Modelos de Distribucion y Divers Id Ad de Arboles en La Amazonia Usando Sensores Remotos

8/3/2019 Modelos de Distribucion y Divers Id Ad de Arboles en La Amazonia Usando Sensores Remotos

1/18

Modeling distribution of Amazonian tree species and diversity using remotesensing measurements

Sassan Saatchi a,, Wolfgang Buermann b, Hans ter Steege c, Scott Mori d, Thomas B. Smith e

aJet Propulsion Laboratory, California Institute of Technology, Pasadena, CA 91109 USAb Center for Tropical Research, Institute of the Environment, University of California, Los Angeles, Los Angeles, CA 90095 USA

c Institute of Environmental Biology, Section Plant Ecology and Biodiversity, Utrecht University, Sorbonnelaan 14, 3584 CA Utrecht, The Netherlandsd New York Botanical Garden, 200th Street and Kazimiroff Blvd., Bronx, NY 10458 USA

e Center for Tropical Research, Institute of the Environment and Department of Ecology and Evolutionary Biology, University of California, Los Angeles,

Los Angeles, CA 90095 USA

Received 26 December 2006; received in revised form 8 January 2008; accepted 12 January 2008

Abstract

The availability of a wide range of satellite measurements of environmental variables at different spatial and temporal resolutions, together withan increasing number of digitized and georeferenced species occurrences, has created the opportunity to model and monitor species geographicdistribution and richness at regional to continental scales. In this paper, we examine the application of recently developed global data productsfrom satellite observations in modeling the potential distribution of tree species and diversity in the Amazon basin. We use data from satellitesensors, including MODIS, QSCAT, SRTM, and TRMM, to develop different environmental variables related to vegetation, landscape, andclimate. These variables are used in a maximum entropy method (Maxent) to model the geographical distribution of five commercial trees and toclassify the patterns of tree alpha-diversity in the Amazon basin. Maxent simulations are analyzed using binomial tests of omission rates and the

area under the receiver operating characteristics (ROC) curves to examine the model performance, the accuracy of geographic distributions, andthe significance of environmental variables for discriminating suitable habitats. To evaluate the importance of satellite data, we used the Maxentjackknife test to quantify the training gains from data layers and to compare the results with model simulations using climate-only data. For allspecies and tree alpha-diversity, modeled distributions are in agreement with historical data and field observations. The results compare withclimate-derived patterns, but provide better spatial resolution and detailed information on the habitat characteristics. Among satellite data products,QSCAT backscatter, representing canopy moisture and roughness, and MODIS leaf area index (LAI) are the most important variables in almost allcases. Model simulations suggest that climate and remote sensing results are complementary and that the best distribution patterns can be achievedwhen the two data sets are combined. 2008 Elsevier Inc. All rights reserved.

Keywords: Species distribution; Remote sensing data; Maxent; Amazon basin; Tree diversity

1. Introduction

Recent efforts to conserve biodiversity are moving beyond preserving only its pattern, such as particular species orpopulations, to include the many complex processes that pro-duce and maintain biodiversity (Cowling and Pressey, 2001;Crandall et al., 2000). The conservation of regional biodiversity

is inextricably linked with the species that occur in a region, thegenes they contain, and the other biotic and abiotic features thatcomprise the ecosystem (Myers et al., 2000). Under pressure tomake informed management decisions rapidly, conservation

practitioners must increasingly rely on predictive models toprovide them with information on species distributions (Ferrier,2002; Loiselle et al., 2003). In addition, using models to predictspecies distributions have become key elements in documenting

biodiversity on the planet and are critical to understanding theeffect of multiple stresses caused by climate and human-inducedchanges (Fjeldsa & Lovett, 1997; Pimm, 1991).

Available online at www.sciencedirect.com

Remote Sensing of Environment 112 (2008) 20002017www.elsevier.com/locate/rse

Corresponding author. Tel.: +1 818 354 1051; fax: +1 818 393 5184. E-mail address: [email protected] (S. Saatchi).

0034-4257/$ - see front matter 2008 Elsevier Inc. All rights reserved.doi:10.1016/j.rse.2008.01.008
mailto:[email protected]://dx.doi.org/10.1016/j.rse.2008.01.008http://dx.doi.org/10.1016/j.rse.2008.01.008mailto:[email protected]


2/18

The majority of studies in biogeography use species occur-rence or museum collections to map and analyze large-scale

patterns of species distribution and richness (Lovett et al., 2000;Rahbek & Graves, 2001). These studies clearly indicate thatspecies differ in the size of their geographic range. Most species,within the same assemblage, tend to have relatively small

ranges that reflect how they share space (Brown et al., 1996;Gaston, 1998). Range size may depend on a variety ofecological and evolutionary processes and extrinsic factors ofthe physical environment such as soils, nutrients, water, andclimate (Gentry, 1988; Hunter, 2003; Kreft et al., 2006; Smithet al., 2001). Capturing the interplay of these factors is fun-damental to understanding the uneven distribution of diversityon regional and global scales.

In most biogeographic theories, geographic distribution ofspecies and their diversity or richness are conceived in terms ofa multidimensional coordinate system, whose axes are variousresource gradients (e.g. ecological and environmental vari-

ables). This coordinate system defines a hyperspace, and therange of the space that a given species occupies is its niche. Theniche is an abstract characterization of the intra-community

position of the species that depends on time, space, and dif-ferences in resource gradients that cause the species evolution(Whittaker, 1972). Geographic distribution of species and theirdiversity or richness depends on how well their ecological nicheis understood.

It is widely accepted that measurement of environmentalrequirements to quantify the range size and patterns of speciesdistribution and richness is an important step towards thisunderstanding (Woodward, 1987). This generalization is true ata variety of spatial scales, suggesting the importance of mea-

surements of environmental variables at different scales. Forexample, climate variables are of increasing importance as thescale increases from regional to continental to global scales.Currently, there is an increasing urgency among conservation

biologists to quantify the environmental requirements of par-ticular species at finer spatial scales in order to better prioritizeconservation efforts. This has created the need to collect spatialinformation over large regions using remotely sensed measure-ments from airborne or satellite sensors (Turner et al., 2003).In addition, the use of remote sensing data by conservation

biologists has helped frame new and important research ques-tions. Can remote sensing data identify areas of significance to

biodiversity, predict species distributions, and model commu-nity responses to environmental and anthropogenic changes?Answering these questions depends on several assumptions:1) environmental variables, and biophysical properties thatcharacterize species habitat, and drive its distribution are de-tectable by existing remote sensing sensors, 2) there are suf-ficient and spatially representative field observations of species

presence or absence and habitat characteristics, and 3) there aredistribution models capable of extending the field observationsto regional and global scales with the aid of environmentalvariables produced by remote sensing measurements. There has

been an increasing interest in studying these assumptions inrecent years (Turner et al., 2003; Nagendra, 2001; Guisan &Zimmermann, 2000; Peng, 2000).

This paper examines the potential use of recently developedglobal datasets from satellite observations for mapping distribu-tion patterns of tree species and diversity in the Amazon basin.Unlike regions with limited species richness and strong gradientsof climate variables, the Amazon basin has one of the highestspecies diversity and richness in the world, but comparatively

little variations in climatic variables (temperature and rainfall)(Nelson et al., 1990; De Oliveira and Mori, 1999). These regionalcharacteristics limit the use of climate variables to developecological and distribution models. Remote sensing data, on theother hand, provide spatially refined information on landscapeand vegetation heterogeneity over the Amazon basin that can bereadily incorporated in models to predict species distribution anddiversity. These models are either strictly mathematical or basedon certain ecological theories. The detailed discussion or reviewof these models and the ecological theories are beyond the scopeof this paper (Elith et al., 2006; Graham & Hijmans, 2006).

Here, we are interested to model the distribution of five

widespread commercial trees, and tree alpha-diversity (ex-pressed as Fisher's alpha) over the Amazon basin. We use themaximum entropy method (Maxent) (Phillips et al., 2005) thatintegrates remote sensing and geographical point locality data ofspecies in order to model distributions and provides a predictive

probability to assess the contribution of remote sensing datalayers. The paper is organized into three sections: 1) descriptionof species, remote sensing, and climate data,2) description of theMaxent model and simulations used for testing the applicationofremote sensing data, 3) assessment of potential range distribu-tions, and 4) discussion on the contribution and significance ofremote sensing data for characterizing suitable areas of specieshabitat.

