Movement, resting, and attack behaviors of wild pumas are revealed by tri-axial accelerometer ...

12 Pages·2015·1.69 MB·English

by  Yiwei Wang

Wangetal.MovementEcology (2015) 3:2
DOI10.1186/s40462-015-0030-0
RESEARCH Open Access
Movement, resting, and attack behaviors of wild
pumas are revealed by tri-axial accelerometer
measurements
Yiwei Wang1*, Barry Nickel1, Matthew Rutishauser2, Caleb M Bryce3, Terrie M Williams3, Gabriel Elkaim4
and Christopher C Wilmers1
Abstract
Background: Accelerometers are useful tools for biologists seeking to gain a deeper understanding of thedaily
behavior of cryptic species.We describe how weused GPS and tri-axial accelerometer(sampling at64 Hz) collars to
monitor behaviors offree-ranging pumas (Pumaconcolor),which are difficult or impossible to observe in thewild.
We attached collars to twelve pumas inthe Santa Cruz Mountains, CA from 2010-2012. By implementing Random
Forest models, weclassified behaviors inwild pumas based ontraining data from observations and measurements
ofcaptivepuma behavior.
Results: Weappliedthesemodelstoaccelerometerdatacollectedfromwildpumasandidentifiedmobileand
non-mobilebehaviorsincaptiveanimalswithanaccuracyrategreaterthan96%.Accuracyremainedabove95%even
after downsampling ouraccelerometerdatato16Hz.Wewerefurtherabletopredictlow-accelerationmovement
behavior(e.g.walking)andhigh-accelerationmovementbehavior(e.g.running)with93.8%and92%accuracy,
respectively.Wehaddifficultypredictingnon-movementbehaviorssuchasfeedingandgroomingduetothesmallsize
ofourtrainingdataset.Lastly,weusedmodel-predictedandfield-verifiedpredationeventstoquantifyacceleration
characteristicsofpumaattacksonlargeprey.
Conclusion:Theseresultsdemonstratethataccelerometersareusefultoolsforclassifyingthebehaviorsofcryptic
mediumandlarge-sizedterrestrialmammalsintheirnaturalhabitatsandcanhelpscientistsgaindeeperinsightinto
theirfine-scalebehavioralpatterns.Wealsoshowhowaccelerometermeasurementscanprovidenovelinsightsonthe
energeticsandpredationbehaviorofwildanimals.Lastlywediscusstheconservationimplicationsofidentifyingthese
behavioralpatternsinfree-rangingspeciesasnaturalandanthropogeniclandscapefeaturesinfluenceanimalenergy
allocationandhabitatuse.
Keywords:Pumaconcolor,Accelerometer,Behavior,Randomforest,Predation
Background time and space remains limited [2]. For example, when
Oneofthemajorlogistical challenges tostudyinganimal studying the impacts of habitat fragmentation on large
behavior lies in our inability to continuously observe carnivores, it is important to understand how landscape
free-ranging animals [1]. While recent technological ad- variables influence population connectivity and animal
vancements in the design and versatility of bio-logging movement, resting, and hunting patterns [3]. Accurately
devices (e.g., Global Positioning System (GPS) tags) have discerning these behaviors at a fine scale is almost im-
substantially improved our capacity to monitor animals, possible fromlocationdataalone,but criticalforinform-
our ability to document behavior continually through ingconservationmanagement decisions[4].
In the last decade, accelerometer sensors have emerged
as useful tools for remotely monitoring animal behavior
*Correspondence:[email protected] [5,6].Bycontinuouslymeasuringbodymovementandpos-
1EnvironmentalStudiesDepartment,CenterforIntegratedSpatialResearch,
ture, accelerometers allow scientists to infer the behavior
UniversityofCalifornia,1156HighStreet,SantaCruz,CA95064,USA
Fulllistofauthorinformationisavailableattheendofthearticle
©2015Wangetal.;licenseeBioMedCentral.ThisisanOpenAccessarticledistributedunderthetermsoftheCreative
CommonsAttributionLicense(http://creativecommons.org/licenses/by/4.0),whichpermitsunrestricteduse,distribution,and
reproductioninanymedium,providedtheoriginalworkisproperlycredited.TheCreativeCommonsPublicDomain
Dedicationwaiver(http://creativecommons.org/publicdomain/zero/1.0/)appliestothedatamadeavailableinthisarticle,
unlessotherwisestated.
