Quantifying and sustaining biodiversity in tropicalagricultural landscapesChase D. Mendenhalla,b,c,1, Analisa Shields-Estradaa,b, Arjun J. Krishnaswamia,d, and Gretchen C. Dailya,b,e,f,1

aCenter for Conservation Biology, Stanford University, Stanford, CA 94305; bDepartment of Biology, Stanford University, Stanford, CA 94305; cThe NatureConservancy, Arlington, VA 22203; dDepartment of Civil and Environmental Engineering, Stanford University, Stanford, CA 94305; eStanford WoodsInstitute for the Environment, Stanford University, Stanford, CA 94305; and fThe Natural Capital Project, c/o Department of Biology and Stanford WoodsInstitute, Stanford, CA 94305

Edited by Simon A. Levin, Department of Ecology and Evolutionary Biology, Princeton University, Princeton, NJ, and approved September 27, 2016 (receivedfor review May 19, 2016)

Decision-makers increasingly seek scientific guidance on investingin nature, but biodiversity remains difficult to estimate across diverselandscapes. Here, we develop empirically based models for quanti-fying biodiversity across space. We focus on agricultural lands in thetropical forest biome, wherein lies the greatest potential to conserveor lose biodiversity. We explore two questions, drawing fromempirical research oriented toward pioneering policies in CostaRica. First, can remotely sensed tree cover serve as a reliable basisfor improved estimation of biodiversity, from plots to regions?Second, how does tropical biodiversity change across the land-usegradient from native forest to deforested cropland and pasture?We report on understory plants, nonflying mammals, bats, birds,reptiles, and amphibians. Using data from 67,737 observations of908 species, we test how tree cover influences biodiversity acrossspace. First, we find that fine-scale mapping of tree cover predictsbiodiversity within a taxon-specific radius (of 30–70 m) about apoint in the landscape. Second, nearly 50% of the tree cover inour study region is embedded in countryside forest elements,small (typically 0.05–100 ha) clusters or strips of trees on privateproperty. Third, most species use multiple habitat types, includingcrop fields and pastures (to which 15% of species are restricted),although some taxa depend on forest (57% of species are restrictedto forest elements). Our findings are supported by comparisons of 90studies across Latin America. They provide a basis for a planningtool that guides investments in tropical forest biodiversity similarto those for securing ecosystem services.

conservation science | countryside biogeography | ecosystem services |extinctions | species–area relationship

What is the potential of sustaining biodiversity and ecosystemservices in agricultural landscapes? The future of biodiversity

hinges on the answer, given the limited scope for expanding pro-tected areas. Moreover, the generation and delivery of many vitalecosystem services occurs on local to regional scales in social–ecological systems where people make their livelihoods throughcropping, grazing, forestry, and other rural activities. The answeris incomplete, but appears to be “high” (e.g., refs. 1–3). A furtherquestion, however, is how can this potential for conservation intothe Anthropocene can be realized, with land use, other dimen-sions of global change, and rates of extinction intensifying rapidlyworldwide (4–6) and weak institutions for protecting the globalcommons (7)?Efforts to secure biodiversity and ecosystem services in rural

landscapes are expanding and becoming more sophisticated (e.g.,refs. 8 and 9). In the case of ecosystem services, both scientific andpolicy support for targeting investments have advanced rapidly (10,11). In China, for example, 200 million people presently are beingpaid to engage in conservation and restoration activities; since2000 these investments have resulted in many ecosystem serviceimprovements, although not in biodiversity, at a national scale (12,13). By contrast, scientific support for targeting biodiversity in-vestments remains limited by relatively coarse and underdevelopeddata, models, practical tools, and implementation efforts (14). This

limitation is a bit ironic, because ecosystem services science wasopened originally as a pathway for understanding human dependenceon biodiversity and for motivating efforts to stem biodiversityloss (15–17).Here we address two key questions for guiding investments in

ecosystems services to benefit biodiversity. We seek answers toour questions in the context of Costa Rica, a nation remarkablefor its rich biodiversity and for its innovative ecosystem servicespolicies, launched in 1996, that presently pay landowners for ser-vices provided by forests ≥2.0 ha in size (18). The first question iswhether tree cover estimates from satellites can serve as a reliablebasis for improved estimation of biodiversity, on scales rangingfrom plots to regions. The second question is how biodiversitychanges along the land-use gradient from native forest to defor-ested crop and pasture lands, again on scales ranging from plots toregions. We explore biodiversity at three levels: (i) the totalnumber of species across a region (γ-diversity), broadly comparingforested and deforested areas; (ii) the total number of species at apoint or plot on a landscape (α-diversity); and (iii) changes inspecies and abundances across different habitats (β-diversity). Wedraw lessons from an intensively studied region of Costa Rica andthen generalize our findings through an intensive survey of studiesacross Latin America. We focus on the tropical forest biome be-cause it hosts most of the world’s terrestrial biodiversity.

ResultsTo guide investments in biodiversity across previously for-ested agricultural landscapes, we first asked what character-istics of a region can be remotely sensed using satellite imagesto quantify biodiversity across space. We focused on remotelysensing tree cover because it is a good predictor of biodiversityin deforested tropical regions (3, 19). Moreover, we start ourapproach by identifying tree cover at the finest spatial scalesfeasible using satellite images, for two reasons. First, pre-vious studies of bats and birds suggest that biodiversity re-sponds to changes in small areas, such as the area inside acircle with a 50- to 70-m radius (3, 19). Second, resampling oraggregating tree cover from smaller to larger scales is simplewith spatial data. To estimate biodiversity for our regionalcase study, we created a finely detailed tree cover map for Coto

This paper results from the Arthur M. Sackler Colloquium of the National Academy ofSciences, “Coupled Human and Environmental Systems,” held March 14–15, 2016, at theNational Academy of Sciences in Washington, DC. The complete program and videorecordings of most presentations are available on the NAS website at www.nasonline.org/Coupled_Human_and_Environmental_Systems.