2. Species data

2.1. Amazonian tree species

Five widespread and well-documented commercial timbertrees were selected for distribution modeling. The geographicallocations of trees were extracted from the herbarium collectionof the New York Botanical Gardens and included data froma variety of forest types and landscape features in northernSouth America (Fig. 1). The species studied were: Calophyllumbrasiliense (Clusiaceae), Carapa guianensis (Meliaceae), Hura

crepitans (Euphorbiaceae), Manilkara bidentata (Sapotaceae),and Virola surinamensis (Myristicaceae). The data set did notinclude any subspecies with strong distributional or functionalcharacteristics or preferences that might influence the overalldistribution.

C. brasiliense (Clusiaceae), known in Brazil by the commonname of jacareuba, grows as a canopy tree in a variety of soil,slopes, and elevations (up to 1500 m). The tree can reach 45 min height with a straight bole without any buttresses or branchesfor about 2/3 of the height. C. (Clusiaceae) is a tropical genuscomposed of approximately one hundred species. Its naturalgeographical range extends from southern Mexico throughoutCentral America to northern parts of South America (Record& Hess, 1943). It is also found in several Caribbean islands

2001S. Saatchi et al. / Remote Sensing of Environment 112 (2008) 20002017


3/18

(Marques & Joly, 2000). We acquired 96 point localities spreadover the Amazon basin for this tree.

C. guianensis (Meliaceae) is a deciduous or semi-evergreenmedium-size species, up to 35 m tall with a straight cylindrical

bole up to 100200 cm in diameter. The species, with the tradename of andiroba, is found in the West Indies, in the Caribbeanislands, throughout Central America and south to the centralAmazon. We acquired 88 point localities for this tree speciesspread mainly along rivers. The tree establishes itself on richsoils along streams, in periodically inundated swamp forests, inupland forests along the rivers of the Amazon basin, and inelevations ranging from shorelines to 1200 m (Guariguata et al.,2002).

H. crepitans (Euphorbiaceae), known as Acacu in Brazil, isanothertalltree,rangingfrom25to50minheightwithclearbolesof 15 to 30 m and with diameters ranging from 100150 cm (at

times to 200 cm). The tree has a native range in tropical America,but has been naturalized in other parts of the world. Its rangeextends from Central America and the Caribbean islands tonorthern South America, with larger concentrations in Colombia,Ecuador, and northern Peru and within the white water varzeafloodplains along the Amazon River. H. crepitans is also foundextensively in coastal Venezuela and the Guyanas on pure sand ormoist sandy loam and is frequently cultivated as a shade treeelsewhere (Freiberg, 1996). However, for this study, we couldonly acquire 45 point localities for this species.

M. bidentata (Sapotaceae) is a large evergreen forest tree foundthroughout the West Indies, ranging from Mexico throughoutPanama to northern South America, and from Venezuela to Peru,including northern Brazil and the Guyanas. The tree is extremely

shade tolerant and grows from coastal sea levels up to fewhundred meters in elevation. There were 140 point localities forthis species that included the two subspecies of bidentata and

surinamensis.

V. surinamensis (Myristicaceae), known as Ucuba in Brazil,has a variety of commercial and medicinal values and is foundin swampy, fertile and periodically inundated riverbanks, inAmazonian varzea forests, and in degraded and secondaryforests. Its geographical range in the Neotropics extends fromCentral America, Costa Rica, and Panama down to the northernAmazon basin and the eastern coastal region in the Guyanas.The tree grows modestly in the open forest gaps and can attain asize of 30 m in height and 100 cm in diameter (Fisher et al.,1991). The tree canopy has seasonal characteristics. In FrenchGuiana, Ucuba flowers twice a year, in March and September,

but near Manaus flowering extends from August to November

and fruiting from January to July (Howe, 1990; Rodriguez,1972). We found 133 point localities for this species.

2.2. Amazonian tree diversity

The tree diversity data were from a total of 633 plots locatedon a variety of forest types, including terra firme, floodplains,and swamps in the Amazon basin and the Guiana Shields(Fig. 2). The data were primarily from published 1 ha plots ofthe ATDN database (ter Steege et al., 2003), however, a numberof smaller plots with sufficient trees (more than 150 individuals)of a diameter at breast-height, dbhN10 cm or larger ones werealso included in the data set (ter Steege et al., 2003). Tree alpha-diversity, expressed as Fisher's alpha (), a measure which

Fig. 1. Geographic locations of tree species inventory data for five commercial tree species, C. brasiliense, C. guianensis, H. crepitans, M. bidentata, andV. surinamensis.

2002 S. Saatchi et al. / Remote Sensing of Environment 112 (2008) 20002017


4/18

corrects for the sample size, was calculated for each plot usingS a ln 1 N

a

, where S is the number of species, N is the

number of individuals, and is the diversity coefficient (Fisheret al., 1943).

Plots' geographical locations, tree alpha-diversity, and foresttypes were among data sets acquired from the Amazon TreeDiversity Network. Detailed information about the data setsand a partial list of references for the published plot data in theAmazon can be found in ter Steege et al. (2003).

3. Environmental data

3.1. Remote sensing data

We compiled a set of remote sensing data and products fromdifferent earth observing sensors to derive metrics sensitive to

vegetation and landscape variables. The data set included bothoptical and microwave satellite sensors. To quantify spatial andtemporal patterns in canopy structure, we used the monthly 1 kmLAI (Leaf Area Index) data derived from MODIS reflectanceover the five-year period, 20002004 (Myneni et al., 2002). Itis noteworthy to mention that in this study we preferred LAIover NDVI or any other vegetation index because of how itrelates to canopy structure and seasonality and it had under-gone various quality checks before and during LAI algorithmimplementation (Myneni et al., 2002). The MODIS 8-day LAI

products provided the basis for these monthly composites,which improved the data quality by further reducing the impactof clouds and any possible LAI estimation errors. We pro-duced monthly climatological means by averaging values over

5 years (20002004). The climatological composites were thenused to generate five metrics: annual maximum, (Fig. 3a)minimum, mean, standard deviation, and range (difference ofmaximum and minimum). These LAI metrics provide infor-

mation on net primary productivity and vegetation seasonality,both important for characterizing species geographical range.

We also included the MODIS-derived vegetation continuousfield (VCF) product as a measure of the percentage of treecanopy cover within each 1 km pixel resolution (Hansen et al.,2002). The VCF product is generated from the time seriescomposites of MODIS data from year 2001 and is availablefrom the Global Land Cover Facility at the University ofMaryland. The VCF product separates open (e.g., shrub lands,savannas), fragmented, and deforested areas from those of intactold growth forests (Fig. 3b).

As part of the microwave remote sensing measurements, we

included global QSCAT (Quick Scatterometer) data available inthree-day composites at 2.25 km resolution (Long et al., 2001).The three-day data over 5 years (20002004) were used tocreate average monthly composites at 1 km resolution and thenfurther processed to produce four metrics that included annualmean and standard deviation of radar backscatter at both HHand VV polarizations (H: horizontal, V: vertical). QSCAT radarmeasurements are at KU band (12 GHz) and are sensitive tosurface or canopy roughness, moisture, and other seasonalattributes, such as phenological changes. For areas with lowvegetation biomass, such as woodlands and savanna, measure-ments at different polarizations correlate positively with theaboveground biomass (Long et al., 2001; Saatchi et al., 2007).For areas with dense forest, backscatter measurements are

Fig. 2. Geographic locations of tree alpha-diversity (Fisher's alpha) in terra firme and inundated forests. Dots indicate the maximum Fisher's alpha found at onelocation (n =633).



5/18

sensitive to canopy roughness and moisture and contribute tomeasuring differences in forest types and canopy structure. Thelong term (5 years) average of MODIS and QSCAT data andthe metrics used in this study are assumed to approximatelyrepresent the climatological mean of the environmental varia-

bles they represent (Buermann et al., 2002). In this study, weused the annual mean (Fig. 3c) and standard deviation ofQSCAT HH backscatter data over 1 year and excluded the VV

backscatter data because of its high correlation with the HHbackscatter over tropical forests.

Finally, we included the SRTM (Shuttle Radar TopographyMission) digital elevation data, aggregated from a 100-meterresolution to 1 km. In addition to the mean elevation (Fig. 3d),the standard deviation was also included to represent surface

ruggedness or roughness. Overall, seven remote sensing datalayers (2 LAI, 2 QSCAT, 1 VCF, 2 SRTM) were included in thisstudy (Table 1). These layers were chosen after performing acorrelation test and removing highly correlated layers (Buer-mann et al., in press).

3.2. Climate data

A series of climate metrics were obtained from WorldClim(WorldClim version 1.4; Hijmans et al., 2005). These climatemetrics are derived from monthly temperature and rainfall valuesand represent biologically meaningful variables for characterizingspecies distribution (Nix, 1986). The WorldClim data layers in-cluded 11 temperature and eight precipitation metrics, expressing

Fig. 3. A selection of the remote sensing data layers used in this study. The panels show (a) MODIS LAI annual maximum, (b) MODIS percentage tree cover,(c) QSCAT annual mean, and (d) mean elevation from SRTM.



