lable at ScienceDirect

Environmental Modelling & Software 60 (2014) 121e133

Contents lists avai

Environmental Modelling & Software

journal homepage: www.elsevier .com/locate/envsoft

Conservation planning to zone protected areas under optimallandscape management for bird conservation

Yu-Pin Lin a, Chun-Wei Huang a, *, Tzung-Su Ding b, Yung-Chieh Wang a, Wei-Te Hsiao a,Neville D. Crossman c, Szabolcs Lengyel d, Wei-Chi Lin a, Dirk S. Schmeller e, f, g

a Department of Bioenvironmental Systems Engineering, National Taiwan University, Taipei 10617, Taiwanb Department of Forestry and Resource Conservation, National Taiwan University, Taipei 10617, Taiwanc CSIRO Ecosystem Sciences, PMB 2, Glen Osmond, South Australia 5064, Australiad Department of Tisza River Research, Danube Research Institute, Centre for Ecological Research, Hungarian Academy of Sciences, 4026 Debrecen, Bem t�er18/c, Hungarye UFZ e Helmholtz Centre for Environmental Research, Department of Conservation Biology, Permoserstr. 15, 04318 Leipzig, Germanyf Universit�e de Toulouse; UPS, INPT; EcoLab (Laboratoire Ecologie Fonctionnelle et Environnement), 118 route de Narbonne, 31062 Toulouse, Franceg CNRS, EcoLab, 31062 Toulouse, France

a r t i c l e i n f o

Article history:Received 20 December 2013Received in revised form4 June 2014Accepted 11 June 2014Available online

Keywords:Systematic conservation planningHabitat suitabilitySpatial optimizationLandscapeMarxan softwareBird species

* Corresponding author. Tel./fax: þ886 2 3366 346E-mail addresses: yplin@ntu.edu.tw (Y.-P. Lin)

(C.-W. Huang).

http://dx.doi.org/10.1016/j.envsoft.2014.06.0091364-8152/© 2014 Elsevier Ltd. All rights reserved.

a b s t r a c t

This study proposes a two-stage conservation planning approach. Firstly, the Land-Use PatternOptimization-library is used to maximize the suitability of habitats for target species by optimizingconfiguration based on the current landscape. Secondly, the systematic conservation planning tool,Marxan is used to identify protected areas based on the estimated species distributions from the optimallandscape configuration. We compared our conservation plan for three target bird species from ahighland farm with the conservation plan produced using Marxan alone. Our comparison showed theeffectiveness of our approach by selecting a reserve network with higher habitat suitability, betterconnection, and smaller size after relatively minor landscape modification. The proposed approach ad-vances previous reserve site selection algorithms by considering optimal landscape configuration andpotential species distributions for a reserve network design. Our approach yields priority maps to guidethe design of a reserve network as well as identify landscape management for conservation.

© 2014 Elsevier Ltd. All rights reserved.

Software availability

Name: LUPOlib 1.0Programming language: C/CþþDeveloper: Annelie Holzk€amperAvailability: http://www.ufz.de/index.php?en¼17779Name: Marxan Optimized Version 2.43Developer: Matt WattsHardware required: PCSoftware required: X64 Windows OSAvailability: http://www.uq.edu.au/marxan/marxan-software

7., yolatengoseam@gmail.com

1. Introduction

The purpose of conservation planning is to identify cost-effective, representative and complementary biodiversity conser-vation areas for the protection or restoration of species or habitats(Margules and Pressey, 2000; Margules and Sarkar, 2007; Hermosoet al., 2013). Systematic conservation planning (SCP) uses quanti-tative and systematic approaches for design reserve networks thatconserves species according to the conservation planning princi-ples (Margules and Pressey, 2000; Margules and Sarkar, 2007; Kleinet al., 2009). SCP has been applied to terrestrial (Smith et al., 2006;Zhang et al., 2012; Levin et al., 2013; Nackoney andWilliams, 2013),marine (Smith et al., 2009; Delavenne et al., 2012; Levy and Ban,2013), and freshwater ecosystems (Linke et al., 2012; Esselmanet al., 2013). SCP principles have also been applied to the designof restoration priorities in degraded landscapes (Crossman andBryan, 2006; Bryan and Crossman, 2008).

The last two decades have seen a large growth in quantitativesystematic spatial conservation approaches and tools (Ball et al.,

Fig. 1. Land-use of theHighland Experimental Farm, National TaiwanUniversity, and distributions of each target species, the VividNiltava, Brownish-frankedBush-warbler, andGreen-backed Tit. The Vivid Niltava was found mainly in the central forest, with more frequent anthropogenic disturbances; the Brownish-flanked Bush-warbler was found mainly in forestand cropland in the eastern part of the study area. The Green-backed Tit was foundmainly in pristine forest in the northern part of the study area,which is less affected than other partsby human activity. The distributions of the first two species are strongly related to the volume of the foliage in the canopy while the last species favors shrubs (Lin et al., 2011).

Y.-P. Lin et al. / Environmental Modelling & Software 60 (2014) 121e133122

2009). Marxan software (here after referred to as Marxan) (Ball andPossingham, 2000), Zonation (Moilanen, 2007; Lehtom€aki andMoilanen, 2013), Consnet (Ciarleglio et al., 2009) and C-Plan(Pressey et al., 2009) all implement target-based planning as theprimary planning method (Minin and Moilanen, 2012). Popularamong these tools is the use of quick heuristic-based algorithms toperform spatial optimization to achieve conservation goals. Simu-lated annealing (SA) has been used to identify spatial prioritizationfor locating reserved planning units, as in Marxan (Ball et al., 2009).Marxan is now a widely used tool for performing spatial prioriti-zation in SCP (Zielinski et al., 2006; Zhang et al., 2012; Levy and Ban,2013) and to cost-effectively select protected areas that supportconservation targets (Delavenne et al., 2012).

Most approaches to SCP are limited by their binary decisionframework, except the approach described in Watts et al. (2009).Such a framework cannot simultaneously capture the full range of

potential management nor conservation actions as part of SCP(Moilanen et al., 2009; Watts et al., 2009) and cannot considerspatial patterns within the selected areas despite land-use patternsbeing amajor driver of ecosystem functions and services (Crossmanet al., 2013; Labiosa et al., 2013). A genetic algorithm-based spatialoptimization model, the Land-Use Pattern Optimization-library(LUPOlib) (Holzk€amper et al., 2006; Holzk€amper and Seppelt,2007a,b) could be used in combination with existing reserve siteselection algorithms to assess beforehand the spatial patterns oflandscape elements for habitat suitability in protected areas.LUPOlib has been used to solve conservation planning-type prob-lems, such as optimizing the trade-off between ecological andeconomic objectives and optimizing landscape management ac-tions (Holzk€amper and Seppelt, 2007b). We suggest that SCP couldbe enhanced by combining optimal landscape management withconservation planning of reserves. Current approaches to SCP

Fig. 2. Flow chart of the proposed LUPOlib-Marxan approach.

