A practical approach to measuring the biodiversity impacts of land conversion

Further progress in reducing biodiversity loss relies on the improved quantification of the connections between drivers of habitat loss and subsequent biodiversity impacts. To this end, biodiversity impact metrics should be able to report linked trends in specific human activities and changes in biodiversity state, accounting for both the ecology of different species and the cumulative effects of historical habitat losses. These characteristics are not currently captured within a single metric. Here, we develop a globally applicable methodological framework that uses freely and publicly available datasets to quantify the relative impacts of anthropogenic activities on biodiversity. We use species‐specific habitat suitability models to link specific land uses to ensuing changes in the likelihood that local populations of those species will persist. To illustrate our approach, we assess the impacts of soy expansion and other land uses within the Brazilian Cerrado on over 2,000 species of amphibians, birds, mammals and plants for three periods between 2000 and 2014. Our results showed that mammals and plants suffered the greatest overall reduction of suitable habitat. However, among endemic and near‐endemic species—which face greatest risk of global extinction from habitat conversion in the Cerrado—birds and mammals were the most affected groups. While conversion of natural vegetation to grassland and planted pastures were together responsible for most of the biodiversity impact of recent changes, soy expansion (via direct conversion of natural vegetation) had the greatest impact per unit area. The total biodiversity impact of recent land‐use change was concentrated in the southern states of the Cerrado—Minas Geráis, Goiás and Mato Grosso—but the impact on biodiversity of production of soy was greatest within the agricultural frontiers of Bahía and Piauí. The flexibility of our approach to examine linkages between biodiversity loss and specific human activities has clear potential to better characterize the pathways by which habitat loss drivers operate. Its capacity to incorporate species‐specific ecological needs, through a globally applicable methodology, can improve the tangibility of biodiversity loss assessments.


| INTRODUC TI ON
Metrics that allow the impact of human activities on biodiversity to be monitored and reported are essential tools in global conservation efforts. They have revealed unprecedented rates of species extinctions (Butchart et al., 2010), the extent of local population losses (Hill et al., 2016) and the contribution of different sectors (such as agriculture) towards these impacts (Hoekstra & Wiedmann, 2014;Moran & Kanemoto, 2017). With such evidence, habitat loss from land-use change is now recognized as the biggest contributor to biodiversity decline and this information supports multiple international initiatives for the protection of biodiversity on land and at sea (CBD, 2012;Millennium Ecosystem Assessment, 2005;Ramsar, 1971). Yet, while quantifying the scale of global biodiversity loss remains vital, more work is needed to develop tools for identifying and tracking the drivers of land-use change that underlie such trends.
Attributing biodiversity losses to specific drivers has proved challenging. Actionable information requires that biodiversity impact estimates can be derived at scales at which information on anthropogenic activities and drivers are available, and decisions are made (Guerrero, McAllister, Corcoran, & Wilson, 2013). Biodiversity, however, is a multifaceted construct and difficulties in measuring changes in its state still limit the effective guidance of conservation efforts (Sparks et al., 2011). Most common techniques for measuring biodiversity changes provide an estimate of species richness loss, which poses important shortcomings for biodiversity conservation (Hillebrand et al., 2018). Of particular importance is the inability to predict extinction risk for individual species. Due to the absence of species identity, richness metrics fail to incorporate species-specific information such as their distribution and ecological requirements.
The absence of this information limits the ability of the metric to be adapted to different spatial scales (Veach, Di Minin, Pouzols, & Moilanen, 2017). This is required to assess drivers of biodiversity loss as it demonstrates the linkages between changes in the state of biodiversity and specific human activities .
Furthermore, conservation decisions rely on metrics that permit the user to quantitatively differentiate levels of extinction risk among individual species. This requires that the cumulative and nonlinear effects of historical habitat loss are accounted for.
Approaches based on species' area of habitat (AOH; previously known as extent of suitable habitat-ESH; Brooks et al., 2019) show promise in the development of biodiversity impact metrics because they can integrate spatially explicit information on the ecology of individual species with data on the distribution of anthropogenic land use (De Baan et al., 2015;Rondinini et al., 2011). Unlike approaches that estimate potential regional or local loss of species richness, AOH maps are also adaptable to different spatial scales while retaining species-specific information (de Baan, Mutel, Curran, Hellweg, & Koellner, 2013;Rondinini et al., 2011). Specifically, they can quantify the relative change in AOH arising from land conversion, which allows species-specific impacts associated with a particular human land-use change to be calculated (De Baan et al., 2015). AOH is described by the intersection of a species' geographical range with its environmental preferences, measured in terms of variables such as vegetation cover, elevation and proximity to water bodies (Rondinini et al., 2011). AOH is a key determinant of species extinction risk (Blackburn, Gaston, Quinn, Arnold, & Gregory, 1997;Harris & Pimm, 2008), with reductions in AOH affecting the persistence of local populations (Mantyka-Pringle, Martin, & Rhodes, 2012).
An important benefit for applied work of using species-specific AOH in the assessment of biodiversity impact is that it can allow for incorporation of the cumulative, nonlinear effects of habitat loss on species persistence. Impacts of habitat loss are not typically linear because as the area of habitat diminishes the effect of losing each additional hectare of habitat increases (Kitzes & Harte, 2014).
Failure to account for this cumulative effect will underestimate the impacts of current habitat loss on species that have suffered historical habitat loss prior to the land-use change in question. This issue has been addressed by a handful of studies that have considered a non-proportional relationship between the extent of remaining habitat and species' persistence (De Baan et al., 2015;Strassburg et al., 2017). However, these have, so far, had limited taxonomic coverage, and use projected land-use changes rather than direct states of the Cerrado-Minas Geráis, Goiás and Mato Grosso-but the impact on biodiversity of production of soy was greatest within the agricultural frontiers of Bahía and Piauí.
4. The flexibility of our approach to examine linkages between biodiversity loss and specific human activities has clear potential to better characterize the pathways by which habitat loss drivers operate. Its capacity to incorporate species-specific ecological needs, through a globally applicable methodology, can improve the tangibility of biodiversity loss assessments.