Author contributions: C.D.M. and G.C.D. designed research; C.D.M., A.S.-E., and A.J.K.performed research; C.D.M. analyzed data; and C.D.M. and G.C.D. wrote the paper.

The authors declare no conflict of interest.

This article is a PNAS Direct Submission.1To whom correspondence may be addressed. Email: cdm@stanford.edu or gdaily@stanford.edu.

This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.1073/pnas.1604981113/-/DCSupplemental.

14544–14551 | PNAS | December 20, 2016 | vol. 113 | no. 51 www.pnas.org/cgi/doi/10.1073/pnas.1604981113

Brus, Costa Rica (Fig. 1A) by measuring the surface area ofland covered by large tree canopies (≥10 m in diameter). Weidentified all the tree cover across the 934-km2 Costa Ricancanton (Fig. 1B).

When we used our remote sensing technique to estimate treecover, fully 100% of the 46 randomly selected sites across theCoto Brus region agreed with ground observations of habitatclassifications of forest and nonforest. Moreover, measurements

Fig. 1. Using tree cover, we spatially modeled biodiversity across a diverse region of Costa Rica. (A) First, we remotely sensed tree cover at a fine spatialresolution (0.61 m) across the canton of Coto Brus (934 km2). (B) Tree cover is concentrated in protected areas. (C) Nearly 50% of tree cover in Coto Brus isembedded in smaller countryside forest elements (0.05–100 ha), clusters, or strips of trees. We created models that predict biodiversity at a point in space,using the proportion of tree cover within a spatial scale specific to each taxonomic group.

Las Cruces Forest Reserve &

A Total number of plant and vertebrate species from Coto Brus, Costa Rica

countryside forest elements sites onlyBoth

522249137

Crop field & pasture sites only

B

0.80.91.0

ies

Las Cruces Forest Reserve & smaller forest elements only

0 30.40.50.60.7

portio

n of s

peci elements only

Both

0.00.10.20.3

516)

(22)

(40)

263)

(39)

(28)

Prop

Crop field & pasture sites only

Plan

ts (5

NF m

amma

ls

Bats

Bird

s (2

Repti

les

Amph

ibian

s

Fig. 2. (A) Of the 908 species recorded at all sites, 58% were observed only in forest sites, and 15% were observed only in crop field and pasture sites.(B) Taxonomic groups varied in their response to deforestation and use of crop fields and pastures. Overall, the total number of species in each broad habitattype of forest and nonforest, i.e., the γ-diversity of each habitat, was 771 species from all sites in forest elements (n = 101) and 386 species from all sites in cropfields and pastures (n = 81). NF mammals, nonflying mammals.

Mendenhall et al. PNAS | December 20, 2016 | vol. 113 | no. 51 | 14545

SUST

AINABILITY

SCIENCE

COLLOQUIUM

PAPE

R

of canopy cover taken inside each 0.05-ha site across the regionshowed no differences in canopy cover between forest sites insidea protected area (Las Cruces Forest Reserve) (Fig. 1B) andforest sites on private property (mean canopy cover ± SD, 84 ±14% canopy cover, n = 23 sites). Sites in crop fields, mostlycoffee plantations, and pastures had significantly less tree canopycover than sites inside forests (mean canopy cover ± SD, 43 ±34%; n = 23 sites), much of it in the form of banana plants andtrees used for agricultural purposes.According to our remotely sensed tree cover map, nearly 50%

of the tree cover in Coto Brus is embedded in rural agriculturalland, much of it in small and complexly configured “countrysideforest elements” mostly on private property (Fig. 1C). Theseelements comprise trees lining property boundaries, scatteredacross pastures, and in winding, filamentous configurations overcomplex terrain in interconnected forms that typically cannotbe considered isolated forest patches. Collectively, countrysideforest elements cover ca. 264 km2, i.e., 40% of the 654 km2 ofthe land outside the protected areas (Fig. 1B).Taken together, tree cover estimates at a fine scale reveal two

broad habitat types in Coto Brus, where we conducted bio-diversity sampling: (i) forest elements, both inside legally pro-tected areas (public and private) and on legally unprotectedprivate property, and (ii) crop fields and pastures, virtually allon legally unprotected private property. To quantify the totalnumber of species in these two broad habitat types, we sampledforested sites (in both the Las Cruces Forest Reserve and in 2- to100-ha countryside forest elements of various ages and shapes,n = 101 sites) and deforested sites (crop fields and pastures, n =81 sites).We made 67,737 plant and animal observations, for a total of

908 species at all sites. We found 771 species in the forest sites(including both the Las Cruces Forest Reserve and the coun-tryside forest elements sites) and 386 species in crop fields andpastures (Fig. 2A). Only 4% of the 516 plant species were exotic,and all vertebrate species were native to the region. (We ex-cluded humans, horses, cattle, cats, and dogs from all analyses).Overall, 58% of all species were encountered only in forest sites,15% of species were encountered only in crop field and pasturesites, and 27% were observed in both (Fig. 2B). Plant speciesshowed the greatest differences between the two broad habitattypes in total number of species, with 407 species observed only inforest sites, 87 species observed only in crop field and pasture sites,and 22 species observed in both. Of the 392 vertebrate species, weobserved 115 species only in forest sites, 50 species only in cropfield and pasture sites, and 227 species in both (Fig. 2B).In four of the six taxonomic groups sampled (i.e., plants,

nonflying mammals, bats, and birds) the number of species at a site(α-diversity) increased significantly as the amount of tree cover at asite increased. For example, the number of plant species was thesame among forest sites located in the Las Cruces Forest Reserveand countryside forest elements, but sites located in crop fields andpastures contained about half the number of plant species (R2 =0.40, P < 0.001, n = 44) (Fig. 3A and Table S1). Medium- to large-sized nonflying mammal species (>1 kg in size) were the verte-brates most sensitive to deforestation, never venturing into openareas (i.e. with <15% tree cover within 70 m) over nearly a year ofcontinuous camera trapping, even when bait was present (R2 =0.70, P < 0.001, n = 21) (Fig. 3B and Table S2). The number of batspecies decreased by about a third in sites with no tree cover within60 m (R2 = 0.32, P = 0.008, n = 18) (Fig. 3C and Table S3).The number of bird species at a site responded differently to

the proportion of tree cover. The area of tree cover within a 30-mradius correlated best with the number of small bird species ata given point. However, because 24 bird species live exclusivelyin agricultural land and many species thrive in that habitat, es-pecially as tree canopies increase in size (19; Fig. 2), two distinctspecies–area curves related to tree cover emerge (R2 = 0.67, P <

