Finding Fossils in New Ways: An Artificial NeuralNetwork Approach to Predicting the Location ofProductive Fossil LocalitiesROBERT ANEMONE, CHARLES EMERSON, AND GLENN CONROY

Chance and serendipity have long played a role in the location of productivefossil localities by vertebrate paleontologists and paleoanthropologists. We offeran alternative approach, informed by methods borrowed from the geographic in-formation sciences and using recent advances in computer science, to moreefficiently predict where fossil localities might be found. Our model uses an artifi-cial neural network (ANN) that is trained to recognize the spectral characteristicsof known productive localities and other land cover classes, such as forest, wet-lands, and scrubland, within a study area based on the analysis of remotelysensed (RS) imagery. Using these spectral signatures, the model then classifiesother pixels throughout the study area. The results of the neural network classifi-cation can be examined and further manipulated within a geographic informationsystems (GIS) software package. While we have developed and tested thismodel on fossil mammal localities in deposits of Paleocene and Eocene age inthe Great Divide Basin of southwestern Wyoming, a similar analytical approachcan be easily applied to fossil-bearing sedimentary deposits of any age in anypart of the world. We suggest that new analytical tools and methods of the geo-graphic sciences, including remote sensing and geographic information systems,are poised to greatly enrich paleoanthropological investigations, and that thesenew methods should be embraced by field workers in the search for, and geo-spatial analysis of, fossil primates and hominins.

Vertebrate paleontologists andpaleoanthropologists search for fos-sils today in very nearly the sameways that the pioneers in our fieldhave since the late nineteenth cen-tury. We often follow in the tracks ofgeologists or other paleontologists,read the geological and paleontologi-cal literature to determine whereothers have collected the kinds of fos-sils we are interested in, and study ge-ological and topographic maps tolearn where relevant deposits may beexposed at the earth’s surface. Webring field crews of colleagues andstudents to distant locales and walkmany kilometers with eyes trained onthe ground in search of the rare andelusive evidence of fossil riches. Many,perhaps most, new fossil localities areliterally stumbled upon by geologists,paleontologists, and paleoanthropolo-gists, albeit typically as a result ofmonths to years of determined search-ing. Serendipitous discoveries of fa-mous hominins by equally famouspaleoanthropologists are numerousand legendary in our field. Someobvious examples include Don Johan-son’s recovery of Lucy at Hadar1 andLee Berger’s recent finds at Malapa.2,3

While these methods have certainlyled to much successful field work, therole of chance and good luck in deter-mining success or failure in paleoan-thropological field work might bereduced by applying new techniquesand analytical approaches to the prob-lem of predicting where fossils arelikely to be found. If successful, suchapproaches have great potential toincrease the efficiency of paleontologi-cal field work while, at the same time,

ARTICLE

Robert Anemone is Professor of Anthropology at Western Michigan University. Havingtrained in primate functional morphology and vertebrate paleontology at the University ofWashington, he has conducted field and museum research on the anatomy and life historyof living and fossil primates, and mammalian evolution in North America, Europe and Africa.He has been leading field crews to the Great Divide Basin of southwestern Wyoming since1994 in order to collect Paleocene and Eocene primates and other mammals. He is theauthor of Race and Human Diversity: A Biocultural Approach (Prentice-Hall, 2010). E-mail:robert.anemone@wmich.eduCharles Emerson is Associate Professor of Geography at Western Michigan University. Aspecialist in the theory and practice of quantitative spatial analysis, he received a Ph.D. inGeography at the University of Iowa in 1996 and has recently been examining grasslanddegradation in Inner Mongolia. Dr. Emerson is one of the developers of the Image Charac-terization and Modeling System, a software package that measures the fractal dimensionsof remotely sensed images. E-mail: charles.emerson@wmich.eduGlenn Conroy is Professor of Anatomy and Anthropology at Washington University Schoolof Medicine, St. Louis, MO. He is known for his work on nearly every epoch of primate andhuman fossil history. He is author of Primate Evolution and Reconstructing Human Origins,both published by W. W. Norton. E-mail: conroyg@pcg.wustl.edu

Key words: predictive models; GIS; remote sensing; paleoanthropology

VVC 2011 Wiley-Liss, Inc.DOI 10.1002/evan20324Published online in Wiley Online Library (wileyonlinelibrary.com).

Evolutionary Anthropology 20:169–180 (2011)

decreasing the sometimes staggeringlogistical costs such field excursionsoften incur.4

The use of GIS and RS to guidepaleoanthropological field work andto provide new analytical approachesto locating hominin-bearing depositslags well behind the use of thesemethods by archeologists and geolo-gists. Archeologists were early adopt-ers of the tools and techniques of thegeographic sciences5–7 and havemade major advances in using GISand RS to locate new sites8 oruncover previously hidden featuresat known sites,9 analyze the spatialrelations of artifacts within sites,10,11

and develop predictive models forsite location.12 In vertebrate paleontol-ogy, Oheim13 developed an innovativeGIS-based predictive model for locatingdinosaur-bearing deposits in the TwoMedicine Formation of Montana basedon four simple variables: geology, eleva-tion, vegetation cover, and distance toroads. Field testing of the model indi-cated a significant correlation betweenfossil density and the predicted likeli-hood of fossils. Vertebrate paleontolo-gists have also used GIS to analyzeaspects of taphonomy at individual fos-sils sites,14 as well as sampling and di-versity across entire basins.15 As awhole, however, there is little evidenceof sophisticated uses of GIS and RS inthe literature of paleoanthropology orvertebrate paleontology.4

The most successful use of RS inthe search for hominin fossils can befound in the work of Berhane Asfaw,Tim White, and colleagues on thePaleoanthropological Inventory ofEthiopia project.16 Begun in 1988,this project has used a wide range ofRS imagery of largely unexploredareas in the Afar Depression and theMain Ethiopian Rift to identify lith-ologies and geological features ofpotential paleontological and paleo-anthropological interest. The analysisof RS imagery led directly to theidentification of promising depositsat Fejej and Kesem-Kebena. Laterfield expeditions to these areas suc-cessfully identified significant paleo-anthropological resources, including3.7-my-old dental remains attributedto A. afarensis at Fejej17,18 and adiverse Pliocene vertebrate fauna anda rich Acheulean assemblage from

Kesem-Kebena.19,20 While the Paleo-anthropological Inventory of Ethiopiademonstrated a clear ‘‘proof-of-princi-ple’’ for the use of remotely sensed im-agery in locating deposits of paleoan-thropological significance, few paleo-anthropologists have followed its lead.One notable exception is the recentuse of high-resolution satellite im-agery by Njau and Hlusko to locatesmall patches of sedimentary depositsof paleoanthropological and archeo-logical interest in remote regions ofTanzania.21 When survey crews vis-ited these locations, 28 new fossil and/or archeological localities were identi-fied, providing another clear demon-stration of the utility of this approachto the search for fossil hominins andtheir archeological remains.

In some of our earlier work,22,23 wedemonstrated how geospatial infor-mation from paleontological and/orpaleoanthropological investigationscan be analyzed and shared usingGIS databases in combination withGoogle Earth imagery. We encour-aged the paleoanthropological com-munity to embrace these new toolsand techniques from the geographicsciences. A recent symposium organ-ized by Denne Reed and Chris Cam-pisano on the use of bio- and geo-informatic databases in paleoanthro-pology held at the AAPA meetings inApril 2011 suggests much recent pro-gress and a bright future for newtechniques in the analysis, presenta-tion, and storing of geospatial data inpaleoanthropology.

