ebook img

Predicting species distributions across the Amazonian and Andean regions using remote sensing ... PDF

17 Pages·2008·1.07 MB·English
by  
Save to my drive
Quick download
Download
Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.

Preview Predicting species distributions across the Amazonian and Andean regions using remote sensing ...

JournalofBiogeography(J.Biogeogr.)(2008) Predicting species distributions across the ORIGINAL ARTICLE Amazonian and Andean regions using remote sensing data Wolfgang Buermann1*, Sassan Saatchi1,2, Thomas B. Smith1,3, Brian R. Zutta1,3, Jaime A. Chaves1,3, Borja Mila´1,3 and Catherine H. Graham4 1CenterforTropicalResearch,UCLAInstitute ABSTRACT ofEnvironment,UniversityofCalifornia,Los Aim We explore the utility of newly available optical and microwave remote Angeles,CA,USA,2JetPropulsionLaboratory, sensingdatafromtheModerateResolutionImagingSpectroradiometer(MODIS) CaliforniaInstituteofTechnology,Pasadena, California,USA,3DepartmentofEcologyand and QuikSCAT (QSCAT) instruments for species distribution modelling at regional to continental scales. Using eight Neotropical species from three EvolutionaryBiology,UniversityofCalifornia, LosAngeles,CA,USA,4DepartmentofEcology taxonomicgroups,weassesstheextenttowhichremotesensingdatacanimprove predictionsoftheirgeographicdistributions.Fortwobirdspecies,weinvestigate andEvolution,StonyBrookUniversity,Stony Brook,NY,USA the specific contributions of different types of remote sensing variables to the predictions and model accuracy at the regional scale, where the benefits of the MODIS and QSCAT satellite data are expected to be most significant. Location South America, with a focus on the tropical and subtropical Andes and the Amazon Basin. Methods Potentialgeographic distributionsofeightspecies,namelytwobirds, twomammalsandfourtrees,weremodelledwiththemaxentalgorithmat1-km resolutionovertheSouthAmericancontinentusingclimaticandremotesensing data separately and combined. For each species and model scenario, we assess model performance by testing the agreement between observed and simulated distributionsacrossallthresholdsand,inthecaseofthetwofocalbirdspecies,at selected thresholds. Results Quantitative performance tests showed that models built with remote sensing and climatic layers in isolation performed well in predicting species distributions,suggestingthateachofthesedatasetscontainsusefulinformation. However,predictionscreatedwithacombinationofremotesensingandclimatic layers generally resulted in the best model performance across the three taxonomic groups. In Ecuador, the inclusion of remote sensing data was critical in resolving the known geographically isolated populations of the two focal bird species along the steep Amazonian–Andean elevational gradients. Withinremotesensingsubsets,microwave-baseddataweremoreimportantthan optical data in the predictions of the two bird species. Main conclusions Our results suggest that the newly available remote sensing data (MODIS and QSCAT) have considerable utility in modelling the contemporary geographical distributions of species at both regional and continental scales and in predicting range shifts as a result of large-scale land- *Correspondence:WolfgangBuermann,Center use change. forTropicalResearch,UCLAInstituteof Environment,UniversityofCalifornia,Los Keywords Angeles,LaKretzHall,Suite300,619CharlesE. Conservation biogeography, ecological niche characterization, microwave YoungDr.East,LosAngeles,CA90095-1496, remote sensing, MODIS, optical remote sensing, QSCAT, South America, USA. E-mail:[email protected] spatial scale, species distribution modelling. ª2008TheAuthors www.blackwellpublishing.com/jbi 1 Journalcompilationª2008BlackwellPublishingLtd doi:10.1111/j.1365-2699.2007.01858.x W.Buermann etal. MODIS era have minimized the impact of water vapour and INTRODUCTION aerosol effects, which were problematic in previous satellite A wide range of applications in biogeography, including missions (Vermote etal., 1997). Furthermore, improvements evolutionary (Peterson, 2001; Hugall etal., 2002; Graham in retrieval techniques that translate the remotely sensed etal.,2004,2006),ecological(Andersonet al.,2002),invasive surface reflectances to more interpretable biophysical param- species (Peterson & Robins, 2003), conservation (Godown & eters have led to a number of advanced MODIS vegetation Peterson, 2000; Sa´nchez-Cordero & Martı´nez-Meyer, 2000) products.Amongthese,andappliedinthisstudy,aretheleaf and climatechange studies(Berryet al.,2002; Petersonet al., areaindex(LAI),definedastheone-sidedprojectedgreenleaf 2002; Thomas etal., 2004; but see also Buckley & Roughgar- area per unit ground area (Knyazikhin etal., 1998; Myneni den, 2004; Thuiller etal., 2004; Arau´jo etal., 2006), rely on et al., 2002), and the vegetation continuous field (Hansen information concerning the geographical distribution of et al., 2002) products as measures of vegetation density, species and their environmental requirements. In species seasonality and percentage of tree cover. Although the distribution modelling, which is often used in these studies, application of optical satellite data over areas with persistent twosetsofinputdataarerequired:occurrencedataandspatial cloud cover remains a challenge, the frequent revisit time of informationontheenvironmentalcharacteristicsofthespecies MODISathighspatialresolutionallowscircumventionofthis habitats. Whereas significant work has been done on the basic physical limitation to some extent through appropriate complexityandrobustnessofthestatisticalmethodology(e.g. spatial and temporal compositing. Elith etal., 2006) as well as on the quality and quantity of In addition, microwave remote sensing data from active occurrence data (e.g. McPherson et al., 2004; Segurado & space-borne scatterometers [e.g. SeaWinds on board of Arau´jo,2004;Hernandezetal.,2006),relativelylittleattention NASA’s QuikSCAT (QSCAT)] at relatively high spatial