0.001, n = 21) (Fig. 3D and Table S4) for the two different,overlapping bird communities, as described further in aβ-diversity analysis. The number of reptile and amphibian speciesdid not vary significantly across sites, but trends suggest more speciesat sites with more tree cover (Fig. 3 E and F) and also show evidenceof overlapping biological communities in further β-diversity analyses.Taken together, these results show that adding tree cover to atropical forest region generally increases the number of species at apoint or plot.We tested for changes in species and abundances across dif-

ferent habitats, i.e., β-diversity between sites, in two ways, first bycomparing the relative abundances of each species from forestand nonforest sites, and second, by comparing abundance-basedcommunity similarity coefficients at forest and nonforest sites.First, the relative abundances of each species were calculated forsites in the Las Cruces Forest Reserve, forest sites in country-side forest elements, and sites in crop fields and pastures. Afteradjusting for sampling effort across multiple sites, a total of 30%

0123456789

10

Crop fields &pastures (n = 24)

Countryside forestelements (n=12)

Las Cruces ForestReserve (n = 3)

Obse

rved s

pecie

s rich

ness

(mea

n ±SD

)

0123456789

10

Crop fields &pastures (n = 24)

Countryside forestelements (n=12)

Las Cruces ForestReserve (n = 3)

Obse

rved s

pecie

s rich

ness

(mea

n ±SD

Amphibians

020406080

100120140160180200

0 0.2 0.4 0.6 0.8 1Mode

led sp

ecies

richn

ess (

estim

ate ±

SE)

Proportion tree cover within 30 m

Countryside forest elementsCrop fields

Reptiles

0

5

10

15

20

25

30

35

40

0 0.2 0.4 0.6 0.8 1

Mode

led sp

ecies

richn

ess (

estim

ate ±

95 C

I )

Proportion of tree cover within 60 m

Las Cruces Forest ReserveCountryside forest elementsCrop fields

0

20

40

60

80

100

120

140

Crop fields &pastures (n = 9)

Countryside forestelements (n = 27)

Las Cruces ForestReserve (n = 8)

Mode

led sp

ecies

richn

ess (

mean

±SD

)

0

2

4

6

8

10

12

14

0 0.2 0.4 0.6 0.8 1

Obse

rved s

pecie

s rich

ness

Proportion of tree cover within 70 m

Las Cruces Forest ReserveCountryside forest elementsCrop fields & pastures

A B

C D

E F

*Non-flyingmammals

Las Cruces Forest Reserve

Birds

Plants

Bats

±

±±

)Fig. 3. Increasing the proportion of local tree cover generally increases thetotal number of species at a point, i.e., the α-diversity. (A) The number of plantspecies was the same in sites located in forest elements (light and dark green)but was halved at sites in crop fields and pastures (yellow). (B–D) The numberof nonflying mammal species at a point responded strongly to the amount oftree cover measured within 70 m of a site (B), whereas the numbers of bat(C) and bird (D) species at a point on a landscape were explained best by theamount of tree cover measured within 60 m and 30 m of the site, respectively.(E and F) The numbers of reptile (E) and amphibian (F) species did not increasesignificantly in sites with higher amounts of local tree cover, but a trend wasobserved.

14546 | www.pnas.org/cgi/doi/10.1073/pnas.1604981113 Mendenhall et al.

of the 67,737 plant and animal observations were recorded insidethe Las Cruces Forest Reserve, 38% in countryside forest ele-ments, and 32% in crop fields and pastures (Fig. 4).Species varied greatly in their dependence on forest, with 54%

of the 908 species found in multiple habitat types (Fig. 4). Of the516 plant species, we classified 75% as forest dependent because>50% of their population was observed in forest sites, inclusiveof the Las Cruces Forest Reserve (Fig. 4A). Of the 22 nonflyingmammal species, only three species (a rabbit, an opossum, and asquirrel) preferred crop fields and pastures with >50% of theirobservations occurring there; these data suggest that 86% of allnonflying mammal species are dependent on forest elements(Fig. 4B). Finally, assuming that a species is forest dependent if>50% of the individuals are observed in forest sites, inclusive ofthe Las Cruces Forest Reserve, we find that 88% of the 40 batspecies, 61% of the 263 bird species, 59% of the 39 reptilespecies, and 54% of the 28 amphibian species are dependent on

forest elements (Fig. 4 C–F). Note the decreasing order of sen-sitivity to deforestation.Next, to test statistically for changes in species and abundances

across different habitats, i.e., β-diversity, we compared abundance-based community similarity coefficients at forest and nonforest sites.All taxonomic groups differed significantly in their abundance-based community similarity coefficients when categorized into forestsites and nonforest sites. Permutational multivariate ANOVAs(PERMANOVAs) detected the strongest differences in abundance-based community composition between forest sites (including theLas Cruces Forest Reserve and countryside forest elements) andcrop fields and pastures: plants, F1,42 = 9.60, P < 0.001 (Fig. 5A);nonflying mammals, F1,19 = 3.57, P = 0.005 (Fig. 5B); bats, F1,249 =24.65, P < 0.001 (Fig. 5C); birds, F1,607 = 128.76, P < 0.001 (Fig. 5D);reptiles, F1,36 = 6.67, P < 0.001 (Fig. 5E); and amphibians, F1,36 =5.60, P = 0.01 (Fig. 5F). Essentially, shifts in abundance andchanges in species composition in all taxonomic groups from forestto nonforest has resulted in the formation of two overlapping