Y.-P. Lin et al. / Environmental Modelling & Software 60 (2014) 121e133 123

generally assume that biotic and abiotic conditions are static inspace and in time (Levin et al., 2013). The combination we proposewould yield potential species distributions under optimal land-scape scenarios in support of the selection of reserve areas in SCP,thereby overcoming the previous static space-time limitations.

Here, we describe a two-stage planning approach that combinesLUPOlib and Marxan for SCP. We applied the combined LUPOlib-Marxan approach to the three bird species in an agriculturallandscape of Taiwan to design a reserve network consideringchanges in the landscape pattern and habitat suitability. Thereserve sites selected by the combined approach were evaluated bycomparing the valuable areas for target species in optimizedlandscapes with those in the current landscape. The proposed two-stage approach improved habitat suitability of reserve areas ascompared to current approaches and showed potential to advancesubstantially existing reserve site selection methods by consideringtheir spatial patterns. Our combined model provides guidelines forconservation actions that include optimal landscape managementand modification for conservation purposes.

2. Materials and methods

2.1. Study area and target species

The study area was the 50 ha highland experimental farm at National TaiwanUniversity (HEF-NTU) (24�050 N, 121�100 E, altitude 2100 m). The boundary wasextended to 50 m beyond giving a total size of approximately 61.61 ha. At thataltitude breeding bird species in Taiwan are prevalent (Ding et al., 2005). An earlierfield investigation (Lin et al., 2011), collected data on land coverage, the distributionof birds, and other related environmental data in resolution of 10 � 10 m (Fig. 1). Wechose as target species three endemic subspecies with distinct habitat preference,the Vivid Niltava (Niltava vivida vivida), Brownish-flanked Bush-warbler (Cettiafortipes robustipes) and Green-backed Tit (Parus monticolus insperatus). These threesubspecies are identified in Taiwan as requiring conservation (Council of Agriculture,

Taiwan, 2013). The presence data included 162 points of occurrence of the VividNiltava, 598 points of occurrence of the Brownish-flanked Bush-warbler and 138points of occurrence of the Green-backed Tit (Fig. 1).

2.2. Conservation planning approaches

The goal of our SCP was to optimize the landscape to improve landscapestructure for the target species, each of which have different habitat preferences.Two approaches were established for comparison: i) using Marxan only to deter-mine potential reserve areas in the current landscape, and ii) using LUPOlib tooptimize the current landscape and then using Marxan to identify potential reserveareas in the optimal landscape scenario (Fig. 2). We applied the two models to eachspecies (case 1¼ the Vivid Niltava, case 2¼ the Brownish-flanked Bush-warbler, andcase 3 ¼ the Green-backed Tit) and to all three species combined (case 4). Thepurpose of case 4 was to find an optimal solution for all target species by using anobjective function that sums the equallyweightedmeans of habitat suitability acrossspecies (described below). For all cases we assumed that habitat suitability was asurrogate for conservation costs because less suitable habitat would likely requireadditional conservation measures to improve suitability. Therefore, higher habitatsuitability of target species was used in the reserve design network (Tole, 2006;Zielinski et al., 2006; Wang et al., 2008). The difference was assessed between thespatial pattern of landscape elements before and after the optimal landscapemodification influenced the reserve network design.

Habitat suitability maps and the size of selected reserves were used to assess theeffectiveness of both approaches and we followed a five-step procedure (fromBennett et al., 2013) to investigate model performance: i) (re)assessment of themodel’s aim, scale and scope; ii) characterization of the data for calibration andtesting; iii) visual and other analysis to detect under- or non-modeled behavior andto gain an overview of overall performance; iv) selection of basic performancecriteria, and; v) consideration of more advanced methods to handle problems suchas systematic divergence between modeled and observed values to evaluate theproposed LUPOlib-Marxan approach. More details are described in Appendix S1.

2.3. Conservation planning approach using Marxan software

The goal of Marxan is to achieve a set of biodiversity conservation targets atminimal cost andminimal boundary length of the reserve network (Ball et al., 2009).Marxan uses a simulated annealing algorithm to solve the set-covering problem

Y.-P. Lin et al. / Environmental Modelling & Software 60 (2014) 121e133124

(Ball and Possingham, 2000). The main terms in the objective function are planningunit cost, species penalties and the boundary length of the chosen site. The functionof the objective function is:

Minb

ftotal costg (1)

total cost ¼X

unit costþX

specie spenaltiesþ BLMX

boundary length (2)

where unit cost is a cost that is assigned to each planning unit; species penalties isadded upon failure tomeet biodiversity target goals; boundary length is a cost that isdetermined by the total outer boundary length of all selected areas; The BoundaryLength Modifier BLM is a function of the importance of the compactness and con-nectivity of the reserve system (Schill and Raber, 2009). In our study, the cell sizewas 10 m2 and the total number of cells was 4269.

We ran Marxan to delineate reserve areas in the current landscape and theoptimal landscape scenarios. We used two criteria to determine differences betweenthe reserve networks before and after the optimal landscape modification: i) athreshold of lowest habitat suitability of occurrences of target species, and; ii) ahabitat suitability map as a cost map. Firstly, a threshold was set using lowestsuitability of occurrences of target species. Points at which suitability was higherthan the threshold were set to identify possible occurrences (pseudo targets)assumed as actual occurrences of conservation targets following landscape opti-mization scenarios. Not only were the pseudo targets identified, the habitat suit-ability map was used to examine the differences between the reserve networksbefore and after the optimal landscapemodification. To demonstrate the differences,both of the best solution and the selection frequency from the 500 top candidates ofselected reserve areas were conducted for each case. The best solution representedthe lowest cost to create a reserve network that covers 80% of the target species; theselection frequency from the 500 top candidates of selected reserve areas specifiedthe irreplaceability of an important (low total cost) location within the potentialareas for species conservation.