6/18

spatial variations in annual means, seasonality, and extreme orlimiting climatic factors. The climate metrics were developedusing long time series of a global network of more than 4000weather stations from various sources such as the GlobalHistorical Climatology Network (GHCN), the FAO (the United

Nations Food and Agricultural Organization), the WMO (WorldMeteorological Organization), the International Center forTropical Agriculture (CIAT), R-HYdronet, and additionalcountry-based stations. The station data were interpolated tomonthly climate surfaces at 5 km spatial resolution by using athin-plate smoothing spline algorithm with latitude, longitude,and elevation (SRTM) as independent variables (Hijmans et al.,

2005).In addition to the bioclimatic variables interpolated from the

station data, we used remotely sensed precipitation data fromthe sensors onboard the Tropical Rainfall Mapping Mission(TRMM) (Kummerow et al., 1998). The TRMM products wereobtained from the global rainfall algorithm (3B43), combiningthe estimates from the sensors with the global gridded raingauge data from Climate Assessment and Monitoring System(CAMS), produced by NOAA's Climate Prediction Center and/or global rain gauge product, produced by the Global Pre-cipitation Climatology Center (GPCC). The output is rainfall for0.25 0.25 degree grid boxes for each month. Monthly rainfall

data from TRMM covering the tropical region (20N20S) andextended to (50N50S) over a period of 9 years (19982006)were used to develop climatologically averaged precipitationmetrics such as the total annual, driest quarter, wettest quarter,and seasonality (coefficient of variation). While developing theclimatological metrics, we resampled the TRMM data to 5 kmresolution using a cubic-spline routine in order to be compatiblewith the WorldClim data layers. The TRMM measurements aresuperior to precipitation layers in the WorldClim dataset becauseof direct rainfall measurements from space, calibration accuracy,and coverage over areas in the tropics where no ground sta-tions are available. After removing the correlated climate layers(Buermann et al., in press), we used only nine independentclimate variables for the model runs (Table 2).

4. Methodology

4.1. Maxent model

We used the Maxent algorithm, which has been very recentlyintroduced for modeling of species distributions (Phillips et al.,

2005). Maxent is a general-purpose algorithm that generatespredictions or inferences from an incomplete set of information.The Maxent approach is based on a probabilistic framework. Itrelies on the assumption that the incomplete empirical probabilitydistribution (which is based on the species occurrences) can beapproximated with a probability distribution of maximum en-tropy (the Maxent distribution) subject to certain environmentalconstraints, and that this distribution approximates a species

potential geographic distribution (Phillips et al., 2005). The inputdata includes a set of environmental layers for a geographicalregion and a set of species presence data inside that region. Likemost maximum likelihood estimation approaches, Maxent, a

priori assumes a uniform distribution and performs a numberof iterations in which the weights are adjusted to maximizethe average probability of the point localities (also known asthe average sample likelihood), expressed as the training gain(Phillips, 2005). These weights are then used to compute theMaxent distribution over the entire geographic space. As in thecase of the present study, Maxent can be applied to species

presence-only geographic locations and remote sensing data toproduce distributions expressing suitability of each grid cell as afunction of the environmental variables at that grid cell. A highvalue of the function at a particular grid cell indicates that thegrid cell is predicted to have suitable conditions for that species(Phillips, 2005).

Compared to other existing models, Maxent has a number offeatures that makes it very useful for modeling species distribu-tion (Elith et al., 2006; Phillips et al., 2005). These include adeterministic framework and, hence, stability as well as theability to run with presence-only point occurrences, high per-formance with few point localities, better computing efficiencyenabling the use of large-scale high-resolution data layers,continuous output from least to most suitable conditions, andability to model complex responses to environmental variables.Last but not least, the newest Maxent version (2.3) is equippedwith several features aimed at supporting the interpretation ofthe model results. For example, Maxent has a built-in jackknife

Table 2Bioclimatic variables used in Maxent predictions

Bioclimate layer5 km resolution

Layer description and unit

BIO1 Annual mean temperatureBIO2 Mean diurnal range (mean of monthly (max tempmin temp))BIO3 Temperature seasonality (standard deviation100)BIO4 Max temperature of warmest monthBIO5 Min temperature of coldest monthBIO6 TRMM annual precipitationBIO7 TRMM precipitation seasonality (coefficient of variation)BIO8 TRMM precipitation of wettest quarter

BIO9 TRMM precipitation of driest quarter

Table 1Overview of remote sensing data sets used in the Maxent predictions along withtheir native resolution and ecological interpretation

Data record Instrument Ecological variable Nativeresolution

Leaf area index

(LAI)

MODIS Vegetation phenology,

structure, and net primary productivity

1 km and 8 days

Maximum LAILAI rangePercent tree cover MODIS Forest cover and

heterogeneity1 km

Scatterometer-backscatter

QSCAT Surface (canopy)moistureand roughness(forest structure)

2.25 km and 3 days

Annual mean HHAnnual STD HHDEM SRTM Topography and

ruggedness90 m1 km

Rainfall TRMM Monthly rainfall 0.25 0.25 deg.



7/18

option, which allows the estimation of the significance ofindividual environmental data layers in computing the speciesdistributions. It provides statistical measures for model per-formance such as omission rates and the areas under theReceiver Operating Characteristic (ROC) curve (AUC), andresponse curves for each environmental layer showing how the

Maxent prediction depends on a particular environmentalvariable (Phillips, 2005). In this study, we used two approachesto evaluate the model performance: the omission rates and theAUC. In general, a low omission rate of species occurrenceis necessary for potentially predicting the species distributionranges (Anderson et al., 2003).

4.2. Scenarios and quantitative analysis

We developed several experiments to test the contribution ofhigh-resolution satellite data in modeling species distributionand diversity. First, we used Maxent to model the distribution of

five tree species with remote sensing data only at 1 km reso-lution and evaluated the overall performance of the model andthe contribution of each data layer. We compared these resultswith Maxent models derived from climate data only andexamined the spatial details obtained from remote sensing data.

For species diversity, we divided the tree alpha-diversityvalues into different range classes based on the histogram of treealpha-diversity values for available sites, and used Maxentto provide distributions for each class. The selection of thediversity range for each category was based on how the valueswere distributed over the entire 633 sites (Fig. 2). We dividedthe sites into incremental groups to sample the distribution of

point locations and provide ample point locations for modeling.

The initial Maxent run was performed for training data withalpha b20, corresponding to all the low diversity sites in theAmazon. Afterward, the model runs were performed for all thesites with values greater than the Fisher's alpha threshold foreach category (N20, N40, N60, N80, N100, N120, N180). Weused a threshold of 25% for the predictive probabilities obtainedfor each class range and combined the derived distributions in adecision rule approach to develop a classification map of treealpha-diversity for the entire range. The threshold value of 25%allowed the largest predictive area and suitability for each classrange. After evaluating the contribution of remote sensing datalayers to model outputs, we compared the results with a similar

experiment obtained from climate-only data.Utilizing all point localities available for each species pro-duced the final distribution maps of tree species and diversity.Spatial accuracy of the Maxent predictions was tested closelyfollowing the procedures in Phillips et al. (2005). In detail, wecreated 10 random data partitions with 60% of the pointlocalities assigned for training and 40% for testing, and ran eachscenario with each of these 10 data partitions. Model perfor-mance was then tested at fixed thresholds (threshold-dependent)and across all thresholds (threshold-independent). In thethreshold-dependent case, we evaluated extrinsic (test) omis-

sion rates, defined as the fraction of test localities that fallinto pixels outside the predicted area, at the 10% cumulative

probability threshold. The proportional predicted area is also

provided as the fraction of all the pixels predicted as beingsuitable for the species. A one-tailed test, as a measure to assesswhether the omission rate is lower or higher than random, wasused to determine whether the model could significantly predictthe test localities.

In the threshold-independent test, we analyzed the area under

the ROC curve (AUC) for both training and the test datasets andestimated how significantly each model prediction differed fromrandom using a ties-corrected MannWhitney-U test (Phillipset al., 2005). The ROC curve provides a quantitative representa-tion of the tradeoffs between omission (sensitivity) and commis-sion error (1-specificity). The sensitivity represents the absenceof the omission error, and the quantity 1-specificity representsthe commission error (Cantor et al., 1999). The ROC curve isobtained by plotting the sensitivity on the y axis and 1-specificityon the x axis for all possible thresholds (Swets, 1988). The areaunder the ROC curve is an important metric to measure the model

performance. The larger the AUC, the highest is the sensitivity

rate and the lower is the 1-specificity rate. An AUC equal to 1.0represents an ideal diagnostic test because it achieves both 100%sensitivity and100% specificity. If AUC is 0.5, it indicates that thetest has50% sensitivity and50% specificity rates, suggesting highomission and commission errors (Cantor et al., 1999).

