EDITOR'S CHOICE: Surrounding habitats mediate the trade-off between land-sharing and land-sparing agriculture in the tropics



  1. Two strategies are often promoted to mitigate the effects of agricultural expansion on biodiversity: one integrates wildlife-friendly habitats within farmland (land sharing), and the other intensifies farming to allow the offset of natural reserves (land sparing). Their relative merits for biodiversity protection have been subject to much debate, but no previous study has examined whether trade-offs between the two strategies depend on the proximity of farmed areas to large tracts of natural habitat.

  2. We sampled birds and dung beetles across contiguous forests and agricultural landscapes (low-intensity cattle farming) in a threatened hotspot of endemism: the Colombian Chocó-Andes. We test the hypothesis that the relative biodiversity benefits of either strategy depend partially on the degree to which farmlands are isolated from large contiguous blocks of forest.

  3. We show that distance from forest mediates the occurrence of many species within farmland. For the majority of species, occurrence on farmland depends on both isolation from forest and the proportionate cover of small-scale wildlife-friendly habitats within the farm landscape, with both variables having a similar overall magnitude of effect on occurrence probabilities.

  4. Simulations suggest that the biodiversity benefits of land sharing decline significantly with increasing distance from forest, but land sparing benefits remain consistent. In farm management units situated close to large contiguous forest (<500 m), land sharing is predicted to provide equal benefits to land sparing, but land sparing becomes increasingly superior in management units situated further from forest (1500 m). The predicted biodiversity benefits of land sparing are similar across all distances, provided that sparing mechanisms genuinely deliver protection for contiguous forest tracts.

  5. Synthesis and applications. The persistence of bird and dung beetle communities in low-intensity pastoral agriculture is strongly linked to the proximity of surrounding contiguous forests. Land-sharing policies that promote the integration of small-scale wildlife-friendly habitats might be of limited benefit without simultaneous measures to protect larger blocks of natural habitat, which could be achieved via land-sparing practices. Policymakers should carefully consider the extent and distribution of remaining contiguous natural habitats when designing agri-environment schemes in the tropics.


With growing human demand for food and biofuels, the future of global biodiversity is increasingly placed in the hands of agricultural policymakers (Foley et al. 2005; Green et al. 2005; Sayer et al. 2013). A key question in the debate surrounding agricultural development is whether biodiversity will benefit from land-sharing practices, where ‘wildlife-friendly’ habitat features are integrated throughout the farmed landscape, or land-sparing practices, where land is farmed at higher intensities to allow a portion of land to be preserved as intact natural habitat (Balmford, Green & Scharlemann 2005; Green et al. 2005; Fischer et al. 2008).

Both strategies have important and potentially complementary roles in land-use policy (Fischer et al. 2008; Scherr & McNeely 2008; Perfecto & Vandermeer 2010), but empirical studies of their relative merits for biodiversity remain few (Edwards et al. 2010; Phalan et al. 2011; Chandler et al. 2013; Hulme et al. 2013). In the tropics, lower-intensity agriculture can support surprisingly high biodiversity (Ricketts et al. 2001; Hughes, Daily & Ehrlich 2002; Daily et al. 2003; Ranganathan et al. 2008), particularly under land-sharing practices that promote features such as forest fragments, riverine strips and isolated trees within farmland (Fischer et al. 2008). In some cases, elevated biodiversity can benefit food production by providing biological pest control and other ecosystem services (Perfecto & Vandermeer 2010; Tscharntke et al. 2012). However, land-sharing practices can also reduce per hectare food production, potentially increasing pressure to convert remaining natural habitats to agriculture under growing demand for food and biofuels (Fischer, Lindenmayer & Manning 2006; Dorrough, Moll & Crosthwaite 2007; Ewers et al. 2009; Edwards et al. 2010). The alternative land-sparing strategy may thus be more beneficial for biodiversity (Edwards et al. 2010; Phalan et al. 2011; Hulme et al. 2013) and carbon capture (Gilroy et al. 2014a), though it carries its own risks to agricultural sustainability (Matson & Vitousek 2006; Fischer et al. 2008; Scherr & McNeely 2008).