hasbeenpaidtotheenvironmentaldatasetsusedinecological (2.25 km) and temporal (3days) resolutions with global modelling.Hereweexploretheuseofanarrayofnewsatellite coverage are also becoming widely available. Microwave data, which have seldom been used to predict species radiation probesthedielectric properties ofland and,thus, is distributions and areoften poorly understoodby modellers. particularlysensitivetosurfacemoisture.Furthermore,micro- The availability of satellite observations of the Earth’s waveradiationissensitivetosurfaceroughness,andhenceits surfaceinthepastthreedecadeshasheraldedanewerainthe interpretationissomewhatmorecomplexthanthatofoptical ecological modelling of species habitats. Direct measurements data.However,recentstudieshaveshownthattheintensityof of the Earth’s hydrological and biophysical characteristics, its theradarbackscatterovervegetatedregionsrelatestobiomass geological features and its climate from space have provided density and forest structure (Saatchi etal., 2000, 2007; Long newdatalayerswithspatialandtemporalresolutionsrelevant et al., 2001). In addition, microwave measurements are less to landscape-scale habitat characteristics and ecological pro- sensitivetocloudcoverandpenetratedeeperintothecanopy, cesses(Gould,2000;Nagendra,2001;Turneretal.,2003).The leading to a better discrimination of distinct forest architec- opticalportionoftheelectromagneticspectrumacquiredfrom turesoverdenselyforestedregions.Takentogether,thenewly space-borne passive sensors, such as theAdvanced Very High available optical and microwave remote sensing data layers ResolutionRadiometer(AVHRR),ismostcommonly usedto provideadditionalinformationonsurfaceparameters,suchas infervegetationcharacteristics.Greenleavesharbourpigments, vegetation density and seasonality as well as canopy moisture notably chlorophyll, that absorb strongly at blue and red androughness,withgreatpotentialtodescribetheecologyand wavelengths, whereas a lack of such absorption in the near- geographical distributions ofspecies. infrared results in strong scattering. A popular measure that Inthisstudy,wetestedtheutilityofnewlyavailableremote captures this unique spectral response is the Normalized sensing data for distribution modelling from regional to DifferenceVegetationIndex(NDVI);sometimesreferredtoas continentalscalesforasuiteofspecieswithdifferentecological ameasureof‘greenness’,itiscomputedthroughthenormaliz- characteristics. At these scales, the remote sensing data act as ed difference in surface reflectances at near-infrared and red surrogates for land-cover variables and, hence, are closely wavelengths. NDVI, or similar vegetation indices, are among related to climate variations (Egbert et al., 2002). In total, we the few satellite variables that have been used extensively in modelledthedistributionsofeightNeotropicalspecies,namely ecological niche modelling (e.g. Fuentes etal., 2001; Osborne two birds, two mammals and four trees, over the South etal.,2001;Luotoet al.,2002;Zinneret al.,2002;Parraet al., American continent at 1-km spatial resolution. Because the 2004; Roura-Pascual et al.,2006). usefulness of remote sensing data in distribution modelling ThroughtherecentlaunchoftheNASAsatellitesTerraand may be dependent on a species’ ecology (Hernandez etal., Aqua with the Moderate Resolution Imaging Spectroradiom- 2006; McPherson & Jetz, 2007), these species were chosen to eter (MODIS) on board, global optical measurements of capture a variety of ecological traits such as range size and vegetation at finer spatial resolution (250–500m) with high habitat specificity. Specifically, for widespread species in revisit times (1–2 days) and improved data quality have relatively homogeneous climates, such as throughout the becomewidelyavailable(Justiceet al.,1998).Inparticular,the Amazon basin, we expect remote sensing data to be particu- use of narrower spectral bands from the visible to middle larly useful in providing additional constraints on the species infrared and advances in atmospheric corrections in the ecologicalniches.Incontrast,forAndeanspeciesthatresidein 2 JournalofBiogeography ª2008TheAuthors.Journalcompilationª2008BlackwellPublishingLtd Application of remotesensingdatain speciesdistribution modelling narrowclimateregimes,thebenefitofremotesensingdatamay 4.9g), is found in wet to humid montane forests along the be more limited. Furthermore, the utility of remote sensing Andean slopes at elevations ranging from 1000 to 3000 m data is likely to be scale-dependent, increasing in importance (Fjeldsa˚ &Krabbe,1990;DelHoyoet al.,1999).Followingthe fromcontinentalscaleswhereabroadspeciesrangeisdepicted orientation of the Andes, this species occupies an impressive to regional and landscape scales where vegetation and land- north-to-south range from northern Venezuela to north- scape featuresneedto beincorporated. western Argentina. Seven subspecies of A. melanogenys are Models were developed using bioclimatic data from the recognized, occupying various regions of the northern Andes WorldClim data base (Hijmans et al., 2005) and remote (Del Hoyo et al.,1999). sensing data in isolation and combined, and tested by The localities of four Amazonian tree species were taken analysing the area under the receiver operating characteristic from a recent distribution modelling study on Amazon tree curve (AUC). For two bird species, the wedge-billed wood- diversity (Saatchi et al., 2008). The species analysed were creeper (Glyphorynchus spirurus) and the speckled humming- jacareuba (Calophyllum brasiliensis; 88 point localities), andi- bird(Adelomyiamelanogenys),withtheformershowingavery roba (Carapa guianensis; 73), balata´ (Manilkara bidentata; widespread distribution throughout the rainforests of the 127), and ucuu´ba (Virola surinamensis; 113). The four trees Amazon basin and Andean foothills and the latter a very differ markedly in habitat specificity and range size (Saatchi narrow distribution in the Andean montane