Modeling the spatio-temporal dynamics and interactions of households, landscapes, and giant panda...

Ecological Modelling 183 (2005) 47–65

Modeling the spatio-temporal dynamics and interactions ofhouseholds, landscapes, and giant panda habitat

Marc A. Lindermana,∗, Li Ana, Scott Bearera, Guangming Hea,Zhiyun Ouyangb, Jianguo Liua

a Department of Fisheries& Wildlife, Michigan State University, East Lansing, MI 48824, USAb Department of Systems Ecology, Research Center for Eco-Environmental Sciences, Chinese Academy of Sciences, Beijing, China

Received 17 November 2003; received in revised form 15 June 2004; accepted 12 July 2004

Abstract

Human modification of land-cover has been a leading cause of floral and faunal species extirpation and loss of local and globalbiodiversity. As natural areas are impacted, habitat and populations can become fragmented and isolated. This is particularlyevident in the mountainous areas of southwestern China that support the remaining populations of giant pandas (Ailuropodamelanoleuca). Giant panda populations have been restricted to remnants of habitat from extensive past land use and land-coverchange. Households are a basic socio-economic unit that continues to impact the remaining habitat through activities such asfuelwood consumption and new household creation. Therefore, we developed a spatio-temporal model of human activities andtheir impacts by directly integrating households into the landscape. The integrated model allows us to examine the landscapef plication,w useholds,t g habitatw ution ofh he currentl©

K

pB

ndoft-also

0

actors influencing the spatial distribution of household activities and household impacts on habitat. As an example ape modeled household activities in a giant panda reserve in China and examined the spatio-temporal dynamics of ho

he landscape, and their mutual interactions. Human impacts are projected to result in the loss of up to 16% of all existinithin the reserve over the next 30 years. In addition, we found that accessibility largely controls the spatial distribousehold activities and considerable changes in management and household activities will be required to maintain t

evel of habitat within the reserve.2004 Elsevier B.V. All rights reserved.

eywords:Landscape; Households; Giant panda; Habitat; Model; Human impacts

∗ Corresponding author. Present address: Department of Geogra-hy, University of Louvain, Place Pasteur 3, 1348 Louvain-la-Neuve,elgium. Tel.: +32 10472871; fax: +32 10472877.E-mail address:[email protected] (M.A. Linderman).

1. Introduction

Appropriation of natural areas through urban aagricultural expansion has drastically altered muchthe land surface (Vitousek et al., 1997; Rutledge eal., 2001). Modification of habitat through less intense land use such as fuelwood collection has

304-3800/$ – see front matter © 2004 Elsevier B.V. All rights reserved.doi:10.1016/j.ecolmodel.2004.07.026

48 M.A. Linderman et al. / Ecological Modelling 183 (2005) 47–65

resulted in drastic changes in natural systems (Liu et al.,2001). These changes have enormous implications forecosystem processes, biodiversity, and species persis-tence (Ceballos and Ehrlich, 2002). This is particularlyrelevant for the conservation of the giant panda (Ail-uropodamelanoleuca). Habitat destruction and poach-ing have reduced the wild population to approximately1000 pandas (Schaller et al., 1985). Many studies havebeen conducted on panda biology and behavior (e.g.Schaller et al., 1985). Most of the studies are empiri-cal or field based. There have been also a number ofmodeling studies, which have simulated panda pop-ulation dynamics (Zhou and Pan, 1997), panda rela-tionships to bamboo dynamics (Reid et al., 1989; Wuet al., 1996; Carter et al., 1999), and household prefer-ences and characteristics related to panda habitat (An etal., 2001, 2002). However, few studies have examinedthe factors influencing the spatio-temporal dynamicsof households, their impacts on giant panda habitat,and their mutual interactions (Liu et al., 1999). To bet-ter understand household impacts on giant panda habi-tat, we developed a model in which the interactionsbetween households, the landscape, and giant pandahabitat could be studied and based on the analyses pro-vided practical information for conservation and man-agement planning.

Much of human land-cover change is carried out atthe household level as households are basic decisionand consumption units (Liu et al., 2003). The rapidincrease in the number of households increases thedh foree turals inter-a ionsf tom eiri ped,i ,1 npS tiono notcM reas-i andt ngd iven

the complexity of the models and human–environmentsystems (Couclelis, 2001). Cellular models, discreetin time and space, allow for simplified modelingrelationships and provide a structured environment inwhich various interactions and levels of detail can bestudied (Benenson, 1999).

The overall goal of the model, Household AndLandscape Integration Model (HALIM) was to developa generalized modeling approach in which spatio-temporal household processes could be integratedinto realistic landscapes. For this study, we used ageneric spatially explicit cellular model to examine theinteractions of households and panda habitat throughtheir mutual relationships with the landscape. Usinga generic cellular framework facilitated the use of de-tailed digital data to accurately describe the landscapeand household characteristics while providing a meansto integrate inherently different natural and householdprocesses. Furthermore, this flexibility provides apractical and accessible framework in which varyingaspects and complexity of socio-economic and naturalsystems and their interactions can be integrated.

As a preliminary study, we used HALIM to evaluatethe spatio-temporal effects of landscape-level house-hold activities on giant panda habitat in southwest-ern China by integrating households, forest cover, andwildlife habitat through their mutual relationships withthe landscape. This allowed us to examine the individ-ual spatio-temporal dynamics and the various interac-tions between the landscape, household activities, andw ret holdc tiono fac-t eser vari-o gated tifiedi rac-t

2

2

a-t

emand for more resources (Liu et al., 2003). Couplingousehold activities with natural processes is theressential to accurately model human impacts on naystems, to increase our understanding of humanctions with landscapes, and to provide viable opt

or mitigating future impacts. Various approachesodeling spatially explicit human activities and th

mpacts on natural systems have been develoncluding statistical techniques (Mertens and Lambin997), agent-based models (Berger, 2001; An et al., iress), and cellular approaches (Balzter et al., 1998).tatistical models have provided detailed informaf the spatial dynamics of systems, but are oftenonducive to generic frameworks (Lambin, 1994).ore complex agent-based approaches allow inc

ngly detailed human interactions with each otherhe environment in which they live. However, buildiescriptive agent-based models is often difficult g

ildlife habitat. Our specific aims of this study weo examine the influence of landscape-level househaracteristics on the quantity and spatial distribuf panda habitat and to determine the landscape

ors influencing these household activities. Using thesults, we examined possible consequences ofus policy scenarios, provided suggestions to mitiamage to the remaining panda habitat, and iden

mportant landscape, household, and habitat inteions for future modeling efforts.