A key gap in current knowledge is whether the biodiversity benefits of land sharing and land sparing depend partially on the proximity of surrounding natural habitats (Norris 2008; Batáry et al. 2011; Kleijn et al. 2011). Some species may persist in agricultural landscapes only when large contiguous natural habitat tracts are present in adjacent areas (Daily et al. 2003; Brosi, Daily & Ehrlich 2007; Tscharntke et al. 2008), either due to source–sink dynamics (Pereira & Daily 2006) or periodic movements between natural and farmed habitats (Evelyn & Stiles 2003; Sekercioglu et al. 2007). For these species, farm-based conservation activities may be insufficient to maintain stable populations unless tracts of natural habitat are also protected. This protection could be achieved via land sparing practices, but not land sharing. Failure to account for the links between natural habitat tracts and farmland biodiversity could lead to spurious inferences about the relative merits of each approach for biodiversity conservation.

Here, we examine how proximity to large forest tracts influences farmland biodiversity under land-sharing and land-sparing scenarios, using birds and dung beetles as model communities. We use field data from the Chocó-Andes of South America, a zone straddling two of the most highly threatened hotspots of global biodiversity (Myers et al. 2000; Jenkins, Pimm & Joppa 2013). We model community dynamics across farmland sites with varying levels of within-farm woodland habitat, assessing how these dynamics are influenced by increasing distance from large contiguous forests. We use the resulting models to predict whether species would benefit from land-sharing or land-sparing agriculture, exploring how these benefits vary as the distance between farmland and contiguous forest increases.

Materials and methods

Study Areas

We sampled three study areas in the departments of Antioquia, Risaralda and Chocó, Colombia (Fig. S1, Table S1, Supporting information). We focus on cattle farming as the dominant land-use in the region, accounting for more than 95% of farmed land at each site, mirroring wider land-use patterns throughout the Colombian Andes (Etter et al. 2006). The study areas span an altitudinal range of 1290–2680 m above sea level, a range typified by subtropical and submontane cloud forest (Armenteras, Gast & Villareal 2003). Each site straddles the interface between farmland and large contiguous tracts of forest (>1 000 000 ha; Fig. S1, Supporting information), dominated by primary forests with some secondary forest cover (age range 6–30 years).

We sampled bird and dung beetle communities at points arrayed within 400 × 400 m squares allocated in proportion to the relative cover of different habitat types across the landscape, summing to 38 squares in contiguous forest (23 in primary, six in secondary aged 15–30 years, nine in younger secondary) and 20 squares in farmland (see Fig. S1, Supporting information). We made no distinction between primary and secondary forests in our analyses. Squares were located randomly within habitats, applying a minimum spacing of 300 m for squares in different habitats and 400 m for squares within the same habitat. Inside each square, we placed individual sampling points at regular intervals along randomly situated triangular transects, with sufficient spacing between points to allow community independence: 100 m for dung beetles (Larsen & Forsyth 2005) and 200 m for birds (Pearman 2002; Hill & Hamer 2004). Random transect placement ensured that individual points were situated randomly with respect to microhabitat variation. All sampling was carried out from January to March and June to July 2012, corresponding with the relatively dry period in the region.

Habitat Variables

Farmland squares incorporated varying levels of remnant woodland habitat, including fragments (size range 0·1–27 ha), riparian corridors, hedgerows and other wooded features (e.g. gardens). We classed these features collectively as ‘wildlife-friendly habitat’, and visually mapped their distribution in a 100 m radius around each farmland sampling point, along with the limits of all grazed pasture and any other non-pasture habitats (Phalan et al. 2011). We used digitized copies of these maps to calculate an index of wildlife-friendly habitat cover W at each point, calculated thus:

display math(eqn 1)

where, for an area of radius r, Fr is the proportion of wildlife-friendly habitat cover, and Pr is the proportion of pasture cover. Other habitats (e.g. roads, buildings) are excluded from consideration in the index. We selected the radius r for each taxonomic group based on previous evidence for the grain on spatial community turnover in tropical forests: 100 m for birds (Pearman 2002; Hill & Hamer 2004) and 50 m for dung beetles (Larsen & Forsyth 2005). All forest sampling points were assigned a value of = 1.