forests, we et al., 2008). Amongst them, C. brasiliensis shows the widest provide a detailed analysis on model performance. This distributionthroughoutnorthernSouthAmericaandisfound includes omission tests at selected thresholds, evaluations of in the rain forests of the Amazon basin and the Andean model predictive power as well as of the contributions of lowlands up to elevations of 1500 m, as well as in the variousenvironmental variablestothepredictions, andvisual woodlands of south-eastern Brazil. Carapa guianensis, on the inspections of the predicted maps. Given that the greatest other hand, has the most specific habitat requirements and is benefit of the newly available optical and microwave remote found along streams, in the periodically inundated swamp sensinglayersisexpectedattheregionallevel,wealsoprovide forests, and in upland forests along the rivers of the Amazon an in-depth analysis of the predictions for the two focal bird Basin, and at elevations ranging from shorelines to 1200 m. species in Ecuador, a country with highly variable microcli- Manilkara bidentata is quite common throughout the rain mates and ecological gradients along the steep slopes of the forests of the Amazon basin and along the coastal regions of Andes. northern Brazil and Guyana, but is restricted to lower elevations. Virola surinamensis is found in swampy, fertile and periodically inundated riverbanks, in Amazonian va´rzea DATA AND METHODS forests,andindegradedandsecondaryforests.Comparedwith M. bidentata, its range is more restricted to the northern Species andenvironmental data Amazon basin. From Phillips et al.(2006), we obtained point localities for Study species two Neotropical mammals: the brown-throated three-toed For this study, point localities for eight Neotropical species sloth, Bradypus variegatus (77 point localities), and a small wereobtained.ThespeciesanalysedweretwocommonSouth montane muridrodent,Microryzomys minutus(85).Bradypus Americanbirdsthatwereselectedforfurtherdetailedanalysis, variegatusiswidelydistributedinthedeciduous,evergreenand the wedge-billed woodcreeper (Glyphorynchus spirurus) and montaneforestsoftheAmazonbasinandinthevicinityofthe thespeckledhummingbird(Adelomyiamelanogenys),aswellas tropical and subtropical Andes. Its geographic range extends two Neotropical mammals and four Amazonian trees, which fromHondurastonorthernArgentina.Microryzomysminutus, have been used in recent distribution modelling studies in contrast, shows a much narrower distribution throughout (Phillips et al., 2006; Saatchi et al., 2008). In the following, thetropicalandsubtropicalAndeanwetmontaneforests,with brief descriptions of ecological characteristics are given for elevational ranges from 1000 to 4000 m and a geographic eachoftheselectedspecies,withmoredetailforthetwofocal extent from Venezuela to Bolivia. bird species. Additional information about the six non-focal Inthecaseofthetwofocalbirdspecies,weobtainedprecise species is provided intherespectivereferences. point-locality data by sampling each bird species using mist Glyphorynchus spirurus is a widespread Neotropical bird nets during multiple field surveys throughout Ecuador species (length, 130–160 mm; mass, 10.5–21g), which is between 1999 and 2006 (22 sites for G. spirurus, 26 for frequent in the understorey of lowland and submontane A.melanogenys).Inaddition,weacquiredoccurrencedatafor tropical forests throughout the Amazon basin and Andean these two bird species from other sources (see Table S1 in foothills (Del Hoyo et al., 2003). There are 13 subspecies of Supplementary Material) leading to a total of 78 unique G.spirurusknownintheregion,withrangesspanningfromthe georeferenced locations for G. spirurus and 180 for A. southernAmazonbasintosouth-easternMexico,andalongthe melanogenys (Fig. 1). In compiling this data set, we screened AtlanticcoastalforestsofBrazil(DelHoyoetal.,2003). therecordstoincludeonlygeoreferencedpointsobtainedfrom Adelomyiamelanogenys,oneofthesmallerspeciesofSouth recent surveys (within the last 10years), thereby ensuring a American hummingbirds (length, 85–90mm; mass, 4.2– broadly overlapping time frame with the acquisition of the JournalofBiogeography 3 ª2008TheAuthors.Journalcompilationª2008BlackwellPublishingLtd W.Buermann etal. (a) Glyphorynchus spirurus (b) Adelomyia melanogenys Figure1 Pointlocalitiesfor(a)the wedge-billedwoodcreeper(Glyphorynchus spirurus;78records)and(b)thespeckled hummingbird(Adelomyiamelanogenys;170 records)utilizedinthisstudy. remotesensingdata.Themodellingstudywasconcentratedon derived from atmospherically corrected MODIS surface the distribution of each of the two bird species pooled across reflectances (Vermote etal., 1997) over the 5-year period subspecies because ofthelimited numberof occurrences. 2000–2004.TheaccuracyoftheMODISLAIproducthasbeen evaluated against ground measurements of leaf area in a host of vegetation types across the globe (see online supporting Climate data information in Myneni etal., 2007), and was generally found AseriesofbioclimaticmetricswereobtainedfromWorldClim tobeexactwithinLAIvaluesofroughly0.6.Foreachyear,we version 1.4 (Hijmans etal., 2005). These metrics are derived created monthly composites by averaging the 8-day LAI from monthly temperature and rainfall climatologies and product.EventhoughtheMODISalgorithmisequippedwith represent biologically meaningful variables for characterizing improved cloud masking (Platnick et al., 2003), effects from species range (Nix, 1986). Eleven temperature and eight subpixelcloudinessonLAIestimatesoverareaswithpersistent precipitationmetricswereused,expressingspatialvariationsin cloud cover, such as within the Amazon basin and along the annual means, seasonality and extreme or limiting climatic Amazonian–Andean elevational gradients, were occasionally factors.Themonthlyclimatologiesweredevelopedusinglong observed. To reduce these effects along with any natural time series of a global network of weather stations from interannualvariabilitypresentinthedata,wecreatedmonthly various sources, such as the Global Historical Climatology climatologies by averaging the 5 years of data (2000–2004). Network(GHCN),theUnitedNationsFoodandAgricultural The climatological monthly composites were then used to Organization (FAO), the World Meteorological Organization generate five metrics: annual maximum (Fig. 2a), minimum, (WMO), the International Center for Tropical Agriculture mean,standard deviation,andrange(differenceofmaximum (CIAT), and additional country-based station networks. The andminimum).TheseLAImetricsprovidespatialinformation station data were interpolated to monthly climate surfaces at on net primary productivity andvegetation seasonality. 