves forests is a special form of it) Drivers of tree-cover transition are space and time dependent and knowledge of past drivers in any particular landscape cannot be directly extrapolated to the future; yet, there may be predictability in the succession of drivers (driver-change hypothesis) Six hypotheses on the consequences of tree-cover transitions Land-use types that are part of the tree-cover transition differ in effectiveness of 'provisioning' and 'environmental' goods and services, labour absorption and profitability (trade-off hypothesis, ASB Matrix; Tomich et al. 2005) Tree cover of all types and at all stages is positively associated with buffer functions in ecological, social and economic senses, with the spatial pattern and degree of integration linked to human resilience and adaptive capacity in the face of climate and market variability (integration, buffer and resiliency hypothesis) Appreciation of tree cover and its associated ecosystem services varies with gender, wealth, cultural backgrounds, ecological knowledge and exposure to extreme events, leading to a diversity of opinions, preferences for the status quo and possible changes in tree cover ('diversity of stakes' hypothesis; includes gender specificity) Feedback mechanisms from beneficiaries of (certain types of) tree cover to the drivers or agents can take multiple forms (rules, incentives, suasion, investment in value chains and technology) and need to be evaluated in the interaction between instruments rather than as specifically targeted approaches ('no silver bullet' hypothesis) Dynamics of tree-cover changes can be influenced by multistakeholder negotiation-support processes that recognize multiple knowledge, perceptions, stakes, powers and influences (negotiation-support hypothesis; includes gender specificity) Public discourse on aspects of tree-cover transition and the relevance of interventions follows a policy-issue cycle, with different opportunities for knowledge-based analyses to support and influence the emergence of transparent, effective, efficient and fair solutions, linking political will to actionable knowledge (impact-pathway hypothesis) macroeconomic drivers. It can also help in seeking counterfactuals for quantitative impact studies, which is essential for finding evidence relevant to policymakers. The forest transition theory (Figure 2) as used by the FTA is a good candidate for use as both a ToP and a ToC. It provides a logically appealing account of the diversity of situations (ToP) as well as plausible directions of historical change (ToC) and addresses the livelihoods, landscapes and governance dimensions of land-cover change. Its use as ToIC, however, remains unclear. FTA has confronted this theory of facts at scale by grounding extrapolation in the sentinel landscapes.
Bias in scientific evidence can stem from any one of several steps in a research cycle. Pantropical or global research for sustainable development that examines (1) scientific theories of endogenous change and (2) development-oriented ToIC typically involves a staged selection (moving downwards in Figure 3) of places for observation (within an explicit ToP), application of methods and collection of data sets.
Moving upwards in Figure 3 to derive conclusions to be used as 'evidence' requires a staged reintroduction of stratum weights and bias corrections. The first steps make use of the variability within the data set, with many statistical techniques at our disposal to assist in interpolation. The confidence intervals for conclusions derived from any stratified sampling depend on variability of data within a stratum plus uncertainty in estimates of stratum weights. Beyond the locations in which observations were made, however, there is an issue of 'external validity' and 'extrapolation domains'. These validity steps can be approached in two ways: by the (1) process (e.g. strict randomization with replication) used to derive the next level in the hierarchy or by (2) analysis of representativeness within global data sets. With existing global data sets on human population, forest and tree cover, elevation, climate, watersheds and soils plus national statistics on trade, economies and human development index (HDI), for example, the bias involved in a specific combination of case studies can be estimated and corrected for.
Here, we restrict our analysis to the first steppost-selection checks of representativenessfor the three portfolios of tropical landscapes, testing that means, variance and parameter distributions in the global data are conserved in the subsets, similar to what would be expected for a random sub-sampling of the pantropical domain.

The sentinel landscapes portfolio
The selection of landscapes for the portfolio (Table 2) was a complex process in which existing research networks, access to historical data, geographic balance, institutional agendas of the main FTA partners and insights of the researchers all played a role. Although the process of selection could not rely on any explicit and formal method of assessing pantropical representativeness, the researchers reflected on an analysis of representativeness in several dimensions of the PEN and ASB data sets. FTA partners were invited to propose landscapes and all were evaluated for their internal validity as well as potential contribution to the global set. Two landscapes that had not been selected in a first round were added subsequently (Nile-Congo and the Miombo) to ensure a better mix of African landscapes. Table 2 lists the names and locations of the 10 landscapes in the portfolio.