Finally, we compared the distributions obtained from remotesensing data with Maxent distributions derived from climatevariables. We repeated the same experiments with nine inde-

pendent bioclimatic variables and compared the final distribu-tions for tree species and tree alpha-diversity to illustrate thesignificance of spatial information in satellite observations andto explain the complementary habitat characteristics obtainedfrom remote sensing data and products.

5. Results

5.1. Distribution of tree species

Maxent models for distribution of five tree species weregenerated using the georeferenced locations and remote sensingdata layers at 1 km spatial resolution (Table 1) excluding theTRMM precipitation metrics. We used two indicators to examinethe performance of the model: extrinsic omission evaluated at afixed threshold and the threshold-independent area under theROC curve (AUC) (Table 3). The indicators were obtained using

40% of the point locality data as test localities with the remainderused for training. For all species and all data partitions, the AUCvalues were significantly better than random (0.5). The AUCsin the training and test cases generally showed only smalldifferences, suggesting little overfitting in the Maxent predictions.At the 10% fixed cumulative probability threshold, the extrinsicomission rates were small, associated with reasonable fractions of

predicted areas, again suggesting meaningful model predictions.The overall performance of the model for all five species werehigh, indicating that the Maxent-derived distributions were aclose approximation of the probability distribution that representsthe reality.

Maxent models for distribution of five tree species weregenerated using the georeferenced locations and remote sensing



8/18

data layers at 1 km spatial resolution (Table 1) excludingthe TRMM precipitation metrics. We used two indicators toexamine the performance of the model: the fraction of predictedarea and extrinsic omission rate as threshold-dependent testsand the area under the ROC curve (AUC) as the threshold-independent test (Table 3). The indicators were obtained usingapproximately 40% of the training data as test localities forevaluating the performance statistics. For all species, the AUCvalues were significantly better than random (0.5), and withone-tailed pb0.001. This result was obtained for both thetraining and the test data, with the small difference in AUCvalues suggesting a robust performance of the Maxent algo-

rithm to capture the variations in environmental variables over point localities. All omission tests were calculated at 10%threshold value. At this threshold, the fractional predicted areashows the fraction of all the pixels that are predicted suitable forthe species. For all species, the extrinsic omission rates weresmall, suggesting that only a small fraction of the test locationsfell into pixels not predicted as suitable for species. The overall

performance of the model for all five species was high, im-plying that the Maxent-derived distributions were a close ap-proximation of the probability distribution that represents thereality.

The spatial distributions of tree species in terms of the

predictive probability were segmented in five categories torepresent the ranges of habitat suitability (Fig. 4). As de-scribed earlier, the value assigned to a pixel is the sum ofthe probabilities of that pixel and all other pixels with equalor lower probability multiplied by 100 to give a percentage.Theoretically, any pixel with probability greater than 1% isconsidered suitable for the species habitat. However, here weare mainly interested in areas with higher probability (N20%).

For C. brasiliense (Fig. 4a), the areas in central Amazonhave the highest cumulative probabilities (N20%). Floodplainsof the central Amazon, dominated by close canopy varzeaforests, the tidal varzea of the Amazon estuary in the state ofPara, and southern basins of Tapajos, Itiri, and Xingu rivers, allfall in the N50% probability. C. brasiliense is considered to be

one of the most exploited timber species in the varzea forestsand is on the verge of extinction in these regions due to un-sustainable logging practices (Higuchi et al., 1994). The modelalso predicts areas in fragmented forests along the Atlanticcoast of Brazil in southern Bahia, the northwestern Amazon inColombia, Ecuador, and Peru, and some areas of the Guiana

Shields as suitable habitats for the species (Fisher & DosSantos, 2001). The results of the jackknife test of variableimportance showed the highest gain (N0.3) for the QSCATmean backscatter data, suggesting areas with high moisture in afloodplain and terra firme forests as the suitable habitat forC. brasiliense (Fig. 5a). Other variables, such as the maximumLAI and the percent tree cover with moderate gains (0.10.2),were the next contenders in defining the habitat.

Maxent prediction for C. guianensis pointed to the centralAmazon, the states of Amazonas and Para, the Guiana Shields,the northern coast of Venezuela, and the Atlantic Coastal forestsand varzea floodplains as regions of highest probability for the

geographical range of the species (Fig. 4 b). Remote sensingvariables with the highest gains were QSCAT mean (0.49) backscatter and SRTM elevation (0.6). A close examinationofFig. 5b shows areas delineated by elevation less than 300 mand high moisture along the floodplains and coastal regionswere suitable habitats. In addition, the Colombian Pacific Coastregion (Choc) (Lellinger and Sota, 1978), a stretch of landmainly between the Pacific Ocean and Cordillera Occidental ofthe Andes, from west of the mouth of the Atrato River nearPanama to Mataje River in the south, bordering northwesternEcuador, was also predicted as the suitable range for C.

guianensis (known as tangare in the region) (Gentry, 1982).This result clearly shows the strength of remote sensing data and

the Maxent model for predicting species range, in particular inareas where no training data were available.

H. crepitans had the lowest numbers of point localities amongthe five species and they were scattered mainly in the westernAmazon. Maxent predicted areas outside the central Amazon asthe suitable habitat. Areas with the highest probabilities (N50%)were in the western lowlands of Peru and Ecuador, in southernBolivia, along the Beni river basin, and in the eastern regions ofthe Brazilian Amazon and the coastal regions of Surinam,Guyana, and French Guiana (Fig. 4c). Along the Atlantic coastof Brazil, the model predicts small regions in southern Bahia as

potentially suitable habitat for H. crepitans. Similarly, narrow

regions in varzea forests along the Rio Solimoes and itstributaries are delineated as potential habitat. In general, H. crepitans is considered a semi-evergreen species found inseasonal forests which, along with other emergent trees,undergoes foliage reduction during the dry season (Conditet al., 2000; Schongart et al., 2002). Among remote sensingdata,QSCAT mean and standard deviation, maximum and range ofLAI were selected as variables with high gains for defining thespecies range. Mean QSCAT and maximum LAI both reachedN0.6 gains through the jackknife analysis, suggesting forestswith canopy moisture, roughness, and leaf area as potentialhabitat (Fig.5c). The standard deviation of QSCATand the rangeof LAI had gains N0.3, and both pointed to forests with seasonalcanopy characteristics. Topography, on the other hand, was not

Table 3Results of threshold-dependent omission tests and threshold-independent ROCtests for five tree species, including fractional predicted area, test omission rates,and the area under the ROC curve (AUC)

Species name andnumber of occurrences

Threshold-dependent test Threshold-independenttest

Fractionalpredictedarea

Test omissionrate AUC test (training)

(a) C. brasiliense (96) 0.533 0.106 0.756 (0.821)(b) C. guianensis (88) 0.356 0.153 0.843 (0.914)(c) H. crepitans (45) 0.289 0.212 0.832 (0.921)(d) M. bidentata (140) 0.321 0.095 0.853 (0.907)(e) V. surinamensis (133) 0.377 0.142 0.833 (0.896)

Test omission rates were calculated at the 10% threshold level. Values representaverages from 10 separate random training/test data partitions. For eachpartition, statistical evaluations of test omission rates (one-tailed binomial) andtest AUC (MannWhitney-U) indicated that the predictions were significantlybetter than random (pb0.001; individual p-values not shown).



9/18

Fig. 4. Maxent Prediction of potential geographic distribution of five tree species made using all occurrence records and the remote sensing data at 1 km resolution.The predictive probability values ranging from 0 to 100 are depicted by colors. (a) C. brasiliense, (b) C. guianensis, (c) H. crepitans, (d) M. bidentata, and(e) V. surinamensis.



10/18

Fig. 5. Results from the Maxent jackknife test of the importance for remote sensing variables used for five tree species. The graphs depict the training gains when avariable is used in isolation, when the variable is excluded, and when all variables are utilized. The gain is a measure of how much better the Maxent probabilitydistribution fits the distribution of occurrence data. A variable has useful information when the gain is high and it is used in isolation and has unique information if itreduces the gain most when it is excluded.



11/18

an important indicator for the habitat. Visual inspection of thepredicted areas at 1 km resolution confirmed that the distributioncompared well with the seasonal characteristics ofH. crepitansobserved in field experiments (Freiberg, 1996; Schongart, et al.,2002).