1-km spatial resolution using a thin-plate smoothing spline WealsoincludedtheMODIS-derivedvegetationcontinuous algorithm with latitude, longitude, and elevation as indepen- field (VCF) product as a measure of the percentage of tree dent variables (Hijmans et al.,2005). canopy cover within each 500-m pixel (Hansen etal., 2002). This data set was produced from time-series composites of MODIS data of the year 2001. Preliminary validation results Remote sensing data over the conterminous United States (Hansen et al., 2002) Remotesensingdataobtainedfromavarietyofsatellitesensors suggestthattheVCFproductcanreliablyseparatemoreopen provided measurements of environmental variables directly (e.g. shrublands, savannas), deforested and, to some extent, related to species habitat characteristics (Turner et al., 2003). fragmented areas from those of closed forests (Fig. 2b). For We compiled a large data set covering a diverse range of thisstudy,the500-mnativetreecoverdatawereaggregatedto surfaceparameters,namelyvegetationdensityandseasonality, 1 km. moisture, roughness, andtopography. In addition to these optical remote sensing data, we To quantify spatial and temporal vegetation patterns, we included microwave QSCAT data available in 3-day compos- used the MODIS 8-day LAI product (Myneni etal., 2002) ites at 2.25-km resolution (Long et al., 2001). The 3-day data 4 JournalofBiogeography ª2008TheAuthors.Journalcompilationª2008BlackwellPublishingLtd Application of remotesensingdatain speciesdistribution modelling (a) (b) 0 1 3 5 7 0 10 30 50 >75% LAI (MODIS) TREE (MODIS) (c) (d()d) <–12 –10 –8.5 –7.5 >–7 15 50 100 500 2000 >4000 QSCAT(dB) Topography (m) Figure2 Aselectionoftheremotesensingdatalayersusedinthisstudy.Thepanelsdepict(a)MODISLAI(leafareaindex)annual maximum,(b)MODISpercentagetreecover,(c)QSCATannualmean,and(d)meanelevationfromSRTM. JournalofBiogeography 5 ª2008TheAuthors.Journalcompilationª2008BlackwellPublishingLtd W.Buermann etal. of the year 2001 with complete data coverage were used to from the northern half of South America in order to capture create average monthly composites at 1-km resolution and both covariance at local levels close to the bird vicinities and then further processed to produce four metrics that included more general covariance over larger spatial scales. Various annualmeanandstandarddeviationofradarbackscatterwith criteriawereusedtodecidewhichlayersofcorrelatedpairsto horizontal and vertical polarizations. QSCAT backscatter retain for further analysis (with Pearson’s correlations of the measurementsarecalibratedwithhighaccuracy[c.0.1decibel orderof0.9orlarger).These includedkeeping layersthatare (dB);Tsaietal.,1999],andthelong-termaveragesensurethe morecommonlyusedindistributionmodelling(WorldClim), reductionofanypossiblehigh-frequencynoisefromcompos- that exhibit larger contrast/variance over the study area iting the data or from atmospheric disturbances (e.g. rain (QSCAT), and that havethehighestdata quality (LAI). events) while preserving information on the backscatter FortheWorldClimmetrics,significantcorrelationsamonga variability. The QSCAT radar measurements, at wavelengths number of the original 19 data layers were observed, and a of c. 2cm, are sensitive to surface canopy roughness, surface subset of nine bioclimatic variables was selected for further canopy moisture, and other seasonal attributes, such as analysis, namely annual mean temperature, mean diurnal deciduousness of vegetation (Frolking et al., 2006). For low- temperature range, temperature seasonality, minimum temper- density vegetation cover, such as woodlands, shrublands, and ature of coldest month, maximum temperature of warmest grassland savannas, measurements at various polarizations month, annual mean rainfall, rainfall seasonality, rainfall of correlate positively with vegetation biomass and surface coldest quarter and rainfall of warmest quarter. High correla- moisture (Saatchi et al., 2007). For areas covered with dense tions were also evident among the five LAI metrics and the forests, the QSCAT data are sensitive to large-scale variations four QSCAT metrics that were produced for this study, and, incanopy roughnessand moistureconditions (Fig. 2c). based on the criteria given above, we selected the LAI annual Finally, we used the Shuttle Radar Topography Mission maximum(Fig. 2a)andLAIannualrangelayersaswellasthe (SRTM) digital elevation data, aggregated from the native QSCAT annual mean (Fig. 2c) and QSCAT seasonality layers 90-m resolution to 1km (Fig.2d). In addition to mean withhorizontal polarization for further analysis. elevation, the standard deviation based on the 90-m data was As a result, the final reduced data set used in this study included asan indicator ofsurface ruggedness. converged to a total of 16 environmental layers: nine climate Overall,10continuousremotesensingdatalayers(5LAI,4 (fivetemperatureandfourrainfall),twoLAI,twoQSCAT,two QSCAT,1VCF)andthetwoelevationdatalayers(meanand topography(meanandstandarddeviation),andonetree-cover standarddeviation)representing variousvegetation andland- layer. It should be noted that the model algorithm (maxent) scape features were included in this analysis. In Table 1, we used in this study is largely robust