A B

C D

E F

Fig. 4. Plots A–F show changes in abundances for each species across dif-ferent habitats (i.e., a measure of β-diversity) for 908 species. A total of67,737 individual plants and animals were recorded in sites distributed acrossthree broad categories: (i ) Las Cruces Forest Reserve (dark green);(ii) countryside forest elements (light green); and (iii) crop fields and pastures(yellow). Most species used multiple habitat types, including crop fields andpastures. The forest-dependency rank ranges from forest avoidance (on theleft side of the x axis) to forest dependence (on the right side of the x axis)and was determined by comparing relative abundance in the Las CrucesForest Reserve (dark green) with that in crop fields and pastures (yellow). Allproportions of populations account for differences in sampling effort amongthe three habitat types.

Dime

nsion

2

Dimension 1

Dime

nsion

2

Dimension 1

Birds

Dime

nsion

2

Dimension 1

Dime

nsion

2

Dimension 1

Dime

nsion

2

Dimension 1

Bats

Reptiles Amphibians

Dime

nsion

2

Dimension 1

Las Cruces Forest ReserveCountryside forest elementsCrop fields & pastures

A B

C D

E F

Plants Non-flyingmammals

Stress = 27.5; n = 44 sites 516 species; 16,257 individuals

Stress = 34.8, n = 251 mist nets40 species; 4,020 individuals

Stress = 28.1, n = 609 mist nets263 species; 40,008 individuals

Stress = 19.9; n = 20 sites22 species; 2,975 observations

Stress = 23.3, n = 38 sites39 species; 685 observations

Stress = 24.5, n = 38 sites28 species; 981 observations

Fig. 5. Significant shifts in community composition across different habitats(a measure of β-diversity) occur among all taxonomic groups. Shown in A–Fare nonmetric multidimensional scaling plots that depict abundance-basedcommunity-similarity coefficients among all sites (or individual mist nets forthis analysis in C and D). Closer proximity represents greater similarity insites’ community composition. All taxonomic groups were significantly dis-similar in community composition when split into two broad habitat types:forest [including the Las Cruces Forest Reserve (dark green points) andcountryside forest elements (light green points)] and crop fields and pastures(yellow points).

Mendenhall et al. PNAS | December 20, 2016 | vol. 113 | no. 51 | 14547

SUST

AINABILITY

SCIENCE

COLLOQUIUM

PAPE

R

biological communities in our study region—a larger, more bio-diverse community associated with forest and a smaller, less bio-diverse community associated with agricultural land (Figs. 2 and 5).

To generalize our findings beyond the study region in Costa Rica,we conducted a literature survey of countryside biogeographicstudies of plants, nonflying mammals, bats, birds, reptiles, andamphibians in the tropical forest biome across Latin America.Of 857 reviewed articles, 19 (from Brazil, Costa Rica, Guate-mala, Mexico, and Panama) met our criteria for study design andreporting of data (Table S5). We were able to extract data for 90study comparisons of biodiversity change across different habitatsrepresenting the six taxonomic groups that we focused on in CostaRica (Dataset S1). The 90 study comparisons tested for differencesin the number of species (α-diversity) at sites in countryside forestelements or at sites in crop fields and pastures compared with thenumber of species at sites in minimally altered forest (>100 ha insize). In total, 29% of the 90 study comparisons reported a signif-icant loss of species in crop fields and pastures compared withminimally altered forest, 68% reported no change, and 7% reporteda significant increase in the number of species (Fig. 6A).Moreover, of the 29 study comparisons that tested for shifts in

abundances and/or changes in species composition between forestedand deforested areas, 25 reported significant changes in β-diversityusing various metrics. Finally, data published in 12 studies (20–31) offive vertebrate taxa (i.e., nonflying mammals, bats, birds, reptiles, andamphibians) supported the argument that two overlapping biologicalcommunities form with deforestation—one community associatedwith forest and a smaller, less biodiverse community associatedwith agricultural land (Fig. 6B). (Tests to determine generaltrends between taxonomic groups were not permitted because ofdata deficiency.)

DiscussionProtected areas span only 13% of the global land surface, and inreality these areas are only partially protected (32), suggestingthat the future of terrestrial biodiversity depends in large part onthe effect of agricultural use of unprotected lands (33–36). Wedeveloped an approach for quantifying the relationship betweenlocal tree cover and biodiversity (Fig. 3). Nearly 50% of the ruralagricultural countryside of Coto Brus, our study region, is cov-ered by trees, much of it in small forest elements of varied ages,shapes, and sizes along rivers, steep terrain, and property bound-aries (Fig. 1B). These forest elements, together with the agriculturalland in which they are embedded, support significant dimensions ofbiodiversity, whereas the adjacent crop fields and pastures supporta different community (2, 3).Despite the biodiversity found in deforested habitats, pro-

tected areas remain critical because most species at risk for ex-tinction in the tropical forest biome require expansive forests.For example, 9% of 908 species in our study were completelydependent on the local protected area, occurring only in the LasCruces Forest Reserve. Moreover, several local extinctions haveoccurred in the region because of habitat loss and poaching (37).Our approach for modeling biodiversity across a tropical agri-

cultural region can inform conservation policies by quantifyingbiodiversity using remotely sensed tree cover to scale from plots toregions. By selecting specific spatial scales to measure the area oftree cover for our α-diversity models that predict the number ofspecies at a site, our approach accounts for forest configuration(38) and operationalizes the universal species–area relationship(39). Finally, the region we use as a case study has >10% naturalhabitat for most species; therefore large effects on biodiversityresulting from habitat fragmentation are not expected (40). Ofcourse, our approach requires testing and adjustment based on theecological contexts in which it might be applied. Moreover, ourapproach relies on high-quality inputs that will come from im-provements in tree cover mapping and remote sensing. To repli-cate our approach, global estimates of tree cover are needed toguide investments in biodiversity and ecosystem services, especiallyin the tropical forest biome where biodiversity is concentrated andfarms are small family businesses.