Wangetal.MovementEcology (2015) 3:2 Page2of12
and energy expenditure of the instrumented individual Table1Total2secondobservations(N)ofcaptivepuma
[7]. Accelerometers have been used to study a wide range behaviorsclassifiedbymobilityclassandbehaviorclass
of behavior and physiology research topics, including for- Mobility Behavior N
aging,reproduction,activity,energybudgets,andlocomo- Yes Lowaccelerationmovement 564
tion[7,8].Whileaccelerometryhasbeenusedsuccessfully
Yes Highaccelerationmovement 50
to differentiate behaviors across a variety of taxa, most
No Resting 1167
accelerometer-based behavioral studies have focused on
No Eating 284
marine animals including marine mammals, sea turtles,
sharks,andseabirds[8]. No Grooming 77
Few studies have used accelerometers to document be-
havioral budgets of wild terrestrial animals. Wilson et al. removed six individuals from analyses entirely and
[9]recentlyusedacombinationofcomprehensivetri-axial extracted 4-26 days of accelerometer data from each of
accelerometer measurements and extremely fine scale the remaining individuals (Table 2).
GPS data to describe second-by second hunting behavior
in cheetahs (Acinonyx jubatus). Another study on chee- Mobilitymodelclassificationofcaptivepumaactivity
tahs and one on oystercatchers (Haematopus ostralegus) Our mobility model, which segregated puma activity into
also indicated that our understanding of wildlife behavior mobile and non-mobile periods, correctly classified move-
isenhancedbyevenlimitedaccelerometerdatacompared ments 96.17% of observed movement behaviors. The
toGPS dataalone[5,10].Bothof thesestudiesused short model identified Amp M (the amplitude of the dominant
temporal segments (i.e., a few seconds) of accelerometer frequencyforthemagnitudeofthemeasurements;referto
measurementseveryfewminutes,calibratedwithfieldob- Table3for variablenamesanddescriptions),DFZ(domin-
servationstopredictbehaviorbetweensuccessiveGPSlo- antfrequencyoftheZ-axis),andSDM(standarddeviation
cations.However,manyanimalscannotbeobservedeasily ofthemagnitude)asthemostimportantvariablesforpre-
inthewildandsamplingbehaviorforafewsecondsevery dicting mobility. We observed little loss of predictive
few minutes does not necessarily reflect the animal’s power down to 16 Hz for our mobility model when we
primaryactivityduringthattimesegment. down-sampled our accelerometer measurements from
Herewedescribetheuseofaccelerometermeasurements 64Hzto32,16,8,4and2Hz(Figure1).
and observations of captive animals to predict behavior in
wildpumas(Pumaconcolor).Likemostlargefelids,pumas Behaviormodelclassificationofcaptivepumaactivity
are cryptic animals and infrequently observed in their Our second and more comprehensive behavioral model
natural environment [11],makingit difficulttodocument categorizedthepuma’smovementbehaviorsintolowaccel-
the fine-scale behavioral patterns of this species. Our pri- eration (e.g., walk) and high acceleration (e.g., trotting and
mary objectives were to use continuous accelerometer running) movements, and the non-movement behaviors
measurements (sampling at 64 Hz), in combination with into resting, eating, and grooming behaviors. This behavior
periodic GPS readings, to distinguish different behaviors
in free-roaming pumas (Puma concolor), describe the
Table2TotalaccelerometerandGPSdatagatheredfrom
accelerometer signatures of puma predation events, and free-rangingpumasfrom2010-2011
determinethesufficientaccelerometersamplingfrequency
PumaID Sex Accelerometerdayssampled GPSAvailable
forcategorizingthesebehaviors.