to covariance among present a brief overview of the original remote sensing variables, and that data reduction was performed mainly to products/data from which the various metrics were derived, improve interpretation. along with corresponding ecological interpretations, native spatial resolution andorbiting frequency. Modelling approach Data reduction maxent To facilitate interpretation of the environmental information Weusedthemaxentalgorithm(version2.1),whichhasbeen andhowitrelatestoaspecies’habitat,wereducedthetotalof veryrecentlyintroducedforthemodellingofspeciesdistribu- 31datalayers(19WorldClim +10remotesensing +2SRTM) tions (Phillips etal., 2006). maxent is a general-purpose to a set of less correlated variables. Covariance within the algorithm that generates predictions or inferences from an subsetsofclimateandremotelysensedvariableswasestimated incompletesetofinformation.Themaxentapproachisbased by computing cross-correlation matrixes based both on the onaprobabilisticframework.Themainassumptionisthatthe birds’ point localities and on 1000 points randomly drawn incomplete empirical probability distribution (which is based Table1 Overviewoftheremotesensingdatasetsusedinthisstudy.Foreachremotesensingdatalayer,nativespatialandtemporal resolutionsaswellasanecologicalinterpretationaregivenalongwiththerespectivemodelscenariosinwhichthesedatalayerswere included.Modelscenarioacronymsrefertomodelrunswithremotesensingandelevationdata(RSE),climateandelevationdata(CLIME), andclimate,remotesensingandelevationdatacombined(CLIMERS). Datarecord Instrument Ecologicalvariable Resolution Scenario Leafareaindex(LAI) MODIS Vegetationdensity,seasonalityandnetprimaryproductivity 1kmand8day RSE,CLIMERS Percentagetreecover MODIS Forestcoverandheterogeneity 500m RSE,CLIMERS Scatterometer-backscatter QSCAT Surfacemoistureandroughness(foreststructure),seasonality 2.25kmand3day RSE,CLIMERS Digitalelevationmodel SRTM Topographyandruggedness 90m RSE,CLIME CLIMERS 6 JournalofBiogeography ª2008TheAuthors.Journalcompilationª2008BlackwellPublishingLtd Application of remotesensingdatain speciesdistribution modelling on the species occurrences) can be approximated by a elevation data only (RSE), with two QSCAT, two LAI, one probability distribution of maximum entropy (the maxent tree-cover and two topography layers from the reduced distribution)subjecttocertainenvironmentalconstraints,and subset, and (3) climate, elevation and remote sensing data that this distribution approximates a species’ potential geo- combined (CLIMERS) using all 16 layers of the reduced graphic distribution (Phillipset al., 2006). subset (see Data reduction). For visualization, predicted maps Like most maximum-likelihood estimation approaches, the weregeneratedwiththethreepredictorsetsCLIME,RSEand maxent algorithm a priori assumes a uniform distribution CLIMERS using all available point localities of a given and performs a number of iterations in which the weights species. Thecorrespondingmaps forthesixnon-focalspecies associated with the environmental variables, or functions (two mammals and four trees) are provided in Figs S1–S6 in thereof, are adjusted to maximize the average probability of the Supplementary Material. the point localities (also known as the average sample Thespatialaccuracyofthepredictionswastestedfollowing likelihood), expressed as the training gain (Phillips, 2006). Phillips etal. (2006). For all eight species we created 10 These weights are then used to compute the maxent random data partitions with 60% of the point localities distribution over the entire geographic space. Consequently, assigned for training and 40% for testing, and ran the three this distribution expresses thesuitability of each grid cellasa scenarios with each of these 10 data partitions. Model function of the environmental variables for that grid cell. A performance was then tested across all thresholds (threshold- high value of the function (in units of cumulative probabil- independent)and,inthecaseofthetwofocalbirdspecies,at ity) for a particular grid cell indicates that the grid cell is fixedthresholds (threshold-dependent). predicted to have suitable conditions for the species in In the threshold-dependent case, we evaluated extrinsic question (Phillips, 2006). omissionrates,definedasthefractionoftestlocalitiesthatfall maxent has several characteristics that make it highly into pixels outside the predicted area, both at constant suitable for the task of modelling species ranges (Phillips proportional predicted area, defined as the fraction of all the etal., 2006). These include a deterministic framework; the pixels that are predicted as suitable for a species, and at the ability to run with presence-only point occurrences; a high ‘balance’ threshold. Whereas the former allows a direct performance with few point localities (Hernandez et al., comparison of the extrinsic omission rates among the three 2006); better computing efficiency, enabling the use of modelscenarios,thelatterbalancestwomeasuresofqualityof large-scale high-resolution data layers; continuous output a binary prediction, namely intrinsic (training) omission and from least to most suitable conditions; and the ability to proportional predicted area (Phillips, 2006). For all model model complex responses through a number of distinct scenariosanddatapartitions,wetestedwhethertestpointsfell feature classes (e.g. functions of environmental variables). into areas predicted present more often than expected at As a consequence, in a recent large model intercompari- randombyapplyingaone-tailedbinomialtestontheextrinsic son project with 15 other algorithms, maxent’s perfor- omission rate and proportional predicted area (Anderson mance was generally rated among the highest (Elith etal., et al.,2002).Intheconstantproportionalpredictedareacase, 2006). atwo-tailedWilcoxonsignedranktestwasutilizedtoevaluate There are several aspects of maxent (2.1) that support the statistical significance in