. Methods

.1. Study area

Our field study was conducted in the Wolong Nure Reserve in southwestern China (Fig. 1), located

M.A. Linderman et al. / Ecological Modelling 183 (2005) 47–65 49

Fig. 1. Wolong Nature Reserve lies in the Qionglai Mountains between the Tibetan plateau and Sichuan basin.

between 102◦52′ and 103◦24′E and 30◦45′ and31◦25′N. Wolong is one of the largest reserves (cov-ering approximately 2,00,000 ha) dedicated to giantpanda conservation and is estimated to contain about10% of the remaining wild panda population (Zhanget al., 1997). Approximately 40% of the reserve is cur-rently forested. Elevations range 1200–6525 m creatingseveral climatic zones and consequently high biolog-ical diversity. The distribution of overstory vegetationin the reserve is related to the elevation.

Most forests in the reserve were logged (either clearcut or selectively cut) from 1916 until the reserve wasestablished in 1975, reaching peak intensity between1961 and 1975 (Schaller et al., 1985). Commerciallogging typically resulted in relatively large clearcutsdistributed throughout the reserve. Logging has been

officially banned in the reserve since 1975; however,to varying degrees illicit logging does continue (M.Linderman, personal observation). Other humanactivities have also been a major contribution to forestloss and, consequently, to the spatial distribution ofhabitat (Liu et al., 1999, 2001).

In 2001, approximately 4440 local residents inabout 1000 households resided within the reserve.The majority of these residents are farmers with theprimary economic activities consisting of farmingmaize and vegetables, raising livestock such as pigsand yaks, and collecting wild herbs. A householdusually relies on fuelwood for heating, cooking, andlivestock feed preparation (An et al., 2001). Selectivelogging for household fuelwood collection typicallychanges the species composition in the overstory and


reduces canopy cover until all overstory vegetation isremoved. Since 1974, immigration and new householdcreation have largely been dictated by local policy withimmigration restricted by marriage and new householdcreation limited or inadvertently encouraged throughvarious policies. Households have traditionally fo-cused on subsistence agriculture, but increasing accessto markets has provided some cash crop opportunities.

2.2. Data and model parameterization

Several sources of data were used as model input orused to parameterize and validate the model. Satellitedata and topographic maps were resampled to a pixelsize corresponding to the landscape grid and used todescribe the abiotic features and the distribution ofhousehold activities and vegetation throughout thereserve. Socio-economic and demographic data werecollected from local government agencies and ourhousehold surveys conducted from 1998 to 2001 (An etal., 2001) to determine fuelwood collection, householdlocations, and household creation rates. Literature onpanda behavior and landscape analyses of habitat wasused to parameterize the habitat sub-model (Schalleret al., 1985; Ouyang et al., 1996; Liu et al., 2001).

Abiotic information was derived from topographicmaps of the reserve. A Digital Elevation Model (DEM)was interpolated from digitized 100-m contours. Slopeand aspect data were derived from the DEM. Informa-tion on the distribution of forests was obtained fromt and1 The1 uireda ellitep uthD , andt o ac-c eent for-e seeL

ousf tionp ail-a ragev ativei iffi-c

gest that much of the reserve was commercially loggedfrom 1916 until 1975. Measurements taken in the late1990s indicated much of the lower altitude forests tobe well below old-growth volumes. Average volumesfor broadleaf forests below 2600 m were approximately80 m3/ha. It is likely that these forests were the first tobe harvested in the first half of the century and haveregrown to current volume levels.

Based on regrowth data for the broadleaf forestsin Wolong, we estimated the average volume for1965 to be approximately 45 m3/ha. Stand volumefor subalpine conifers was on average approximately300 m3/ha. Subalpine stand volume was high enoughsuch that variations in estimates would not significantlyinfluence the model results. Forest regrowth was in-cluded in the model to allow for previously logged re-gions to regenerate and the addition of biomass andregrowth in selectively logged cells. Separate regrowthmodels were developed for each forest type based onspecies composition, stand age, and altitudinal zone.Model parameters were derived from over 30 plots dis-tributed throughout the reserve (Liu et al., 1999), andapproximation of species regrowth and yield modelswas derived from the data of the Sichuan Departmentof Forestry (Yang and Li, 1992).

A household survey was conducted from 1998 to2001 and included 220 of the households within thereserve (An et al., 2001). Households were queriedon fuelwood use, fuelwood collection, agricultural ac-tivity, household creation, and other associated socio-e nals waso in-f ther tiono veysa holdm In-c , ando tely0 en tob eena h yearb ately2

holdt nos1 ery

he classification of four dates (1965, 1974, 1987,997) when remote sensing data were obtained.965 data are Corona stereo-pair photographs acqs part of the Corona photo-reconnaissance satroject (USGS Eros Data Center, Sioux Falls, Soakota). The 1974 data are Landsat MSS images

he 1987 and 1997 data are Landsat TM images. Tount for the spectral and spatial differences betwhe data, each image was visually interpreted intost and non-forest areas (for classification detailsiu et al., 2001).

Uncertainty in the 1965 stand volume of the variorest types posed the most difficult parameterizaroblem. While basic coverage information was avble from satellite photographs, data on the aveolume throughout the reserve were scant. Quantitnformation dating back nearly 40 years is either dult to obtain or non-existent.Schaller et al. (1985)sug-

conomic and demographic information. Additioocio-economic and demographic informationbtained from local government records. Census

ormation was obtained from each township withineserve. Local governments also maintain informan land allocated to each household. From the surnd census data, it was found that each houseaintains on average 0.7 ha of agricultural land.

luding the area of the physical house, garden areather buildings, the typical total area is approxima.8 ha. Therefore, the scale of the model was chose 90 m× 90 m (0.81 ha). New households have bdded to the reserve at a rather steady number eacetween 1965 and 1997. On average, approxim4 new households were created each year.

We measured the location of each househrough the use of field measurements or Iko-m resolution satellite imagery. Ikonos imag


acquired in 2000 by SpaceImaging was georeferencedwith ground control points measured using a GlobalPositioning System with sub-meter accuracy (TrimblePathfinder Pro XRS receiver and Community BaseStation). We then identified households in the imagesand recorded the location. We used all households cre-ated on or before 1965 to create the initial distributionof households to correspond to the initial 1965 forestcover information.

Fuelwood use was calculated based on a survey ofover 50 households and physical measurements of an-nual use (An et al., 2001). The volume of wood variedbetween 8 and 30 m3 and averaged 15 m3. A base an-nual volume of wood used by each household in themodel was then 15 m3. We derived preference for fuel-wood collection and household creation sites by com-paring DEM and slope coverages, and house locationsand fuelwood sites. Distances between household lo-cations to fuelwood collection sites varied from 50 mto over 5 km. The average distance for 100 householdssurveyed was approximately 500 m. Households pre-ferred to collect fuelwood in flat areas (<20◦ slope)and had a decreasing probability relative to elevation.