(3)

0

20

40

60

80

100

Decline Same Increase

Prop

ortio

n of s

tudy c

ompa

rision

s

Change in species richness compared to minimally altered forest (> 100 ha in size)

Countryside forest elementsCrop fields & pastures

A

Minimally altered forest (> 100 ha in size) & countryside forest elements sites only

Crop field & pasture sites only

397235100

Both

B Total number of vertebrate species from 12 other studies across Latin America

(15)

(25)

(3)

(11)

(33)

0

20

40

60

80

100

Decline Same Increase

Prop

ortio

n of s

tudy c

ompa

rision

s

Change in species richness compared to minimally altered forest (> 100 ha in size)

Fig. 6. Across Latin America, the biodiversity of the tropical forest biome isexperiencing change and loss much in the same way as we report for Coto Brus,Costa Rica. (A) A survey of comparisons of species richness representing 3,041species of plants, nonflying mammals, bats, birds, reptiles, and amphibiansthroughout Latin America revealed that sites located inside countryside forestelements usually support the same number of species as sites located in minimallyaltered forest. Comparisons of species richness in minimally altered forests and incrop fields and pastures revealed greater losses in species richness, but 58%of thecomparisons between those habitat types were still statistically indistinguishable.(B) Community shift statistics from 29 studies and data from 12 studies of speciesoccurrences throughout Latin America in three broad habitats for species ofnonflying mammals (47 species), bats (62 species), birds (460 species), reptiles (67species), and amphibians (96 species) confirmed that deforestation is leading tothe formation of a novel, less biodiverse community that overlaps with a larger,more diverse biological community associated with tropical forests.

14548 | www.pnas.org/cgi/doi/10.1073/pnas.1604981113 Mendenhall et al.

In practice, conservation applications of the predictive bio-diversity models we present can be used to insulate protected areasand to diversify farmland in once-forested landscapes. Specifically,our models predicting the number of species at a site (α-diversity;Tables S2–S4) coupled with results that describe the kinds ofspecies at a site (β-diversity) support arguments that increasing treecover around existing forest reserves will lead to small gains in thenumber of species, but the species gained tend to be rare, unique,and at risk for extinction because of deforestation (2, 3, 41, 42) Ourresults also suggest that conservation efforts to diversify farmlandby increasing tree cover in and around crop fields and pasturesleads to large gains in the number of species, but the species gainedtend to be common. Our results provide a theoretical basis foraligning conservation strategies to slow extinctions and for boostingsome ecosystem services by buffering remaining forests with wild-life-friendly farmland and regenerating forests where politicallyfeasible, given demands to farm the planet (43).Our work supports valuing farming systems in Costa Rica

and Latin America—including coffee plantations, pastures, andremnants of forests— holistically for the provision of biodiversitybenefits as well as crops, forage, timber, carbon sequestration,water quality, scenic beauty, and other ecosystem services. In ourstudy, remotely sensed tree cover serves as a reliable basis forestimating biodiversity, on scales from plots to regions. Specifi-cally, we find that increasing the amount of tree cover on privateproperty leads to increases in the richness of plant, nonflyingmammal, bat, and bird species at the plot level. Moreover, treecover on private property contributes to regional biodiversity bysupporting the larger, more biologically diverse communitiesassociated with tropical forests. Our survey of Latin Americanbiodiversity and similar work on other continents, e.g., in an>2,000-y-old tropical agricultural system in the Western Ghatsof India (44), demonstrate that it is possible to sustain patternsof relatively high biodiversity after deforestation when farmingecosystems are heterogeneous, incorporate elements of naturallyoccurring habitats, and grow perennial crops.In grasslands and other biomes not dominated by forests or

situated in the tropics, sustaining the conservation values of ag-ricultural lands will depend on successfully identifying the fea-tures of agricultural lands linked with native biodiversity andlocal ecosystem services. In the context of Costa Rica, with anestablished program of payment for ecosystem services that hasincentivized forest conservation and restoration on privateproperty for 20 y, our approach of spatially linking tree cover attaxon-specific spatial scales to biodiversity patterns can be usedas a basis for developing spatial planning tools to target invest-ments in biodiversity and align values with goals to protect water,store carbon, and provide scenic beauty. We share a replicableapproach for reducing extinctions caused by tropical de-forestation and demonstrate it in the context of an existing policythat pays farmers for ecosystem services and biodiversity on theirproperties.

MethodsStudy Area. This research is being conducted in the rugged, hilly countrysideof Coto Brus, a canton of Costa Rica, across agricultural landscapes spanning700–1,350 m elevation. The region, originally tropical premontane wetforest, was heavily deforested in the 1960s and early 1970s (45). Today, thecountryside comprises crop fields (mostly of coffee and diversified gardens),pasture, and trees scattered through crop fields and pastures, in strips bor-dering streams and property lines, and in forest elements of variable sizes(typically 0.05–100 ha). The landscape and biogeography are described morefully elsewhere (33, 37).

Tree Cover Estimates. We created a finely detailed tree cover map by handdigitizing aerial photographs in Google Earth for the 934-km2 canton of CotoBrus. We used photographs taken in 2014 by commercial QuickBird satellitesat a 0.61-m spatial resolution. We collected ground-truth data at 46 ran-domly selected plots measuring 500 m2 in three major habitat types: (i) nine

sites in the Las Cruces Forest Reserve, the last large forest tract at mid-elevation and thus serving as a protected regional baseline; (ii) 14 sites inforest elements of various ages and sizes; and (iii) 23 sites in crop fields andpastures. In each site we surveyed vegetation to check whether the GoogleEarth tree-cover map matched the tree cover observed in the field.