differences in overall extrinsic omis- interpretation of the model results. For example, maxent sionratesbetweentheCLIMEandRSEscenarios.Aone-tailed has a built-in jackknife option through which the impor- Wilcoxon signed rank test was utilized to test if decreases in tance of individual environmental data layers can be overall extrinsic omission rates in the CLIMERS scenarios estimated. It also provides response curves showing how relative to those in the CLIME and RSE scenarios were the prediction depends on a particular environmental statistically significant. variable (Phillips, 2006). For all model runs in this study, In the threshold-independent evaluation, we tested the we used the maxent default settings for regularization and performance of the maxent scenarios using receiver oper- in selecting the feature classes (functions of environmental ating characteristic (ROC) analysis and pseudo-absence in variables). These include linear, quadratic, product, threshold place of absence localities (Phillips etal., 2006). In this and hinge features, depending on the number of point case, the AUC is typically used as a measure of model localities (Phillips, 2006). performance. We employed a ties-corrected one-tailed Mann–Whitney U-test (AccuROC; Vida, 1993) to test if a particular prediction in terms of AUC was significantly Scenarios and quantitative analysis better than random. A two-tailed Wilcoxon signed rank test To evaluate the merit of the set of newly available satellite was used to test whether differences in overall AUC in the data in species distribution modelling, we ran maxent with CLIME and RSE scenarios were significant. Finally, a one- climate and satellite layers in isolation and combined. Three tailed Wilcoxon signed rank test was also used to assess scenarios were evaluated: (1) maxent runs with climate and whether increases in overall AUC in the CLIMERS scenarios elevation data only (CLIME); these include the five temper- relative to those in the CLIME and RSE scenarios were ature, four precipitation and two topography layers from the statistically significant at levels P<0.005, P<0.05 and final reduced subset (see above), (2) remote sensing and P<0.2. JournalofBiogeography 7 ª2008TheAuthors.Journalcompilationª2008BlackwellPublishingLtd W.Buermann etal. nosignificant newinformation (seealsoDatareduction)does RESULTS not leadtoimproved model performance. Quantitative results Threshold-dependent tests Threshold-independent tests Forthetwofocalbirdspeciesinthisstudy,modelperformance For all eight Neotropical species in this study, AUC values was also evaluated at two selected thresholds: the optimized generated using the extrinsic test data were highly statistically ‘balance’ threshold and at constant proportional predicted significant (P<0.001, Mann–Whitney U-test) for all model area.Inthethreemodelscenarios,RSE,CLIMEandCLIMERS, scenariosanddatapartitions.Thissuggeststhateachpredictor with 10 random data partitions, the ‘balance’ thresholds setalonecontainsusefulinformationformodellingthespecies’ ranged from 0.5 to 6.0 (in units of cumulative probabilities) distributions.AUCtestsonmodelperformanceindicatedthat for G. spirurus, and from 1.5 to 4.5 for A. melanogenys. For for six species scenarios with climate and elevation variables G.spirurus,thebinarymapscorrespondingtothesethreshold (CLIME)performedbetterthanscenarioswithremotesensing values were broadly consistent with its known distribution and elevation variables (RSE), whereas for two species these (DelHoyoet al.,2003),butforA.melanogenysthesethreshold two scenarios performed equally well (Table 2). For seven values were somewhat too low, resulting in overprediction species, however, scenarios with climate, remote sensing and (Del Hoyo et al., 1999). However, to facilitate statistical elevation variables combined (CLIMERS) performed better analysis in a consistent manner, we evaluated model perfor- thanthecorrespondingscenarioswithasubsetofthepredictor mance at the ‘balance’ threshold for both bird species. At the variables (Table 2). These results are suggestive of a generally ‘balance’ threshold, the binomial test on extrinsic omission improved discrimination of suitable vs. unsuitable habitats rate and predicted area was highly significant (P<0.001, across taxonomic groups when all environmental data layers one-tailed) for allscenarios anddata partitions. areused incombination. A more detailed look at extrinsic omission and predicted To test whether increases in model performance in the area showed that the combination of climate, remote sensing CLIMERS scenarios were not simply an artifact of the larger and elevation data resulted in more accurate and spatially number of environmental data layers used, we performed more explicit predictions for both bird species (Table 3). For additionalmodelrunswiththefullsetofdatalayerswithinthe G.spirurus,theCLIMERSscenarioachievedthelowestoverall climate subset (19 WorldClim and 2 elevation), the remote extrinsicomissionrateswitharelativelylowoverallpredicted sensing subset (10 remote sensing +2 elevation) and with all area. For A. melanogenys, the situation is somewhat different. environmentallayers(19WorldClim +10remotesensing +2 BoththeCLIMEandRSEscenariosshowedextrinsicomission elevation).Modelperformanceforscenarioswithreducedand rateslowerthanthoseinCLIMERS,buttheseloweromission full sets of variables within the various predictor sets were rates were only realized at the cost of a notably larger overall generallystable,documentingthataddingmorevariableswith predictedarea.Independentevidenceofoverpredictioninthe Table2 ComparisonofmodelperformanceacrossallthresholdsforeightNeotropicalspeciesproducedinthemodelscenariosRSE, CLIME,andCLIMERS(seeTable1).Foreachscenario,medianAUCandrange(brackets)calculatedonextrinsictestdataandbasedon10 randomdatapartitionsareshown. Birds Mammals G.spirurus A.melanogenys B.variegatus M.minutus RSE 0.877(0.861–0.896) 0.975(0.970–0.980) 0.834(0.790–0.873) 