At the heart of our currentresearch lies a simple question about

the taphonomic nature and othercharacteristics of productive fossillocalities that distinguish them fromdeposits that fail to produce fossils.While the question may be simple,the answer is certainly complex, andthe subject of much active researchby paleontologists, geologists, andpaleoanthropologists. A multitude ofdeterministic factors play roles inthe formation of fossil deposits.These include geological factors,such as tectonic, erosional, and geo-morphological ones, and environ-mental factors, such as depositionalenvironments and climatic ones, aswell as random factors that mayhave influenced the death, preserva-tion, and exposure of individual fos-sil organisms.24,25 The approachesdeveloped in Ethiopia by Asfaw andcolleagues,16,17,19,26 and in Tanzaniaby Njau and Hlusko21 both involvevisual identification on RS imageryof geological deposits and featuresthat are thought to be of the correctage and lithology to bear theremains of fossil hominins and theirtool kits. Our approach to this prob-lem is different, being neitherstrictly taphonomic nor geologicalin nature. Since detailed geologicalmaps are readily available for ourresearch area in Wyoming, we al-ready know where some of the fos-sil-bearing units are exposed.Instead, the problem we addresshere is determining where to focusthe efforts of collecting surveyswithin sedimentary deposits that areextensively exposed over an area ofsome tens of thousands of squarekilometers. Therefore, we do notseek to determine, for example,whether the mammalian fossils inthe Great Divide Basin are preferen-tially found in certain geological for-mations, depositional environments,or sediment types, for we alreadyknow the answers to these ques-tions. Rather, our model seeks todetermine if the spectral signaturesof productive localities can be dis-tinguished from those of nonproduc-tive deposits through classificationof multi-spectral remotely sensedimagery using an analyticalapproach known as an artificial neu-ral network (ANN) and geospatialanalysis using GIS software.

At the heart of ourcurrent research lies asimple question aboutthe taphonomic natureand other characteristicsof productive fossillocalities that distinguishthem from deposits thatfail to produce fossils.

170 Anemone ARTICLE

DIGITAL IMAGE CLASSIFICATION

The launch of the Landsat 1 satel-lite in 1972 ushered in the age of dig-ital earth imaging. The SPOT serieswas developed by a French, Belgian,and Swedish consortium in the

1980s. Since then, many other gov-ernmental and private sector satel-lites, such as Ikonos, Orbview, andQuickbird, have been continuouslyproviding imagery of the Earth’s sur-face.27 The sensors aboard the satel-

lites vary in capability, but they gen-erally measure reflected or emittedelectromagnetic radiation in wave-lengths ranging from the ultravioletto the infrared. The continuousrange of wavelengths are segmented

into discrete groups, called bands, sothat a sensor will record the amountof radiation in the blue, green, red,and near-infrared bands, for exam-ple. The sensor converts the meas-

ured amount of radiation comingfrom a patch on the earth surface toa series of digital numbers, which,after processing and correction, forma set of brightness values for a par-ticular area, called a picture element,

or pixel.28 Pixels are arranged inrows and columns to form an imagein which each pixel has values corre-sponding to the intensity of reflectedor emitted radiation in the spectralbands measured by the sensor.29

Different earth surface materialshave different reflectance properties,so that, for instance, a material thatreflects more green than blue or redradiation (and much more nearinfrared radiation, which is outsidethe human eye’s sensitivity) wouldmost likely be green vegetation.Areas with different types of landcover, the predominant material cov-ering the surface, would thereforehave characteristic spectral signa-tures, or combinations of typical re-flectance values observed for theset of spectral bands measured by asensor.

Differences between these signa-tures are often subtle. The problemof identifying a unique spectral sig-nature for a particular land-covertype involves a multidimensionalanalysis of each pixel, with thedimensions corresponding to thespectral bands measured by a given

sensor. The process of taking thecontinuous range of brightnessmeasurements in the various bandsand grouping them together in nomi-nal categories is known as imageclassification.30,31 Unsupervised clas-sification uses statistical clustering ofthe digital numbers corresponding tothe brightness measured in the vari-ous bands to establish spectralclasses, which are then related in apost hoc fashion to the appropriateground cover. Supervised classifica-tion uses the locations of known landcovers to derive the measured bright-ness values for these training sites todevelop a spectral signature for thatland cover. The classification algo-rithm then finds other pixels thatmatch this signature.There are many different methods

for performing image classification,such as the ISODATA and k-meansunsupervised methods and the paral-lelpiped, maximum likelihood, andneural network approaches to per-forming supervised classification.31

These differ in speed and complexity,and each involves different assump-tions about the statistical character-istics of the image. In this investiga-tion, the neural network method waschosen because it is relatively robustand can handle nonnormally distrib-uted data.

THE NEURAL NETWORK MODEL

Artificial neural networks imitatethe complex biological networks ofneurons, axons, dendrites and syn-apses in the brain and can be usedto solve a variety of scientific prob-lems that may involve classification,prediction, and decision-making.32

The back-propagation multilayer per-ceptron model33 that is used hereconsists of input nodes, one or morehidden node layers, and a layer ofoutput nodes (Fig. 1). In a neuralnetwork, the input nodes are gener-ally the spectral bands correspondingto a particular training pixel; the out-put nodes are the land-coverclasses.34 The nodes in each layer areconnected to nodes in the otherlayers by weights. The input thateach node receives, such as a seriesof brightness values for all bands, isa weighted sum the output of which

Figure 1. A simplified version of the ANN model. For our analysis, the input layers includeLandsat 7 bands 1-5 and 7 from the ETMþ sensor; the outputs include the 10 spectralland-cover classes.

ARTICLE Finding Fossils in New Ways 171

‘‘fires’’ if the sum exceeds a thresholdvalue as defined by an activationfunction. The process is analogous tothe ways in which neural inputs aresummed at synapses, where thresh-olds determine if excitatory or inhibi-tory impulses are passed down theaxon to the next neural synapse. Aninitially random pattern of weights isfed forward through the network ofinterconnected nodes and the outputis compared to the known land-coverclasses. Before another iterationthrough the network is performed,errors are back-propagated throughthe network of nodes and adjust-ments to the weights are made. Thepatterns are fed forward and errorsare back-propagated iteratively untilthe error reaches an acceptable anduser-specified magnitude. In thisfashion, the neural network istrained to recognize the patterns ofspectral reflectance values; that is,the spectral signatures that corre-spond to the land covers of interest.In our model,87 the input nodes

are six visible and infrared bands ofelectromagnetic radiation from theLandsat 7 Enhanced Thematic Map-per Plus sensor (ETMþ) obtainedfrom the USGS EROS Data Center35

based on images taken on August 8and September 2, 2002. These dateswere chosen because the imageswere cloud-free and were obtainedbefore the partial failure of theETMþ sensor in 2003. The Septem-ber image was projected to the Uni-versal Transverse Mercator Zone 12,World Geodetic System 84 coordi-nate system to match the Augustimage. Each scene was calibratedusing the instrument gain values,and the elevation and azimuth of thesun at the time of acquisition to yieldpercent reflectance images. The pan-chromatic bands (band 8) were usedto increase the spatial resolutionfrom the original 28.5 m to 14.25 musing a principal componentsapproach in which the first principalcomponent of the 28.5 m visible andreflective infrared bands (bands 1 – 5and 7) was replaced by the high-reso-lution panchromatic band and backtransformed to yield a pan-enhancedmultispectral image. The imageswere then joined together to form acontinuous image of the GreatDivide Basin.