Biodiversity. To measure biodiversity, we identified and counted the abun-dances of plants and all vertebrate taxa: nonflying mammals, bats, birds,reptiles, and amphibians. We sampled plants andwildlife using standardized,systematic techniques appropriate for each taxon over extensive time periods(e.g., 1–5 y), with substantial trapping effort (i.e., ∼95,000 mist net hours and∼15,000 trap nights), and at spatially independent sites across Coto Brus (i.e.,18–44 sites per taxa). We used sampling in the protected Las Cruces ForestReserve as the regional baseline of minimally altered forest after sampling inLa Amistad International Park (25 km away) revealed no differences in thetotal number of species recorded at that sampling site (i.e., in α-diversity)compared with sites in the Las Cruces Forest Reserve. We also conductedsampling in a suite of countryside forest elements of different ages and sizesand in deforested crop fields, mostly coffee plantations, and in pastures usedfor grazing. Biodiversity sampling of animals was conducted and approvedunder Stanford University’s Institutional Animal Care and Use Committee(IACUC), Assurance Number: A3213-01, Protocol: 26920.

Understory Plant Biodiversity. We used ground surveys to sample understoryplants, identifying and counting all plants that fell inside a 1-m2 frame, in-cluding trees, shrubs, vines, forbs, grasses, and domestic plants. We selectedsites randomly using Global Information System (GIS) software and repli-cated the 1-m2 sampling systematically 20 times in a total of 44 sites, eachmeasuring 500 m2. We surveyed plants in four major habitat types: cattlepastures (three sites), coffee plantations (six sites), countryside forest ele-ments (2- to 100-ha forest; 27 sites), and the Las Cruces Forest Reserve (eightsites). Understory plant biodiversity was aggregated for each site by poolingthe 20 replicates of the 1-m2 frame inside each site. Plant species richnessvalues for each site were modeled using techniques that account for de-tection biases (46).

Bird Biodiversity. We sampled birds through constant-effort mist netting,using 20 12 × 2.6 m, 38-mm mesh ground-level mist nets in sites that covered3–5 ha. We conducted sampling between January 25 and May 12 for 6 y(2007–2012) at 21 sites in three major habitat types: coffee plantation (ninesites), countryside forest elements (2- to 100-ha forest; nine sites), and in theLas Cruces Forest Reserve (three sites). We primarily captured passerines andother small birds, species that comprise the majority of the avifauna in thearea. We fitted each captured bird with a unique aluminum leg ring,recorded the mist net location of capture, and released the bird at thecapture site. Bird species richness values for each site were modeled usingtechniques that account for detection biases (46).

Bat Biodiversity.We used constant-effort mist netting to sample bats, placing20 12 × 2.6 m 38-mmmesh ground-level mist nets in a site that covered 3–5 ha.We conducted sampling between January 24 and March 28 for 4 y (2007–2010) at 18 sites in three major habitat types: coffee plantations (six sites),countryside forest elements (2- to 100-ha forest; nine sites), and the LasCruces Forest Reserve (three sites). We sampled birds at all sites where batswere sampled. We primarily captured fruit bats; these species comprise themajority of the bats in the area (3). We fitted all captured bats with a uniquecollar, noted the net location of capture, and released bats at the capturesite. Bat species richness values for each site were modeled using a Bayesianoccupancy model that accounts for detection biases.

Nonflying Mammal Biodiversity.We sampledmedium-sizednonflyingmammals(>1 kg in size) using both passive and baited camera traps. Camera trap pro-tocols involved a motion-triggered camera trap fastened to a tree or post at 21sites in four major habitat types of cattle pasture (three sites), coffee plantation(six sites), countryside forest elements (2- to 100-ha forest; nine sites), and theLas Cruces Forest Reserve (three sites). We operated camera traps without baitfor nearly 1 y (September 2014–July 2015) and operated camera traps with bait(two bananas and three chicken gizzards, replenished daily) for a total of 14d at each site. All cameras were at ground level. Species excluded from allanalyses were humans, horses, cattle, cats, and dogs. Observed species richnessof nonflying mammals at each site was used for analyses.

Reptile and Amphibian Biodiversity. We sampled reptiles and amphibiansusing two complementary techniques: diurnal and nocturnal visual-encounter

Mendenhall et al. PNAS | December 20, 2016 | vol. 113 | no. 51 | 14549

SUST

AINABILITY

SCIENCE

COLLOQUIUM

PAPE

R

surveys and drift-fence trapping with pitfall and funnel traps. Six surveys forreptiles and amphibians took place during the wet and dry seasons over 3 y(2002–2004) at 39 sites in four major habitat types: cattle pasture (12 sites),coffee plantation (12 sites), countryside forest element (2- to 100-ha forest;12 sites), and the Las Cruces Forest Reserve (three sites). We identified cap-tured individuals but did not mark them and released all animals at thecapture site. Observed reptile and amphibians species richness values foreach site were used for further analyses.

Statistical Analyses. To quantify how biodiversity changes along the land-usegradient from protected native forests to deforested agricultural plots, weexplore biodiversity at three levels: (i) the total number of species in a regionof forest and nonforest (the γ-diversity of each habitat), (ii) the total numberof species at a point on a landscape (α-diversity), and (iii) similarity in speciesand abundances across different habitats (β-diversity).

To quantify changes in the total number of species at a point on alandscape, i.e., the α-diversity, we compared observed or modeled speciesrichness estimates for each site and tested for differences among sites withdifferent amounts of local tree cover. Plant and bird species richness wasmodeled using Chao species richness estimates, and bat species richness wasestimated using a Bayesian occupancy model (46). We measured tree coverat each site in two ways. In the first, we used three habitat categories of LasCruces Forest Reserve, countryside forest elements, and crop fields andpastures. In the later, more refined approach, we used the proportion oftree cover within a taxon-specific area, defined as the spatial scale at whichspecies respond most strongly to a habitat variable (47) and hypothesized tobe inclusive of configuration (38). We compared all linear models at differ-ent spatial scales using corrected Akaike Information Criteria (AICc). Includingspatial structures in linear mixed effects models did not significantly influenceresults and was not done in final analyses.