K E Y W O R D S
agriculture, area of habitat, Brazilian savannah, habitat suitability models, soybean, specieslevel impact observations of habitat conversion. They have also been focused at particular spatial scales; given that land-use data and decisions operate across a range of scales, it would be helpful to develop these approaches so that they are applicable across multiple scales.
Here, using a nonlinear and spatially explicit approach, we describe a biodiversity impact metric designed to provide information on changes in local population persistence, which can be both linked to specific human activities and adapted to scales relevant to different datasets and levels of decision-making. We use freely and publicly available datasets to generate AOH models and develop them in four important ways. First, we account for various levels of information on the ecology of individual species, including the estimated importance of breeding and non-breeding ranges. Second, we incorporate historical losses to estimate the cumulative and nonlinear impacts of further losses. Third, we calculate the marginal value of spatial units of species' habitat so that the impact metric can be adapted across different spatial scales. Fourth, we use crop-specific as well as general land-use maps to estimate the most important drivers of loss. We illustrate the details and power of our approach using the example of the cultivation of a globally important crop, the soybean Glycine max, in the Brazilian Cerrado, a globally important savanna that hosts 5% of the world's species (Strassburg et al., 2017). We use species-specific AOH models for 2,009 species from four taxonomic groups: amphibians, birds, mammals and plants.

| MATERIAL S AND ME THODS
Our approach to calculating a biodiversity impact metric involves four main steps (Figure 1): (a) mapping species' AOH at different time F I G U R E 1 Schematic of method stages specifying input data required, analyses to be performed and resulting outputs points of interest; (b) calculating the proportional losses in AOH due to land conversion, and estimating the resulting change in likelihood of persistence for each species; (c) mapping these marginal changes in AOH, and aggregating them across species and at different scales and (d) measuring the relative contribution of specific human activities to species impacts. Below we detail these steps for soy production in the Brazilian Cerrado.