The image was classified using theANN function in the ENVITM imageanalysis program. The output nodesare represented by ten spectral land-cover classes (Table 1), each ofwhich exhibits similar responses inthe six Landsat bands. In order forthe neural network to be trained torecognize the multivariate spectralcharacteristics of the desired out-puts, the user must first identify asample of representative pixels fromeach output type in the image to beanalyzed. The ANN algorithm thendetermines the spectral characteris-tics of each output class andattempts to identify this same spec-tral signature in other parts of thebasin. As it proceeds, each pixel in

the image is classified into one of theoutput classes, with an associatedprobability that is determined byhow closely the spectral signature ofthe output pixels matches those ofthe training pixels for that land cov-erage class.GPS coordinates of fossiliferous

localities were gathered during sev-eral field seasons between 1994 and2010.23,36 The Landsat image of thebasin was segmented into spectrallyhomogeneous polygons around eachlocality by growing a region of inter-est, using a 0.5 standard deviationaround the mean spectral responsesfor each band. This allows the train-ing site to expand beyond the spotGPS measurement to include an areahaving a relatively homogeneousspectral response pattern.28 In theGreat Divide Basin, fossils are gener-ally found in eroded sandstones orsoft mudstones overlain by a protec-tive cap of sandstone. Productivelocalities are generally in areas withsandstone outcrops that have theappropriate near-surface geology.Thus, the spectral signature for thelocalities land-cover class is actuallythe complex mix of brightness valuesthat result from the sunlit andshadow areas of the outcrops, theeroded mudstone and sand at thebase of the outcrops, and any othermaterials that may be found withinthe roughly 203 m2 area (14.25 msquared) corresponding to a trainingpixel.Our model requires us to train the

classifier to recognize other land-cover types that are not of interest inorder to minimize confusion withthe locality class. Training sites forthree types of soil spectral classesbased on appearance (red, light, anddark soils), forests (mainly located athigh elevations at the margin of the

TABLE 1. Spectral and InformationalLand-Cover Classes

Spectral ClassInformational

Class

Localities LocalitiesWhite Soil/Rock BarrenDark Soil/Rock BarrenRed Soil/Rock BarrenGrassland ScrublandSagebrush ScrublandCropland ScrublandMixed Scrub ScrublandForest ForestHerbaceous Wetland Wetland

TABLE 2. Classification Results of the Artificial Neural Network

Land Cover Class Barren Forest Localities Scrub Wetland Total User’s Accuracy

Barren 2411 19 1035 114 20 3599 66.99%Forest 0 2402 23 193 18 2636 91.12%Localities 35 6 5525 22 1 5589 98.85%Scrub 12 238 408 1429 50 2137 66.87%Wetland 0 5 0 6 11 22 50.00%Total 2458 2670 6991 1764 100 13983Producer’s Accuracy 98.09% 89.96% 79.03% 81.01% 11.00%

172 Anemone ARTICLE

basin), and four types of vegetatedscrubland (sagebrush, grassland,brushland, and mixed) were obtainedusing visual interpretation of high-re-solution aerial photos from GoogleEarth. The National Land CoverDataset37 for this area showed sev-eral wetlands in low-lying areas,training or accuracy assessment siteswere also obtained for these loca-tions. We used roughly 75% of theidentified pixels for each land-cover-age class to train the neural network,holding back 25% for post hoc accu-racy testing of the resulting classifi-cation (Table 2).With the six input nodes, two hid-

den layers, and ten output spectralclasses, the model converged after236 iterations to reach a user-speci-fied root mean square (RMS) errorof 0.1. The results of the neural net-

work can then be displayed graphi-cally (Fig. 2) and analyzed in a vari-ety of ways within a GIS to visualizethe predicted land-coverage class ofeach pixel in the image. Since we areprimarily interested in how well themodel predicts the identity of local-ities, for the most part we ignoredthe predicted locations of the otherland-cover classes, focusing solely onthe locality class of outputs.

We recognize not only that termslike area, site, or locality have beenused in many different ways by pale-ontologists and paleoanthropologists,but the confusion that can resultfrom these differing uses. For ourpurposes in this paper, a productivelocality can be defined as a locationon the landscape where a concentra-tion of mammalian fossils, typicallynumbering in the low tens to low

thousands, has been recovered as aresult of surface collecting or quarry-ing. Productive localities in the GreatDivide Basin range from hundreds tothousands of square meters in arealextent and, while they may reflectsome degree of time-averaging, aregenerally considered as representinga single time horizon. They tend tobe derived from a single sedimentarylithology, typically mudstones orsandstones, and are separated fromother productive localities by unfos-siliferous areas of exposed or unex-posed sediments.

GEOSPATIAL ANALYSIS WITH GIS

Our GIS analysis begins with thecreation of a digital elevation model(DEM) of the entire Great Divide Ba-sin (Fig. 2). The DEM is essentially arasterized height map in which Xand Y coordinates represent the geo-graphic coordinates and Z representselevation. We obtained the data forour DEM from the National Eleva-tion Database (NED)38 and created itusing ArcGIS software. Using Arc-GIS, we resampled the NED datafrom an original pixel size of 10 m to14.25 m in order to match the pixelsize on our Landsat images. We alsoderived a hillshade layer from theDEM. This basically creates a shadedrelief image with shadows resultingfrom an assumed sun in the north-west, thus giving the viewer an ideaof the topography. This hillshadeimage forms the backdrop for thegraphical outputs of our ANN.

IMAGE CLASSIFICATION ANDRESULTS

Table 2 presents the results ofpost-hoc accuracy testing for theANN’s classification of the GreatDivide Basin. These results measurehow well the neural network modelidentified the 25% of known pixels ineach of the five land-cover classesthat were held back for this purposeduring training. For accuracy assess-ment purposes, the three spectrallydistinct bare rock or soil classes (red,light, and dark brown) were com-bined into a single ‘‘Barren’’ informa-tional class; four different types of

Figure 2. A. The DEM of the Great Divide Basin is a simple rasterized XYZ coordinate mapin which the Z dimension represents altitude. B. Applying a hillshade is an arbitrary pro-cess, easily accomplished within GIS software packages, that greatly improves theappearance of a DEM.

ARTICLE Finding Fossils in New Ways 173

vegetative cover (sagebrush, grass,cropland, and mixed scrub) werecombined into a single ‘‘Scrubland’’

informational class, yielding fiveinformational classes of land coverfor comparison (Table 1). Vertical

columns in accuracy assessmentresults (Table 2) represent groundtruth,39 meaning that the total num-ber of pixels known to exist in eachclass can be found running from leftto right near the bottom of the Table(for example, 2,458 pixels wereknown to be barren, 100 were knownto be wetland, and so on). Horizontalrows represent the classified pixelsfor each land-cover class; the totalnumber of pixels predicted to fall ineach class can be found from top tobottom along the right side of the ta-ble (for example, 3,599 pixels werepredicted to be barren and 22 werepredicted to be wetland). The overallresults are encouraging: the modelcorrectly classified 84.21% of the pix-els in all land-cover classes, with aKappa coefficient (a more conserva-tive measure of correctly classifiedpixels) of 77.44%. For the localitiescolumn, the model correctly classi-fied 5,525 of the 6,991 actual localitypixels to yield a ‘‘producer’s accu-racy’’ (a measure of errors of omis-sion in which pixels that are actuallylocalities, for example, are incor-rectly classified as something else) of79.03%. Of the 5,589 pixels predictedto represent localities, 5,525 wereknown to be localities, for a ‘‘user’saccuracy’’ (a measure of errors ofcommission in which pixels thatwere predicted to be localitiesactually belonged to another land-cover class) of 98.85%. Each pre-dicted pixel comes with an associ-ated probability, which is not repre-sented in these tabular data. The bestway to explore this is through an ex-amination of the graphical output ofthe ANN, known as the classifiedimage.The classified image (Fig. 3)

presents the predicted identity of allpixels in the entire basin for eachof the five output land-coverclasses, regardless of the probabilityassociated with each pixel. It formsthe starting point for all GIS-basedanalyses. The next step in the anal-ysis is to apply some reasonableconstraints to the classified imageto create a ‘‘Rule Image’’ for thosepixels for which highest classifiedprobability suggests they may repre-sent localities. We used a cut pointof 95% probability associated with

Figure 3. Classified Image. This image of the Great Divide Basin classifies each pixel intoone of the five land-cover classes used in this study. All red pixels are predicted to repre-sent localities, but their associated probability will vary from low to high.