M. bidentata distribution is widespread in the Amazon basinwith the highest predicted probability in the Central regionof Brazil, especially in the state of Amazonas, Para (heavilylogged), Amapa, and Roriama, and along the northeasternAtlantic coast, the Amazon estuary, French Guiana, Surinam,

and Guyana (Fig. 4d). Maxent predicts areas in upland terrafirme and floodplain forests as preferred habitats for M. bidentata. The distribution extends to western SouthAmerica, the state of Acre in Brazil, northern Peru, Ecuador,and Colombia. The range also covered areas in the westernChoc region of Colombia to southern Panama (Faber-Langendoen & Gentry, 1991). Several remote sensing variableshelp define the geographical range ofM. bidentata. The traininggains obtained from the jackknife statistics showed meanQSCAT (gain N0.5), standard deviation (gain N0.3), MODIS

percent tree cover (gain N0.4), SRTM elevation (gain N0.4),and maximum LAI (gain N0.4) are among the important re-

mote sensing variables contributing to the predicted distribution(Fig. 5d). In general, closed canopy forests (high percent treecover), moist (high mean QSCAT), with low seasonality (lowQSCAT standard deviation) and medium LAI with almost noseasonal variations were the best indicators for the habitat. Therange was primarily limited to areas of low elevation and smallvariations in topography.

V. surinamensis is also predicted to be widespread in thecentral Amazon, extending from the eastern Atlantic coast andthe Guiana Shields to the western regions of Peru, Ecuador,Colombia, and areas along the lowland Andes to southernridges in Bolivia (Fig. 4e). Areas with high predictive

probability are predominantly along the Amazon River flood-plains, coastal forests extending to Venezuela, and in the Choc

region of western Colombia to Panama (Fisher et al., 1991;Gentry, 1975; Howe, 1990; Rodriguez, 1972). Areas in thesouthern Amazon region of Brazil along the transitional forestsare also predicted as suitable habitat, but with lower probability(b20%). Analysis of Maxent results showed seasonality as lessimportant in characterizing the distribution. This is mainly due

to the fact that the point localities of the species covered a widerange of landscapes with a wide range of seasonality. However,the canopy moisture from QSCAT backscatter (gain N0.5),maximum LAI from MODIS (gain N0.4), and the elevationfrom SRTM (gain 0.4) contributed significantly in definingthe species range (Fig. 5e). One of the most apparent features inthe distribution is the high probability of prediction (N50%) inthe Amazon floodplains.

5.2. Distribution of tree alpha-diversity

Using the remote sensing data layers, we ran the Maxent

model for nine categories of Fisher's alpha (Fig. 6). Thenumbers of point localities for each category were sufficient forthe model runs without encountering problems associated withover- or under-predictions. In fact, in all cases, the model

performance determined by the area under the ROC curve wassignificantly (pb0.001) better than random (Table 4). Similarly,the training omission rates were small and the fractional

predicted area over the entire environmental space coveringnorthern South America was reasonably large. The results inTable 4 provided confidence in Maxent prediction of spatialdistribution of tree alpha-diversity for each category. The resultsfrom dividing the point localities randomly in training (60% of

points) and testing (40% of points) reduced the AUC values

about 5% on the average, suggesting a reliable model per-formance under more constrained conditions. The threshold-dependent omission tests also provided low omission rates andrelatively large fractional predicted area (Table 4).

After combining the distribution maps derived for all ninescenarios, a classification map of Fisher's alpha was produced

Fig. 6. Results from Maxent predictions of tree alpha-diversity made fromremote sensing data and inventory plots (n =633). Classification map of treealpha-diversity produced from Maxent predictions using 25% probability

threshold for each class range.

Table 4Results from the threshold-dependent omission tests and threshold-independentROC tests for eight range classes of tree alpha-diversity (expressed as Fisher'salpha), including fractional predicted area

Fisher'salpha-

diversityclass

Numberof point

localities

Threshold-dependent test Threshold-independenttest

Fractionalpredicted area

Test omissionrate

AUC test (training)

b20 118 0.345 0.118 0.861(0.897)N20 515 0.278 0.128 0.877(0.909)N40 370 0.266 0.106 0.886(0.915)N60 249 0.226 0.117 0.901(0.932)N80 176 0.221 0.141 0.907(0.940)N100 112 0.207 0.106 0.911(0.933)N120 79 0.184 0.080 0.935(0.951)N180 28 0.144 0.038 0.952(0.970)

Test omission rates were calculated at the 10% threshold level. Values representaverages from 10 separate random training/test data partitions. For eachpartition, statistical evaluations of test omission rates (one-tailed binomial) and

test AUC (Mann

Whitney-U) indicated that the predictions were significantlybetter than random (pb0.001; individual p-values not shown).



12/18

at 1 km resolution. The resulting map is a potential distributionof ranges of tree alpha-diversity over the basin (Fig. 6). Giventhe Maxent performance and the distribution of sampling sitesover the basin, any extrapolation of diversity outside the mainrange of the sampled data and over other ecosystems or biomassis not allowed. Therefore, the analysis of the range map is

restricted to the Amazon basin.The largest swath of high tree diversity is in terra firme forest

of the western Amazon, stretching from the foothills of theAndes in southern Colombia, Ecuador, and Peru into the centralAmazon in Brazil. The Fisher's alpha associated with thisregion exceeded 180. The map also revealed many smaller areasof high tree diversity (alpha N180) in areas east of the Rio Negroand north of Manaus, in northeastern Brazil in the state ofAmapa, the southern basins of Tapajos and Xingu rivers in thestate of Para, areas along southern and eastern Guiana Shields,and outside the basin in the Atlantic coastal forests in Bahia.

In contrast, areas in the central Amazon between the Rio

Negro and Solimoes and their tributaries are dominated byforests with lower tree alpha-diversity (alpha b120). This regionis dominated by extensive river systems, varzea and igapofloodplains, and depending on their proximity to rivers and theirsediment load, and topography, the forests contain a large rangeof diversity. Further south, in transitional deciduous and semi-deciduous forests of Brazil and in Chiquitano dry forests ofBolivia, the diversity drops to its minimum (alpha b40). Notethat the separation of the high diversity Amazon basin from thesurrounding woodland and grassland savanna (cerrado) andhigher elevation Andean vegetation types is an artifact resultingfrom Maxent's extrapolation over the environmental space. Asthere were no point localities sampling these biomes in our

database, model predictions for these regions are not warranted.The jackknife training gains for the remote sensing variables

were performed for all individual runs and showed almost thesame results with slight variability in gain values. We showresults for alpha N20 to demonstrate the significance of eachvariable (Fig. 7). The most important variables were vegetationcanopy roughness and moisture from QSCAT, percent treecover from MODIS, and mean and seasonality of LAI fromMODIS. The highest gain was achieved for the QSCAT mean

backscatter (gain N0.7), maximum LAI (gain N0.6), and percenttree cover (gain N0.5). SRTM elevation, although important forthe overall performance of the model, had a low gain compared

to other variables (gain b0.3). LAI range and the standarddeviation of QSCAT both had relatively high gains, suggestingseasonality as important variables for the distribution of areaswith high diversity. An examination of the model responsecurves to input variables suggested that, in general, high mois-ture and low seasonality were associated with areas of high

diversity.To further demonstrate the role of remote sensing data in

separating the tree diversity over the basin, we plotted theFisher's alpha for all 633 plots with respect to the mean andstandard deviation of backscatter from QSCAT and the SRTMstandard deviation (Fig. 8). In each case, we developed anenvelope based on exponential functions to show the relation-ship between the maximum (equivalent to 90th quantile) treealpha-diversity and the remote sensing variable. The envelopesshowed three important trends: 1) QSCAT backscatter were

positively correlated (R2 =0.73, pb0.001) with the maximumalpha-diversity suggesting areas with higher canopy moisture

and roughness associated with higher diversity (Fig. 8a),2) QSCAT standard deviation of backscatter was negativelycorrelated (R2 =0.81, pb0.001) with maximum tree alpha-diversity indicating higher tree diversity was associated withareas of less seasonality in high stability in moisture (Fig. 8b),and 3) maximum tree alpha-diversity was negatively correlated(R2 =0.62, pb0.001) with standard deviation of SRTMsuggesting areas with less variations in elevation, primarilylowlands, were associated with higher tree diversity (Fig. 8c).

5.3. Comparison with climate derived models

We performed the comparison of remote sensing results with

distributions derived from nine bioclimatic layers for all fivetree species and the Fisher's alpha scenarios. Here, we showdistributions from V. surinamensis and the tree alpha-diversityto summarize the results from bioclimatic variables (Fig. 9). Forall tree species, the distributions derived from remote sensingdata were superior to climate data mainly because of the verycoarse spatial resolution of the climate data and the limitedvariations of temperature and precipitation over the lowlandAmazonian forests where species data were collected. Forexample, the V. surinamensis distributions from climate (Fig. 9a)and remote sensing data (Fig. 4e) have similar patterns for com-

parable predictive probabilities. However, there are two distinct

Fig. 7. Example of the Maxent jackknife test gains for the importance of remote sensing variables for tree alpha-diversity N20 (n =515).