Sources of data
The extent of the study is the terrestrial areas of the tropics, between 23°26′13.8″N and 23°26′13.8″S. The  Global statistical data at country level were derived from • HDI at country level: UNDP, 2014. Muthayya et al. (2013) showed that the 'hidden hunger index' is tightly correlated with HDI; we did not include this index in addition to the HDI.

Preprocessing of raw data
All global spatial data were clipped to the tropics as defined above and reprojected into Mollweide projection to accommodate the semi-global nature of the data. We kept uniform cell size of all the raster data according to the highest resolution data layers, that is 300 × 300, through resampling procedures. Similarly, all the presence point data and the sites were transformed into the same projection system. As a proxy of human well-being, we used the HDI 2014 at country level. The statistical data were converted into a raster map according to the political boundaries of countries. For PEN sites, 5 km surrounding the centre of villages was considered as the extent. For ASB sites, 35 km was used.

Analyses
We conducted three main analyses, which were mostly spatially explicit. Beyond that, extraction from spatial data into tabular formats was conducted in a GIS environment and further analyses were conducted using statistical software.

Agro-ecological zonation.
The first main factor we adopted in characterizing the agro-ecological zones was the aridity-humidity index, that is the ratio of precipitation and potential evapotranspiration at an annual time scale. The higher the index, the more humid the area [Zomer et al. (2016) renamed the index from 'aridity' to 'aridity-humidity']. We modified the CGIAR-CSI (2012) classification by adding a water-tower category (Table 3).
Considering the importance of spatial interconnectivity in transporting water from more elevated areas to the surroundings, the concept of a 'natural water tower' is crucial in landscape characterization. In identifying water towers, we needed to take into account variations in elevation within basins (the tower) in combination with humidity (the water). An area is considered a water tower if (1) it is either humid or (per-) humid, that is the aridity-humidity index is >0.65 and as such expected to generate river flow; and (2) the aridity-humidity index relative to elevation was >2.77, with relative elevation defined from the percentiles of elevation within each basin.

An operational forest-transition typology derived from global data sets.
Expanding on existing analyses of forest-transition theory at national scale (Köthke et al. 2013(Köthke et al. , 2014, we changed the unit of analysis from country to sub-basin scale, which aligned with ecosystem services and local land governance within nested subnational, national and global systems. We conducted a cluster analysis of subbasins based on two main factors: (1) land-use/cover composition and (2) forest configuration.
Land-use/cover composition was derived from global land-cover data of 2006, with simplified legends for closed forest, medium forest, open forest, shrub land, grassland, cropland, mixed cropland and trees, sparse vegetation, urban and bare land. The fraction of each type in the terrestrial part only was calculated for each sub-basin to quantify land use/cover composition with multiple variables. We adopted the forest-configuration approach coined by Chomitz et al. (2007) as an important proxy for anthropogenic activities with regards to forests. From the global land-cover map, we produced a forest map by reclassifying land-use/cover types into 'forest' and 'non-forest' classes. We then delineated pixels of forest as polygons of forest patches and calculated the areas and proximities. Based on this, we defined pixels into three types: (1) 'forest core', if the pixel belonged to a forest patch with an area of 20,000 ha or larger; (2) 'forest edge', if a forested pixel was located less than 10 km from the closest forest core area; and (3) 'forest mosaics', for the remaining forested pixels. For each sub-basin, we then calculated the proportions of the three types of forest configurations.
For each ecological zone, a K-means cluster analysis was conducted for sub-basins based on the information on land-use/cover composition and forest configuration. The results were then combined to group tropical sub-basins into six typologies to be used as a proxy of the stages of forest transitions.

Similarity modelling of four commodities.
Four pantropical commodities prominent in global trade (coffee, cacao, rubber and oil palm) are important elements in tree-cover transitions, both as drivers of change (deforestation) and as influences on re-emerging tree cover. We used a maximum entropy model to derive the areas most similar to where the crops were known to exist; this method is suitable for presence data modelling at the regional level (Phillips et al. 2006). MaxEnt software, the engine of the model, was run on a Q-GIS platform and integrated into Land-use Planning for Multiple Environmental Services software (Dewi et al. 2015) to reduce human error and increase efficiencies in data preprocessing and the iterative processes for producing the final model.
Our independent layers were (1) elevation, (2) agroecological zones and (3) climate. From the literature, we selected four variables within the WorldClim data set specific to the commodity species to be modelled based on the importance within the climatic envelopes of the four commodities, according to the literature (Ovalle Paterson et al. 2015;Ray et al. 2015;Schroth et al., 2016) and personal communication with experts ( Table Supp2).
The presence data points were collected from many sources, described in the supplementary information below. The most limited presence data we had was for oil palm, which mostly came from Southeast Asia. The modelling exercises were conducted separately for each region and each commodity species.