Figure 4. Rule Image. This image of the Great Divide Basin includes in red those pixels thathad a >95% of belonging to the locality class, and had a slope >5%. This represents ourcurrent best estimate of parts of the basin that may include localities with high probabil-ity, and high priority for ground truthing in upcoming field seasons.

174 Anemone ARTICLE

the classification as localities inorder to focus only on those pixelsthat were predicted to be localitieswith a high degree of confidence.We also constrained the predictedpixels to those with a slope ofgreater than 5% in order to ignoreareas of little or no vertical reliefthat may spectrally resemble local-ities but lack active erosional surfa-ces where fossils tend to be found.The rule image is then displayed asa layer on top of a hillshaded DEMof the entire GDB. The results canbe seen in Figure 4, where the redpixels classified as localities can beseen clustered in several parts ofthe GDB.

TESTING THE MODEL IN THEBISON BASIN

Directly to the north of the centralpart of the Great Divide Basin lies asmall isolated area containing sedi-mentary deposits of the Fort UnionFormation known as the Bison Ba-

sin (Fig. 5). The existence of fossilmammals of Tiffanian age (middleto late Paleocene) has been knownthere since the 1950s as a result ofearly work by USGS geologists and,particularly, the Smithsonian Insti-tution paleontologist C. L. Gazin.40

During the past ten years, fieldcrews from the Carnegie Museum ofNatural History under the directionof K. C. Beard have returned to theBison Basin to work at its threemost productive localities: WestEnd, Saddle, and Ledge. The geo-graphic and chronological proximityof the Bison Basin vertebrate local-ities to those in the Great Divide Ba-sin suggests that they might allowan interesting test of our predictivemodel. We view this test of ourmodel as a conservative one sincewe trained our ANN to recognizefossil-bearing sediments of theWasatch Formation of Eocene agein the Great Divide Basin. The spec-tral signatures of fossil-bearing unitsin the older Fort Union Formation

might be somewhat different owingto the different lithologies and faciesin that rock unit.The goal of this test was to deter-

mine how well our ANN could pre-dict the presence of fossil-bearingdeposits in the vicinity of the threeknown and productive localities inthe Bison Basin. In effect, we addedthe Bison Basin to our clipped imageand DEM of the Great Divide Basinand ran the ANN classification of thecombined basins in the same manneras before on this slightly larger area.We again generated a rule image forthe output class locality, includingonly those pixels with associatedprobability of greater than 95% andhaving a slope greater than 5%. Theresulting image (Fig. 5) indicates theareas within the Bison Basin whereour model predicts fossil mammalswould be located with the highestprobability. When we overlay the ge-ographic coordinates of the threeproductive Bison Basin localities(supplied to RA by K. C. Beard) uponthis image, the results of our predic-tive model were confirmed by indi-cating not only the general areawhere fossils have been found, butalso the specific locations of thethree known localities (Fig. 5). Wesuggest that an experienced paleon-tological field crew’s chances of ‘‘dis-covering’’ these three productivelocalities on the basis of the outputof our predictive model would begreatly increased compared toanother, equally experienced crewsurveying the exposed deposits intraditional fashion, without the bene-fits of the predictions generated byour model.

CONCLUSION

Our research is a first attempt todevelop a predictive model to aidpaleoanthropologists and vertebratepaleontologists in the critical activ-ity of fossil exploration. We trainedan artificial neural network to rec-ognize ten different categories ofland cover, including known local-ities, in our study area in the GreatDivide Basin of Wyoming. Workingwith satellite images derived fromthe ETMþ sensor on Landsat 7, theANN determined those areas having

Figure 5. Bison Basin. The predictive model that was developed for the Great Divide Basinselected potential localities just north of the GDB in the Bison Basin. Three known produc-tive localities in blue closely match potentially productive localities identified by the ANNmodel. The inset on the upper right side provides a locator map indicating the position ofthe Bison Basin as a red box just north of the central part of the Great Divide Basin.

ARTICLE Finding Fossils in New Ways 175

a high potential of being fossilifer-ous, due to the similarity of theirspectral signatures to those ofknown fossil localities. The resultsof our model’s predictions of thelocation of fossil-bearing localitiesin the nearby Bison Basin areextremely encouraging.

Paleoanthropology has for manyyears been an interdisciplinary fieldin which the tools and techniques ofthe natural sciences, typicallyapplied in collaboration with geolo-gists, chemists and physicists, haveenriched our analyses of fossil pri-mates and hominins. We suggest

that the geospatial sciences haveearned a place in the paleoanthropo-logical tool kit, and that twenty-firstcentury research in paleoanthropol-ogy must increasingly rely on thekinds of sophisticated spatial analy-ses that can only come from collabo-rations with our colleagues in the

Box 1: The Great Divide Basin

The Great Divide Basin (GDB) ofsouthwestern Wyoming is one ofthe many large structural and sedi-mentary basins in the Rocky Moun-tain region of the American Westthat are well known to vertebratepaleontologists.47,48 The namecomes from the fact that the Conti-nental Divide actually splits andencircles the basin, so that whendriving Interstate 80 across south-ern Wyoming one crosses the Con-tinental Divide twice between Lara-mie and Rock Springs. The GDBcontains thousands of meters offossiliferous sedimentary rock ofCretaceous to Eocene age that arevariably exposed throughout itsapproximately 10,000 square kilo-meters.49,50 The impetus for theresearch reported here stems fromthe difficulties of determiningwhere one’s efforts might be bestapplied toward the goal of findingfossils in such a large geographicspace.The first paleontologist to work

in the GDB, in the 1950s and1960s, was the Smithsonian Institu-tion’s C. L. Gazin. A series of Gaz-in’s publications described a hand-ful of latest Paleocene and earliestEocene localities scattered throughthe basin.51–54 The specimens andlocalities Gazin described were suf-ficient to allow him to fill in thepaleontological blank space that theGDB had been before his work, butcompared to the richer and moreproductive basins in the northernand southern Rocky Mountainregions (for example, the Bighornand Wind River Basins in the northand the San Juan and Uinta Basinsin the south), the GDB languishedin obscurity for decades. Beginning

in 1994, the senior author beganleading annual field crews to theGreat Divide Basin with the dualintentions of collecting new fossilsfrom Gazin’s localities and identi-fying new localities in this under-studied basin. After approximately15 summer field seasons in theGDB, we have succeeded in bothof our original goals.36 Followingin Gazin’s footsteps, we have col-lected, identified, and cataloguednearly 10,000 mammalian fossilsfrom approximately 100 localities.We succeeded in locating severalof Gazin’s named localities and

substantially increasing the mam-malian fauna from each. In 2009we located one of the richestWasatchian (early Eocene) mam-mal localities in North Americafrom which we recovered morethan 400 mammalian jaws andapproximately 5,000 isolated teethand postcranial elements.55 Therealization that we located thissuperb new locality purely bychance challenges us in our workto develop better approaches topredicting where productive fossil-bearing localities might be identi-fied in the future.

Box 1. The Great Divide Basin lies astride the Continental Divide in southwestern Wyom-ing. It is one of several large sedimentary basins in the Rocky Mountain West with exten-sive fossil-bearing deposits spanning the Cretaceous through the Eocene.