Farmland squares also spanned a continuum of distances from contiguous forest (range 50–1,550 m, see Fig. S1, Supporting information). We estimated the distance from each farmland sampling point to the nearest contiguous forest edge using a ground-truthed map based on ALOS/PALSAR pantropical cloud-free forest cover data (Shimada, Tadono & Rosenqvist 2010). Due to significant inaccuracies in the remote-sensed data, we also directly mapped the limits of contiguous forest across each site using hand-held GPS units, or by visually mapping forest edges for parts of study areas that were inaccessible. We combined these direct observations with the ALOS/PALSAR map to create a layer estimating the current extent of contiguous forest (i.e. excluding isolated fragments) across each site in ArcMap v 10 (Fig. S1, Supporting information). This layer incorporated both primary and secondary forests, which support similar bird and dung beetle communities within the study region (Gilroy et al. 2014b). All forest sampling locations were assigned a distance value of zero.

Bird and Dung Beetle Surveys

We sampled bird communities using repeat-visit point counts at three sampling points spaced 200 m apart within each square (174 points in total). We visited each point on four consecutive mornings for counts of 10-min duration (06·00–12·00), avoiding conditions of rain or high winds. We varied the routes taken by observers each day to ensure that each point was visited both early and late in the sampling window. We recorded unknown vocalizations using Sennheiser ME66 microphones and Olympus LS11 recording devices, allowing subsequent identification using online reference material (www.xeno-canto.org, recordings deposited in the Colección de Sonidos Ambientales, Instituto Alexander von Humboldt, Colombia). We restricted our analyses to detections within an estimated 100-m radius, excluding records of highly mobile or transient species (e.g. non-breeding trans-continental migrants, large raptors and swifts). All point counts were conducted by experienced observers familiar with the regional avifauna (JJG and DPE).

We used pitfall traps to sample dung beetles (Coleoptera: Scarabaeidae: Scarabaeinae), baiting traps with fresh human dung, which attracts virtually all dung-feeding species (Davis et al. 2001). Dung beetles were sampled in 36 squares (Fig. S1, Supporting information), with single traps placed at five points spaced 100-m apart in each square (180 points in total). Traps were collected at 24-h intervals across 4 days and were rebated after 2 days. Species determinations were made by F. Edwards, C. A. Medina, A. González and J. S. Cardenas using the reference dung beetle collection at the Instituto Alexander von Humboldt, Colombia, where all specimens were deposited.

Statistical Analysis

Modelling species responses to forest distance and wildlife-friendly habitat cover

We used Bayesian hierarchical models to estimate the shape and magnitude of species responses to wildlife-friendly habitat cover and distance from contiguous forest, controlling for variation in detection probability and site-level random effects via a state-space formulation (Dorazio & Royle 2005; Dorazio et al. 2006; Zipkin, DeWan & Royle 2009). To minimize model uncertainty, we excluded bird species detected at fewer than three sampling points from analysis (88 species, see Table S2, Supporting information), leaving a total community of 243 species. We divided these into species detected in forest (212 species) and those never detected in forest (31 species), and modelled the responses of each community separately. As almost all dung beetles species were encountered in forest (27 of 28), we modelled dung beetles as a single community.