Behavioral studies have described panda habitat asa function of forest cover, slope, and altitude (Schalleret al., 1985; Ouyang et al., 1996; Liu et al., 2001).Therefore, we determined habitat suitability using amultiplicative combination of the three factors (forestcover, altitude, and slope) available for the 30-yeartime span (Liu et al., 2001). Because non-foresteda giantp ulti-p nda theo

2

S(a -l lingo nd-s geo-r thati l ands s thefl ugh

Fig. 2. A conceptual flow schematic diagram of the model.

sub-models and individual modeling aspects of Markovchains, cellular automata, percolation models and oth-ers according to the process being modeled.

HALIM includes four sub-models: fuelwood collec-tion, household creation, forest regrowth, and pandahabitat. The resulting impacts of the distribution ofhousehold activities are integrated directly into giantpanda habitat models and allow model predictions tobe measured in terms of changes to landscape indicesof panda habitat. The sub-models and their interactionsare shown inFig. 2. Household activities and forest dy-namics are influenced by the abiotic characteristics ofthe landscape. Each of the household activities influ-ences the spatial distribution of forest cover. The forestregrowth sub-model allows for forest re-establishmentand annual growth of non-climax forests. Finally, thesuitability of giant panda habitat is determined fromforest cover along with abiotic factors (Liu et al., 2001).

The landscape was divided into a regular latticecomposed of 90 m× 90 m grid cells. For this model, theprobability of the initiation of most sub-model events(e.g. fuelwood collection, household creation, etc.) oc-

reas are considered unsuitable habitat for theanda, forest/non-forest classifications were mlicative factors of 1 or 0, respectively. Slope altitude multiplicative factors were proportional tobserved use by pandas.

.3. Model description

Our model (HALIM) was developed using SELESpatially Explicit Landscape Event Simulator) (Fallnd Fall, 2001; Fall et al., 2001). SELES is a high

evel programming language that facilitates modef the temporal and spatial dynamics of gridded lacapes. SELES also allows the incorporation ofeferenced raster data, the definition of systemsnteract on gridded landscapes, and the temporapatial dynamics of these systems. SELES provideexibility to incorporate these various systems thro


curring at each grid cell was determined by the pixelvalues (data layer values) of the cell and, dependingon the sub-model, surrounding cells. The number ofsub-model events is determined by the sub-model pa-rameters with the location of the event stochasticallydetermined by a relative cell probability (e.g. a cell witha probability of 0.5 has twice the probability of the oc-currence of a landscape event compared to a cell witha 0.25 probability, but does not have a 50% probabil-ity of an event occurrence). Depending on the processof interest the model also allows for landscape eventsto spread to neighboring cells (e.g. if a cell does notcontain sufficient fuelwood for the annual collectionof a household’s fuelwood needs, necessary fuelwoodcollection can take place in a neighboring cell).

The sub-models are described below along with ex-amples of the parameters and probability functions:

• Fuelwood collection– It was assumed that house-hold residents collect fuelwood based on availabil-ity, accessibility, and previous fuelwood collectionactivity. Typically, fuelwood is collected around thehousehold. As these areas are diminished, foragingextends to the neighboring areas characterized byeasy accessibility (Liu et al., 2001). Many residentshave been forced to travel several kilometers to col-lect annual stocks of fuelwood (An et al., 2001).Accessibility is characterized in this model by thedistance to collection site, slope, and elevation andis defined as a cost function relative to the distance

ilityath

s a

eter-elyestel-of

stedofn a

• e-on

For

example, past trends have been relatively stable.Policies, however, have been shown to affecthousehold creation. Therefore, a range of householdcreation rates about the past trend was examined.Each new household was presumed to establish itsown agriculture land, clearing the forest area or oc-cupying previously deforested area. The location ofeach new household was dependant on suitable agri-culture land and proximity to transportation routesand other households. The household sub-modelwas, therefore, determined by three parameters:distance–cost factor to transportation, abioticfactors, and proximity to other households. ThepreciseX andY coordinates of the actual residencewere not included in this model. Rather, households,including the physical residence, agriculture land,garden area, and various other buildings, werepresumed to occupy cells of the landscape. Suitableagriculture areas are based on abiotic factors: slope,aspect, and elevation. While agriculture activityoccurs on slopes up to 40◦, low-slope areas arepreferred. Preference for low-elevation areas wasalso considered. For example, based on survey dataprobabilities for household placement based solelyon elevation were measured as:

P(household|e) =

{0.00 (e > 2500)}{0.08 (2250< e ≤ 2500)}{0.82 (1750< e ≤ 2250)}{1.00 (e ≤ 1750)}

sht.

andithinjor

• st,andtheand

zoneer-ed

pixelity

ion.pre-

to roads and main paths and topographic variab(i.e. slope and elevation difference along the pto the cell location). The probability function walinearly decreasing function of increasing cost:

P(fuelwood|cost)=(

1 −(

Cost

Maximum cost

))

Forest cover and average yield per hectare dmined availability. Households are also more likto return to the same cell location, if sufficient forvolume exists, or neighboring cells of previous fuwood extraction. Therefore, a higher probabilitycollection was assigned to cells previously harveand to neighboring cells. The overall probabilityfuelwood extraction for each forested cell is themultiplicative combination of these factors.Household creation– The number of new housholds each year was predetermined basedpotential policy and socio-economic impacts.

In areas of higher elevation (e), preference wagiven to slopes facing south to maximize sunligHouseholds were also more likely to develop ladjacent to previously established houses and wshort distances (typically less than 2 km) of matransportation routes.Forest cover– Four forest categories (non-foreevergreen broadleaf, deciduous broadleaf,subalpine conifer) were identified throughoutreserve based on remote sensing, elevation,species distribution (Schaller et al., 1985). Initialstand volume was estimated for each elevationbased on approximate time and intensity of commcial logging activity. Each forested cell was assumto increase in biomass and each non-forestedhad a probability to re-establish based on proximto other forest pixels and time since deforestatRegrowth models were derived for each of the


dominant species within each elevation zone frompublished and empirical data (Yang and Li, 1992).Regrowth is calculated based on species and approx-imations of logistic regrowth curves of total volume.An example of the calculation is given below:

V (t, Van, Vmax) = {0.0 (t < tlag)}{V + Van(t>tlag andV<Vmax)}

where t is the time since harvest,tlag is a nor-mally distributed lag time since harvest untilre-establishment,Van is the annual volume incre-ment, andVmax is the maximum volume accordingto species type. Upper asymptotic limits on volumewere controlled by stand maximum values ratherthan time due to concurrent fuelwood collection.