176 Anemone ARTICLE

geographical and geospatial scien-ces. We envision a paleoanthropol-ogy that is critically engaged in thespatial analysis of the raw materialsof our science at a number of hier-archical levels, ranging from thescatter of bones and artifacts atindividual sites to the distribution offossil sites within landscapes,between sedimentary basins, andacross continents. While we recog-

nize that the use of hand-held GPSreceivers has become ubiquitousamong field workers in manybranches of biological anthropology,we encourage our colleagues tomove beyond static and descriptiveuses of GPS and GIS that simply re-cord the geographic coordinates oflocalities or specimens. Spatial dataneed to be incorporated intodynamic models in order to answer

a variety of important questions ofpaleoanthropological interest.41,42

Some obvious candidates includemigration and colonization scenar-ios for past human populations,43

effects of climate change on presentand past distributions of primatetaxa and communities,44 and prob-lems of dispersals, origins, andextinctions of fossil primates andother mammals.45,46

Box 2: The Landsat Program

As a joint effort of the NationalAeronautics and Space Administra-tion (NASA) and the United StatesGeological Survey (USGS), theLandsat Program56 has acquiredspace-based images of the earth’ssurface continuously since thelaunch of Landsat 1 on July 23,1972. In so doing, it has created alongitudinal record of natural aswell as human-induced changesacross the global landscape. It alsoprovides a valuable resource to sci-entists working in many differentfields, including agriculture and for-estry, geology and geography, globalclimate change, and anthropology.Of the seven Landsat missions to

date, four have been decommis-sioned (Landsat 1-4), one failed atlaunch (Landsat 6), and two are stilloperational in 2011 (Landsat 5 and7). Landsat 5 is still collecting im-agery 27 years after its launch in1984, well in excess of its originallyplanned three-year mission. Advan-ces in scanning technology have ledto significant improvements in spa-tial resolution and spectral band-width across the generations ofLandsat satellites. For example, theMultispectral Scanner (MSS) car-ried aboard Landsat 1-3 achieved aspatial resolution of 80 m in fourspectral bands ranging from visiblegreen to near infrared wavelengths.The Thematic Mapper (TM) sensoron Landsat 4 and 5 achieved a spa-tial resolution of 30 m in 7 spectralbands. Landsat 7 carries theEnhanced Thematic Mapper Plus(ETMþ) sensor that collects multi-spectral images in 8 bands spanningthe visible to thermal infrared wave-

lengths in 185 km wide scenes witha spatial resolution of 28.5 m.Launched on April 15, 1999 andplaced in orbit at an altitude of 705km, Landsat 7 completes approxi-mately 14 orbits each day to acquirecomplete coverage of the Earth’ssurface every 16 days. The scan linecorrection device in the ETMþ sen-sor failed in 2003, thus limiting theutility of more recent Landsat 7 im-agery.Perhaps the most remarkable fea-

ture of the remotely sensed imageryprovided by the Landsat program isthat the entire archive of images isnow available over the Internet(http://glovis.usgs.gov/) at no chargeand with no restrictions to users.The consistency and completenessof the Landsat archive make it aninvaluable resource for scientistsinterested in detailing long-termchanges to the earth’s surface, aswell as aiding in relief efforts bydocumenting before and afterimages of areas that have experi-enced natural or man-made disas-ters. All in all, Landsat images con-

stitute an extremely valuable andunder-used resource for vertebratepaleontologists and paleoanthropol-ogists.Although there are other earth-

imaging satellites currently in orbit,the imagery they provide is gener-ally not free, and in some cases, isquite expensive. The Ikonos, Orb-View, and Quickbird satellites areprivately owned and provide im-agery with spatial resolution as highas 0.4 m. Several foreign govern-ments also operate remote sensingsatellites, including France (SPOT),India (IRS), and a joint Chinese-Brazilian program (CBERS). TheAqua and Terra satellites operatedby the U.S. National Aeronauticsand Space Administration contain anumber of sensors, including AS-TER, which is most similar to theEuropean SPOT series of satellites,and MODIS, which has pixel sizesranging from 250 m to 1000 m.29

The kind of analyses describedhere could also be performed usingimagery from any of these othersources.

BOX TABLE 1. Landsat 7 ETMþ Sensor Band Designations

Spectral Bands Use

1. Blue-green Bathymetric mapping: distinguishes soil from vegetation,deciduous from coniferous vegetation

2. Green Emphasizes peak vegetation, useful for assessing plant vigor3. Red Emphasizes cultural features4. Reflected IR Emphasizes biomass content and shorelines5. Reflected IR Discriminates moisture content of soil and vegetation;

penetrates thin clouds6. Thermal IR Useful for thermal mapping and estimated soil moisture7. Reflected IR Useful for mapping hydrothermally altered rocks associated

with mineral deposits8. Panchromatic Includes visible through infrared with 15-m resolution for

pansharpening multispectral images.

ARTICLE Finding Fossils in New Ways 177

Box 3. Other Uses of GIS and RS in Biological Anthropology

Biological anthropologists haveused innovative new methods andtechniques derived from the geo-graphic sciences to study primates intheir natural habitats, the relationshipbetween diet and dental morphologyand microwear, and the migrationsout of Africa of Plio-Pleistocene homi-nins.4 Green and Sussman57 demon-strated how satellite imagery couldreveal a multi-decadal record of defor-estation in the eastern forested zoneof Madagascar and highlighted theconservation implications of thistrend for the future of the island’smostly endemic fauna and flora.Smith, Horning, and Moore58 createda large GIS database to study habitatdegradation and deforestation inwestern Madagascar, and evaluatehow well the system of naturereserves was succeeding in safeguard-ing the local habitat and fauna.Recently, Irwin, Johnson, andWright59 used Landsat imagery andfield censuses of 10 lemur populationsto develop a multi-faceted geospatialanalysis of the current ranges, avail-able habitats, and threats to the con-tinued existence of Madagascar’slemuriform fauna, mostly as a resultof human activity. As a part of studiesof global and environmental change,the analysis of RS images within GISdatabases will continue to play a lead-ing role in documenting habitat deg-radation and its effects on fauna andflora in many parts of the world.The development of miniaturized

GPS-enabled collars and their use instudies of animal movement andranging behavior (that is, wildlife te-lemetry) has led to a revolution inthe ecological study of animal behav-ior.60–62 Dozens of different terres-trial and marine organisms havebeen successfully tracked and studiedby ecologists and animal behavio-rists, including sea turtles,63 Africanelephants,64 and penguins.65 A pilotstudy66 demonstrated the feasibilityof studying baboon ranging patternswith an automated, GPS-enabled col-lar in the open, savannah habitats ofAmboseli, Kenya. The use of suchdevices, and even hand-held GPSreceivers, to track primates in closedcanopy forest has not yet proven tobe entirely feasible due to the diffi-culty of obtaining the required GPSsatellite signals when the sky isobscured by dense foliage.67–69

Following Denne Reed’s70 initialuse of three-dimensional (3-D) imag-ing and GIS software to study toothmorphology in 1997, several workershave applied the tools of GIS to theanalysis of tooth morphology anddental microwear with the goal ofunderstanding dietary differencesamong a wide variety of living andfossil primates.4 Using differentapproaches to collect either land-mark data70 or three-dimensionalpoint clouds derived from 3-D scan-ners,71,72 a DEM of the tooth occlusalsurface can be generated and ana-lyzed using GIS software. PeterUngar of the University of Arkansas,along with a long list of collabora-tors, students, and post-docs, hasapplied the tools of geospatial analy-sis to the functional morphology ofteeth and patterns of dental micro-wear.73,74 These approaches havebeen widely used in assessing micro-wear on extant primates of knowndiet, and in using the patterns thatemerge to reconstruct the diet ofextinct primates, including hominins.One notable success of this approachwas the solution of the ‘‘worn toothconundrum,’’ allowing, for the firsttime, the study of how wear-relatedchanges to tooth occlusal surfacesinfluence dental function over anindividual’s lifetime.75 Ungar’s inno-vative work constitutes a majorimprovement on the older techniquesof counting scratches and pits onscanning electron micrographs todetermine whether an animal hasbeen fed primarily on hard or softobjects. His team has recently devel-oped a promising new approach thatuses the confocal microscope to gen-erate a 3-D model of tooth occlusalsurfaces that is then analyzed by ascale-sensitive fractal analysis76–79 toquantitatively characterize and com-pare patterns of microwear. Sincethis technique does not rely on land-mark data, it can be applied to teethat any wear stage, and since it doesnot involve subjective identificationof scratches and pits, it should pro-vide reproducible results with lowerror rates.76,78