2.4. Two-stage conservation planning approach

We used the spatial land-cover pattern optimization model, LUPOlib, to identifythe land-cover pattern that provides optimal ecological improvements. SinceLUPOlib was developed as a patch-based optimization tool, the decision variables inLUPOlib were specified as a vector of the composition of patches from the candidatelandscape. Our objective was to maximize the suitability of the habitat for the targetspecies; therefore, the fitness function was given by,

Maxa.

(X3i¼1

WiHSIi

)(3)

where HSIi is the average habitat suitability index for the i th target species; a.

denoted the vector of decision variables, which specified the spatial composition ofland-cover patternMx,y,Wi is the weighting importance of the ith target species. Thecases described in Section 2.2 were specified by j, j ¼ 1 to 3, withWi ¼

�1; if i ¼ j0; if isj

;in case 4, Wi ¼ 1; i ¼ 1 to 3.

The optimal landscape modification considered orchard, cropland, coniferplantation and broadleaf plantation as replaceable, while pristine forest, buildingsand water were not able to be converted due to the land-use policy of HEF-NTU. Arule was implemented that required maintaining 72.5% of the total area of orchardsand cropland for income to support HEF-NTU.

We used logistic regression to construct a habitat suitability model, HSIi (e.g.Crossman and Bass, 2008). The regression was performed on raster data, with thedependent variable being presence data of the target species, and the independentvariables being the landscape metrics that quantified the influence of land-coverpatterns of the study area, and the distances of each cell to the roads and tobuildings. A forward stepwise logistic regression and a Receiver Operating Charac-teristic (ROC) analysis was completed using the Statistical Package for the SocialScience (SPSS) for Windows (SPSS Inc., IL, USA). The ROC value was used to evaluatethe performance of habitat suitability models.

Additionally, since habitat suitability was summed across all species in case 4,the most common species may drive the solution (Westphal et al., 2007). To equalizethe importance of all species in case 4, the corresponding habitat suitability wasnormalized to their maximum andminimum suitability values in the iteration of theoptimization process (Holzk€amper et al., 2006). The function of the habitat suit-ability model is (Holzk€amper et al., 2006; Holzk€amper and Seppelt, 2007a):

HSIi ¼Xxmax

x

Xymax

yhsi

�Mx;y

��Cells; (4)

hsi�Mx;y

� ¼ exp�b0 þ

Xn

k¼1bk � dfkðx; yÞ

�.h1þ exp

�b0 þ

Xn

k¼1bk � dfkðx; yÞ

�i(5)

where hsiðMx;yÞ2½0;1� is the habitat suitability index at location (x,y) in a candidateland-cover pattern, Cells is the number of locations (4269 cells with a consistent size

of 10 m2), b0 and bk are the intercept and coefficient of the habitat model, respec-tively, and dfk(x,y) are the driving factors, which are the landscape metrics and thedistance to human-dominated land cover (buildings and roads).

Landscape metrics have been used to understand the relationship betweenspatial patterns and ecological processes at a landscape scale (Zaragozí et al., 2012).We used class area (cal), largest patch index (lpi), sum of the edge lengths betweentwo land-use types (esl;m), and patch cohesion (cohl) to describe landscape patterns.The landscapemetrics of a sub-landscapewithin the species' territory size of each cellwere calculated (Holzk€amper et al., 2006). The sub-landscape represented a territoryof target species. Amovingwindow analysis was used to calculate landscape metricsof each sub-landscape as habitat variables (Holzk€amper et al., 2006). Therefore, thevalue of habitat suitability of each cell was calculated based on the landscapestructure within the range of species' territory (see Appendix S2 for more details).

LUPOlib uses a genetic algorithm which we built using the following steps:

1. Possible solutions for landscape patterns, representing a population of n chro-mosomes, were randomly generated. Since LUPOlib is a patch level-basedoptimization model, the chromosomes were represented as a vector of thecomposition of landscape patterns (Fig. 3) (Holzk€amper and Seppelt, 2007b).

2. The average habitat suitability index (HSIi) was calculated for the objective score(fitness) of each possible solution.

3. The genetic operations of selection, crossover and mutation were repeated untila proportion of current solutions were replaced by improved solutions. Initially apopulation of 100 individual genomes was chosen from a population for laterbreeding. Then the crossover and mutation operator carried on exchanges ofpatches with different land-use types (Fig. 3). Here, the crossover probabilityand the mutation rate were 0.5 and 0.01, respectively.

4. Steps 2e3 were repeated until the required convergence criteria were satisfied.The iteration converged when the genetic algorithm found a solution within1000 iterations, or the deviation between the 100 previous best-of-generationfrom the current best-of-generation was less than 0.01% (such that theconvergence ratio equals to 0.9999).

3. Results

3.1. Comparison of planning cases

All the results of the Marxan-only and the LUPOlib-Marxan ap-proaches were compared for four cases (Figs. 4e7). Species distribu-tiondataover the current landscapewere shownfor each case (panelsa). The configuration of optimal landscape scenarios from LUPOlibshowed different preferences of landscape structure for the targetspecies in the four cases (panels e). TheMarxan-only results (panels bto d) with the combined LUPOlib-Marxan results (panels f to h) werecompared according to the differences of selected locations beforeand after the optimal landscape modification. For cases 1 through 4,the best solution (panels b and f) and irreplaceability based on theselection frequency (panels c and g) were analyzed for comparison.The LUPOlib-Marxan results suggest that themore aggregate reservesites were selected in each of the four cases (Figs. 4e7). Furthermore,the LUPOlib-Marxan approach selected smaller reserve areas inmostof the cases. The areas selected by theMarxan-only approach and theLUPOlib-Marxan approachwere, 3.59 ha and 1.30 ha (case 1), 4.20 haand 4.79 ha (case 2), 3.38 ha and 1.10 ha (case 3), and 12.89 ha and4.79 ha (case 4), respectively (Figs. 4e7).

In case 1 (Vivid Niltava), the Marxan-only approach suggestedthat the best locations with higher irreplaceability were mostly inthe center of the study area (Fig. 4b and c), where some coniferplantations were surrounded by a mosaic of croplands and build-ings. In contrast, the combined LUPOlib-Marxan approach sug-gested that the best location (Fig. 4f) was in the southern part of thestudy area where orchards and croplands should be replanted as aconifer plantation to form a reserve area. Furthermore, both thesouthern and central parts of the study area, where orchards shouldbe converted to conifer plantations, have higher irreplaceabilitythan the Marxan-only approach (Fig. 4g).