| Mapping species' area of habitat at different points in time
We produced AOH maps for all amphibians, birds and mammals whose geographical ranges intersect the Cerrado boundary (IBGE, 2004), and for which habitat preferences were available.
Following Rondinini et al. (2011), we clipped these ranges to exclude unsuitable habitat, based on their habitat preferences as coded against the IUCN habitats classification scheme (IUCN, 2018). For each species, we retained only the habitats listed in level 2 of the scheme that were coded as 'suitable'. To calculate and map AOH for each species, we used national land-cover maps for the years 2000, 2010(IBGE, 2014250-m resolution). These were matched to the IUCN habitat classification scheme following a conversion table that allows the reclassification of habitat preferences into land cover categories ('crosswalk'; see Supporting Information Section 1 and Table S1).
Then, using altitudinal preferences from the IUCN Red List and the Shuttle Radar Topography Mission elevation model (USGS, 2006), we excluded land at unsuitable altitudes within species ranges. The elevation map was produced by resampling (averaging) to 250 m the elevation model, originally at 1 arc second resolution (approximately 90 m). For migratory species, AOH was mapped separately for resident, breeding and non-breeding ranges, to reflect seasonal differences in species' habitat requirements. This resulted in a total of 234 AOH maps for amphibians, 846 for birds and 288 for mammals, each at 250-m resolution (the resolution of the best available land cover maps for Brazil with which land-use change can be quantified consistently; IBGE, 2015).
We also calculated AOH for 641 plant species whose ranges intersect the Cerrado. AOH maps for each year were generated by clipping plant geographical ranges (from Martinelli & Moraes, 2013) so they only included those land-cover categories classified as natural by the Brazilian Institute of Geography and Statistics (IBGE, 2014), and excluded other semi-and non-natural categories (Table S1).
For each species, we produced baseline AOH maps against which subsequent proportional losses of AOH were assessed. For the vertebrate groups, we used a map of original vegetation cover for the Brazilian Cerrado, which showed the distribution of natural vegetation prior to large-scale cultivation (c. 16th century; IBGE, 2004).
This reclassification was also performed using a conversion table (Table S1) and was restricted to each species' geographical range.
For plants, the entire geographical range intersecting the study region was considered its baseline AOH.

| Calculating proportional loss of species' AOH, and estimating changes in species' local likelihood of persistence
For each species, we calculated the proportion of its baseline AOH remaining at each subsequent point in time-years 2000, 2010, 2012 and 2014. We translated the change in AOH between years into an associated change in likelihood of persistence. The impact on population persistence of losing a given amount of AOH increases as total AOH decreases, resulting in a concave relationship between remaining AOH and local persistence (Kitzes & Harte, 2014). Although the detailed form of this relationship has yet to be investigated, we followed other studies (Balmford, Green, Onial, Phalan, & Balmford, 2019;Strassburg et al., 2017;Thomas et al., 2004) in converting changes in AOH into changes in population persistence using a power-law function with an exponent <1.
Remaining proportions of AOH were used to derive a nonlinear 'persistence score', P, which captures the cumulative effect of habitat loss on the likelihood of the species' persistence in the study region: where E is the remaining proportion of the original AOH, and z is the extinction coefficient. While Equation (1)  The change in a species' likelihood of persistence between two points in time was then calculated as ΔP, the corresponding difference in persistence score values: where E t0 and E t1 are the remaining proportions of AOH at t 0 and t 1 , respectively.
For migratory species, an overall ΔP mig score was calculated from ΔP scores derived separately for the species' breeding and non-breeding AOH. To estimate the overall change in a migratory species' persistence score, we assumed a multiplicative effect of changes in both parts of a species' range, as previously suggested by empirical (Lockwood, 2004) and theoretical studies (Iwamura et al., 2013): where P b and P nb are the persistence scores within the breeding and non-breeding ranges, respectively. For further discussion of the conservation implications of this approach, see Supporting Information Section 2 and Figure S2.
We used a range of z-values to assess the effects of plausible variation in the extinction coefficient (see Supporting Information Section 3 and Figure S3). Our qualitative conclusions concerning the relative role of human activities on estimated biodiversity impacts were not dependent on the choice of a particular value of z. We therefore adopted a z-value of 0.25 in the main text, based upon its ability to predict proportions of species becoming extinct or threatened as a result of habitat loss (Brooks & Balmford, 1996;Brooks, Pimm, & Oyugi, 1999).
We also considered increases in species' persistence scores due to gain of suitable grid cells (e.g. through reversion of converted land to natural habitat). However, reversion is currently on such a limited scale in the Cerrado that incorporating these gains had a minor impact on the results for most of the groups, and was therefore not considered in the main text ( Figure S4).