• Habitat suitability– The final habitat classificationwas a categorized suitability measure of fourclasses termed highly suitable, suitable, marginallysuitable, and unsuitable (Liu et al., 2001). Theimpacts from household activities are reflectedin the habitat suitability model as impacts fromfuelwood activity and agriculture development.Measures of panda habitat quantity and suitabilityallow analysis of the temporal and spatial dynamicsof, the influence of household characteristics on,and future giant panda habitat.

Landscape events (e.g. fuelwood collection, forestregrowth) occurred on an annual time frame. The firstlandscape event in the model each year is the estab-lishment of new households and associated agriculturald nualf re-g ilityo

2

ereb itiald off pho-t wasb 965.T for3 singd ughv newh rac-

teristics and the relative influence of each individualparameter on the model output. Validation was donethrough comparison of model output over this time tomeasured habitat and household distributions.

We conducted sensitivity analyses for the householdand fuelwood collection sub-models. We examined thesensitivity of the household sub-model to each of itscomponents (abiotic, proximity, and cost function) bycomparing scenarios excluding components or varyingparameter estimations and the measured householddistribution in 1997. This was done because we wantedto show the overall influence each function had on theselection of new households and because some func-tions could not be varied systematically (e.g. abioticinfluences were based on conditional probabilities). Wemeasured accuracy and calculated landscape metricsbased on the average of 20 simulations. We also con-ducted systematic analyses of sensitivity of individualparameters for the fuelwood sub-model, such as thepropensity to return to previous fuelwood collectionsites and distance to fuelwood collection sites. Sinceparameterization of stand volumes for broadleaf forestsbelow 2600 m contained relatively large uncertainty,several average stand volumes for the broadleaf forestswere tested, including 30, 45, 60, 75, and 90 m3/ha.

The accuracy of the predicted distribution of house-holds was measured through comparison of predictedlocations of households in 1997 to measured locations.Precise cell-by-cell prediction, however, was not theintention of this model. Foremost, the model is stochas-t tiala ar-e olde nt ofh larlyc rcento (1,2

redb ts tof ex-a s andt to ad (i.e.a arec ouse-h ma-t he

evelopment. Each household then collects its anuelwood volume. At the end of the year, forestrowth occurs for each forested cell and the suitabf panda habitat is updated.

.4. Model validation and sensitivity analyses

Model validation and sensitivity analyses wased on simulations started in 1965 with the inistribution of forest based on the classification

orest/non-forest categories from the 1965 Coronaographs. The original distribution of householdsased on all households established prior to or in 1he sensitivity and validation simulations were run2 years to correspond to the latest remote senata available (1997). We measured sensitivity throarying individual parameters such as the rate ofousehold creation, fuelwood use, and forest cha

ic. In addition, households do not occupy all potengricultural areas within the reserve. This leads toas with similar probabilities available for househstablishment. However, as the spatial arrangemeouseholds may have an impact on habitat, particurucial secondary habitat, we also examined the pef predicted households falling in close proximity, and 3 cells) of measured households.

Impacts from fuelwood collection were measuy comparison of predicted and measured impac

orest cover and habitat. Again, we did not expectct correspondence between the model prediction

he measured distributions. Collections sites are,egree, stochastically chosen both by the models with households, not all potential fuelwood siteshosen) and households (i.e. some degree of hold decisions is unpredictable regardless of infor

ion available). In addition, the natural variability of t


forests was not fully captured in the visual classifica-tions (i.e. the visual interpretation of forest distributiondid not include all forest gaps and edge complexity ata 90-m resolution) and illicit logging activities not in-cluded in the model make a direct accuracy assessmentdifficult.

To minimize the effect of natural and other influ-ences on the accuracy assessments, we limited analysesto regions within 5 km of the current household distri-bution. This distance corresponds to the approximatemaximum distance residents travel to collect fuelwood.Within the 5 km buffers, we used three validationmethods: visual appraisals of multitemporal data;direct comparison to a supervised classification; and acomparison between landscape indices. We comparedpredicted fuelwood impacts on forest cover to visualclassifications of forest cover from 1974 to 1997satellite imagery (Liu et al., 2001). We comparedmeasurements of the distribution of households anddigital classifications of forest cover as measured in1997 to final outputs from the model. Digital clas-sification of the 1997 forest cover was possible withextensive ground sample data and provided a moredetailed snapshot of the distribution of forest cover.Accuracy is reported as the percentage of predictedcells that correspond to measured cells (e.g. predictednon-forest versus measured non-forest cells). Thisignores possible omission errors and was used becauseof the difficulty in distinguishing natural variabilityand human impacts (e.g. illicit logging) on forest coverf heh ionsa forc rors.

eent andl , andc gb ndn cyc them de-t ure.T alsor pacto restc hes,m

(Baker and Cai, 1992) describing patch compactness,and connectivity between patch centroids (Forman andGodron, 1986) that describes clustering of patches.

2.5. Household impacts

To examine the relative influences of differenthousehold conditions on the landscape, a variety ofscenarios were run from 1965 until 2030. Each sce-nario was started using 1965 land-cover and house-hold data. From 1965 to 1997, we based the modelparameters on measured values. We then varied modelparameters for 1997–2030 to examine the impacts ofpossible changes. These scenarios represent situationswhere new policies were introduced after 1997. Param-eters we examined included fuelwood consumption perhousehold and the household growth rate (or immi-gration/emigration rate). The length of the simulationswas chosen based on the reliability of the model overthe previous 32 years and to permit sufficient time tocompare various scenarios and predict future impacts.We compared model scenarios based on impacts to gi-ant panda habitat as deforestation from fuelwood andhousehold construction removed habitat.

These scenarios included changes in fuelwoodconsumption levels of 30, 15, 10, 5, and 0 m3/year/household and household growth rates of 36, 24,12, 0,−12, and−24 new households created or re-moved each year after 1997, as well as combinations ofthese parameters. We chose these levels to reflect possi-b newp s, re-s city.F ndr ub-s ningf re-s on.E form ok-i en-e ture( ono ducef mi-g ten-t ver,t lts to

rom household activities even within 5 km of touseholds. Visual comparisons of model predictnd measured forest cover change are shownomparison between commission and omission er

In addition, comparisons were made betwhe quantity of forest area and disturbed areasandscape metrics such as patch size, shapeomplexity. Given the difficulty in distinguishinetween timber logging, fuelwood collection, aatural variability in forest cover, simple accuraomparisons of the model predictions relative toeasured landscape (particularly those from the

ailed classification) do not provide a complete picthe impacts measured from simulations wereeported as the landscape indices relative to the imf interest (e.g. household distribution and foover). Indices used include total number of patcean patch size, corrected perimeter to area (p/a) ratio

le future household characteristics resulting fromolicies and management efforts such as subsidietrictions, and/or increased accessibility to electrior example, efforts to limit fuelwood collection aeclaim agriculture land were initiated in 2000. Sidies have been offered in exchange for maintaiorests. The administration has also attempted totrict the location and quantity of fuelwood collectilectricity prices are also currently unaffordableost local farmers, particularly for heating and co

ng purposes. Affordable and consistent alternativergy sources may influence fuelwood use in the fuAn et al., 2002). Each of these or the combinatif these changes may provide an incentive to re

uelwood use. In addition, efforts to encourage eration out of the reserve are being instituted po

ially decreasing the number of households. Howehere is an increasing preference by younger adu


establish new households, and in response to subsidyopportunities, new households have actually recentlyincreased at much higher rates than in the past. There-fore, to reflect the possible range of values, we chosefuelwood consumption levels ranging from the currentmaximum known household consumption (double thecurrent average) to no fuelwood use. We also exam-ined household creation rates varying from a 50% in-crease in household establishment to a net emigrationof households to reflect policy influences on householdcreation over the next 30 years.