In case 2 (Brownish-flanked Bush-warbler), the Marxan-onlyapproach suggested that the best locations with higher irreplace-ability were the three major zones with croplands and pristine for-ests in-between (Fig. 5b and c). The combined LUPOlib-Marxan

Fig. 3. The crossover operator exchanges two vectors of the composition of landscape patterns based on patch topology of decision units. Area units are identified in an area-ID map.In this example patches of cropland and grassland are assigned to area units; urban area are excluded from crossover operator (Holzk€amper and Seppelt, 2007a,b).

Y.-P. Lin et al. / Environmental Modelling & Software 60 (2014) 121e133 125

approach suggested that the best zone for the Brownish-flankedBush-warbler was in the south-eastern part of the study areawhere orchards should be converted to croplands and broadleafplantations (Fig. 5f). On the other hand, there were two major irre-placeable habitats for the Brownish-flanked Bush-warbler (Fig. 5g).

In case 3 (Green-backed Tit), the Marxan-only approach showedthe importance of the pristine forest in the north for the Green-backed Tit (Fig. 6b and c). Similarly, the combined LUPOlib-Marxan approach suggested that the best solution and irreplace-able habitats were in the pristine forest (Fig. 6g).

In case 4 (all species), the protected areas were similar to theoverlapping areas of cases 1 to 3 when using the Marxan-onlyapproach (Fig. 7b and c). Unlike the fragmented areas selected bythe Marxan-only approach, the combined LUPOlib-Marxanapproach suggests that the best zone for restoration was theeastern part of the study area (Fig. 7f) where orchards should beconverted from broadleaf plantations into a homogeneous forestedlandscape. The higher irreplaceability of the central and south-eastern parts of the study area represented high potential asreserve areas for all target species (Fig. 7g).

3.2. Preferred spatial patterns of target species to be reserve areas

The results of logistic regression showed that the habitat suit-ability models satisfactorily represent the suitability of each loca-tion for each target species with all of the ROC values exceeding 0.7(Table 1; for model performance see Appendix S1.). The resultsrevealed the strength of positive or negative effect (favor ordisfavor) of driving factors on habitat suitability. Model resultsindicated that fragmented forests and conifer plantations werepreferred by the Vivid Niltava, but buildings and the edge betweenhuman-dominated land-cover types, such as orchards and croplandwere avoided. Model results showed that the Brownish-flankedBush-warbler avoided pristine forests, conifer plantation and

edges of conifer plantation, but prefer edges in cropland and areasdistant to roads and buildings. Model results for the Green-backedTit favored pristine forest but avoided conifer plantation (Table 1).

Based on the habitat suitability models, habitat suitability mapsbefore (panel d of Figs. 4e7) and after (panel h of Figs. 4e7) theoptimal landscape modification for each case were generated.Comparing the Marxan-only and the combined LUPOlib-Marxanapproaches, the latter efficiently increased the mean habitat suit-ability for all target species in the cases (Table 2) due to the optimallandscape modification (panel e of Figs. 4e7), in comparison to theoriginal landscape (panel a of Figs. 4e7). In case 1, the mean habitatsuitability for the Vivid Niltava was almost doubled from 0.03449 to0.07127. In case 2, the mean habitat suitability for the Brownish-franked Bush-warbler was improved from 0.07515 to 0.21374. Incase 3, although the suitability of most of the study area for theGreen-backed Tit could be improved, themost suitable area remainsthe northern forest (Fig. 6d and h). Finally, the results showed thesummation of all habitat suitability in case 4 (Fig. 7d and h) and thesuitability for all target species was improved (Table 2). However,the variation in the habitat requirements among the three targetspecies resulted in a compromise among the habitat suitabilityimprovements compared to cases 1 to 3. Additionally, not only forthe whole study area but also for the selected best zones, thecombined LUPOlib-Marxan outperformed the Marxan-onlyapproach in selecting suitable habitats with a higher mean habitatsuitability, identifying optimal zones of reserve sites.

3.3. Land-cover patterns with optimal habitat suitability for targetspecies

The total area to be converted in the optimal landscape scenariowas less than 12 ha in all four cases (Table 3). The minor modifi-cation resulted in improved suitability of habitat for each targetspecies. The best modifications (panel e of Figs. 4e7) of the current

Fig. 4. The results of two approaches (panels b to d for the Marxan-only approach; panels e to h for the LUPOlib-Marxan approach) for case 1 (Vivid Niltava): a) species distribution data over the current landscape; e) best modificationof the current landscape; b and f) best solutions for reserve area selection; c and g) irreplaceability of reserve areas; d and h) habitat suitability maps.

Y.-P.Linet

al./Environm

entalModelling

&Softw

are60

(2014)121

e133

126

Fig. 5. The results of two approaches (panels b to d for the Marxan-only approach; panels e to h for the LUPOlib-Marxan approach) for case 2 (Brownish-franked Bush-warbler): a) species distribution data over the current landscape; e)best modification of the current landscape; e) best modification of the current landscape; b and f) best solutions for reserve area selection; c and g) irreplaceability of reserve areas; d and h) habitat suitability maps.

Y.-P.Linet

al./Environm

entalModelling

&Softw

are60

(2014)121

e133

127

Fig. 6. The results of two approaches (panels b to d for the Marxan-only approach; panels e to h for the LUPOlib-Marxan approach) for case 3 (Green-backed Tit): a) species distribution data over the current landscape; e) bestmodification of the current landscape; b and f) best solutions for reserve area selection; c and g) irreplaceability of reserve areas; d and h) habitat suitability maps.

Y.-P.Linet

al./Environm

entalModelling

&Softw

are60

(2014)121

e133

128

Fig. 7. The results of two approaches (panels b to d for the Marxan-only approach; panels e to h for the LUPOlib-Marxan approach) for case 4 (all the target species): a) species distribution data over the current landscape; e) bestmodification of the current landscape; b and f) best solutions for reserve area selection; c and g) irreplaceability of reserve areas; d and h) habitat suitability maps.

Y.-P.Linet

al./Environm

entalModelling

&Softw

are60

(2014)121

e133

129

Table 1Results (b values) of logistic regression for habitat suitability models.