Tests of representativeness of ASB, PEN and sentinel landscape sites
Representativeness of the three data sets (ASB, PEN and the sentinel landscapes) was judged from correspondence in the fractions of areas and human AE: product of aridity-humidity index and relative elevation within basin. Source: Modified from the CGIAR-CSI (2012) classification of the aridityhumidity index (rainfall/potential evapotranspiration).
populations between pantropical data and the three sampling frames. Beyond these fractions, the relationships between ecological zones and/or forest-transition stages and properties, such as HDI, distance to protected areas and tree-crop distribution, were evaluated to estimate the bias in each.

Forest-transition stages
The six forest-transition stages grouped sub-basins across the tropics according to forest fraction, forest configuration (core, edge and mosaic forest) and the logarithm of population density (4A). In stage 1, core forest covered around 80% of the total area of the sub-basins inhabited by human populations with densities below 1 km −2 . As population density increased, the fraction of core forest tended to decrease while the fraction of non-forest increased. Each stage was well separated from others with regards to human population density. Based on the fractions of core forest and non-forest areas, stage 1 was clearly separated from stages 2 and 3, and further with stages 4-6, while the last two groups were separated from each other but not within each other. Fractions of edge forest and mosaic forest did not discriminate the stages and did not seem to correlate with human populations at sub-basin level, indicating the two variables are merely intermediate states between core forest and non-forest. The further stages were associated with increments in the logarithm of human population density and core forest and non-forest configurations but also with more specific characteristics of land cover (Figure 4(b)). The fraction of medium broadleaved forest and the fraction of tree cover decreased almost monotonically from stages 1 through 5 and increased again from stages 5 to 6, while the fraction of mixed tree cover followed exactly the opposite trend, suggesting an inflection point at stage 5. The fraction of cropland consistently increased and the fraction of shrub land was low at both ends and plateaued at stages 2-5. The initial increase in shrub land can imply logging as a driver of deforestation and/or non-permanent land uses that left degraded forest behind. The reduction of shrub land in the later stages implies that (planted) tree cover catches up with deforestation while further expansion of agriculture shifts to such land rather than remaining forest. Similar to the characterization of stages of forest configuration above, stage 1 was clearly separated from others with regards to fraction of tree and forest covers, while stages 2 and 3 were only barely separated based on the fraction of tree cover only. Stages 4-6 were quite well separated with regards to most variables. Figure 4. Fraction of forest and tree cover (a) and forest configuration (b) across typologies at subwatershed level; absolute (c) and relative (d) distribution of land across forest transition stages for the various ecological zones 5.2. Agro-ecological zones and human population density Distribution of human populations across the agroecological zones is presented in Figure 5. Compared to the average for the tropics as a whole, human population densities were more evenly distributed in the semi-arid to humid zones and more concentrated in the (per-) humid zone ( Figure 5). In the (hyper-) arid zone, this concentration was even stronger, with most people living in cities.
Scatterplots of (log) human population densities and tree cover at sub-basin level showed a clear downward trend but the highest population densities were associated with approximately 20% tree cover ( Figure 6). This level of tree cover is still possible in (peri-) urban areas while primarily agricultural areas can have around 10% tree cover (Zomer et al. 2016). Agro-ecological zones clearly influenced the interactions between human population densities and tree cover ( Figure 6 (c,d)). Tree cover in the (hyper-) arid and semi-arid zones was much lower compared to those in other zones with similar levels of human population densities.

Tree crops
The relationship between human population density and the potential distribution of the four major tree crops (Figure 7) reflected the agro-ecological zones in which they primarily occurred. Relatively high population densities were associated with coffee and oil palm areas, and lower ones with rubber and cacao, even though the climatic requirements of oil palm overlap with those of cacao and rubber but not with those of arabica coffee.
Coffee (Coffea arabica) was associated with watertower zones with relatively high human population densities (Figure 7, 8) while oil palm (Elaeis guineensis) was effectively restricted to the (per-) humid zone, which it shares with rubber (Hevea brasiliensis) and cacao (Theobroma cacao). Similar areas suitable for the commodities were compared to the actual harvested areas, taken from FAO (2013) statistics for cacao, rubber and oil palm, and from Bunn et al. (2015) for coffee. The results were that conditions similar to where arabica is now grown feature in 3.56 million km 2 or about 57 times the current harvest area of 0.0625 million km 2 (Figure 9). For cacao, the potential area is 3.96 million km 2 or 40 times the current harvest area of 0.100 million km 2 . Rubber could be grown on 5.47 million km 2 or 50 times the current harvest area of 0.110 million km 2 . And oil palm could expand to 1.84 million km 2 or 19 times the current harvested area of 0.180 million km 2 . Climate change will involve shifts in the geographic boundaries for all tree crops but future expansion is still most likely in areas with similar conditions.