13/18

differences: 1) finer resolution remote sensing data allows

prediction of range patterns along geomorphological features inthe central Amazon and topographical and vegetation featuresalong the Guiana Shields. These features disappear in climateresults.2) Models fromclimate data underpredict thespeciesrangein southern margins of the Amazon along the transitional forestsand overpredict in savanna regions such as Roraima, the GranSavanna and areas in cerrado Brazil. The jackknife test (Fig. 9b)shows several temperature and rainfall variables such as thetemperature mean diurnal range (BIO2), seasonality (BIO3),temperature of the coldest month (BIO5), total annual precipita-tion (BIO6), and mean precipitation of the driest month (BIO9)as important habitat characteristics. If predicted correctly, thesevariables are complementary to remote sensing data (Fig. 5e)where canopy moisture and seasonality (QSCAT), maximum leaf

area (MODIS), and low elevation (SRTM) are the dominanthabitat characteristics.

The distribution for Fisher's alpha classes derived fromclimate data (Fig. 9c) has general patterns similar to thoseobtainedbyremotesensingdata(Fig.7). Visual comparison of thetwo diversity maps reveals three regions of high diversity within

the Amazon basin: 1) the western Amazon basin, includingwestern and northwestern Brazil, northern Peru, Ecuador, and theeastern Colombian Amazon, 2) the central Amazon basin, in-cluding areas east of the Rio Negro and north of Manaus towestern Para, and 3) areas in northeastern Brazil and part of theGuiana Shields. Except for the western Amazon basin, the pat-terns do not necessarily cover the exact geographical regions. Ingeneral, the distribution from remote sensing data is spatiallyrefined and shows patterns delineated by geomorphological andgeological features of the Amazon basin, whereas, the climate-derived distribution is coarse in resolution and is distinguished

primarily by patterns of precipitation. Similarly, areas of high

diversity (alphaN

180) in the Guiana Shields appear continuousin the remote sensing results, with the patterns following thegeological and vegetation gradients, but patchy and discontinuousin the climate results. The southwestern region of the Amazon,including southern Peru and Bolivia, also appear different in thetwo distributions. The climate results show low diversity withinthe dense old growth forests of lowland Peru caused by tem-

perature and precipitation seasonality. In contrast, the remotesensing results show higher tree alpha-diversity (alpha N120) inthe lowland old growth forests of Peru and northern Bolivia. Ingeneral, the distribution of tree alpha-diversity from the remotesensing data has much smoother variations within the Amazon

basin than the climate results. The Maxent model results from the

climate data underpredict the tree alpha-diversity in most areas ofthe central Amazon and create patterns different from observedtree diversity in the region (ter Steege et al., 2003). The jackknifetest highlights these points by choosing rainfall of the driestquarter (BIO9), total rainfall (BIO6), and temperature range(BIO2), seasonality (BIO3), and minimum temperature of thecoldest month (BIO5) as important variables in predicting areas ofhigh diversity (ter Steege et al., 2006). Given the complementaryinformation in remote sensing and climate layers, it is expectedthat the best distribution for diversity may be produced fromcombined climate and remote sensing data (Prates-Clark et al.,2008).

6. Discussion and conclusion

6.1. Contribution of remote sensing data

For both tree species and tree alpha-diversity, remote sensingdatasets provided meaningful and significant contributions indefining the distribution range and spatial patterns. We sum-marize these contributions in two areas: 1) improving the spatialresolution and, therefore, providing landscape-level details on

potential habitat characteristics, and 2) adding to the pool ofenvironmental variables beyond climate surfaces and henceimproving the definition of habitat properties and ecologicalniche.

Fig. 8. Distribution of tree alpha-diversity of inventory plots as a function ofselected remote sensing variables. Exponential functions are used as envelopesto show the general trends of maximum diversity (equivalent to 90th quantile)with respect to (a) QSCAT annual mean backscatter, (b) QSCAT annual standarddeviation of backscatter, and (c) standard deviation of SRTM data.



14/18

A subset of the tree alpha-diversity maps produced from 1 kmremote sensing data (Fig. 6) and 5 km climate data (Fig. 9c) overthe central Amazon in Amazonia and the Para states of Brazil

demonstrated this point (Fig. 10). With the 1 km remote sensingdata (Fig. 10a), landscape features due to forest fragmentation,differences in forest cover (terra firme and inundated), andlandscape geomorphology were readily delineated and reason-able differences in their tree alpha-diversity were observed. Incontrast, the climate-derived map (Fig. 10b), although showinga general pattern, does not provide reasonable prediction anduseful patterns of tree alpha-diversity in this region. In general,climate surfaces interpolated from station data or derived fromcoarse resolution satellite measurements cannot capture land-scape-scale variations in diversity. It is also important to note thatunlike potential species distribution influenced by environmen-tal factors, diversity depends on the size of the area sampled,climate, past history, and local influences, such as soil, geology,

and nutrients. Therefore, diversity is very much a local orregional property of a forest and cannot be readily extrapolatedto other regions. Nevertheless, as demonstrated in this study,

quantification of landscape heterogeneity from high-resolutionsatellite observations can readily improve our understanding ofthe biogeography and biodiversity of the lowland Amazonianrainforests from typical postulated or observed distribution

barriers such as unfavorable past climates, mountains, rivers,and river floodplains (Tuomisto et al., 1995).

Furthermore, remote sensing data provide measurementsdirectly related to forest structure, species composition, gapfraction, and the overall health of the ecosystem that cancollectively improve our understanding of suitable habitats forspecies. Maxent offers response curves for the input environ-mental variables that allow examining how the predictions ofsuitable habitatdepend on each variable. To demonstrate this, weexamined the QSCAT response curves, as one of the most

Fig. 9. Maxent predictions derived from bioclimatic variables at 5 km resolution and corresponding Maxent jackknife test results for the variable importance.(a) V. surinamensis distribution, (b) jackknife test forV. surinamensis. (c) classification of tree alpha-diversity, and (d) jackknife test for alpha N20 class (n =515).



15/18

important variables in predicting the distribution ofM. bidentataand V. surinamensis (Fig. 11). The response curves were derivedfrom Maxent runs with QSCAT used in isolation in order toavoid interferences with other variables. QSCAT values overmost of the point localities (shown as diamonds) and plotted aslog of sample frequency, are scattered within a small range of thevariable between 10.0 dB and 6.0 dB. These are typicalQSCAT values measured over tropical forests. Within thisrange,

there is a major difference in the response curves between thetwo species. These differences force Maxent to choose thresh-olds to separate suitable areas and predict different distribution

patterns, Areas outside of this range, although included in theoverall environmental space, will not contribute to defining thespecies range. We expect the combination of climate and remotesensing data and the multiscale analysis may provide the bestdistributions for species range and diversity.

Fig. 11. Maxent response and sample frequency forM. bidentata and V. surinamensis as a function of QSCAT annual mean backscatter. The response curves illustratehow the contribution to the raw Maxent prediction depends on a particular environmental variable (Phillips et al., 2005). The Maxent response curves were derived

from Maxent runs with QSCAT used in isolation to avoid interferences with other variables. The sample frequency (plotted as diamonds) shows the number of pointlocalities that fall in a certain QSCAT interval.

Fig. 10. Comparison of Maxent predictions of tree alpha-diversity classification over the Atlantic Coastal Forests of Bahia from (a) 1 km remote sensing data and(b) 5 km bioclimatic variables.



16/18

6.2. Maxent predictions

For those species whose distribution patterns are determinedprimarily by evolutionary processes and gene flow and less byenvironmental variables, especially those detected by climate,landscape, or vegetation, predictive models are not useful. In

general, models are based on algorithms that extrapolate thesensitivity developedby thetraining data from point localities andthe environmental variables to a larger space. A true statistical testfor any model prediction is how significantly and consistently it

performs better than a random prediction. Maxent has internalroutines that provide threshold-dependent and threshold-inde-

pendent tests of its performance by examining the omission rateand the area under the ROC curve (AUC) for a set of randomlyselected test localities. In case of low sensitivity to environmentalvariables for a set of point localities, both tests will show poor

performance of the algorithm. In all cases studied here, consistentand significantly better than random performance were achieved.