Driving factors Vivid Niltava Brownish-flanked Bush-warbler Green-backed Tit

Class area of forest (ca3) 3.59 �4.169 2.305Class area of coniferplantation (ca5) 4.278 �5.173 �3.575Class area of broadleaf plantation (ca6) e e e

Patch cohesion for pristine forest (coh3) e 0.025 e

Patch cohesion for conifer plantation (coh5) e 0.01 e

Patch cohesion for broadleaf plantation (coh6) e 0.012 e

Sum of edge length between building and orchard (es1;2) �0.117 e e

Sum of edge length between building and cropland (es1;4) e �0.046 �0.105Sum of edge length between orchard and cropland (es2;4) �0.265 e e

Sum of edge length between orchard and conifer f plantation (es2;5) �0.094 �0.082 e

Sum of edge length between forest and cropland (es3;4) e 0.064 �0.089Sum of edge length between cropland and conifer plantation (es4;5) e �0.057 0.14Sum of edge length between cropland and broadleaf plantation (es4;6) e �0.08 e

Largest patch index (lpi) �0.022 �0.01 e

Distance to building (disto_b) 0.003 0.008 e

Distance to raod (disto_r) e 0.022 �0.028Constant �4.001 �3.589 �3.521

ROC 0.706 0.766 0.749

Table 2Comparison of habitat suitability for target species using the combined LUPOlib-Marxan approach and the Marxan-only approach.

HSIa(HSIBestb) Case 1c Case 2d Case 3e Case 4f

Marxan-only approach 0.03449 (0.05836) 0.07515 (0.17395) 0.02895 (0.08652) 0.43330 (0.58839)LUPOlib-Marxan approach 0.07127 (0.15908) 0.21374 (0.28460) 0.04319 (0.11962) 0.69833 (0.86162)

a Average value of Habitat Suitability Index of whole study area.b Average value of Habitat Suitability Index of the selected areas (best solution).c Case study for the Vivid Niltava.d Case study for the Brownish-franked Bush-warbler.e Case study for the Green-backed Tit.f Case study for all the target species in case 1, 2, and 3.

Y.-P. Lin et al. / Environmental Modelling & Software 60 (2014) 121e133130

landscape varied across target species because of the wide range ofhabitat preferences of the species.

In case 1, orchards and croplands were converted to coniferplantations for the benefit of the Vivid Niltava (Table 3). Most of theconverted patches were located at the edge of areas with frequenthuman activities, such as the boundaries between buildings andorchards (Fig. 4e).

In case 2, orchards and conifer plantations were converted tocreate edges between forests and croplands for the benefit of theBrownish-franked Bush-warbler (Fig. 5e). However, the number ofcroplands decreased because the edges between conifer planta-tions and croplands in the center of the study areawere replaced bybroadleaf plantations (Table 3).

In case 3, the Green-backed Tit preferred a pristine forest with acanopy with a high foliage volume rather than conifer plantations.

Table 3Comparison of landscape composition for target species using the combinedLUPOlib-Marxan approach with different target species.

Area size(ha) Current landscape Case 1b Case 2c Case 3d Case 4e

Water body 0.4 0.4 0.4 0.4 0.4Built up 6.01 6.01 6.01 6.01 6.01Orchardsa 5.66 3 3.04 8.94 3.08Pristine Forest 15.42 15.42 15.42 15.42 15.42Croplanda 7.34 3.4 4.75 3.4 3.4Conifer plantationa 7.34 14.28 1.4 0.88 5.54Broadleaf plantationa 0.51 0.17 11.66 7.63 8.83

a Replaceable units.b Optimal landscape modification case for the Vivid Niltava.c Optimal landscape modification case for the Brownish-franked Bush-warbler.d Optimal landscape modification case for the Green-backed Tit.e Optimal landscape modification case for all the target species in case 1, 2, and 3.

The Green-backed Tit was rarely observed in open areas such asedges of croplands (Table 1). Accordingly, the croplands and coniferplantations in the central and southern parts of the study areawereconverted to broadleaf plantations (Fig. 6e).

In case 4 the result showed the optimal win-win situation for alltarget species (Fig. 7e). In the optimal landscape scenario, the areaof orchards, cropland, and conifer plantations was reduced, whilethat of broadleaf plantations was increased (Table 3). With respectto the configuration of the optimal landscape scenario, croplands inthe central part of the study area were converted to conifer plan-tations to increase the habitat area for the Vivid Niltava. Coniferplantations, croplands and orchards in the eastern and southernparts of the study area were converted to broadleaf plantations toincrease the habitat for the Brownish-franked Bush-warbler andthe Green-backed Tit.

4. Discussion

There are many differences between the protected areas iden-tified using theMarxan-only approach and the areas for reservationidentified using the combined LUPOlib-Marxan approach. Here, wedemonstrate that effective conservation planning should take intoaccount optimal landscape management with respect to habitatsuitability. We were able to show the applicability of using theLUPOlib model together with Marxan to consider future changes inthe distributions of species from an optimal land-use pattern inresponse to natural and anthropogenic environmental changes(Pressey et al., 2007; Levin et al., 2013). Additionally, our combinedLUPOlib-Marxan approach provides priority maps to guide morespecific conservation actions, in which the spatial patterns areoptimally modified to design an environment that appears suitable

Y.-P. Lin et al. / Environmental Modelling & Software 60 (2014) 121e133 131

for target species with various habitat preferences. Therefore, notonly for selected zones but also for thewhole study area, the habitatsuitability for the target species would be significantly improvedusing our LUPOlib-Marxan approach (all Wilcox-oneManneWhitney p-values were less than 0.001 (Appendix S3)).

The combined LUPOlib-Marxan approach provides moreaggregate reserve sites with comparatively higher suitability for thetarget species. The consistent results showed that the combinedapproach can be applied to the conservation of species withdifferent preferred features. Despite the combined approachselecting a larger reserve site in case 2, the combined approachgenerally has higher efficiency for management by selectingsmaller reserve areas for preserving the same proportion of targetspecies after minor landscape modifications. Aggregate reservesites may have resulted due to the fact that the species' suitability ofthe neighboring planning unit will be promoted simultaneously,since the combined approach can optimize the landscape patternswithin a given range for each cell (planning unit) based onsuitability.

A number of studies (e.g. Smith and Shugart, 1987; Rodenhouseet al., 1997; Marshall and Cooper, 2004) have shown that birds inbetter quality habitat may have smaller range requirements leadingto a support of larger populations in the same or even smallerhabitat area (see also Lin, 2006). Although our results conclude thatthe overall area of selected reserves is smaller, the proposed solu-tions have better habitat structure and quality and can thereforesustain larger populations. However, the sizes of converted areaswhich are less than 12 ha is still around 1/5th of the total size of ourstudy area. For a project at larger scale, we suggest that the cost oflandscape modification should be taken into account by adjustingthe size and the number of land-use types which can be convertedbased on precise cost accounting.