3. Results

3.1. Model validation and sensitivity

To examine the overall influence of the householdsub-model parameters (e.g. topography, distance totransportation, and proximity to other households),several variations of the household sub-model werecompared. We could not do a typical sensitivity testfor this sub-model as some of the parameters wereempirical look-up tables. Therefore, to examine theinfluence of each parameter, model outputs werecompared for several combinations of sub-modelparameters. For example, the household sub-modelincluding all three hypothesized parameters (abi-otic, distance, and proximity) (Fig. 3a) resulted inapproximately the same number of patches and sim-i sub-m ize,a ea-s -e holds( acedi s ofe ndt tesyT s themwo ingo izear olds(

Accuracy in terms of predicted household locationsagreeing with measured cell locations of household dis-tribution varied from 20 to 27% (Table 2). Incorpo-rating all of the parameters hypothesized to influencehousehold placement resulted in an accuracy of 27, 68,and 82, and 88% for predicted households within 0,1, 2, and 3 cells from measured households (Table 2).This suggests that the model was predicting householdsessentially within the same areas as those measuredto also contain households. Not including the distancefunction yielded the lowest accuracy of 63% for pre-dicted households within 3 cells of measured house-holds. The accuracy was 80% when a preference tocreate new households next to existing households wasnot included. Excluding the selection based on abioticfactors (i.e. slope and elevation) achieved an accuracyof 81% within 3 cells.

Sensitivity analyses conducted for each of the fu-elwood parameters showed influences from variationsin the distance and proximity factors (Table 3). Relax-ing the tendency for households to collect fuelwoodfrom previously cleared areas led to more fragmenta-tion and is reflected in the landscape metrics. Variationof the proximity factor three times more likely to re-turn to previous sites resulted in 35% fewer patchesand 54% larger patch sizes. Reducing the proximityfactor three times resulted and 52% more patches and34% smaller patch size (Table 3). In addition, perimeterand connectivity indices show increasing clustering asthe proximity factor is increased. Varying the distancec ingt dis-p is iss allerp n-n ara patchn ches.P r var-i reasei

ndv ucedf inp sus sedp itial

lar p/a ratio as the measured households. Thisodel also led to a 44% larger mean patch snd slightly higher connectivity compared to the mured distribution (Table 1). Excluding abiotic prefrences resulted in 71% more patches of houseTable 1) and caused some households to be pln regions of atypical topographic relief (e.g. areaxtreme slope) (Fig. 3b). Excluding the distance aopographic variation from main transportation rouielded a wide distribution of households (Fig. 3c).he number of patches was more than three timeeasured distribution. Mean patch size andp/a ratioere both considerably lower (Table 1). And, the lackf a proximity factor resulted in decreased clumpf households (low connectivity), smaller patch snd an increase in the number of patches (Table 1)elative to the measured distribution of househFig. 3d).

ost factor by 20% resulted in similar results. Eashe influence of the distance factor generated moreersed impacts occurring in smaller patches. Theen in the patch characteristics with more and smatches and decreasedp/a ratios and diminished coectivity (Table 3). Increased probability of using nereas conversely increased patch size, decreasedumber, and increased connectivity between patatch size varied by 17.9–33.7% and patch numbe

ed by 24.1 and 20.5% for a 20% decrease and incn the cost factor, respectively (Table 3).

Trends in deforestation relative to initial staolume were decreasing area of impact and redragmentation since more volume was availablereferred collection areas (Table 3). While the outputsing each of the five initial volumes shown inFig. 5doeemingly conform largely to expectations, increaeripheral impacts occur at both increased in


Fig. 3. Comparisons of the influence of the three multiplicative factors contained within the household sub-model. Accuracy of each scenariois shown relative to the measured households with corresponding predicted households and measured households shown in black, incorrectlypredicted households are shown in dark gray, actual households where no households were predicted are shown in white: (a) shows the predictedhousehold distribution in 1997 including all factors relative to the actual distribution; (b) is without abiotic preferences; (c) without cost factors;and (d) without proximity influences.

Table 1Landscape characteristics of the measured households in 1997 (Households 1997) compared to model scenarios

Number of patches Mean patch size (ha) p/a ratio Connectivity

Households 1997 94.00 40931 1.49 0.046All parameters 110.35 59101 1.50 0.053No proximity factor 261.00 24905 1.41 0.015No abiotic factor 161.90 40229 1.46 0.034No cost factor 280.60 23152 1.29 0.009

Values are averages of 20 simulations for each scenario.


Table 2Accuracy of the predicted household locations for the model scenarios relative to the household locations in 1997

Cells

0a 1b 2b 3b

No cost factor 20.6± 1.3c 47.3± 2.4 57.1± 2.6 63.0± 2.4No proximity factor 21.2± 1.1 54.3± 1.7 70.6± 1.3 79.8± 1.4No abiotic factor 22.4± 1.4 55.8± 2.4 71.8± 2.0 81.2± 2.3All parameters 27.4± 0.7 67.9± 1.5 82.5± 1.9 88.3± 1.9

a Accuracy as measured as predicted household locations occurring at measured household locations (titled 0).b Predicted locations within 1, 2, and 3 cells (labeled 1, 2, and 3, respectively) of measured household locations.c Uncertainties represent one standard error of the accuracies of the 20 simulations conducted for each scenario.

volumes and decreased volumes. Landscape metricsand overall model accuracy also follow this trend(Table 3). The lowest number of patches occurredwhen the initial forest stand volumes was 45 m3/ha.Decreasing stand volume caused larger overall habitatloss, particularly the core area nearest to households;however, smaller peripheral impacts were morecommon. As initial stand volume was increased, theoverall impact was diminished, however small pocketsof impact emerged where more continuous impactspreviously existed. These trends are clearly shown inthe decreasing patch perimeter and mean patch size.