2 F 26.72 Yes
5 M 14 Yes
Results
Behavioralmeasurements 7 F 9.42 Yes
We paired observations of captive pumas performing 16 M 16 Yes
activitiesincluding resting, feeding,moving,andgrooming 17 M 14 Yes
with accelerometer measurements to build a classification 28 F 4 Yes
algorithmtocategorizethosebehaviorsinthewildanimals.
4 M 8.23** Yes
Using 2 captive pumas (1 male and 1 female), we docu-
11 F 0.66 No
mented 2142 discrete behavioral observations, including
walking, grooming, resting, feeding, and fast movement 13 F 20.35* Yes
and their corresponding accelerometer measurements 20 F 11.2 No
(Table 1). From 2010-2011, we outfitted 12 wild pumas 27 M 0 Yes
(5males,7females)withGPS collarsandaccelerometer
30 F NA** Yes
sensors (Table 2). Due to mechanical or software failure
*onlytwoaxesaccuratelymeasured.**date/timeinaccuratelyrecorded.
in the onboard accelerometer sensors or SD cards, we Bolded,italicizedIDsindicateindividualsusedinanalyses.
Wangetal.MovementEcology (2015) 3:2 Page3of12
Table3Labelsandexplanationsofparametersextracted model predicted captive puma resting, low, and high accel-
fromaccelerometerdataandusedinRFmodels erationmovementbehaviorswith96.8%,93.8%,and92%ac-
predictingpumabehavioralclassification[12,13] curacy, respectively. The model predicted feeding behavior
Parameter Label Definition with 65.7% accuracy but failed to detect grooming (0%)
Axes X,Y,Z X,Y,Zaxes (Table 4). This behavior model identified SDM, DFZ, and
SDZ as the most important variables for classification.
Magnitude M Squarerootofthesumsof
squaresoftheaccelerationin
theX,YandZaxes Modelpredictionsofwildpumaactivity
Dynamicbody ODBAX, Meanofdynamic We found a strong positive relationship between our
acceleration(ing) ODBAY, accelerationvaluealongX,Y,
model predictions of percent mobility and the distance
ODBAZ, andZaxes
traveled by wild pumas between successive 15-minute
Overalldynamicbody ODBA SumofODBAX,ODBAY, GPS points (β=5.245 standard error=0.174, p<0.001),
acceleration ODBAZ
which further supports the model’s overall classification
Dominantpower AmpX, Amplitudeofdominant
accuracy.Random effectswerenot significant.
spectrum AmpY,Amp frequency
Z,AmpM Our24-houractivitybudgetsforindividualpumasshow
Dominantfrequency(Hz) DFX,DFY, Frequencyatdominant a general pattern of decreased movement activity in the
DFZ,DFM powerspectrum daytime and increased activity at night (Figure 2). Males
StandardDeviationof SDX,SDY, StandardDeviationof were more active than females (t=13.37, p=0) and more
dynamicbody SDZ,SDM dynamicaccelerationand often spent over 50% of their hourly increments moving
accelerationand magnitudeinwindow
(15% for males versus 1% for females). Males were more
magnitude
activenocturnally(6PM to6AM; t =15.22, p=0)whereas
females did not display a significant nocturnal movement
preference(t=0.59,p=0.28).However,evenforpumasof
thesamesex,individualsexhibitedconsiderablevariability
intheiractivity.Forexample,2Fand7F,bothfemaleswith
kittens, moved very little even during dawn and dusk,
whereas 28F, a female without kittens, was much more
active and even moved regularly throughout the daytime.
A territorial male, 5M, exhibited one peak in activity
around midnight whereas the two other males experi-
enced a decrease in activity around midnight. When de-
tailed GPS information was available, the actual distance
traveled by pumas strongly corroborated the proportion
oftimewepredictedmovement(Figures3and4).
We tested whether puma predation events, identified
from GPS data, were defined by clusters of high acceler-
ation movements. We observed a sustained cluster of high
acceleration movements, as identified by the behavior
model,inthetimeperiodbetweenoneGPSsamplinginter-
valbeforethestartofapredationeventthroughtheendof
the first quartile of the event (Figure 5, Additional file 1:
Figure S1). When compared to other clusters of high
Table4Cross-validationofactual(rows)andpredicted
(columns)behaviorsofcaptiveanimalsascategorizedby
thebehaviormodel
Feed Groom Rest High Low Percentaccurate
Feed 179 0 67 0 38 63.7
Groom 20 0 54 0 1 0
Figure1Theaccuracyofpredictionsbythemobilitymodel
Rest 26 1 1130 0 10 96.8
remainshighuntilthedataissampledbelow8Hz(topgraph).