Prior to analysis, we centred and standardized our two predictor variables (linear forest distance D and the wildlife-friendly cover index W) to ensure that parameter estimates (β) give robust measures of the relative effect of each on occurrence probability (Schielzeth 2010). Correlation between the two variables was minimal for both taxonomic groups (dung beetles R2 = 0·25; birds R2 = 0·09). We modelled occurrence probabilities for each species i at each sampling point j via a logit link function incorporating linear, quadratic and interaction terms for W and D:

display math(eqn 2)

The model includes a parameter representing the change in species occurrence at sites within contiguous forest (β6), plus a random intercept (α) indexed by species and study area s to account for additional site-level variation in occurrence probabilities (over-dispersion). In addition, we compare three reduced model parameterisations: (i) linear effects only, (ii) linear effects plus the interaction between W and D, and (iii) linear and quadratic effects but no interaction (Table S3, Supporting information). We identify the best model in each case using the Deviance Information Criterion, calculated at the community level for each taxonomic group (Spiegelhalter et al. 2002).

We fit each model to an observation matrix xi,j,k indicating the number of times each species i was detected at sampling point j on occasion k over a fixed number of sampling events (K). We use these data to estimate true occurrences zi,j, which are binary indicators of whether or not species i occurs at point j, by specifying the observed data as the sum of K Bernoulli trials where θi,j,k is the detection probability for species i on visit k:

display math(eqn 3)

For birds, we assume that detection probabilities vary in relation to the timing of each visit, as most birds are easier to detect at times closer to dawn. We therefore model detection probabilities θ for each species using a logit function:

display math(eqn 4)

controlling for over-dispersion via a random intercept λ indexed by species i and study area s. For dung beetles, sampling was continuous over each trap day, so detection probabilities are simply modelled as a function of species and study area effects on the logit scale. Following Dorazio & Royle (2005), we specify a joint bivariate normal distribution for the occurrence model intercept ui and detection model intercept λi:

display math(eqn 5)

where Σ is a 2 × 2 matrix specifying the variance components among species for occurrence and detection (σ2u and σ2λ; Dorazio & Royle 2005; Kéry & Royle 2008).

We incorporate hierarchical structuring at the community level by specifying all model parameters as random effects drawn from ‘hyper-parameter’ distributions (Kéry & Royle 2008), such that species-specific coefficients (β1, β2 etc.) are drawn from distributions that represent the full spectrum of variation across the community. To ensure that posterior parameter estimates reflect the data rather than model assumptions, we assign diffuse uniform (0·1) priors for hyper-parameter means and inverse-gamma (0·1, 10) priors for hyper-parameter variances (Dorazio et al. 2006). We fit the models using WinBUGS version 1.4 (Spiegelhalter et al. 2003), sampling the posterior distribution of each parameter for 50 000 iterations following a burn-in of 30 000 iterations.

Simulating land-sparing and land-sharing strategies

To evaluate the benefits of each strategy, we use fitted models to simulate community dynamics for a hypothetical landscape that is partially developed for agriculture (Fig. 1). In each simulation, the developed portion is divided into management units (analogous to individual farms) spaced at increasing distance from the edge of the remaining contiguous forest (Fig. 1). In each unit, a fixed proportion of land g is grazed, and the remainder is either made up of wildlife-friendly habitats spread throughout the unit (land sharing, Fig. 1d) or a protected area located within the contiguous forest (land sparing, Fig. 1e). For robust comparisons, food production levels must be equal for each strategy (Phalan et al. 2011), involving two assumptions: (i) that the rate of cattle production per hectare is constant for grazed pastures within our study areas, and (ii) that within-farm wildlife-friendly habitats do not contribute to yield (noting that most forest patches in our study areas were fenced off with barbed wire). Food production then depends solely on the proportion of land that is grazed, rather than covered by other habitats.

Figure 1.

Examples of land-sharing and land-sparing strategies in tropical agriculture. The land-sharing approach preserves small areas of wildlife-friendly habitat within farmland (a), whilst land sparing protects larger blocks of contiguous forest (b) by maximising food production within farmed areas (c). We simulate these strategies by building hypothetical landscapes comprising sets of management units (hatched boxes) that are each composed of n sites (circles, = 5 in this example). The farmed portions of each management unit are arrayed at increasing distances from a contiguous forest block. Land-sharing units are exclusively made up of farmland that retains some wildlife-friendly habitat (d), whereas land-sparing management units consist of intensive farmland sites paired with spared sites within the contiguous forest (e, where white text and arrows indicate pairing between spared and farmed sites).