| Mapping the marginal value of land-use change and its biodiversity impacts across different scales
Although our assessment is not global, we wanted to capture the global conservation consequences of Cerrado habitat loss.
Therefore, species' ΔP-values were weighted by the proportion of their global geographical range that fell within the study region. This assigns more weight to impacts on those species restricted to the biome.
To make the weighted ΔP spatially explicit, we next divided this by the number of cells converted over a given time interval (e.g. 2012-2014), thus capturing the marginal value of the loss of suitable habitat for each cell for that specific time period. The marginal value of land-use change permits estimation of the contribution of one spatial unit of AOH to a given species' likelihood of persistence. By estimating the marginal value of each spatial unit of land-use change, we can then map this as a continuous metric in a gridded landscape and assess the spatial distribution of the overall impact of habitat loss on species persistence. The resulting maps have the flexibility to be combined across species and to be aggregated across any scale of interest. Thus, for a period of time t 0 → t 1 , the marginal value of the loss of AOH within cell j (belonging to a total set of converted cells R), for the weighted persistence score of species k, MV t0→t1,j,k , can be represented as follows: where R is the total number of cells converted from suitable to unsuitable for that species in the period t 0 → t 1 , and w is the weight of species k. MV-values were then assigned to a gridded map, detailing the cells that turned into unsuitable habitat within the corresponding time interval. The resulting distribution maps of marginal loss values for individual species were then overlaid and values summed across species to obtain, for each cell, an aggregated biodiversity impact metric of land-use change. Using maps of administrative boundaries (e.g. municipalities, states), cell-level impact values were then aggregated to give totals for administrative units of interest.

| Measuring the relative contribution of specific human activities to biodiversity impacts
We overlaid the aggregated biodiversity impact maps with land-use conversion maps to attribute impact estimates to soy cultivation and non-soy land-uses. To this end, we combined the soy-expansion maps with IBGE land-conversion maps to distinguish soy expansion from non-soy expansion (see Supporting Information Section 5 and

| Assessing species-level impacts
To illustrate changes in the persistence score at species level,

| Capturing biodiversity impact variation across taxonomic groups
On Areas of particular conservation concern can be further ex-

| Attributing biodiversity impacts to specific land uses
For 2000-2014, our results show that different types of land conversion vary in their impacts across taxonomic groups ( Figure 5).
For instance for birds, mammals and plants, while conversion to grassland comprised on average ~50% of their converted AOH, this was responsible for <25% of the total biodiversity impact for each group (grey line- Figure 5). In contrast, conversion to planted

| D ISCUSS I ON
Four criteria shaped the design of our biodiversity impact metric. We required it to (a) be able to capture the status of different components of biodiversity, while including species-specific aspects such as the distribution and ecological needs, (b) have the ability to account for historical habitat losses to estimate the cumulative impacts of further losses, (c) be scalable so it can be adapted to resolutions at which information on human activities and drivers is available and (d) allow biodiversity impacts to be linked to specific human activities. Below we highlight strengths and limitations of our approach in relation to both these criteria and existing biodiversity impact metrics.

| Capturing the status of different components of biodiversity
Our results illustrate that species' AOH maps provide a practical way to incorporate species-specific information in the assessment of biodiversity loss. Such accounting is an important first step in capturing biodiversity's multiple dimensions within a single metric. Here, we have focused on the ecological dimension, from which we have incorporated two key ecological attributes: global distribution and