Two other early adopters andinnovators in the application ofapproaches derived from the geo-graphic and imaging sciences to theanalysis of dental morphology,diet, and wear are the University of

Helsinki’s Jukka Jernvall and Mon-ash University’s Alistair Evans.80,81

Working independently and to-gether, these scholars and their col-leagues have advanced the science invarious ways, including use of thelaser confocal microscope to gener-ate DEMs of dental surfaces71 andthe development of homology-freecomparisons of dental morphologythat allow quantification of the com-plexity of occlusal morphology (thatis, orientation patch complexity, orOPC).82 A particularly interestingstudy scanned fossil horse teeth tostudy the transition to hypsodontyand grazing with increasing aridityin the European Miocene.83 Three-dimensional scans of tooth occlusalsurfaces were used to create DEMs;these were analyzed with GIS soft-ware to measure the slope of theenamel surfaces, which could berelated to degree of hypsodonty. Inthis way, the authors were able torelate evolutionary changes in dentalmorphology (hypsodonty) to globalclimate change (cooling and aridity)and diet (grazing).Recently, Doug Boyer has been

applying innovative approaches toanalysis of the 3-D shape of theteeth of living and fossil primatesand other mammals, the goal beingto improve our ability to determinedietary preferences of fossil mam-mals. Building on the earlier workof Ungar, Jernvall, and others,Boyer84 has compared extant arch-ontans with respect to the reliefindex of molar shape (a ratio oftooth crown 3-D area to 2-D areathat expresses the degree of hypso-donty and tooth shape complex-ity).74 The results suggested thatvariability in relief index is bettercorrelated with diet than with phy-logeny, and that it can distinguishamong primate insectivores, foli-vores, and frugivore/gumnivores.84

More recently, Boyer has scannedthe molar surfaces of three speciesof Plesiadapis to reveal interestingdifferences in both molar relief andcomplexity that may relate to intra-specific dietary differences.85 Boyerand colleagues86 have used a prom-ising new measure of dental surfacecomplexity known as Dirichlet nor-mal energy (DNE) in the analysis of3-D occlusal shape and diet inextant primates.86

ACKNOWLEDGMENTS

We thank John Fleagle for the op-portunity to contribute this paperand for his editorial suggestions, aswell as the comments of the reviewers,all of whom greatly helped to improvethe final version of this paper. Our fieldwork in the Great Divide Basin isunder federal permit 287-WY-PA95 (toRA) and administered by the WyomingBureau of Land Management. The per-manent repository for our fossil collec-tions is the Carnegie Museum ofNatural History in Pittsburgh, PA. Bi-son Basin field work by Chris Beard isalso permitted by the Wyoming Bureauof Land Management under federalpermit 042-WY-PA98. RA thanks Mr.Walt Worthy, the Office of the VicePresident for Research, and the FRA-CAA fund at Western Michigan Univer-sity for financial support. CE acknowl-edges the financial support of the LuciaHarrison Endowment at WesternMichigan University. We also thankour colleagues Justin Adams, AaronAddison, Chris Beard, Wendy Dirks,Brett Nachman, Richard Stucky, andJohn Van Regenmorter.

REFERENCES

1 Johanson D, Taieb M. 1976. Plio-Pleistocenehominid discoveries in Hadar, Ethiopia. Nature260:293–297.

2 Berger LR, de Ruiter DJ, Churchill SE,Schmid P, Carlson KJ, Dirks PHGM, Kibii JM.2010. Australopithecus sediba: a new species ofHomo-like Australopith from South Africa. Sci-ence 328:195–204.

3 Dirks PHGM, Kibii JM, Kuhn BF, SteiningerC, Churchill SE, Kramers JD, Pickering R, {JF:all aus needed.] Farber DL, Meriaux AS, Her-ries AIR and others. 2010. Geological settingand sge of Australopithecus sediba from South-ern Africa. Science 328:205–208.

4 Anemone R, Conroy G, Emerson C. n.d. GISand paleoanthropology: incorporating newapproaches from the geospatial sciences in theanalysis of primate and human evolution.Yearbk Phys Anthropol. In press.

5 Kvamme K. 1999. Recent directions anddevelopments in geographic information sys-tems. J Archaeol Res 7:153–201.

6 McCoy M, Ladefoged T. 2009. New develop-ments in the use of spatial technology inarchaeology. J Archaeol Res 17:263–295.

7 Sever T, Wiseman J. 1985. Conference onremote sensing: potential for the future. NASAEarth Resources Laboratory, NSTL, MS.

8 Johnson J, Madry S, Sever T. 1988. Remotesensing and GIS analysis in large scale surveydesign in north Mississippi. SoutheasternArchaeology 7:124–131.

9 Chase A, Chase D, Weishampel J, Drake J,Shrestha R, Slatton K, Awe J, Carter W. 2011.Airborne LiDAR, archaeology, and the ancient

Maya landscape at Caracol, Belize. J ArchaeolSci 38:387–398.

10 Nigro J, Limp W, Kvamme K, de Ruiter D,Berger L. 2002. The creation and potentialapplications of a 3-dimensional GIS for theearly hominin site of Swartkrans, South Africa.In: Burenhult G, Arvidsson J, editors. Archaeolog-ical informatics: pushing the envelope CAA2001.Proceedings of the 29th Conference, Gotland, April2001 ed. Oxford: Archaeopress. p113–124.

11 Nigro J, Ungar P, de Ruiter D, LR B. 2003.Developing a geographic information system(GIS) for mapping and analysing fossil depositsat Swartkrans, Gauteng Province, South Africa.J Archaeol Sci 30:317–324.

12 Mehrer M, Westcott K, editors. 2006. GISand archaeological site location modeling. BocaRaton, FL: CRC Press.

13 Oheim KB. 2007. Fossil site prediction usinggeographic information systems (GIS) and suit-ability analysis: The Two Medicine Formation,MT, a test case. Palaeogeogr, Palaeoclimatol,Palaeoecol 251:354–365.

14 Jennings DS, Hasiotis ST. 2006. Tapho-nomic analysis of a dinosaur feeding site usinggeographic information systems (GIS), Morri-son formation, southern Bighorn Basin, Wyom-ing, USA. Palaios 21:480–492.

15 Chew A, Oheim K. 2009. Using GIS to deter-mine the effects of two common taphonomicbiases on vertebrate fossil assemblages. Palaios24:367–376.

16 Asfaw B, Ebinger C, Harding D, White T,WoldeGabriel G. 1990. Space-based imagery inpaleoanthropological research: an Ethiopianexample. Natl Geo Res 6:418–434.

17 Asfaw B, Beyene Y, Semaw S, Suwa G,White T, WoldeGabriel G. 1991. Fejej: a newpaleoanthropological research area in Ethiopia.J Hum Evol 21:137–143.

18 Fleagle JG, Rasmussen DT, Yirga S, BownTM, Grine FE. 1991. New hominid fossils fromFejej, Southern Ethiopia. J Hum Evol 21:145–152.

19 WoldeGabriel G, White T, Suwa G, SemawS, Beyene Y, Asfaw B, Walter R. 1992. Kesem-Kebena: a newly discovered Paleoanthropologi-cal research area in Ethiopia. J Field Archaeol19:471–493.