We consider a range of production levels spanning the observed variation in pasture cover across our sampling sites (= 0·2–0·8 in increments of 0·1). For a given value of g, we simulate management units consisting of = 30 sites, corresponding to an area of 92 ha for birds and 24 ha for dung beetles (based on point sampling radii of 100 and 50 m respectively). For land-sharing (Fig. 1d), all n sites are farmed with a level of wildlife-friendly habitat cover dictated by the production level in question (i.e. W1,..,n = 1 − g). For land sparing, y sites are fully grazed farmland (i.e. W1,..,y = 0, where y = g ∙ n), and the remaining ny sites are assigned as a protected area within the contiguous forest (Fig. 1e). For both strategies, all farmed sites are assigned a distance from the edge of contiguous forest (250–1500 m, in 250 m increments, Fig. 1). Following Phalan et al. (2011), we also evaluate a set of intermediate strategies that combine features of land sparing and land sharing, involving incremental increases to the value of W (i.e. more wildlife-friendly habitat within farmland) and decreases to the value of y (i.e. less spared forest), ensuring that g remains constant (see Table S4, Supporting information).

We use fitted versions of the best occurrence model selected for each community (Table S3, Supporting information) to predict occurrence probabilities for 1000 replicate management units under each scenario, sampling the full range of uncertainty in parameter posterior distributions. For land-sparing simulations, we assume that spared forest sites have similar species occurrence probabilities to sampling locations within contiguous forest (i.e. we assume there are no fragmentation or edge effects). For each replicate, we sum the occurrence probabilities across each management unit to generate a species-level index of occurrence. The optimal land allocation scenario for a given species at a given distance from forest is the one that maximizes this index, based on the mean across all replicates.

Predicting winners and losers from agricultural conversion

Following Phalan et al. (2011) we define ‘losers’ as species that, for a given strategy and production level, have a lower mean occurrence index than an equivalent landscape covered entirely by contiguous forest (also generated from 1000 simulations). Losers are therefore species that sometimes or always decline in occurrence under agricultural conversion (Phalan et al. 2011). Winners are classed as species that increase in farmland relative to forest across all production levels and strategies.

Estimating species richness under each strategy

We examine variation in the total number of species expected to persist under land sharing and land sparing by simulating community dynamics across larger hypothetical management units (each made up of 100 sites, totalling 314 ha per landscape for birds and 78·5 ha for dung beetles). For each management unit, we generate a predicted occurrence matrix Ž with dimensions NL, where N is the total number of species in the community and L is the number of sites (Dorazio et al. 2006; Zipkin, DeWan & Royle 2009). Each element of Ž is computed as a single Bernoulli trial with probability generated using fitted versions of the best model selected for each community. The number of species occurring in each landscape is the sum of all elements of Ž for a single set of random draws. We repeat these calculations to generate 1000 random samples of Ž for each scenario, calculating medians and 95% confidence intervals for total species richness.


Species Responses to Distance from Forest and Wildlife-Friendly Habitat

For both dung beetle and bird communities, the best occurrence models included linear and quadratic terms for distance from forest and wildlife-friendly habitat cover, but did not include interaction terms (Table S3, Supporting information). Both predictor variables exerted a similar overall magnitude of influence on species occurrence probabilities in farmland (Fig. 2). Across the dung beetle community, responses to distance from forest (Fig. 2a) were typically more curvilinear than responses to wildlife-friendly habitat cover (Fig. 2b), with many species showing steep declines in occurrence within 500 m of contiguous forest. This was reflected in the high proportion of dung beetle species with strongly negative quadratic terms for distance from forest, relative to wildlife-friendly habitat cover (Fig. S2, Supporting information). Only one dung beetle species responded positively to increasing distance from forest (Fig. 2a), whereas three responded negatively to increasing wildlife-friendly habitat (Fig. 2b).