Additionally, the complexity of the conservation planningproblemwill increase when the number of target species increases.The increasing complexity may result in a compromise among thehabitat suitability improvements of the species with disparatepreferences by providing their preferred features in different parts.Some habitats, such as pristine forest, are not included in thereserve area of case 4. The pristine forest may be selected onlywhen the weighting importance of the species which prefer pris-tine forest is higher than others. The combined LUPOlib-Marxanapproach identifies an optimal landscape scenario for reserve sitedesign that meets the distinct preferences of the three target spe-cies. However, adding also more species, especially those depend-ing heavily on pristine forest, may change that picture considerably.

The combined LUPOlib-Marxan approach provided prioritymaps and guidelines for implementing conservation actions as aplanning product for end users. However, the model providesadditional information for conservation activities, such as optimallandscape maps and habitat suitability maps. The planning andimplementation process is more detailed and hands-on comparedto simpler conservation actions to maintain important features ofthe existing landscape (Pierce et al., 2005; Smith et al., 2006; Knightet al., 2009), typical in SCP using Marxan-only approaches.Compared with the Marxan-only approach, additional habitatsuitability maps and optimal landscape modification maps providemore complementary outcomes for taking habitat structure intothe consideration of reserve network design. The opportunities forhands-on conservation action in the study area requires the sup-port of willing landowners, but if implemented provides a goodexample of local-scale conservation action. Nevertheless, we sug-gest that the time needed and the short-term impact of the con-servation actions on species should be monitored in futurepractices. Implementation of combined LUPOlib-Marxan approachsolutions can involve both establishment of reserve sites through

both restoring and changing land cover to be suitable for targetspecies and also maintaining preferred features in the areas that donot require modification. For example, in case 3, the combinedLUPOlib-Marxan approach identified preserving the pristine forestin the north of the study area for the Green-back Tit. That result wassimilar to that of the Marxan-only approach. Furthermore, theexample also showed the flexibility of the combined approach byadjusting the weighting of each target species to explore theessential areas for SCP.

In order to assess the performance of our combined LUPOlib-Marxan approach, the five-step procedure for model evaluationdescribed in Bennett et al. (2013) was conducted (Appendix S1). Asour approach consisted of two optimization sub-models (i.e. theLUPOlib model and Marxan), we suggest that the visual methodscan be used not only for examining quantities and residuals of thehabitat suitability model (i.e. logistic regression), but also forexamining the convergence of the optimization model. Further-more, the spatial data is often spatially autocorrelated. In our study,although the p-values of theWilcoxoneManneWhitney test for theVivid Niltava (0.24), Brownish-flanked Bush-warbler (0.66), andGreen-backed Tit (0.61) indicated that there are no significant dif-ferences between habitat suitability estimated by logistic modelsand auto-logistic models (Augustin et al., 1997), we suggest thatauto-logistic models can be used in future studies.

The combined LUPOlib-Marxan approach has potential to beapplied across different scales and domains (Robson, 2014). Previ-ous studies have indicated that spatial optimization methods canbe used for landscape modificationwith conservation purposes at aregional scale (Holzk€amper et al., 2006; Holzk€amper and Seppelt,2007a; Westphal et al., 2007). While our LUPOlib-Marxanapproach can be used across various scales, a limitation of areafor landscape modification should be set based on the land priceand the willingness of landowners. The economic cost could be setnot only for constraints but also for the objective function of theLUPOlib-Marxan approach. For potential end users, the LUPOlib-Marxan approach could be used to restore landscape structure forcritically endangered species and habitats (Pierce et al., 2005) andareas with high vulnerability.

5. Conclusion

This study combined optimal landscape and spatialoptimization-related conservation tools, LUPOlib, with Marxan toprovide an approach for conservation that involves priority mapsand basic guidance for implementing restoration actions. Theapproach delivers an alternative design of reserve network inwhich the landscape feature is considered to establish suitableenvironments for target species. The case studies underscored thekey effectiveness (e.g., more suitable, more aggregate, smaller sizeof the reserve network after minor landscape modification) of thecombined approach for landscape management in SCP. Moreover,not only does the combined LUPOlib-Marxan approach help policy-makers identify the boundaries of areas to be modified in theimplementation of a conservation policy, the approach also pro-vides suggestions regarding conservation-related actions in sup-port of the generation of suitable environments for target species.However, having willing landowners is fundamental.

Acknowledgments

The authors would like to thank the Ministry of Science andTechnology of the Republic of China, Taiwan, for financially sup-porting this research under Contract No. NSC101-2923-I-002-001-MY2. DSS received funding through the projects EU BON andSCALES. The EU BON (project no. 308454, Hoffmann et al., 2014)

Y.-P. Lin et al. / Environmental Modelling & Software 60 (2014) 121e133132

and SCALES projects (no. 226852, Henle et al., 2010) were fundedby the European Commission (EC) under the 7th Framework Pro-gramme (no. 226852).

Appendix A. Supplementary material

Supplementary material related to this article can be found athttp://dx.doi.org/10.1016/j.envsoft.2014.06.009.

References

Augustin, N.H., Mugglestone, M.A., Buckland, S.T., 1997. An autologistic model forthe spatial distribution of wildlife. J. Appl. Ecol. 33, 339e347.

Ball, I.R., Possingham, H.P., 2000. Marxan (v. 1.8.6): marine reserve design usingspatially explicit annealing. A manual prepared for The Great Barrier ReefMarine Park Authority. Retrieved from: http://www.marineplanning.org/pdf/marxan_manual_1_8_2.pdf.

Ball, I., Possingham, H.P., Watts, M.E., 2009. Marxan and relatives: software forspatial conservation prioritization. In: Moilanen, A., Possingham, H.P.,Wilson, K.A. (Eds.), Spatial Conservation Prioritization. Oxford University Press.,Oxford, pp. 185e195.

Bennett, N.D., Croke, B.F.W., Guariso, G., Guillaume, J.H.A., Hamilton, S.H., Jakeman, A.J.,Marsili-Libelli, S., Newhama, L.T.H., Norton, J.P., Perrin, C., Pierce, S.A., Robson, B.,Seppelt, R., Voinov, A.A., Fath, B.D., Andreassian, V., 2013. Characterising perfor-mance of environmental models. Environ. Model. Softw. 40, 1e20.