To quantify changes in abundances across different habitats, theβ-diversity, we calculated the relative abundance for each species in each habitattype and adjusted by sampling effort in space and time. Additionally, wetested statistically for changes in species and abundances across differenthabitats (β-diversity) by testing for differences among sites using abundance-based community similarity coefficients (46, 48). For each taxonomic groupwe aggregated captures and counts of all species by site and, for finer res-olution in this analysis, by specific mist net for birds and bats. Each site or

mist net was labeled by habitat type (i) Las Cruces Forest Reserve, (ii) countrysideforest elements, or (iii) crop fields and pastures (Fig. 5). Using PERMANOVA tests,we tested for significant differences and clustering of abundance-basedcommunity similarity coefficients by habitat types.

Survey of Latin American Countryside Biogeography Studies. To gather evi-dence to generalize our approach for modeling biodiversity on agriculturallands, we performed a survey of biogeographic studies in other agriculturallandscapes where contiguous tropical forest once existed. We created asearch term to target countryside biogeography studies in Latin America thatconducted richness sampling of plants, nonflying mammals, bats, birds,reptiles, and/or amphibians (Table S5). All 857 studies returned by the searchterm were evaluated for inclusion by two independent readers to extractdata (Dataset S1). We considered data only from studies with at least threesites in countryside forest elements or three sites in crop fields and/or pas-tures. Moreover, we required that studies consider at least three sites inminimally altered forest. We defined sites as minimally altered forest if theywere located in continuous forest or in large forest fragments (>100 ha insize). Countryside forest elements were defined as small clusters or strips oftrees (0.05–100 ha in size), secondary forests, riparian forest remnants, livefences, and agricultural fields left to fallow at least 5 y ago. Sites in cropfields and pastures were located on land that is currently being cultivatedfor growing crops or rearing animals and included land that had been usedfor agricultural purposes within 5 y.

ACKNOWLEDGMENTS. We thank Federico Oviedo Brenes, Jeisson FigueroaSandí, Paula Mesen Cabezas, and dozens of field assistants for help withsurveys of tree cover and biodiversity; Luke Frishkoff and Hannah Frank formodeling bat species richness estimates; Ania Wrona and Leithen M’Goniglefor help with tree cover calculations; the dozens of landowners inCosta Rica who graciously allowed us to conduct field work on their prop-erty; and Paul R. Ehrlich, Gerardo Ceballos, and Rakan A. Zahawi for pro-ductive collaboration in the supportive environment of the Las CrucesBiological Station of the Organization for Tropical Studies. This work wassupported by the Moore Family Foundation, the Winslow Foundation, theHaberfeld family, and Peter and Helen Bing. C.D.M. was supported by theBing Fellowship in Honor of Paul Ehrlich, a National Science FoundationGraduate Research Fellowship, and a NatureNet Science Fellowship throughThe Nature Conservancy.

1. Rapidel B, DeClerck F, Le Coq JF, Beer J, eds (2011) Ecosystem Services fromAgriculture and Forestry: Measurement and Payment (Earthscan, London).

2. Frishkoff LO, et al. (2014) Loss of avian phylogenetic diversity in neotropical agri-cultural systems. Science 345(6202):1343–1346.

3. Mendenhall CD, Karp DS, Meyer CFJ, Hadly EA, Daily GC (2014) Predicting biodiversitychange and averting collapse in agricultural landscapes. Nature 509(7499):213–217.

4. Tscharntke T, et al. (2012) Global food security, biodiversity conservation and thefuture of agricultural intensification. Biol Conserv 151:53–59.

5. Ceballos G, et al. (2015) Accelerated modern human-induced species losses: Enteringthe sixth mass extinction. Sci Adv 1(5):e1400253.

6. Steffen W, Broadgate W, Deutsch L, Gaffney O, Ludwig C (2015) The trajectory of theAnthropocene: The Great Acceleration. The Anthropocene Review 2(1):81–98.

7. Levin SA (2000) Fragile Dominion: Complexity and the Commons (Perseus Publishing,New York).

8. Lipper L, Sakuyama T, Stringer R, Silberman D, eds (2009) Payment for EnvironmentalServices in Agricultural Landscapes: Economic Policies and Poverty Reduction inDeveloping Countries (Springer, New York).

9. Guerry AD, et al. (2015) Natural capital and ecosystem services informing decisions:From promise to practice. Proc Natl Acad Sci USA 112(24):7348–7355.

10. Kareiva P, Tallis H, Ricketts TH, Daily GC, Polasky S, eds (2011) Natural Capital: Theoryand Practice of Mapping Ecosystem Services (Oxford Univ Press, Oxford, UK).

11. Sharp R, et al. InVEST +VERSION+ User’s Guide (2015) The Natural Capital Project.Available at data.naturalcapitalproject.org/nightly-build/invest-users-guide/html/.Accessed May 1, 2016.

12. Daily GC, et al. (2013) Securing natural capital and human well-being: Innovation andimpact in China. Acta Ecol Sin 33:669–676.

13. Ouyang Z, et al. (2016) Improvements in ecosystem services from investments innatural capital. Science 352(6292):1455–1459.

14. Pereira HM, et al. (2010) Scenarios for global biodiversity in the 21st century. Science330(6010):1496–1501.

15. Ehrlich PR, Mooney HA (1983) Extinction, substitution, and ecosystem services.Bioscience 33:248–254.

16. Barbier EB, Burgess JC, Folke C (1994) Paradise Lost? The Ecological Economics ofBiodiversity (Earthscan, London).

17. Daily GC, ed (1997) Nature’s Services: Societal Dependence on Natural Ecosystems(Island, Washington, DC).

18. Asamblea Legislativa de la República de Costa Rica (1996) Ley Forestal 7575. pp. 1–36.19. Mendenhall CD, Sekercioglu CH, Brenes FO, Ehrlich PR, Daily GC (2011) Predictive

model for sustaining biodiversity in tropical countryside. Proc Natl Acad Sci USA108(39):16313–16316.