Figure 2.

Species response curves for dung beetles (a & b) and forest birds (c & d) showing the influence of distance from contiguous forest (a & c) and proportionate cover of wildlife-friendly habitat (b & d) on occurrence probabilities in farmland. Each grey line represents the response of a single species, using parameter means derived from a hierarchical community model (holding other variables constant). The black line shows the mean response across the whole community (derived from mean hyper-parameter values).

The majority of bird species showed a consistent negative effect of distance from forest (Fig. 2c), and a curvilinear response to wildlife-friendly habitat cover, with occurrence probabilities dropping off rapidly for most species at low levels of cover (< 0·3; Fig. 2d), corresponding with strongly positive quadratic terms (Fig. S2, Supporting information). A minority of species from the forest bird community were positively influenced by distance from forest (9% of forest species) and negatively influenced by wildlife-friendly habitat cover (13% of forest species), mirroring responses from the community of non-forest species (n = 31 species; Fig. S3, Supporting information).

Land-Sharing vs. Land-Sparing Strategies

The influence of distance from forest on farmland biodiversity translated into a clear change in the relative benefits of land sharing and land sparing across distances (Fig. 3). This effect was most marked for dung beetles, where land sharing was marginally superior to land sparing at 500 m from contiguous forest, but became increasingly inferior at greater distances (Fig. 3a, at = 0·2). Dung beetles were increasingly classed as ‘losers’ in agriculture further away from forest, rising from 63% of species at 500 m to 93% at 1500 m from contiguous forest (means across production levels; Fig 3a,b). Amongst those species classed as losers, land-sharing or intermediate strategies were superior for 57% of dung beetle species at 500 m, but for just 12% of species at 1500 m (means across production levels; Fig 3a,b). Of dung beetles classed as ‘winners’ from agriculture, the proportion that benefitted from land-sharing or intermediate strategies also fell dramatically with distance from forest (from 90% at 500 m to 57% at 1500 m; Fig. 3a,b).

Figure 3.

Classification of dung beetle (a & b) and bird (c & d) communities according to whether species are ‘winners’ or ‘losers’ from agricultural development, and whether they benefit from land-sparing, land-sharing or intermediate land management strategies, over a range of distances from contiguous forest. In each panel, upper rows show species classed as winners, and the lower rows show losers. We simulated hypothetical land management units for a range of production levels (g), equating to the proportion of land that is grazed [lowest (a & c) and highest (b & d) extremes shown].

The number of bird species classed as losers in agriculture increased from 74% of species at 500 m to 78% at 1500 m (means across production levels; Fig. 3c,d). Amongst those species, land sparing was the optimal strategy at all distances from forest (Fig. 3c,d), and the relative superiority of land sparing became more pronounced with increasing distance from forest (Fig. 3c,d). For example, at 500 m from forest, 65% of avian losers were more abundant under land-sparing strategies, but this rose to 87% at 1500 m (means across production levels; Fig. 3c,d). Of birds classed as ‘winners’ from agriculture, more species benefited from land sharing than land sparing at all distances and production levels (Fig. 3c,d). For both dung beetles and birds, the proportion of species benefiting under intermediate strategies decreased at higher production levels (Fig. 3).

Total Species Richness Across Strategies

The peak in dung beetle species richness was predicted to occur in land-sharing management units within 500 m of contiguous forest (Fig. 4a,b), where 96% of dung beetle species were predicted to persist at the lowest production level, compared with 93% for land sparing (Table S5, Supporting information). However, in management units at 1500 m from forest, land sharing only supported 36% of dung beetles species, whereas 93% were still supported by land sparing (= 0·2, Table S5, Supporting information). At the highest production level and furthest distance from forest (= 0·8, distance = 1500 m), species richness equated to 29% of the entire dung beetle community for land sharing, whereas 86% of species persisted under land sparing (Table S5, Supporting information). The benefits of land-sharing agriculture for dung beetle biodiversity were therefore highly sensitive to distance of farmland from forest, whereas benefits of land sparing were relatively insensitive to forest distance.