However, without absence data (not available in most biologicaldata), there seems to be no source of negative instances to ac-curately measure the predictions over the rest of the environ-mental space. Phillips et al. (2005) developed non-parametrictests using the MannWhitney-U-statistic and a sample of 10,000

pixels drawn randomly from the study region to examine towhat extent the prediction differs from random (AUC of 0.5).Although, this approach is not as rigorous as tests with absencedata, it provides more confidence in model predictions. Fur-thermore, it provides enough confidence in distribution patternsderived from remote sensing data. Following a similar approach,we concluded that in examples used in our study, Maxent

predictions were significantly different than random and its

performance was close to optimal. These tests suggest that modelpredictions derived from remote sensing data provide meaningfuland reasonable distribution patterns. With the predictions relatedto tree alpha-diversity, the results also confirmed that thedistribution is only valid within the biomes sampled by the

presence data and anyextrapolation of diversity to other biomesisnot valid.

6.3. Future work

Spatial, spectral, and temporal diversity of recent satelliteobservations and improvement of algorithms to derive ecolo-

gically important variables has enhanced our capabilities inconservation biology in different areas: 1) to directly map in-dividual species over relatively large and spatially contiguousunits, 2) to map homogeneous associations dominated by fewspecies or diversity indicators, and 3) to develop environmentalrequirements for species range and diversity. Results obtainedfrom this study and similar studies using satellite data suggestthat nature conservation in Amazonia and other tropical re-gions can benefit from recognizing ecological heterogeneity atlandscape scales and the use of methods and datasets that cancarefully distinguish these heterogeneities (Nagendra, 2001;Tuomisto et al., 1995). Interpolation and extrapolation of fieldinventory data on species presence, richness, and diversity on aspatial scale depend on how well the environmental variables

and ecological heterogeneities are characterized on that scale.Analysis of satellite images over Amazonian forests has shownthat in addition to spatial information, the spectral data can beused to separate structurally and floristically distinct biotopeswithin the vegetation types already known (Chambers et al.,2006; Lucas et al., 2004).

These results also call for new areas of research in usingsatellite observations in biogeography and the conservation of

biodiversity. First, the use of remote sensing data in predictivemodels requires new rules and protocols to be developed inorder to improve the assessment of distribution patterns fromregional and continental to landscape scales. Finer-resolutionenvironmental data may not improve model predictions, as theymay be incompatible with the spatial scale of the inventory datafrom natural history museums and herbaria and they mayintroduce unwanted and additional statistics in the input dataand thus impact the performance of predictive models. Furtherresearch is also required to determine to what degree the spatial

resolution of satellite observation can help or limit theidentification of species' environmental requirements andtheir ecological niche.

Research is also needed regarding the utility of spectral dataor remote sensing products to quantify and map ecologicallyimportant features on landscapes. Currently, remote sensing-

based habitat characterization is mainly based on the relation between spectral data and the structure, chemistry, andheterogeneity of vegetation within a pixel resolution. It is notclear whether habitat suitability for most species can be definedin terms of these variables. Currently, specieshabitat relation-ships are defined by eco-region classifications based on climate,geology, and natural barriers. In particular, for mobile taxa like

birds or butterflies, the specieshabitat relationships are notwell defined or quantifiable. New approaches are required toextend these relationships to physical environmental variablesdetectable by satellite observations. In addition, research is alsorequired to explore whether modeling distributions for a numberof species or communities or developing relationships betweenspectral data and species richness and diversity may enhanceand improve the utility of satellite data.

Detection and assessment of changes in landscapes, such asdeforestation and land use change, have been explored exten-sively in environmental sciences using the temporal diversity ofremote sensing data. In addition, availability of time series

satellite observation in the past three decades has also providedinformation on both the stability and the dynamics of ecosystemsand natural habitats. In this study, we used time series data fromMODIS and QSCAT to develop climatological metrics. However,further research is required to improve the utility of these datasetsin predictive models and to understand the impact of land use

patterns and interannual variability present in time series remotesensing data on species distribution.

Acknowledgement

We thank Donat Agosti for his help with acquiring the treespecies data and Ana Paula Giorgi for her help with the GISlayers. The tree species data was provided by the herbarium of



17/18

the New York Botanical Garden and the tree diversity data isfrom the Amazon Tree Diversity Network. However, this workwould not have been possible without decades of dedicatedand methodic fieldwork of numerous researchers across SouthAmerica. We would also like to thank the reviewers who pro-vided us with important suggestions and criticisms of the original

manuscript. This work was performed at the Jet PropulsionLaboratory, California Institute of Technology, and the UCLACenter forTropical Research, Institute of the Environment, under acontract from the National Aeronautics and Space Administration.

References

Anderson, R. P., Lew, D., & Peterson, A. T. (2003). Evaluating predictivemodels of species' distributions: Criteria for selecting optimal models.

Ecological Modelling, 162, 211232.Brown, J. H., Stevens, G. C., & Kaufman, D. M. (1996). The geographic range:

Size, shape, boundaries, and internal structure. Annual Review of Ecologyand Systematics, 27, 597623.

Buermann, B., Saatchi, S., Zutta, B. R., Chaves, J., Mil, B., Graham, C. H.,

Smith, T. B. (in press). Application of remote sensing data in predictivemodels of species' distribution. Journal of Biogeography.

Buermann, W., Wang, Y., Dong, J., Zhou, L., Zeng, X., Dickinson, R. E., Potter,C. S., & Myneni, R. B. (2002). Analysis of a multiyear global vegetationleaf area index data set. Journal of Geophysical Research, 107, 4646.doi:10.1029/2001JD000975

Cantor, S. B., Sun, C. C., Tortolero-Luna, G., Richards-Kortum, R., & Follen, M.(1999). A comparison of C/B ratios from studies using receiver operatingcharacteristic curve analysis. Journal of Clinical Epidemiology, 52, 885892.

Chambers, J. Q., Asner, G. P., Morton, D. C., Anderson, L. O., Saatchi, S. S.,Esprito-Santo, F. D. B., Palace, M., & Souza, C., Jr. (2006). Regionalecosystem structure and function: Ecological insights from remote sensingof tropical forests. Trends in Ecology & Evolution, 22(8), 414423.

Condit, R., Watts, K., Bohlman, S. A., Prez, R., Hubbell, S. P., & Foster, R. B.(2000). Quantifying the deciduousness of tropical forest canopies under

varying climates. Journal of Vegetation Science, 11, 649658.Cowling, R. M., & Pressey, R. L. (2001). Rapid plant diversification: Planning

for an evolutionary future. Proceedings of the National Academy of Sciences, 98(10), 54525457.

Crandall, K. C., Bininda-Emonds, O. R. P., Mace, G. M., & Wayne, R. K.(2000). Cosidering evolutionary processes in conservation biology. Trendsin Ecology and Evolution, 15(7), 290295.

De Oliveira, A. A., & Mori, S. (1999). A central Amazonian terra firme forest. I.High tree species richness on poor soils. Conservation Biology, 8,12191244.

Elith, J., Graham, C. H., Anderson, R. P., Dudk, M., Ferrier, S., Guisan, A., et al.(2006). Novel methods improve prediction of species' distributions fromoccurrence data. Ecography, 29, 129151.

Faber-Langendoen, D., & Gentry, A. H. (1991). The structure and diversity ofrain forests at Bajo Calima, Choco Region, Western Colombia. Biotropica,

23, 211.Ferrier, S. (2002). Mapping spatial pattern in biodiversity for regional conservation

planning: Where to from here? Systematic Biology, 51(2), 331363.Fisher, R. A., Corbet, A. S., & Williams, C. B. (1943). The relation between the

number of species and the number of individuals in a random sample of ananimal population. Journal of Animal Ecology, 12, 4258.

Fisher, E., & Dos Santos, F. A. M. (2001). Demography, phenology and sex ofCalophyllum brasiliense (Clusiaceae) trees in the Atlantic forest. Journal ofTropical Ecology, 17, 903909.

Fisher, B. L., Howe, H. F., & Wright, S. J. (1991). Survival and growth of Virolasurinamensis yearlings: Water augmentation in gap and understory. Oeco-logia, 86, 292297.

Fjeldsa, J., & Lovett, J. C. (1997). Biodiversity and environmental stability. Biodiversity and Conservation, 6, 315323.

Freiberg, M. (1996). Spatial distribution of vascular epiphytes on three emergent

canopy trees in French Guiana. Biotropica, 28(3), 345355.

Gaston, K. J. (1998). Species-range size distributions: Products of speciation,extinction, and transformation. Philosophical Transactions of the RoyalSociety B, 353, 219/230.

Gentry, A. H. (1975). Additional Panamanian Myristicaceae. Annals of the Missouri Botanical Garden, 62(2), 474479.

Gentry, A. H. (1982). Phytogeographic patterns as evidence for a Choc refuge.In G. T. Prance (Ed.), Biological diversification in the tropics (pp. 112136).