Fig. 4shows a multitemporal comparison of the pre-dicted 32-year simulation of household activity and themeasured forest cover within 5 km of all households.There appears to be a good correspondence betweenthe model outputs and measured forest distribution.The basic trends in forest cover are comparable be-tween measured and predicted distribution of forest

Table 3Sensitivity of individual factors used within the fuelwood sub-model

Factor Parameter Number of patches Mean patch size (ha) p/a ratio Connectivity index

Proximity* 0.33 125.2 75.8 1.668 0.7191 192.2 49.2 1.606 0.3363 291.5 32.7 1.538 0.170

Distance* 0.8 145.8 65.8 1.630 0.5461 192.2 49.2 1.606 0.3361.2 231.6 40.4 1.587 0.277

I 3

Vicative

cover, though some differences from natural and otheractivities are apparent. In addition, the model was suc-cessful in capturing the basic trend in the distribution ofhouseholds based only on the initial 1965 distributionof households.

Accuracy and sensitivity analyses were done todetermine the overall validity of the model and theinfluence of individual parameters. The accuracy ofpredicted impact sites relative to measured impact alsoreflects more concentrated impacts as initial volume isincreased (Fig. 5). As fuelwood activity is focused oncore areas near households, model accuracy increases.At an initial stand volume of 30 m3/ha, the overallprediction accuracy is approximately 55%. As the vol-ume increased to 90 m3/ha, model accuracy increasedto 64% (Table 3). The increase in accuracy is largelya result of smaller areas being affected only nearhouseholds and decreased influence of stochasticity inchoosing distant fuelwood sites.

nitial volume (m /ha) 30 211.445 192.260 258.775 265.990 246.3

alues in bold represent hypothesized values.∗ The proximity and distance coefficients are unitless multipl

51.9 1.567 0.36549.2 1.606 0.33633.6 1.540 0.21230.3 1.502 0.16130.5 1.502 0.167

factors.


Fig. 4. Comparisons between visual classifications of satellite data from 1965, 1974, 1987, and 1997 and predicted forest cover due to householdactivities of corresponding years.


Fig. 5. Differences between predicted forest cover due to fuelwood collection compared to the digital classification at various starting volumesfor low-elevation forests. Forest/forest and non-forest/non-forest categories represent agreement between predicted and measured forested andnon-forested cells, respectively. The non-forest/forest category represents areas where the model predicted non-forest and the digital classificationwas forest. Forest/non-forest is the opposite case: (a–e) with starting volumes of 30, 45, 60, 75, and 90 m3/ha, respectively.


Table 4The influence of household characteristics on habitat over 65 years(1965–2030) relative to a baseline scenario of 0 new households peryear and 0 m3/year of fuelwood consumed after 1997

Household growthrate (householdsper year)

Fuelwoodconsumption(m3/year)

Changein totalhabitat(%)

Change in habitat< 2600 m ofelevation (%)

0 0 0.00 0.0024 0 −0.06 −0.1824 5 −1.34 −3.7924 10 −2.61 −7.3624 15 −3.32 −9.3324 30 −6.06 −15.84

−24 15 −1.84 − 5.17−12 15 −2.12 − 6.16

0 15 −2.77 −7.7412 15 −3.21 −8.9924 15 −3.32 −9.3336 15 −4.31 −11.7412 10 −2.26 −6.41

3.2. Household impacts

Projected household impacts on panda habitat areshown inTable 4. Current levels of household creationand fuelwood consumption caused nearly an additional10% habitat loss below 2600 m of elevation comparedto conditions in which no additional households andfuelwood collection occurred after 1997. Acrossthe entire reserve, an additional 3% of habitat waslost compared to no new household impacts after1997. Levels of household fuelwood consumptionwere systematically varied from 0 to 30 m3/year toexamine the influence of fuelwood consumption onhabitat loss. An increase in fuelwood consumptionafter 1997 to 30 m3/year would result in a nearly 70%increase in loss of habitat from the current level of15 m3/year. Over 6% of the reserve and nearly 16%of the low-elevation forest would be further impactedby doubling the consumption of fuelwood. Reducingfuelwood consumption by two-thirds reduced the lossof habitat below 2600 m of elevation by 59% comparedto baseline scenarios. Forest re-establishment willonly play a limited role over the next 30 years asre-establishment times are typically 30–50 years. Inthe next 30 years, habitat loss may largely be dictatedby fuelwood consumption and increases in volumeof current stocks. Therefore, a near cessation in

fuelwood collection over the next 30 years is requiredto maintain levels of habitat as measured in 1997.

New housing development did not have the sameinfluence on the total habitat co-opted by householdsas fuelwood consumption levels did. A 50% increasein the number of new household starts resulted in a26% increase in low-elevation habitat loss relative tobaseline scenarios. Cessation of new housing develop-ment following 1997 still led to the loss of nearly 3%of the entire reserve and 8% of low-elevation habitatcompared to scenarios with no new households andno fuelwood consumption following 1997. And a netremoval of 24 households per year (the same numberpreviously being added per year) only resulted in a 45%reduction in habitat loss compared to baseline scenar-ios. As seen from a 50% increase in household creationwith no fuelwood collection, increased population andresulting household creation contributed little to habi-tat loss because considerable areas around householdsare already cleared of forest cover. Modest reductionin both future new housing development and fuelwoodconsumption (12 households per year and 10 m3/year)led to approximately 30% less habitat loss relative tocurrent levels of new housing and fuelwood consump-tion.

4. Conclusions and discussion

HALIM was developed to examine the relationshipo enceo top ac-t n bes reasw mored ticss ande odm onap iono ow-e andf e ism ateo ntof ofp ured

f households to the landscape, to assess the influf the landscape on household activities, androvide a practical framework in which the inter

ions between households and the landscape caimultaneously studied. The study does point out ahere further analyses are needed. For example,etailed information on the biophysical characterisuch as total available biomass, growth rates,fficiency in the conversion of biomass to fuelwoight contribute to the model. Except for the Corhotographs used for this study, very little informatn the state of the forest in 1965 was available. Hver, comparing projections of household creationuelwood collection from 1965 to a time when therore detailed information permitted a better estimf forest conditions in 1965 and provided insight i

actors contributing to habitat loss. Comparisonsredicted forest loss from 1965 to 1997 to meas


forest conditions in 1997 for several scenarios of theaverage starting volume of low-elevation forest furthersuggests that these forests were already at relativelylow volumes. The lower forest volume potentiallymagnified household impacts on the forests since1965. It is possible that large-scale logging occurredconcurrently with household fuelwood collectionfrom 1965 until 1975 or later. While timber activitiescontinued after 1975, researchers did not note anylarge-scale commercial logging in the reserve from1983 to the 1990s. Forest loss after 1975 until 1997was likely due to a combination of fuelwood collectionand fine-scale timber activities, and exacerbated byalready low-stand volumes from previous large-scaleactivity. As these forests are increasingly lost, fuel-wood activities are moving to higher elevation forestswith increasing losses of core habitat.