Figure 4.

Patterns of dung beetle (a & b) and bird (c & d) species richness across hypothetical land-sharing and land-sparing management units, showing how species richness varies with increasing distance from contiguous forest. Box whisker plots show median, interquartile and 95th percentile ranges from 1000 randomizations under each land allocation scenario, at low (a & c) and high (b & d) production levels. Production levels equate to the proportion of land (g) that is grazed in management unit, rather than covered by wildlife-friendly habitats (land sharing) or a contiguous forest reserve (land sparing).

For birds, species richness was lower under land sharing than land sparing in all scenarios (Fig. 4c,d; Table S5, Supporting information). Land-sharing strategies were again highly sensitive to distance from forest, with 81% of all bird species being supported at 500 m from forest at low production levels (= 0·2), but only 53% of species at 1500 m (Table S5, Supporting information). Similar patterns were evident when the forest bird community was considered in isolation (Table S5, Supporting information). At the highest production level (= 0·8), land-sharing supported only 44% of forest bird species at 1500 m from contiguous forest, whereas the equivalent land-sparing scenario supported 92% of forest bird species (Table S5, Supporting information).


For the majority of species in our study, distance from contiguous forest was an important predictor of farmland occupancy. As a consequence, our models predict that the benefits of land-sharing agriculture decline markedly across a relatively short gradient of isolation from natural habitat. Birds and dung beetles are considered to be good indicators of wider biodiversity responses to environmental change (Lawton et al. 1998; Barlow et al. 2007), representing a broad range of dispersal abilities. The close similarity of responses to forest distance across both taxonomic groups (Fig. 2) hints at the potential generality of this pattern across other mobile taxa. In tropical landscapes, failure to account for the effects of isolation from large habitat tracts might lead to the adoption of inappropriate land management strategies, as well as misleading inferences about the value of land-sharing agriculture. Our results underline the need to consider landscape-scale habitat distributions when designing, comparing and implementing agricultural policies for biodiversity protection (Norris 2008; Batáry et al. 2011; Kleijn et al. 2011).

Distance from contiguous forest was particularly important for biodiversity under land-sharing scenarios, showing a consistent negative trend regardless of the level of food production across the landscape (Figs 3, 4). Large tracts of forest surrounding our study sites apparently act as population sources, or provide important resources, for many species utilizing agricultural habitats in the region. Mobile species might incorporate both farmland and contiguous forest within their individual home ranges, and consequently depend on resources found within forest for some or all life stages (Ricketts et al. 2001; Sekercioglu et al. 2007). Such species may be unable to persist when farmland is isolated from contiguous forest, even if high levels of wildlife-friendly cover are maintained (Ricketts et al. 2001; Luck & Daily 2003; Tscharntke et al. 2008). Alternatively, farmland could be a sink for some species if intrinsic population growth is insufficient for population maintenance without immigration from surrounding contiguous forests (Pereira & Daily 2006). Both mechanisms are likely to play important roles in determining community dynamics within farmland (Ricketts et al. 2001; Daily et al. 2003), particularly in areas adjacent to large forest tracts.

In contrast to land-sharing, land-sparing scenarios showed relatively consistent biodiversity benefits as the distance between intensively managed farmland and contiguous forest increased. This is because under land-sparing, the farmed portion of each management unit is always paired with a protected area within the contiguous forest. As such, our simulations assume that the farmed and spared portions of individual management units can be widely separated in space (see Fig. 1). Our results suggest that even at higher food production levels, these contiguous forest ‘reserves’ consistently provide greater benefits for biodiversity than their equivalent concessions under land sharing (i.e. incorporating areas of wildlife-friendly habitat within farmland).