New York: Columbia University Press.Gentry, A. H. (1988). Tree species richness of upper Amazonian Forests. Pro-ceeding of National Academy of Sciences , 85, 156169.

Graham, C. H., & Hijmans, R. J. (2006). A comparison of methods of speciesranges and species richness. Global Ecology and Biogeography, 15,578587.

Guariguata, M. R., Claire, H. A., & Jones, G. (2002). Tree seed fate in a loggedand fragmented forest landscape, Northeastern Costa Rica. Biotropica, 34,405415.

Guisan, A., & Zimmermann, N. E. (2000). Predictive habitat distribution modelsin ecology. Ecological Modelling, 135, 147186.

Hansen, M. C., DeFries, R., Townshend, J., Sohlberg, R., Dimiceli, C., &Carroll, M. (2002). Towards an operational MODIS continuous field ofpercent tree cover algorithm: Examples using AVHRR and MODIS data.Remote Sensing of Environment, 83, 303319.

Higuchi, N., Hummel, A. C., Freitas, J. V., Malinowski, J. R. E., & Stokes, R.(1994). Explorao Florestal nas Vrzeas do Estado do Amazonas: Seleode rvore, Derrubada e Transporte. Proceedings of the VII Harvesting andTransportation of Timber Products (pp. 168193). Curitiba, Brazil: IUFRO/UFPR.

Hijmans, R. J., Cameron, S. E., Parra, J. L., Jones, P. G., & Jarvis, A. (2005).Very high resolution interpolated climate surfaces for global land areas.

International Journal of Climatology, 25, 19651978.Howe, F. (1990). Survival and growth of juvenile Virola surinamensis in Panama:

Effects of herbivory and canopy closure. Journal of Tropical Ecology, 6,259280.

Hunter, J. T. (2003). Factors affecting range size differences for plant species onrock outcrops in eastern Australia. Diversity and Distributions, 9, 211 /220.

Kreft, H., Sommer, J. H., & Barthlott, W. (2006). The significance of geographicrange size for spatial diversity patterns in Neotropical palms. Ecography,

29, 2130.Lellinger, D. B., & Sota, E. R. de la (1978). The phytogeography of the

pteridophytes of the Departamento del Choc, Colombia. NationalGeographic Society Research Reports 1969 projects (pp. 381387).

Loiselle, B. A., Howell, C. A., Graham, C. H., Goerck, J. M., Brooks, T., Smith,K. G., & Williams, P. H. (2003). Avoiding pitfalls of using species distributionmodels in conservation planning. Conservation Biology, 17, 15911600.

Long, D. G., Drinkwater, M., Holt, B., Saatchi, S., & Bertoia, C. (2001). Globalice and land climate studies using scatterometer image data. EOS,Transaction of American Geophysical Union, 82(43).

Lovett, J. C., Rudd, S., Taplin, J., & Frimodt-Mller, C. (2000). Patterns of plantdiversity in Africa south of the Sahara and their implications forconservation management. Biodiversity and Conservation, 9, 3342.

Lucas, R., Held, A. A., Phinn, S. R., & Saatchi, S. S. (2004). Tropical forests. InS. Ustin (Ed.), Remote Sensing of Earth Sciences: Manual of Remote

Sensing, Vol. 4. John Wiley & Sons.Kummerow, C., Barnes, W., Kozu, T., Shiue, J., & Simpson, J. (1998). The

Tropical Rainfall Measuring Mission (TRMM) sensor package. Journal ofAtmospheric and Oceanic Technology, 15, 809817.

Marques, M. C. M., & Joly, C. A. (2000). Seed germination and growth ofCalophyllum brasiliensie (Clusiaceae), a typical species of flooded forests.Acta Botanica Brasilica, 14(1), 113120.

Myers, N., Mittermeier, R. A., Mittermeier, C. G., da Fonseca, G. A. B., & Kent,J. (2000). Biodiversity hotspots for conservation priorities. Nature (London),403, 853858.

Myneni, R. B., Hoffman, S., Knyazikhin, Y., Privette, J. L., Glassy, J., Tian, Y.,et al. (2002). Global products of vegetation leaf area and fraction absorbedPAR from year one of MODIS data. Remote Sensing of Environment, 83,214231.

Nagendra, H. (2001). Using remote sensing to assess biodiversity. International

Journal of Remote Sensing, 22, 23772400.

http://dx.doi.org/10.1029/2001JD000975http://dx.doi.org/10.1029/2001JD000975


18/18

Nelson, B. W., Ferreira, C. A. C., da Silva, M. F., & Kawasaki, M. L. (1990).Endemism centres, refugia and botanical collection density in BrazilianAmazonia. Nature, 345, 714716.

Nix, H. A. (1986). A biogeographic analysis of Australian elapid snakes. In:Atlas of Australian ElapidSnakes (pp. 415).Canberra: Bureau Flora Fauna.

Peng, C. (2000). From static biogeographical model to dynamic globalvegetation model: A global perspective on modeling vegetation dynamics.

Ecological Modelling, 135, 3354.Phillips, S. (2005). A brief tutorial on Maxent. AT& T Research (from http://www.cs.princeton.edu/~schapire/maxent/tutorial/tutorial.doc ).

Phillips, S., Anderson, R. P., & Schapire, R. E. (2005). Maximum entropymodelling of species geographic distributions. Ecological Modelling, 190,231259.

Pimm, S. L. (1991). The balance of nature? Ecological issues in conservation ofspecies and communities. Chicago: The University of Chicago Press.

Prates-Clark, C. D. C., Saatchi, S. S., & Agosti, D. (2008). Predictinggeographical distribution models of high-value timber trees in the AmazonBasin using remotely sensed data. Ecological Modelling, 211(3-4),309323.

Rahbek, C., & Graves, G. R. (2001). Multiscale assessment of patterns of avianspecies richness. Proceedings of the National Academy of Sciences of theUnited States of America, 98, 4534 /4539.

Record, S. J., & Hess, R. W. (1943). Timbers of the new world. New Haven(USA): Yale Univ. Press.

Rodriguez, W. A. (1972). A ucuuba de vrzea e suas aplicaes. ActaAmazonica, 2(2), 2947.

Saatchi, S., Houghton, R., Avala, R., Yu, Y., & Soares, J. -V. (2007). Spatialdistribution of live aboveground biomass in Amazon Basin. Global Change

Biology, 13, 816837.Schongart, J., Piedade, M. T. F., Ludwigshausen, S., Horna, V., & Worbes, M.

(2002). Phenology and stem growth periodicity of tree species in Amazonianflood plain forests. Journal of Tropical Ecology, 18, 581597.

Smith, T. B., Schneider, C. J., & Holder, K. (2001). Refugial isolation versus

ecological gradients: Testing alternative mechanisms of evolutionary diver-gence in four rainforest vertebrates. Genetica (Dordrecht), 112113, 383398.Swets, J. A. (1988). Measuring the accuracy of diagnostic systems. Science,

240, 12851293.ter Steege, H., Pitman, N., Philips, O., Chave, J., Sabatier, D., Duque, A., et al.

(2006). Continental-scale patterns of canopy tree composition and functionacross Amazonia. Nature, 443(28), 444447.

ter Steege, H., Pitman, N., Sabatier, D., Castellanos, H., Van Der Hout, P., Daly,D. C., et al. (2003). A spatial model of tree alpha-diversity and tree densityfor the Amazon. Biodiversity and Conservation, 12, 22552277.

Tuomisto, H., Ruokolainen, K., Kalliola, R., Linna, A., Danjoy, W., & Rodriguez,Z. (1995). Dissection Amazonian biodiversity. Science, 269, 6366.

Turner, W., Spector, S., Gardiner, N., Fladeland, M., Sterling, E., & Steininger,M. (2003). Remote sensing for biodiversity science and conservation.Trends in Ecology and Evolution, 18, 306314.

Whittaker, R. H. (1972). Evolution and measurement of species diversity.Taxon, 21, 213251.

Woodward, F. I. (1987). Climate and plant distribution. Cambridge UniversityPress 190 pp.

http://www.cs.princeton.edu/~schapire/maxent/tutorial/tutorial.dochttp://www.cs.princeton.edu/~schapire/maxent/tutorial/tutorial.dochttp://www.cs.princeton.edu/~schapire/maxent/tutorial/tutorial.dochttp://www.cs.princeton.edu/~schapire/maxent/tutorial/tutorial.doc

Modelos de Distribucion y Divers Id Ad de Arboles en La Amazonia Usando Sensores Remotos

Documents

Transcript of Modelos de Distribucion y Divers Id Ad de Arboles en La Amazonia Usando Sensores Remotos