0.976(0.971–0.981) CLIME 0.889(0.865–0.909) 0.975(0.956–0.983) 0.823(0.779–0.875) 0.985(0.984–0.988) CLIMERS 0.898(0.880–0.918) 0.978(0.971–0.984) 0.888(0.826–0.914) 0.984(0.982–0.986) Trees C.brasiliensis V.surinamensis C.guianensis M.bidentata RSE 0.736(0.704–0.802) 0.805(0.754–0.874) 0.849(0.811–0.885) 0.837(0.812–0.863) CLIME 0.759(0.715–0.780) 0.848(0.799–0.881) 0.900(0.881–0.927) 0.888(0.867–0.909) CLIMERS 0.789(0.755–0.829) 0.851(0.808–0.897) 0.905(0.869–0.927) 0.900(0.886–0.917) ToassessstatisticalsignificanceinthecomparisonofmodelperformancebetweentheCLIMEandRSEscenarios,atwo-tailedWilcoxonsignedrank testwasused,andmedianAUCvaluesinunderlined-bold(P<0.005),bold(P<0.05)anditalics(P<0.2)areindicativeofastatisticallysignificant bettermodelperformanceinthecorrespondingscenario.Aone-tailedWilcoxonsignedranktestwasusedtotestwhetherincreasesinAUCsinthe CLIMERSscenariorelativetotheCLIMEandtheRSEscenarioswerestatisticallysignificant.Inthiscase,medianAUCsintheCLIMERSscenariosare onlymarkedinunderlined-bold(P<0.005),bold(P<0.05)anditalics(P<0.2)whendifferencesforbothpairwisecomparisons(RSE-CLIMERS andCLIME-CLIMERS)arebelowtherespectivesignificancelevel. 8 JournalofBiogeography ª2008TheAuthors.Journalcompilationª2008BlackwellPublishingLtd Application of remotesensingdatain speciesdistribution modelling Table3 Comparisonofmodelperformance G.spirurus A.melanogenys atselectedthresholdsforG.spirurusand A.melanogenys.Thetableshows(a)pro- Area Omission Area Omission portionalpredictedareaandextrinsicomis- sionratesatthe‘balance’threshold,and (a)‘Balance’threshold (b)extrinsicomissionratesatconstant RSE 0.367(200–419) 0.061(032–129) 0.141(102–171) 0.022(014–042) proportionalpredictedareaproducedinthe CLIME 0.458(340–561) 0.071(000–161) 0.123(098–177) 0.029(014–042) modelscenariosRSE,CLIMEandCLIMERS. CLIMERS 0.382(317–450) 0.058(000–097) 0.084(068–118) 0.039(014–056) Foreachscenario,themedianandrange (b)Constantproportionalpredictedarea (brackets;onlydecimalshown)valuesbased RSE 0.126(065–186) 0.144(097–194) on10randomdatapartitionsareshown. CLIME 0.124(065–194) 0.121(097–194) CLIMERS 0.109(065–143) 0.089(056–120) For(b),theconstantproportionalpredictedareasare0.295(5%threshold)forG.spirurusand 0.038(10%threshold)forA.melanogenys,andstemfromtheCLIMERSscenarioswithallpoint localitiesused. RSE and CLIME scenarios for A. melanogenys comes from twobirdscanbeestimatedthroughthetraininggainswhenthe visible inspection at the continental scale, where large areas variable ofinterestisusedinisolationandexcludedfrom the werepredicted outside its known range (seebelow). whole set of variables in the maxent runs (Fig. 3). For G. Because this omission test is highly sensitive to the spirurus, this test indicated that the data layers with the most proportional predicted area (Anderson et al., 2003), we also usefulinformationbythemselvesaretheQSCATannualmean comparedextrinsicomissionratesacrossthethreescenariosat followed by rainfall of the warmest quarter. The latter also a constant proportional predicted area close to the observed appears to have the most information that is not shared with ranges of the two bird species (Table 3). For G. spirurus, the other variables. For A. melanogenys, the layers with the extrinsicomissionratesintheCLIMEandRSEscenarioswere most important information by themselves are the two similar. For A. melanogenys, a smaller overall extrinsic topography layers (mean and standard deviation) followed by omissionrateinCLIMErelativetoRSEwasevident,although the maximum temperature of the warmest month and temper- it was not statistically significant (P=0.232; two-tailed atureseasonality,whereasmeantopographyandQSCATannual Wilcoxon signed rank test). Based on one-tailed Wilcoxon mean seem to harbour a significant portion of information signed rank tests, the decreases in overall extrinsic omission that is not contained in the other variables. QSCAT annual ratesfromCLIMEtoCLIMERS(P=0.200forG.spirurusand mean was also generally the most important remote sensing P=0.010 for A. melanogenys) as well as from RSE to variable in the predictions of the additional six non-focal CLIMERS (P=0.077 for G. spirurus and P=0.002 for species (see Fig.S7). To test whether differences in native A.melanogenys)weregenerally statistically significant. spatialresolutionsoftheenvironmentalvariables(seeTable 1) introducebiasesincomparisonsoftheirpredictiveability,we repeated the analysis with all environmental variables coars- Model predictive power, variable importance and response ened to the native 2.25-km QSCAT resolution. In this curves comparison, we found no significant changes in variable To diagnose variations among the three scenarios in model importance,suggestingthatdifferencesinnativedatagrainsize predictivepowerforthetwofocalbirdspeciesinthisstudy,we among variables playaminor role. monitored the likelihood of the training and test point maxent also computes response curves showing how the localities(trainingandtestgain)underthemaxentprobabil- predictions depend on the variables, which greatly facilitates itydistribution(Phillips,2006).Forbothspecies,thisexercise theinterpretationofaspeciesecologicalnicheanditsdefining indicatedastepwiseincreaseintrainingandtestgainsfromthe or limiting environmental factors. The responses of the most RSE to CLIME to CLIMERS scenarios, respectively. In detail, important environmental variables in the predictions for the the overall training/test gains (based on 10 random data twofocalbirdspeciesinthisstudygenerallyagreewellwiththe partitions) increased for G. spirurus from 1.45/1.18 (RSE) to correspondingsamplehistograms(Fig. 4).ForG.spirurus,the 1.53/1.41 (CLIME) to 1.85/1.64 (CLIMERS), and for ecologicalnicheappearstobedefinedbyQSCATannualmean A. melanogenys from 2.89/2.85 (RSE) to 3.00/2.90 (CLIME) values of >)8.0dB and rainfall of warmest quarter between to 3.34/2.99 (CLIMERS). These findings demonstrate the roughly 500 and 1200 mm (Fig. 4a,b). Such environmental