Our simulations assume that these ‘spared’ lands exclusively fall within large contiguous tracts of forest, rather than forests that have been fragmented. If spared lands are subject to edge effects or other fragmentation impacts, the relative biodiversity benefits are likely to be reduced (Laurance et al. 2011). Fragmentation could also disrupt the flow of benefits from natural habitat into land-sharing farmland, potentially resulting in an ever-diminishing pool of biodiversity within agricultural landscapes. Both strategies are therefore likely to be negatively impacted by fragmentation, though the magnitude of these impacts is difficult to predict. It is important to note that land sparing has greater potential to prevent fragmentation, if implemented effectively (Fischer et al. 2008; Phalan et al. 2011). Land sharing, on the other hand, can mitigate the impacts of fragmentation for some taxa by facilitating dispersal across the agricultural matrix (Ricketts et al. 2001; Daily et al. 2003; Brosi, Daily & Ehrlich 2007). The fragmentation issue also highlights the ambiguity of defining land sharing and sparing across spatial scales, given that larger fragments could be defined arbitrarily as belonging to one strategy or the other (Chandler et al. 2013).

The superiority of land sparing in our study relies partially on an assumption that food production increases in direct proportion to pasture cover, such that productivity is maximized when the landscape is 100% pasture. However, wildlife-friendly habitats can provide benefits to agricultural productivity, including soil and watershed protection (DeFries & Rosenzweig 2010; Schroth & McNeely 2011), as well as shading for livestock (Bailey 2005). Resulting nonlinearities in the relationship between food production and pasture cover could make land sparing less feasible, if production gains from intensification fall short of the level needed for land offsets. This is a common criticism of the land-sparing paradigm, reflecting doubts that intensification can genuinely assist in delivering large-scale habitat protection in practice (Ramankutty & Rhemtulla 2012). Many authors also fear that intensification under land sparing will negatively affect both social dynamics and ecosystem services (Matson & Vitousek 2006; DeFries & Rosenzweig 2010; Lambin & Meyfroidt 2011; Ramankutty & Rhemtulla 2012). Although these issues could push the trade-off in favour of land sharing, it is important to note that biodiversity benefits under land sharing would still be partially contingent on the presence of natural habitats in surrounding areas, given the strong dependence on contiguous forest proximity shown by many species.

Policy Recommendations

Our results predict major biodiversity losses if forested landscapes are converted wholesale to low-intensity agriculture, even if significant wildlife-friendly habitat cover is retained via land-sharing practices. Hence, in regions that still support large tracts of natural habitat, biodiversity interests may be best served by land-sparing policies, provided that they can genuinely deliver protection for contiguous blocks of habitat. This is likely to require mechanisms that pair spared forest reserves with farmland areas that are separated in space, facilitating the protection of pristine natural habitat blocks that are less vulnerable to future encroachment. We see the development of such ‘off-farm’ sparing schemes as an urgent priority for tropical conservation (Fischer et al. 2008; Lambin & Meyfroidt 2011). Land-sharing practices, in turn, can provide important benefits for biodiversity, including facilitating dispersal between forest tracts (Zipkin, DeWan & Royle 2009; Tscharntke et al. 2012). Both strategies should therefore have important roles to play in future agricultural development, and mixed strategies may be more realistic in practical terms (Fischer, Lindenmayer & Manning 2006; Matson & Vitousek 2006). Nevertheless, our results underline the critical importance of contiguous forests for tropical biodiversity, indicating that their protection should remain a top priority.


We thank staff at the Instituto de Investigación de Recursos Biológicos Alexander von Humboldt, particularly F. Forero for logistical support and A. González and J. Stephens-Cardenas for dung beetle identification. For field access permissions, we thank Fundación Colibri (Reserva Mesenia-Paramillo), Fundación ProAves (Reserva Las Tangaras) and L. Tapasco (Cerro Montezuma). We thank Y. Tapasco, O. Cortes, F. Prada, G. Suarez and many local assistants for help with data collection. Funding was provided to T.H. and D.P.E. by the Research Council of Norway, grant number 208836. This is publication #3 of the Biodiversity, Agriculture and Conservation in Colombia/Biodiversidad, Agricultura, y Conservación en Colombia (BACC) project. The authors declare no conflicts of interest.