In addition, most decisions such as consumptionlevel, propensity to use alternative energy sources, em-igration rates, and new household formations are madeat the household-level and are not explicitly modeledin this study. Increasingly complex models can bedeveloped within the framework and the influence ofhousehold-level socio-economic information is beingexamined. In addition, other economic and behavioraldrivers can be incorporated. However, using landscape-level household factors linked to the landscape alreadyprovided considerable insight into human impacts andpotential mitigation strategies. The model provided in-sight into the historical trends and ecological conditionso nge,t ns ofl mento hipsb wardi ayp ms,b delh d bec

le top andp oldsi ds;p ion.U dela oft of

households. Fuelwood collection also is only based ona few landscape variables: distance to roads, previousfuelwood collection locations, slope, and elevation.Again, the model captures the trend in householdreductions in forest cover. The simplicity (e.g. fourhousehold creation factors) and success of the modelsuggest a core set of landscape-level characteristicshas a considerable influence on the spatial distributionof household activities.

HALIM also provided a means to examine the roleof household characteristics on possible future impactsto giant panda habitat. Households were present in thereserve prior to the establishment of the current majortransportation routes. New roads and the introductionof mechanized transportation have likely led to growthin agricultural activity along these routes and increasedaccess to forests near roads away from households. Inaddition, as the reserve is situated in a mountainousarea, topography plays a significant role in shaping thespatial distribution of household activities. Farmingrequires relatively flat land and easy access to trans-portation. In comparison, fuelwood collection is lessdependent on the quality of collection sites than thecost factor of the distance to roads, the slope, elevationchange, and overall accessibility of the location ofcollection sites.

Also, considerable changes in fuelwood consump-tion and/or household creation rates are required tomaintain the current area of forest. While an increase inhousing development itself led to only small decreasesi re-s newh aredl elyt vens um-b eateri es-tt saryt houldf inga oldsw our-a

ac-t cedl olds

f the reserve, the driving factors of land-cover chahe potential consequences of household alteratioand-cover on panda habitat, the spatial arrangef these impacts, and the intricate relationsetween households and landscapes. The trend to

ncorporating household-level data into models mrovide more detailed information of these systeut the necessity of such data to practically moousehold impacts at the landscape level shoulonsidered.

Using landscape-level data, the model was abredict household activities relatively accuratelyarsimoniously. The placement of new househ

s explained by only four factors: distance to roaroximity to other households; slope; and elevatsing only these four factors; however, the moccurately predicts household creation nearly 90%

he time within 3 cells of the measured distribution

n forest area, even limited fuelwood consumptionulted in relatively large habitat losses. As mostouseholds are being constructed on previously cle

and, the placement of new households is not liko directly cause further loss of forest. However, emall amounts of fuelwood required for the large ner of households already in the reserve has a gr

mpact on forest cover. These results are similar toimates as measured byLiu et al. (1999)who showedhat relatively high rates of emigration were neceso restore habitat and suggested that most efforts socus on reducing fuelwood collection and providlternative energy sources for the current househhile providing viable means and incentives to encge emigration.

HALIM provides a basic framework that has prical application for human-dominated or -influenandscapes. The model incorporates househ


directly into landscapes alongside naturally occur-ring dynamics and examines the influences of thelandscapes on household activities. In addition, themethod used is flexible enough to allow the integrationof additional human and landscape componentssuch as the more detailed socio-economic informa-tion discussed above and other natural processessuch as household impacts on understory bamboodynamics. This approach provides a useful meansto better understand and predict impacts of house-holds on wildlife habitat and interactions with thelandscape.

Acknowledgements

We thank the Wolong Nature Reserve Administra-tion for their logistic support of our fieldwork and rec-ognize the assistance from Jinyan Huang, Jian Yang,

Appendix A

Sub-model probability functions and description of parameters and factors

Sub-model Parameter Factors

H ab):P(

al

e-

F

Hemin Zhang, and Shiqiang Zhou. Comments fromDr. Sven Jorgensen and two anonymous reviewerswere very helpful for us to improve the paper. Inaddition, we gratefully acknowledge financial supportfrom the National Aeronautics and Space Adminis-tration (NASA Earth Science Fellowship Award), theNational Science Foundation (CAREER Award andBiocomplexity in the Environment Grant), NationalInstitutes of Health (National Institute of Child Healthand Human Development, R01 HD39789), AmericanAssociation for the Advancement of Sciences/The JohnD. and Catherine T. MacArthur Foundation, MichiganState University (All-University Research InitiationGrant, Global Competence Fund, Institute for Inter-national Agriculture), the National Natural ScienceFoundation of China, the Chinese Academy of Sci-ences, Ministry of Science and Technology of China(G2000046807), and China Bridges International(North America).

ousehold (h):P(h|ab,t, p)

Local abiotic factors (P(s)P(e)P(a)Transportation(t): P(h|t) =clamp(1−cost/max)

Proximity toexistinghouseholds (p):P(h|p) = P(d)

uelwood (f):P(f|a, d, p)

Availability (a):P(f|a) = P(v)

h|ab) = (SeeAppendix B)

Cost = distance× impedance

Distance = horizontal + verticdistance

Impedance =f(slope)

Max = 2000 m (maximum houshold distance)

Distance factor (d):

{1.0 (d < 90 m)}

{0.1 (d < 200 m)}

{0.01 (d > 2000 m)}

Volume (v):

{1 (v > 0 m3)}

{0 (v = 3 m3)}


Appendix A (Continued)


Cost functionfrom householdto collection site:P(f|d) = clamp(1− cost/max)

Cost = distance× impedance

Distance = horizontal + vertical distance

Impedance =f(slope, road)

Max = 9000 m (maximum fuelwood collection distance)

Proximity topreviouscollection site (d)

Distance factor (d):

{1.0 (d≤ 90 m)}

{0.1 (d > 90 m)}

Forest cover:P(c|g, r) Growth (g):P(g|v) = P(v)

Volume (v):

{1 (v < maximum, m3)}

{0 (v = maximum, m3)}

Re-establishment(r): P(r|a) = P(cutage)P(e)P(p)

Cut age: normal temporal Pdf(cut age, 10.0, 2.0)

Elevation (e):

{1 (e≤ max species elevation)};

{0 (e> max species elevation)}

Proximity (p):

{1 (p < 1/2 max species re-establishment distance)};

{0.5 (p < 1 max species re-establishment distance)};

{0.1 (p > 1 max species re-establishment distance)}

Habitat Suitability Slope, elevation, aspect, and forest cover

Appendix B

Empirically derived probabilities of household location from abiotic factors


Local abiotic factors,P(h|ab) =P(s)P(e)P(a) Slope (s): P(ab|s) = P(s) {0.0 (s> 50◦)}

{0.09 (s> 40◦)}

{0.23 (s> 30◦)}

{0.63 (s> 20◦)}

{0.86 (s> 10◦)}

{1.0 (s≤ 10◦)}


Appendix B (Continued)


Aspect (a), P(ab|a) = P(a) {0.14 (a > 315◦)}

{0.24 (a > 270◦)}

{0.26 (a > 225◦)}

{0.35 (a > 180◦)}

{1.0 (a > 135◦)}

{0.56 (a > 90◦)}

{0.30 (a > 45◦)}

{0.14 (a≤ 45◦)}

Elevation (e), P(ab|e) = P(e) {0.00 (e> 2500)}

{0.08 (2250 <e≤ 2500)}

{0.82 (1750 <e≤ 2250)}

{1.00 (e≤ 1750)}

Appendix C. Supplementary data

Supplementary data associated with this article canbe found, in the online version, atdoi:10.1016/j.ecolmodel.2004.07.026.