complementaryinformationcontentoftheremotesensingand conditionsarecharacteristicofdensehumidforests,wherethe climate data layers in describing the habitat characteristics of species is known to occur (Ridgely & Tudor, 1994; Del Hoyo each species, resulting in a model with the highest predictive et al., 2003). Adelomyia melanogenys, on the other hand, power. appears to prefer mid-elevations and variable terrain typically The absolute and relative importance of individual envi- extending from 1000 to 3000 m a.s.l. (Fig. 4c,d). In addition, ronmental variables as predictors of the distributions of the this hummingbird species is adapted to a relatively narrow JournalofBiogeography 9 ª2008TheAuthors.Journalcompilationª2008BlackwellPublishingLtd W.Buermann etal. Min. temperature of coldest month Max. temperature of warmest month m Temperature seasonality Cli Mean diurnal temperature range World TReaminpfearlal otuf rceo aldnensuta ql umaertaenr Rainfall of warmest quarter Rainfall seasonality Rainfall annual mean LAI annual range e LAI annual maximum Satellit Topography standaTrdTo rpdeoeegv rciaaoptviohenyr QSCAT seasonality QSCAT annual mean 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.80.0 0.5 1.0 1.5 2.0 2.5 3.0 Training gain Training gain Without variable With only variable With all variables Figure3 ResultsofajackknifetestofvariableimportanceforG.spirurusandA.melanogenysintheCLIMERSscenariousingallpoint localities.Barsrepresenttraininggain,ameasureoftheaveragelikelihoodofthepointlocalitiesunderthemaxentprobabilitydistribution, whenaparticularvariableisusedinisolation,excluded,andwhenallvariablesareusedinthepredictionsofthedistributionsofthetwobirds. Variableswithhighestgainswhenusedinisolationcontainthemostusefulinformationbythemselves,whereasvariablesthatleadtolarge decreasesingainswhenomittedfromthepredictorsetcontaininformationthatisnotpresentintheothervariables(seealsoPhillips,2006). Resultsofjackknifetestsofvariableimportanceareprovidedforthesixnon-focalspeciesinFig.S7. maximumtemperatureofthewarmestmonthrangeof18–32(cid:2)C, Hoyoet al.,2003).Incontrast,A.melanogenysoccupiedavery but can tolerate a wide range of temperature seasonality restricted region along the western and eastern flanks of the (Fig. 4e,f). A closer inspection suggests that this extended Andes,anditsnorth-to-southextentfromnorthernVenezuela range in the response to temperature seasonality results from to northern Argentina wascorrectly identified (Fig. 5d-f). localadaptationsamongthepopulationsofA.melanogenysin A closer look at the predicted potential geographic distri- the Andes. In comparison with the tropical populations butionsforG.spirurusandA.melanogenysproducedwiththe (diamonds,Fig. 4f),whichresideinregionsoflowtemperature threemaxentscenariosrevealedsomedistinctcharacteristics. seasonality (<10(cid:2)C), the extratropical populations at the For both bird species, all three scenarios agreed well in their southern margin (crosses, Fig. 4f) of the species distribution spatial predictions of more suitable conditions (>20% cantoleratemuchwiderseasonalswingsintemperatureofup cumulative threshold). The CLIMERS scenario agreed best to 30(cid:2)C. with the known ranges of the two bird species (Fig. 5c,f), whereastheCLIMEandRSEscenariospredictedextensivebut not coinciding areas outside the known ranges of each bird Visualinterpretations (Figs5a,b and 5d,e for G. spirurus and A. melanogenys, respectively). Similar patterns of overpredictions were also Continental scale evidentintheCLIMEandRSEscenariosfortheadditionaltwo The output of the maxent model consists of a continuous mammalandfourtreespeciesthatwereusedinthisstudy(see range with values ranging from 0 to 100 (in units of Figs S1–S6). cumulative probabilities) indicating the least suitable to the most suitable conditions for the taxa under consideration Ecuadorian Andes (Phillips etal., 2006). A visual inspection of the predicted potentialgeographicdistributionsofthetwofocalbirdspecies For Ecuador, predictions for the two focal bird species with in this study showed a broad agreement with their known remote sensing data layers (RSE and CLIMERS) at plausible distributions (Ridgely & Tudor, 1994; Del Hoyo etal., 1999, cumulative thresholds of 2.5–5% (G. spirurus) and 10% 2003)whenlowthresholdsofc.2.5–5%(G.spirurus)and10% (A. melanogenys) were more in agreement with their known (A. melanogenys) were applied (Fig. 5). The good match ranges (Fig. 6). In the case of G. spirurus, these predictions between known ranges and predicted distributions also showed more fine-grained features than in the CLIME suggests that the sampling was adequate for predicting the scenarios, in particular in the Central Coast region and the current distribution ofthetwobird species. western Andean foothills (Fig. 6b–d), where large-scale con- Predictions for G. spirurus covered a vast area in the tinued deforestation has led to fragmented landscapes and northernhalfoftheSouthernAmericancontinentthatclosely habitatloss(Sierra etal.,2002).ForA.melanogenys,RSEand followed the spatial extent of rain forest biomes, including CLIMERS predictions showed narrower distributions along thoseintheAmazonbasinandtheAndeslowlands(Fig. 5a–c). boththecoastandthewesternAndeanmontanecloudforests In addition, maxent correctly predicted the known disjunct thanthoseintheCLIMEscenario(Fig. 6b–d),againmatching populationofG.spirurusintheBrazilianAtlanticForest(Del more consistently human-altered landscape patterns (Sierra 10 JournalofBiogeography ª2008TheAuthors.Journalcompilationª2008BlackwellPublishingLtd

Description:
focal bird species along the steep Amazonian–Andean elevational gradients. Within remote information concerning the geographical distribution of species and their . forests of the Amazon basin and along the coastal regions of northern .. performance of the maxent scenarios using receiver oper-.
See more

The list of books you might like

Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.