Bryan, B.A., Crossman, N.D., 2008. Systematic regional planning for multipleobjective natural resource management. J. Environ. Manage. 88, 1175e1189.

Ciarleglio, M., Barnes, J.W., Sarkar, S., 2009. ConsNet: new software for the selectionof conservation area networks with spatial and multi-criteria analyses. Ecog-raphy 32, 205e209. http://dx.doi.org/10.1111/j.1600-0587.2008.05721.x.

Council of Agriculture, Executive Yuan, Taiwan, 2013. Enforcement Rules of theWildlife Conservation Act.

Crossman, N.D., Bass, D.A., 2008. Application of common predictive habitattechniques for post-border weed risk management. Divers. Distribut. 14,213e224.

Crossman, N.D., Bryan, B.A., 2006. Systematic landscape restoration using integerprogramming. Biol. Conserv. 128, 369e383.

Crossman, N.D., Bryan, B.A., de Groot, R.S., Lin, Y.P., Minang, P.A., 2013. Land sci-ence contributions to ecosystem services. Curr. Opin. Environ. Sustain. 5,509e514.

Delavenne, J., Metcalfe, K., Smith, R.J., Vaz, S., Martin, C.S., Dupuis, Coppin, L.,Carpentier, A., 2012. Systematic conservation planning in the eastern EnglishChannel: comparing the Marxan and Zonation decision-support tools. ICES J.Mar. Sci. 69, 75e83. http://dx.doi.org/10.1093/icesjms/fsr180.

Ding, T.S., Yuan, H.W., Geng, S., Lin, Y.S., Lee, P.F., 2005. Energy flux, body size anddensity in relation to bird species richness along an elevational gradient inTaiwan. Global Ecol. Biogeogr. 14, 299e306. http://dx.doi.org/10.1111/j.1466-822x.2005.00159.x.

Esselman, P.C., Edgar, M., Breck, J., Hay-Chmielewski, E.M., Wang, L., 2013. Riverineconnectivity, upstream influences, and multi-taxa representation in a conser-vation area network for the fishes of Michigan, USA. Aquat. Conserv. Mar.Freshwater Ecosyst. 23, 7e22.

Henle, K., Kunin, W., Schweiger, O., Schmeller, D.S., Grobelnik, V., Matsinos, Y.,Pantis, J., Penev, L., Potts, S.G., Ring, I., Simil€a, J., Tzanopoulos, J., van den Hove, S.,Baguette, M., Clobert, J., Excoffier, L., Lengyel, S., Marty, P., Moilanen, A.,Porcher, E., Storch, D., Steffan-Dewenter, I., Sykes, M.T., Zobel, M., Settele, J.,2010. Securing the conservation of biodiversity across administrative levels andspatial, temporal, and ecological scales e research needs and approaches of theSCALES project. GAIA 19, 186e193.

Hoffmann, A., Penner, J., Vohland, K., Cramer, W., Doubleday, R., Henle, K.,K€oljalg, U., Kühn, I., Kunin, W.E., Negro, J.J., Penev, L., Rodríguez, C.,Saarenmaa, H., Schmeller, D.S., Stoev, P., Sutherland, W.J., Tuama, �E.�O.,H€auser, C., 2014. The need for a biodiversity policy support process e Buildingthe European contribution to a global Biodiversity Observation Network (EUBON). Nat. Conserv. 6, 49e65.

Hermoso, V., Januchowski-Hartley, S.R., Pressey, R.L., 2013. When the suit does notfit biodiversity: loose surrogates compromise the achievement of conservationgoals. Biol. Conserv. 159, 197e205.

Holzk€amper, A., Lausch, A., Seppelt, R., 2006. Optimizing landscape configuration toenhance habitat suitability for species with contrasting habitat requirements.Ecol. Model. 198, 277e292. http://dx.doi.org/10.1016/j.ecolmodel.2006.05.001.

Holzk€amper, A., Seppelt, R., 2007a. Evaluating cost-effectiveness of conservationmanagement actions in an agricultural landscape on a regional scale. Biol.Conserv. 136, 117e127. http://dx.doi.org/10.1016/j.biocon.2006.11.011.

Holzk€amper, A., Seppelt, R., 2007b. A generic tool for optimising land-cover pat-terns and landscape structures. Environ. Model. Softw. 22, 1801e1804. http://dx.doi.org/10.1016/j.envsoft.2007.02.008.

Klein, C., Wilson, K., Watts, M., Stein, J., Berry, S., Carwardine, J., Smith, M.S.,Mackey, B., Possingham, H., 2009. Incorporating ecological and evolutionaryprocesses into continental-scale conservation planning. Ecol. Appl. 19, 206e217.

Knight, A.T., Cowling, R.M., Possingham, H.P., Wilson, K.A., 2009. From theory topractice: designing and spatial prioritization approaches to better implementconservation action. In: Moilanen, A., Possingham, H.P., Wilson, K.A. (Eds.),Spatial Conservation Prioritization. Oxford University Press, Oxford,pp. 249e259.

Labiosa, W.B., Forney, W.M., Esnard, A.-M., Mitsova-Boneva, D., Bernknopf, R.,Hearn, P., Hogan, D., Pearlstine, L., Strong, D., Gladwin, H., Swain, E., 2013. Anintegrated multi-criteria scenario evaluationweb tool for participatory land-useplanning in urbanized areas: the Ecosystem Portfolio Model. Environ. Model.Softw 41, 210e222.

Lehtom€aki, J., Moilanen, A., 2013. Methods and workflow for spatial conservationprioritization using Zonation. Environ. Model. Softw. 47, 128e137.

Levin, N., Watson, J.E., Joseph, L.N., Grantham, H.S., Hadar, L., Apel, N.,Perevolotsky, A., DeMalach, N., Possingham, H.P., Kark, S., 2013. A framework forsystematic conservation planning and management of Mediterranean land-scapes. Biol. Conserv. 158, 371e383.

Levy, J.S., Ban, N.C., 2013. A method for incorporating climate change modelling intomarine conservation planning: an Indo-west Pacific example. Mar. Policy 38,16e24.

Lin, C.W., Hsu, F.H., Ding, T.S., 2011. Applying territory mapping method to censusthe breeding bird community composition in a Montane Forest of Taiwan.Taiwan J. Forest Sci. 26, 267e285.