References

An, L., Liu, J., Ouyang, Z., Linderman, M., Zhou, S., Zhang, H.,2001. Simulating demographics and socioeconomic process onhousehold level and their impacts on giant panda habitats. Ecol.Modell. 140 (1–2), 31–49.

An, L., Lupi, F., Liu, J., Linderman, M., Huang, J., 2002. Modelingthe choice to switch from fuelwood to electricity: implications forgiant panda habitat conservation. Ecol. Econ. 42 (3), 445–457.

An, L., Linderman, M., Qi, J., Shortridge, A., Liu, J., in press. Explor-ing complexity in a human-enviornment system: an agent-basedspatial model for multidisciplinary and multi-scale integration.Ann. Assoc. Am. Geogr.

Baker, W.L., Cai, Y., 1992. The r. le-programs for multiscale analysisof landscape structure using the grass geographical information-system. Landscape Ecol. 4, 291–302.

Balzter, H., Braun, P.W., Kohler, W., 1998. Cellular automata modelsfor vegetation dynamics. Ecol. Model. 107 (2–3), 113–125.

Benenson, I., 1999. Modeling population dynamics in the city: froma regional to a multi-agent approach. Discrete Dyn. Nat. Soc. 3(2–3), 149–170.

Berger, T., 2001. Agent-based spatial models applied to agricul-ture: a simulation tool for technology diffusion, resource usechanges and policy analysis. Agric. Econ. 25 (2–3), 245–260.

Carter, J., Ackleh, A.S., Leonard, B.P., Wang, H.B., 1999. Giantpanda (Ailuropodamelanoleuca) population dynamics and bam-boo (sub-family Bambusoideae) life history: a structured popu-lation approach to examining carrying capacity when the preyare semelparous. Ecol. Modell. 123 (2–3), 207–223.

Ceballos, G., Ehrlich, P.R., 2002. Mammal population losses and theextinction crisis. Science 296 (5569), 904–907.

Couclelis, H., 2001. Why I no longer work with agents: a challengefor ABMs of human–environment interactions, in: Parker, D.,Berger, T., Manson, S., McConnell, W.J. (Eds.), Agent-BasedModels of Land-Use and Land-Cover Change, Report and Re-view of an International Workshop, LUCC Report Series No. 6.Irvine, California, October 4–7, 2001.

Fall, A., Fall, J., 2001. A domain-specific language for models oflandscape dynamics. Ecol. Modell. 141 (1–3), 1–18.

Fall, A., Morgan, D.G., Daust, D., 2001. A framework and softwaretool to support collaborative landscape analysis: fitting squarepegs into square holes. Trans. GIS 5 (1), 67–86.

Forman, R.T., Godron, M., 1986. Landscape Ecology. Wiley, NewYork, Chichester, Brisbane, Toronto, Singapore.

Lambin, E.F., 1994. Modelling Deforestation Processes: A Review.European Commission, Luxembourg.

Liu, J., Daily, G.C., Ehrlich, P.R., Luck, G.W., 2003. Effects of house-hold dynamics on resource consumption and biodiversity. Nature421, 530–533.

Liu, J., Linderman, M., Ouyang, Z., An, L., Yang, J., Zhang, H.,2001. Ecological degradation in protected areas: the case of Wo-

http://dx.doi.org/10.1016/j.ecolmodel.2004.07.026

http://dx.doi.org/10.1016/j.ecolmodel.2004.07.026


long Nature Reserve for giant pandas. Science 292 (5514), 98–101.

Liu, J., Ouyang, Z., Yang, Z., Taylor, W., Groop, R., Tan, Y., Zhang,H., 1999. A framework for evaluating the effects of human factorson wildlife habitat: The case of giant pandas. Conservation Biol.13 (6), 1360–1370.

Mertens, B., Lambin, E.F., 1997. Spatial modeling of deforestationin southern Cameroon. Appl. Geography 17 (2), 143–162.

Ouyang, Z., Tang, Y., Zhang, H., 1996. Biodiversity spatial patternanalysis in Wolong. China Man Biosphere, 3.

Reid, D.G., Hu, J., Dong, S., Wang, W., Huang, Y., 1989. GiantpandaAiluropoda melanoleucabehaviour and carrying capacityfollowing a bamboo die-off. Biol. Conservation 49, 85–104.

Rutledge, D.T., Lepczyk, C.A., Xie, J., Liu, J., 2001. Spatio-temporaldynamics of endangered species hotspots in the United States.Conservation Biol. 15 (2), 475–487.

Schaller, G.B., Hu, J., Pan, W., Zhu, J., 1985. The Giant Pandas ofWolong. The University of Chicago Press, Chicago and London.

Vitousek, P.M., Mooney, H.A., Lubchenco, J., Melillo, J.M., 1997.Human domination of Earth’s ecosystems. Science 277 (5325),494–499.

Wu, H., Stoker, R.L., Gao, L.C., 1996. A modified Lotka-Volterrasimulation model to study the interaction between arrow bam-boo (Sinarundinaria fangiana) and giant panda (Ailuropodamelanoleuca). Ecol. Model. 84 (1–3), 11–17.

Yang, Y., Li, C. (Eds.), 1992. Sichuan Forests. China Forestry Press,Yunnan.

Zhang, H., Li, D., Wei, R., Tang, C., Tu, J., 1997. Advances in conser-vation and studies on reproductivity of giant pandas in Wolong.Sichuan J. Zool. 16, 31–33.

Zhou, Z.H., Pan, W.S., 1997. Analysis of the viability of a giantpanda population. J. Appl. Ecol. 34 (2), 363–374.

Modeling the spatio-temporal dynamics and interactions of households, landscapes, and giant panda...

Documents

Transcript of Modeling the spatio-temporal dynamics and interactions of households, landscapes, and giant panda...