20. Almeida-Gomes M, Rocha CFD (2014) Landscape connectivity may explain anuranspecies distribution in an Atlantic forest fragmented area. Landsc Ecol 29(1):29–40.

21. Avila-Cabadilla LD, Stoner KE, Henry M, Añorve MYA (2009) Composition, structureand diversity of phyllostomid bat assemblages in different successional stages of atropical dry forest. For Ecol Manage 258(6):986–996.

22. Bael SA, Zambrano RV, Hall JS (2013) Bird communities in forested and human-modifiedlandscapes of Central Panama: A baseline survey for a native species reforestation treat-ment. International Journal of Biodiversity Science, Ecosystem Services & Management9(4):281–289.

23. Bohn JL, Diemont SAW, Gibbs JP, Stehman SV, Mendoza Vega J (2014) Implications ofMayan agroforestry for biodiversity conservation in the Calakmul Biosphere Reserve,Mexico. Agrofor Syst 88(2):269–285.

24. Cassano CR, Barlow J, Pardini R (2012) Large mammals in an agroforestry mosaic inthe Brazilian Atlantic forest. Biotropica 44(6):818–825.

25. Faria D, Baumgarten J (2007) Shade cacao plantations (Theobroma cacao) and batconservation in southern Bahia, Brazil. Biodivers Conserv 16(2):291–312.

26. Faria D (2006) Phyllostomid bats of a fragmented landscape in the north-easternAtlantic forest, Brazil. J Trop Ecol 22(05):531–542.

27. Garcia S, Finch DM, León GC (1998) Patterns of forest use and endemism in residentbird communities of north-central Michoacán, Mexico. For Ecol Manage 110(1–3):151–171.

28. Marsden SJ, Whiffin M, Galetti M (2001) Bird diversity and abundance in forestfragments and Eucalyptus plantations around an Atlantic forest reserve, Brazil.Biodiversity and Conservation 10(5):737–751.

29. Mendenhall CD, et al. (2014) Countryside biogeography of Neotropical reptiles andamphibians. Ecology 95(4):856–870.

30. Santos-Filho M, Silva DJ, Sanaiotti TM (2008) Edge effects and landscape matrix use bya small mammal community in fragments of semideciduous submontane forest inMato Grosso, Brazil. Braz J Biol 68(4):703–710.

31. Urbina-Cardona JN, Olivares-Pérez M, Reynoso VH (2006) Herpetofauna diversity andmicroenvironment correlates across a pasture–edge–interior ecotone in tropicalrainforest fragments in the Los Tuxtlas Biosphere Reserve of Veracruz, Mexico. BiolConserv 132(1):61–75.

32. Ellis EC, Klein Goldewijk K, Siebert S, Lightman D, Ramankutty N (2010) Anthropo-genic transformation of the biomes, 1700 to 2000. Glob Ecol Biogeogr 19:589–606.

33. Daily GC, Ehrlich PR, Sánchez-Azofeifa GA (2001) Countryside biogeography: Utili-zation of human-dominated habitats by the avifauna of southern Costa Rica. EcolAppl 11:1–13.

34. Rosenzweig M (2003) Win-Win Ecology: How the Earth’s Species Can Survive in theMidst of Human Enterprise (Oxford Univ Press, Oxford, UK).

14550 | www.pnas.org/cgi/doi/10.1073/pnas.1604981113 Mendenhall et al.

35. Gardner TA, et al. (2009) Prospects for tropical forest biodiversity in a human-modi-fied world. Ecol Lett 12(6):561–582.

36. Perfecto I, Vandermeer J, Wright A (2009) Nature’s Matrix: Linking Agriculture,Conservation and Food Sovereignty (Earthscan, London).

37. Daily GC, Ceballos G, Pacheco J, Suzán G, Sánchez-Azofeifa A (2003) Countrysidebiogeography of Neotropical mammals: Conservation opportunities in agriculturallandscapes of Costa Rica. Conserv Biol 17:1814–1826.

38. Fahrig L (2013) Rethinking patch size and isolation effects. J Biogeogr 40:1649–1663.39. Harte J, Smith AB, Storch D (2009) Biodiversity scales from plots to biomes with a

universal species-area curve. Ecol Lett 12(8):789–797.40. Rybicki J, Hanski I (2013) Species-area relationships and extinctions caused by habitat

loss and fragmentation. Ecol Lett 16(Suppl 1):27–38.41. Balmford A, Green R, Phalan B (2015) Land for food and land for nature? Daedalus

144(4):57–75.

42. Laurance WF, et al. (2012) Averting biodiversity collapse in tropical forest protectedareas. Nature 489(7415):290–294.

43. Kremen C (2015) Reframing the land-sparing/land-sharing debate for biodiversityconservation. Ann N Y Acad Sci 1355:52–76.

44. Ranganathan J, Daniels RJ, Chandran MD, Ehrlich PR, Daily GC (2008) Sustainingbiodiversity in ancient tropical countryside. Proc Natl Acad Sci USA 105(46):17852–17854.

45. Zahawi RA, Duran G, Kormann U (2015) Sixty-seven years of land-use change insouthern Costa Rica. PLoS One 10(11):e0143554.

46. Chao A (1987) Estimating the population size for capture-recapture data with un-equal catchability. Biometrics 43(4):783–791.

47. Holland JD, Bert DG, Fahrig L (2004) Determining the spatial scale of species’ responseto habitat. Bioscience 54(3):227–233.

48. Chao A, Chazdon R, Colwell R, Shen TA (2005) A new statistical approach for assessingsimilarity of species composition with incidence and abundance data. Ecol Lett 8:148–159.

Mendenhall et al. PNAS | December 20, 2016 | vol. 113 | no. 51 | 14551

SUST

AINABILITY

SCIENCE

COLLOQUIUM

PAPE

R