Lin, C.W., 2006. Studying Breeding Bird Densities in Meifeng Area by TerritoryMapping (Master dissertation). National Taiwan University. Available fromNational Digital Library of Theses and Dissertations in Taiwan. http://handle.ncl.edu.tw/11296/ndltd/53654031825975833140.

Linke, S., Kennard, M.J., Hermoso, V., Olden, J.D., Stein, J., Pusey, B.J., 2012. Mergingconnectivity rules and large-scale condition assessment improves conservationadequacy in river systems. J. Appl. Ecol. 49, 1036e1045. http://dx.doi.org/10.1111/j.1365-2664.2012.02177.x.

Margules, C.R., Pressey, R.L., 2000. Systematic conservation planning. Nature 205,243e253. http://dx.doi.org/10.1038/35012251.

Margules, C.R., Sarkar, S., 2007. Systematic Conservation Planning. CambridgeUniversity Press, Cambridge.

Marshall, M.R., Cooper, R.J., 2004. Territory size of a migratory songbird in responseto caterpillar density and foliage structure. Ecology 85, 432e445.

Minin, E.D., Moilanen, A., 2012. Empirical evidence for reduced protection levelsacross biodiversity features from target-based conservation planning. Biol.Conserv. 153, 187e191.

Moilanen, A., 2007. Landscape zonation, benefit functions and target-based plan-ning: unifying reserve selection strategies. Biol. Conserv. 134, 571e579. http://dx.doi.org/10.1016/j.biocon.2006.09.008.

Moilanen, A., Possingham, H.P., Polasky, S., 2009. A mathematical classification ofconservation prioritization problem. In: Moilanen, A., Possingham, H.P.,Wilson, K.A. (Eds.), Spatial Conservation Prioritization. Oxford University Press,Oxford, pp. 185e195.

Nackoney, J., Williams, D., 2013. A comparison of scenarios for rural developmentplanning and conservation in the Democratic Republic of the Congo. Biol.Conserv. 164, 140e149.

Pierce, S.M., Cowling, R.M., Knight, A.T., Lombard, A.T., Rouget, M., Wolf, T., 2005.Systematic conservation planning products for land-cover planning: interpre-tation for implementation. Biol. Conserv. 125, 441e458. http://dx.doi.org/10.1016/j.biocon.2005.04.019.

Pressey, R.L., Cabeza, M., Watts, M.E., Cowling, R.M., Wilson, K.A., 2007. Conserva-tion planning in a changing world. Trends Ecol. Evol. 22, 583e592. http://dx.doi.org/10.1016/j.tree.2007.10.001.

Pressey, R.L., Watts, M.E., Barrett, T.W., Ridges, M.J., 2009. The C-plan conservationplanning system: origins, applications, and possible futures. In: Moilanen, A.,Possingham, H.P., Wilson, K.A. (Eds.), Spatial Conservation Prioritization. OxfordUniversity Press, Oxford, pp. 211e234.

Robson, B.J., 2014. State of the art in modelling of phosphorus in aquatic systems:review, criticisms and commentary. Environ. Model. Softw. http://dx.doi.org/10.1016/j.envsoft.2014.01.012.

Rodenhouse, N.L., Sherry, T.W., Holmes, R.T., 1997. Site-dependent regulation ofpopulation size: a new synthesis. Ecology 78, 2025e2042.

Schill, S., Raber, G., 2009. Protected Area Tools (PAT) for ArcGIS 9.3™ Version 3.0 e

USER MANUAL and TUTORIAL (General Technical Report NC-245). The NatureConservancy, Arlington, VA, U.S.. Retrieved from:http://maps.usm.edu/pat/files/PAT_v3_Tutorial.pdf.

Smith, R.J., Goodman, P.S., Matthews, W.S., 2006. Systematic conservation planning:a review of perceived limitations and an illustration of the benefits, using a casestudy from Maputaland, South Africa. Oryx 40, 400e410. http://dx.doi.org/10.1017/S0030605306001232.

Smith, R.J., Eastwood, P.D., Yoshitaka, O., Rogers, S.I., 2009. Developing best practicefor using Marxan to locate Marine Protected Areas in European waters. ICES J.Mar. Sci. 66, 188e194. http://dx.doi.org/10.1093/icesjms/fsn198.

Smith, T.M., Shugart, H.H., 1987. Territory size variation in the ovenbird: the role ofhabitat structure. Ecology 68,, 695e704.

Tole, L., 2006. Choosing reserve sites probabilistically: a Colombian Amazon casestudy. Ecol. Model. 194, 344e356. http://dx.doi.org/10.1016/j.ecolmodel.2005.10.027.

Wang, Y.H., Yang, K.C., Bridgman, K.L., Lin, L.K., 2008. Habitat suitability modellingto correlate gene flow with landscape connectivity. Landsc. Ecol. 23, 989e1000.http://dx.doi.org/10.1007/s10980-008-9262-3.

Y.-P. Lin et al. / Environmental Modelling & Software 60 (2014) 121e133 133

Watts, M.E., Ball, I.R., Stewart, R.S., Klien, C.J., Wilson, K., Steinback, C., Lourival, R.,Kircher, L., Possigham, H.P., 2009. Marxan with Zones: software for optimalconservation based land- and sea-use zoning. Environ. Model. Softw. 24,1513e1521. http://dx.doi.org/10.1016/j.envsoft.2009.06.005.

Westphal, M.I., Field, S.A., Possingham, H.P., 2007. Optimizing landscape configu-ration: a case study of woodland birds in the Mount Lofty Ranges, SouthAustralia. Landsc. Urban Plan. 81, 56e66. http://dx.doi.org/10.1016/j.landurbplan.2006.10.015.

Zaragozí, B., Belda, A., Linares, J., Martínez-P�erez, J.E., Navarro, J.T., Esparza, J., 2012.A free and open source programming library for landscape metrics calculations.Environ. Model. Softw. 31, 131e141.

Zhang, M.-G., Zhou, Z.-K., Chen, W.-Y., Slik, J.W.F., Cannon, C.H., Raes, N., 2012. Usingspecies distribution modeling to improve conservation and land use planning ofYunnan, China. Biol. Conserv. 153, 257e264.

Zielinski, W.J., Carroll, C., Dunk, C., 2006. Using landscape suitability models toreconcile conservation planning for two key forest predators. Biol. Conserv. 133,409e430. http://dx.doi.org/10.1016/j.biocon.2006.07.003.