F I G U R E 5
The proportional contribution of different land-use conversions to the total biodiversity impact 2000-2014 in the Cerrado for four taxonomic groups, plotted against the proportion of the total habitat loss area of each land-use change. Land-use conversions are plotted in order of increasing ratio of proportional contribution to change in persistence score: proportional contribution to loss of AOH (with the ratios shown in parentheses). Higher ratios thus indicate land-use conversions with disproportionately high impacts on our biodiversity footprint metric given the area converted. This is also reflected in a steeper slope. We aggregated IBGE land-use categories as follows: other crop (than soy), planted pasture, mosaic (mosaic-forest, mosaic-crop and mosaic-shrubland), grassland and other habitat preferences. These are particularly relevant as their inclusion captures variation in individual species responses to different land-conversion types.
By combining information on land cover change, individual species' distributions and habitat preferences, this method identifies which biodiversity elements are most affected and where the greatest impacts have occurred (Figures 3 and 4). When information on habitat preferences is unavailable, as it was here for plants, assumptions on habitat requirements need to be made. If such assumptions are generous-such as that species can occur in a wide range of land covers including anthropogenic ones-there is higher chance of incurring errors of commission (assuming a species occurs where it does not) and hence of underestimating species' risk of local extinction (Rondinini, Stuart, & Boitani, 2005). In contrast, more conservative assumptions are prone to errors of omission (incorrectly assuming that a species is absent) and thus of overestimating impact (Rondinini et al., 2005). Under the precautionary principle, widely adopted in assessing biodiversity risk (Dickson & Cooney, 2005;Myers, 1993), conservative assumptions might be more appropriate.
As our study area covers a fraction of the global population of many species, we weighted species' persistence scores by the proportion of their geographical range that intersects the Cerrado. This assigns more weight to impacts on those species While the wide availability of the data used here makes our method practical and accessible, we acknowledge that the variables we use cannot fully capture the ecological complexity of species' responses to habitat changes. For instance, habitat fragmentation and isolation can be important determinants of species occurrence (Rosa, Purves, Carreiras, & Ewers, 2014) and ignoring such landscape-level information can add further error into species' distribution mapping. Even though information on how species respond to fragmentation and edge-effects is currently absent from the IUCN Red List, recent studies have provided insight in how best to model this (Ewers, Marsh, & Wearn, 2010;Pfeifer et al., 2017). Moreover, direct habitat loss is not the only impact of land-use change, which can, in turn, lead to the expansion of invasive species, over-harvesting and pollution. Where land-use change exacerbates these other threats our impact estimates will again be conservative. However, the key characteristics of our metric computations (e.g. being highly spatially explicit, working at species level and across different scales) could provide a flexible framework for incorporating such indirect impacts into this approach in the future.

| Accounting for historical habitat losses
Most studies that assess the impact of land-use change on biodiversity can be considered snap-shots of habitat loss patterns which do not account for historical losses. However, a 50% loss of current habitat for species A and B is not equivalent if species A has lost 20% more of its AOH than species B prior to the assessment. This is because (a) species that have lost a high proportion of their habitat prior to assessment will already have a higher local

| Aggregating biodiversity impact at different spatial scales
To bring these issues into decision-making processes, tools are needed to capture and translate ecological information to the scales at which decisions are made (Guerrero et al., 2013). The results presented here suggest that our proposed method meets these requirements, by capturing relevant ecological information such as species richness, mean historical habitat losses and endemicity ( Figure S7), which can be adapted to different scales of decision-making. Metrics of impact that are adaptable to different scales of threat information are also likely to be useful in evaluating causal connections between biodiversity impact and human activities. We argue that such adaptability to different scales is one of the major shortcomings of biodiversity impact metrics that estimate species richness loss.

| Linking biodiversity impact to specific human activities
Globally, applicable biodiversity impact metrics have traditionally assessed anthropogenic impacts at sector level-in particular agriculture, harvesting, transport, fishery and mining. While this helps identify sectors to target with conservation efforts, it does not provide sufficient detail to design detailed plans to tackle underlying drivers of biodiversity loss (IPBES, 2019). These drivers operate through specific human activities (Moran & Kanemoto, 2017) and, to mitigate their impacts, it is essential to quantify and map the connections between the consumption that drives habitat loss and its biodiversity impact (Lambin & Meyfroidt 2011). In this study, we focused on soy production as the proximate cause of habitat loss, which is influenced by remote drivers such as consumption patterns (Croft, West, & Green, 2018;de Ruiter et al., 2017), production shortages elsewhere (Godfray et al., 2010) and population growth (Dasgupta & Ehrlich, 2013).
The expansion of many worldwide agricultural commodities and their effects on biodiversity through land conversion are determined by local activities and processes, driven by international policies, trade agreements and consumption patterns. The methodological flexibility of our metric will facilitate the assessment of such drivers, and consequently the support of different decisionmaking contexts. Indeed, the different taxonomic and spatial resolution at which this approach works makes it suitable for assessing links between biodiversity impacts and specific human activities with data of different spatial resolution. For instance, if spatial information on specific human activities is only available at a higher administrative scale (e.g. state), aggregated changes in species persistence can still be linked to the human activity considering appropriate proportional relationship between the extent of both the human activity and the habitat converted. The transparency of our method permits its adaptation according to available data while still delivering applicable and practical information on changes in biodiversity state.

DATA AVA I L A B I L I T Y S TAT E M E N T
Full data have not been archived according to the BES data archiving policy due to restrictions from the data owners. Below we indicate where and how to obtain data that were not archived.  (Durán et al., 2020).