20 Wood B. 1992. A remote sense for fossils.Nature 355:397–398.

21 Njau J, Hlusko L. 2010. Fine-tuning paleoan-thropological reconnaissance with high resolu-tion satellite imagery: the discovery of 28 newsites in Tanzania. J Hum Evol 59:680–684.

22 Conroy G. 2006. Creating, displaying, andquerying interactive paleoanthropological mapsusing GIS: an example from the Uinta Basin,Utah. Evol Anthropol 15:217–223.

23 Conroy G, Anemone R, Van Regenmorter J,Addison A. 2008. Google Earth, GIS, and theGreat Divide: A new and simple method for shar-ing paleontological data. J Hum Evol 55:751–755.

24 Bown T, Krause M. 1993. Soils, time, and pri-mate paleoenvironments. Evol Anthropol 2:11–21.

25 Shipman P. 1981. Life history of a fossil: anintroduction to taphonomy and paleoecology.Cambridge, MA: Harvard University Press.

26 Asfaw B, Beyene Y, Suwa G, Walter R,White T, WoldeGabriel G, Yemane T. 1992. Theearliest Acheulean from Konso-Gardula. Nature360:732–735.

27 Lillesand TM, Kiefer RW, Chipman JW.2004. Remote sensing and image interpretation.New York: John Wiley and Sons.

28 Schowengerdt RA. 2007. Remote sensingmodels and methods for image processing. Lon-don: Elsevier.

29 Jensen JR. 2007. Remote sensing of the envi-ronment: an earth resource perspective. UpperSaddle River, NJ: Prentice-Hall.

30 Jensen JR. 2005. Introductory digital imageprocessing: a remote sensing perspective. UpperSaddle River, NJ: Prentice-Hall.

31 Tso B, Mather TB. 2009. Classification meth-ods for remotely sensed data. Boca Raton, FL:CRC Press.

32 Rosenblatt F. 1958. The perceptron: a proba-bilistic model for information storage and orga-nization in the brain. Psychol Rev 65:386–408.

33 Rumelhart D, Hinton G, Williams RD. 1986.Learning internal representation by error prop-agation. In: Rumelhart D, McClelland L, edi-tors. Distributed processing: explorations in themicrostructure of cognition, Vol. 1: Founda-tions. Cambridge, MA: MIT Press.

34 MuraiH,OmatuS. 1997.Remote sensing imageanalysis using a neural network and knowledge-based processing. Int JRemote Sens 18:811–828.

35 USGS. 2011. Landsat 7 scenesL71035031_03120020808 and L71036031_03120020808. U.S. Geological Survey EarthResources Observation and Science (EROS)Data Center accessed 05 May 2011:Availableonline at: edc,usgs.gov.

36 Anemone R, Dirks W. 2009. An anachronis-tic mammal fauna from the Paleocene FortUnion Formation (Great Divide Basin, Wyom-ing, USA). Geol Acta 7:113–124.

37 Homer C, Huang C, Yang L, Wylie B, CoanM. 2004. Development of 2001 national landcoverdatabase for the United States. PhotogrammetricEngineering Remote Sensing 70:829–840.

38 Gesch D, Evans G, Mauck J, Huchinson J,Carswell W. 2009. The national map-elevation.U.S. Geological Survey Fact Sheet 3053.

39 Congalton R, Green K. 2009. Assessing theaccuracy of remotely sensed data: principlesand practices. Boca Raton, FL: CRC Press.

40 Gazin C. 1956. Paleocene mammalian faunasof the Bison basin in south-central Wyoming.Smithsonian Miscellaneous Collections 131:1–57.

41 Bailey GN, Reynolds SC, King GCP. 2011.Landscapes of human evolution: models andmethods of tectonic geomorphology and thereconstruction of hominin landscapes. J HumEvol 60:257–280.

42 Reynolds SC, Bailey GN, King GCP. 2011.Landscapes and their relation to hominin habitats:case studies from Australopithecus sites in easternand southern Africa. J Hum Evol 60:281–298.

43 Holmes K., 2007. GIS simulation of the ear-liest hominid colonisation of Eurasia. Oxford:Archaeopress.

44 Gingerich PD. 2006. Environment and evolu-tion through the Paleocene-Eocene thermalmaximum. TREE 21:246–253.

45 Beard K, Dawson M, editors. 1998. Dawn ofthe age of mammals in Asia. Pittsburgh, PA:Carnegie Museum of Natural History.

46 Beard KC, Dawson MR. 1999. Intercontinen-tal dispersal of Holarctic land mammals nearthe Paleocene/Eocene boundary: paleogeo-graphic, paleoclimatic and biostratigraphicimplications. Bull Soc Geol France 170:697–706.

47 Barrs D, Bartleson B, Chapin C, Curtis B,De Voto R, Everett J, Johnson R, Mlenaar C,Peterson F, Schenk C and others. 1988. Basinsof the Rocky Mountain region. In: Sloss L,editor. The geology of North America: sedimen-tary cover: North American Craton. Boulder, CO:Geological Society of America. p 109–220.

48 Dickinson W, Klute M, Hayes M, Janecke S,Lundin E, McKittrick M, Olivares M. 1988.Paleogeographic and paleotectonic setting of

ARTICLE Finding Fossils in New Ways 179

Laramide sedimentary basins in the centralRocky Mountain region. Geol Soc Am Bull100:1023–1039.

49 Lillegraven J. 1993. Correlation of Paleogenestrata across Wyoming: a users’ guide. In:Snoke A, Steidtmann J, Roberts S, editors. Ge-ology of Wyoming: Geological Survey of Wyom-ing memoir No. 5.p414–477.

50 Love JD. 1961. Definition of Green River, GreatDivide, and Washakie Basins, southwestern Wyom-ing. Bull Am Assoc Petroleum Geol 45:1749–1755.

51 Gazin C. 1952. The lower Eocene Knightformation of western Wyoming and its mam-malian faunas. Smithsonian Miscellaneous Col-lections 117:1–82.

52 [JF: Publishers’ location needed.]Gazin C.1959. Paleontological exploration and dating ofthe early Tertiary deposits in basins adjacent tothe Uinta mountains. In: Williams N, editor.Guidebook to the geology of the Wasatch andUinta Mountains. Intermountain Association ofPetroleum Geologists. p131–138.

53 Gazin C. 1962. A further study of the lowerEocene mammalian faunas of southwesternWyoming. Smithsonian Miscellaneous Collec-tions 144:1–98.

54 Gazin C. 1965. Early Eocene mammalianfaunas and their environment in the vicinity ofthe Rock Springs Uplift, Wyoming. NineteenthField Conference: Sedimentation of Late Creta-ceous and Tertiary Outcrops, Rock SpringsUplift. p171–180.

55 Anemone R, Watkins R, Nachman B, DirksW. 2010. An early Wasatchian mammalianfauna from an extraordinarily rich new localityin the Great Divide Basin, SW Wyoming. J VertPaleo 30(Suppl):138A.

56 Headley R. 2010. Landsat: A global land-imaging project. Washington, DC: US Depart-ment of the Interior.

57 Green G, Sussman R. 1990. Deforestationhistory of the eastern rain forests of Madagas-car from satellite images. Science 248:212–215.

58 Smith A, Horning N, Moore D. 1997. Re-gional biodiversity planning and lemur conser-vation with GIS in western Madagascar. Con-servation Biol 11:498–512.

59 Irwin M, Johnson S, Wright P. 2005. The stateof lemur conservation in south-eastern Madagas-car: population and habitat assessments for diur-nal and cathemeral lemurs using surveys, satelliteimagery and GIS. Oryx 39:204–218.