Representativeness of the three sampling sets
The relative area in the various stages of forest transition differs between the agro-ecological zones, as the operational definition of these stages depends on the specific conditions of the zone (Figure 10). In the (hyper-) arid zone, the forest-transition classification does not apply as there is too little forest; the (per-) humid zone has a larger fraction in the forest-transition stage 1 than any other zone. In the water-tower zone, stage 4 dominates, which is a shift to the right compared with the humid and (per-) humid zones with similar rainfall.
The relative proportions of area in the various ecological zones are fairly well represented in the sentinel landscapes portfolio once the exclusion in the latter of the (hyper-) arid zone is taken into account while for both ASB and PEN, there are strong spatial biases, over representing the (per-) humid zone ( Figure 11). In terms of human population, however, the sentinel landscapes underrepresent the (per-) humid zone and, by implication, the average human population density in the (per-) humid zone in the portfolio differs from that in the tropics as a whole. In this zone, human population density is more heterogeneous than in the other zones: there are both relatively large areas of low population density (stage 1) and high concentrations of people in stage 6.
In terms of forest-transition stages, the sentinel landscapes underrepresent the lowest tree cover, stage 6, but are again less biased than ASB and PEN ( Figure 12) while the representation of human populations is more biased than for area. Because human population densities tend to have a lognormal rather than normal distribution, we considered the logarithm of population density in this discrepancy. Figure 11 shows that sub-watersheds in the sentinel landscapes span the full range of tree cover (0-100%) but might lack the extremes of both very high (cities) and very low (core forest) human population densities. The relationship between ecological zones, human population density, HDI and distance to protected areas within the sentinel landscapes corresponds well with those for the tropics as a whole ( Figure 13) while conclusions based on ASB and PEN sites may be affected by considerable spatial bias. Similarly, conclusions with regards to tree crops derived from sentinel landscapes' data match those for the tropics as a whole much closer than is the case with ASB or PEN data (Fig. Supp1). Table 3 shows interactions between human population density and agro-ecological zones in shaping ES issues and options to address them, along with the priority areas of FTA.