60 Aarts G, MacKenzie M, McConnell B, FedakM, Matthiopoulos J. 2008. Estimating space-useand habitat preference from wildlife telemetrydata. Ecography 31:140–160.

61 Hulbert I, French J. 2001. The accuracy ofGPS for wildlife telemetry and habitat mapping.J Appl Ecol 38:869–878.

62 Cagnacci F, Boitani L, Powell R, Boyce M.2010. Animal ecology meets GPS-based radio-telemetry: a perfect storm of opportunities andchallenges. Philos Trans R Soc B 365:2157–2162.

63 Schofield G, Bishop CM, MacLean G, BrownP, Baker M, Katselidis KA, Dimopoulos P, Pan-tis JD, Hays GC. 2007. Novel GPS tracking ofsea turtles as a tool for conservation manage-ment. J Exp Marine Biol Ecol 347:58–68.

64 Blake S, Douglas- Hamilton I, Karesh W.2001. GPS telemetry of forest elephants in Cen-tral Africa: results of a preliminary study. Afr JEcol 39:178–186.

65 Ryan PG, Petersen SL, Peters G, GremilletD. 2004. GPS tracking a marine predator: theeffects of precision, resolution and samplingrate on foraging tracks of African Penguins.Marine Biol 145:215–223.

66 Markham A, Altmann J. 2008. Remote moni-toring of primates using automated GPStechnology in open habitats. Am J Primatol 70:1–5.

67 Dominy NJ, Duncan B. 2001. GPS and GISmethods in an African rain forest: applicationsfor tropical ecology and conservation. Conser-vation Ecol 5:6.

68 Phillips K, Elvey C, Abercrombie C. 1998.Applying GPS to the study of primate ecology:a useful tool? Am J Primatol 46:167–172.

69 [JF: Title of article strange. Is it correct?]Sprague D, Kabaya H, Hagihara K. 2004. Fieldtesting a global positioning system (GPS) collaron a Japanese monkey: reliability of automaticGPS positioning in a Japanese forest. Primates45:151–154.

70 Reed D. 1997. Contour mapping as a newmethod for interpreting diet from tooth mor-phology. Am J Phys Anthropol (suppl) 24:194.

71 Jernvall J, Selanne L. 1999. Laser confocalmicroscopy and geographic information sys-tems in the study of dental morphology. Palae-ontol Electronica 2:18.

72 Zuccotti L, Williamson M, Limp W, UngarP. 1998. Modeling primate occlusal topographyusing geographic information systems technol-ogy. Am J Phys Anthropol 107:137–142.

73 Ungar P. 2004. Dental topography and dietsof Australopithecus afarensis and early Homo. JHum Evol 46:605–622.

74 Ungar P, Williamson M. 2000. Exploring theeffects of tooth wear on functional morphology:a preliminary study using dental topographicanalysis. Palaeontol Electronica 3:18pp.

75 Ungar P, Simon J, Cooper J. 1991. A semiau-tomated image analysis procedure for the quanti-fication of dental microwear. Scanning 13:31–36.

76 Scott J, Godfrey L, Jungers W, Scott R,Simons E, Teaford M, Ungar P, Walker A. 2009.Dental microwear texture analysis of two fami-lies of subfossil lemurs from Madagascar. JHum Evol 56:405–416.

77 Scott R, Ungar P, Bergstrom T, Brown C,Childs B, Teaford M, Walker A. 2006. Dentalmicrowear texture analysis: technical consider-ations. J Hum Evol 51:339–349.

78 Scott R, Ungar P, Bergstrom T, Brown C, GrineF, Teaford M, Walker A. 2005. Dental microweartexture analysis shows within-species diet variabilityin fossil hominins. Nature 436:693–695.

79 Ungar P, Brown C, Bergstrom T, Walker A.2003. Quantification of dental microwear by tan-dem scanning confocal microscopy and scale-sen-sitive fractal analysis. Scanning 25:185–193.

80 Evans A. 2005. Connecting morphology,function and tooth wear in microchiropterans.Biol J Linn Soc 85:81–96.

81 Evans A, Sanson G. 2003. The tooth of per-fection: functional and spatial constraints onmammalian tooth shape. Biol J Linnaeus Soc78:173–191.

82 Evans A, Wilson G, Fortelius M, Jernvall J.2007. High level similarity of dentitions in car-nivorans and rodents. Nature 445:78–81.

83 Eronen J, Evans A, Fortelius M, Jernvall J.2009. The impact of regional climate on theevolution of mammals: a case study using fossilhorses. Evolution 64:398–408.

84 Boyer D. 2008. Relief index of second man-dibular molars is a correlate of diet among pro-simian primates and other euarchontan mam-mals. J Hum Evol 55:1118–1137.

85 Boyer D, Evans A, Jernvall J. 2010. Evidenceof dietary differentiation among late Paleocene-early Eocene plesiadapids (Mammalia, Prima-tes). Am J Phys Anthropol 142:194–210.

86 Bunn J, Boyer D, Lipman Y, St. Clair E, Jern-vall J, Daubechies I. 2011. Comparing Dirichletnormal surface energy of tooth crowns, a newtechnique of molar shape quantification for die-tary inference, with previous methods in isolationand in combination. Am J Phys Anthropol145:247–261.

87 Emerson C, Anemone R. 2012. An artificialneural network-based approach to identifyingmammalian fossil localities in the Great DivideBasin, Wyoming. Remote Sensing Letters3:453–460.

VVC 2011 Wiley-Liss, Inc.

Books Received

� Agarwal SC, Glencross BA, eds. (2011)Social Bioarchaeology. Malden, MA:Wiley-Blackwell. 449 pp. ISBN:9781444337679. $40.95 (paperback)

� Bateson P, Gluckman P. (2011)Plasticity, Robustness, Developmentand Evolution. New York: CambridgeUniversity Press. 156 pp. ISBN: 978-0-521-73620-6. $45.00 (paperback)

� Bowles S, Gintis H. (2011) A Cooper-ative Species: Human Reciprocityand Its Evolution. Princeton: Prince-ton University Press. 288 pp. ISBN:978-0-691-15125-0. $35.00 (hardback)

� Brown T, Brown K. (2011) Biomo-lecular Archaeology: An Introduc-tion. Malden, MA: Wiley-Blackwell.

312 pp. ISBN: 9781405179607.$79.95 (paperback)

� Falk D. (2011) The Fossil Chronicles:How Two Controversial DiscoveriesChanged Our View of Human Evolu-tion. Berkeley: University of CaliforniaPress. 280 pp. ISBN: 978-0-520-26670-4.$34.95 (hardback)

� Pinhasi R, Stock JT, eds. (2011)Human Bioarchaeology of the Transi-tion to Agriculture. Hoboken, NJ:Wiley-Blackwell. 484 pp. ISBN: 978-0-470-74730-8. $149.95 (hardback)

� Roberts BW, Vander Linden M.(2011) Investigating ArchaeologicalCultures: Material Culture, Variability,and Transmission. New York: Springer.

393 pp. ISBN: 978-1-4419-6969-9.$169.00 (hardback)

� Shipman P. (2011) The Animal Con-nection: A New Perspective on WhatMakes Us Human. New York: W. W.Norton & Company, Inc. 336 pp. ISBN:978-0-393-07054-5. $26.95 (hardback)

� Singer M, Erickson PI, eds. (2011)A Companion to Medical Anthro-pology. Malden, MA: Wiley-Black-well.541 pp. ISBN: 9781405190022.$199.95 (hardback)

� Trivers R. (2011) The Folly of Fools:The Logic of Deceit and Self-Decep-tion in Human Life. New York: BasicBooks. 352 pp. ISBN: 978-0-465-02755-2.$32.50 (paperback)

180 Anemone ARTICLE