Discussion
In response to the first question raised in the introduction, we conclude that the forest-transition concept and associated stages can, in combination with climate and watershed data, be used as a robust typology (ToP) of the range of human populations and agro-ecological zones and tree crops encountered in the tropics. Increases in human population densities up to about 30 km −2 are associated with reductions in tree cover down to about 50% (Figure 12), in line with the regression line Note: (a) for all sub-watersheds differentiated by forest-transition stages; (b) only for sub-watersheds of stages 1-5; (c) for sub-watersheds within (hyper-) arid and semi-arid zones; (d) for sub-watersheds in the remaining agro-ecological zones. established at national scale by Köthke et al. (2013). The highest population densities can have a tree cover of around 20% while the lowest tree cover is found with lower population densities (probably in intensively used agricultural areas). Whether and how these spatial patterns align with qualitative tree-cover transitions (from diverse naturally established to selected trees planted) in rural-urban continuum perspective remains to be tested as part of a ToC and before ToIC can be constructed.
Within this typology, any set of sampling sites can be compared to the global frequency of similar conditions, as answer to question 2. In the comparison as described in Section 4.3, the ASB and PEN data sets were shown to have a rather skewed distribution relative to the tropics at large while the sentinel landscapes had a smaller, but not negligible, spatial bias that addition of more landscapes could reduce. The exclusion of the (hyper-) arid zone is easily acknowledged and justified on the basis of low forest and tree cover as well as low human population densities. Persistent poverty and low HDI, however, call for attention to options for sustainable development in the zone. Per unit of biomass, trees in this zone have great local value.
Based on this analysis, caution is called for as conclusions based on ASB and PEN data are primarily valid within their sample sites only. ASB is primarily concerned with stages 2 and 3 of the forest transition and focused on the humid and (per-) humid tropics. For PEN, the spatial bias to the early stages of the forest transition may colour conclusions about the roles of forest in the    relationship between poverty and environment. While the researchers are generally aware of this, the subsequent use of quantitative conclusions may lose track of the caveats and state presumed general relationships on the way poverty and forest relate to each other.
Our results indicate under-studied areas in several agro-ecological zones in the existing global comparative work, especially in drylands (which constitute 18% of the total area and 3% of the human populations in the tropics). Prioritizing future work in drylands will enrich the research. Further, the upward curve of the forest transition deserves further study. The sentinel landscape portfolio could consider including a better representation of the processes the curve represents. For instance, the review by Hecht et al. (2016) emphasizes the climate relevance, beyond carbon storage, by trees in urban and agricultural environments. Given the global movement to restore forests and landscapes, in the future, this forest-transition stage will deserve more effort, especially in delivering ecosystem services in highly populated areas. The recent surge in interest in a two-way relationship between vegetation and rainfall (Ellison et al. 2017) prompts further scrutiny of the sentinel landscape portfolio to include sources and sink areas for terrestrially recycled atmospheric moisture.
In the type of analysis presented here, we have been restricted to credible global data sets. A challenge is to link these data with country-level characterizations of important drivers for policy and legal frameworks. This is especially true for larger and very heterogeneous countries, such as India, Indonesia and Peru. An additional challenge is to account for the direct and indirect influences of international market forces and regulations and link these drivers to the data. Yet another challenge is to include counterfactuals for global drivers (especially if isolated from their specific temporal context), which could require different samplings strategies. Finally, most of the meta-studies of land-use change have  focused on the consequences for biodiversity and biogeochemical cycles of changes from forest to agriculture; meta-studies of the socioeconomic consequences of the further change to (peri-) urban settings are rare (Van Vliet et al. 2016).
Returning to the question how useful the derived typology is for empirical data collection aimed at usable knowledge, we have so far only addressed the external validity of data (Step 1 in Figure 3) as ex ante analysis within a global ToP for the tropics. Before the data collected in any of the sampling frames can be used for valid conclusions about hypothesessuch as those framed in Table 1 the sampling processes (steps 2 and 3 in Figure 3) need to be assessed for additional biases. The noticeable continental-scale differences in relations between tree cover, climate and human population densities noted by Zomer et al. (2016) suggest major variations at that scale. At the next scale, the diversity of actors and their interactions is the key to understanding the socialecological systems involved (Galudra et al. 2014;Minang et al. 2015). Such actors can, in a first approximation, be grouped into (1) large-scale companies, which directly respond to global markets; (2) medium-scale actors who generate income from other sectors and invest in land acquisition, either through purchasing from smallholders or paying local actors to obtain land in remote areas and provide labour and (3) smallholdersboth indigenous and migrantswho mostly generate their livelihoods from on-farm activities. Whereas for analysis of extrapolation domains, we necessarily rely on global legends in classifying land cover and land use, research that aims to be locally relevant will have to use much richer locally determined typologies of land uses (Beaudoin et al. 2016) and the social and economic actors, and methods to analyse their interactions.
In various contributions to this special issue on eco-certification (Mithöfer et al. 2017), the network of sentinel landscapes across ecological zones was used to provide a balance in the evidence because previous studies tended to focus on areas where certification arose as a response to external concerns over local commodity production, often combining social and ecological issues. With studies on tropical timber, rubber, oil palm, coffee and cacao, the sentinel landscapes provide a sampling frame that includes places where several of the commodities interact outside their core areas, and in which any one commodity might be the dominant source of livelihoods in the interaction with large-scale operations. The synthesis Figure 13. Comparison of key landscape characteristics (human population density, HDI and distance to nearest protected area) for ecological zones and forest-transition stages, as derived for the tropics as a whole and evaluated on the basis of ASB and PEN portfolios.
paper of this special issue will reflect on the difference between conclusions from the FTA-SL's and those from published case studies in this light.
The sentinel landscapes have the potential to enable a further delineations of ToP and articulation with ToC, linking the questions of (1) who? (people, migration histories, land rights, tenure security); (2) what and how? (land-use systems, their profitability, labour absorption and ways of meeting local requirements for food and energy) and (3) where? (spatial organization of landscape elements, for example, aligned with soil toposequences or distances to roads and markets) to the basic three elements of a ToC: (1) so what? (consequences for ecosystem services); (2) who cares? (stakeholders in status quo and those negatively affected) and (3) the opportunities for leverage on the drivers of change . The tools exist for an analysis of these questions across the knowledge systems of local communities, policymakers and researchers ) to allow comparison with baseline data collected from households and spatial sampling points (Vågen et al. 2015). A deeper understanding of how 'induced' change can effectively build on the spontaneous change in these complex socio-ecological systems is needed to make progress across the full spectrum of the SDG.