Thecorrelationbetweenthemobilitymodelpredictionsandthe High 0 0 0 46 4 92
distancetraveleddeclinessteadilyasthedataisdown-sampled
Low 24 0 9 2 529 93.8
(bottomgraph).
Highandlowrepresenthighandlowaccelerationmovements.
Wangetal.MovementEcology (2015) 3:2 Page4of12
Figure2Predictedaveragehourlyactivityacrossa24-hourperiodforallpumas(+1SD).Thesebehavioraldairiesareaveragedover
25.64days(2F),13days(5M),8.42days(7F),13days(17M),16days(16M)and4days(28F).
accelerationmovement,whichmaynotnecessarilybeasso- female puma predation incidents, likely because they were
ciatedwithpredationevents,4ofour6potentialpredation subduing large, adult deer. In comparison, the male puma
events were ranked among the top 10% for cluster size predation incidents we identified consisted of fewer high
(or length) and 3 of 6 for maximum magnitude (Figure 6). acceleration movements because their targeted prey were
We recorded sustained, high acceleration movements for fawnsduringthedatacollectionperiod.
Figure315minutemovementdistancesmeasureddirectlyfromsubsequentGPSlocations(solid)andassociatedpredictedmovement
activitybasedonaccelerometry(dashed)forpumas28Fand16M.Whilewewouldnotexpectactivitylevelstobeperfectlycorrelatedwith
linearmovementdistances,thelevelorcorrespondencebetweenthesetwomeasuresprovidesafield-basedassessmentoftheabilityofaccelerometer
measurementstopredictmovementactivity.
Wangetal.MovementEcology (2015) 3:2 Page5of12
Figure415minuteGPSlocationsandassociatedpredictedmovementactivityfor28F.
From the limited predation incidents we analyzed, we which is characterized by higher acceleration and period-
notedthattheamountoftimeittooktokillpreywasrelated icity, from non-movement activities. The tight association
to the age and size of both the predator and prey species. between footfall frequency and the dominant frequency of
Based on the duration of the high acceleration movements, the accelerometer measurements bodes many promising
fawnswerekilledbymalesinlessthanoneminutewhereas avenuesforcalculatingdailyenergeticexpenditure[15].
females took over two minutes to kill large bucks and an MostcurrentresearchonanimalmovementusesGPSor
unknownpreyspecies. radio-telemetry collars [2], which only allow researchers to
measure locations sporadically throughout each day. How-
Discussion and conclusion ever, with GPS tags sampling at a low temporal resolution,
Our aim in this study was to use accelerometer measure- it is difficult to distinguish between an animal that moves
mentsrecordedoncaptiveanimalsasaproxytoclassifybe- 500metersinastraightlineinashorttimeperiodfroman
havior in wild animals. Using Random Forest models, we animal that is active for longer but meanders only a short
were able to accurately predict periods of non-movement, distanceinanonlinearfashion[17].Incontrast,accelerom-
low acceleration (i.e., stalking, walking), and high acceler- eters take near continuous measurements, thus providing
ationmovements(i.e.,trottingandrunning)inunobservable fine-scale documentation of animal behavior while the in-
wild animals. This insight allowed us to better document strument is activated [1]. The enhanced dataset from con-
pumamovementpatternsandactivitylevelsthroughoutthe tinuously sampling accelerometers can yield detailed
dayandtoidentifyindividualandsexdifferences. information on behavior (e.g. whether the animal is travel-
Our model identified Amp M, DFZ, and SDM as the ingorpotentiallyhunting)andenergyexpenditurebetween
top ranked predictors of puma mobility. The first two successiveGPSfixes[15].
variables arestronglytied tothe periodicityof themove- With continual advancements in biologging technologies,
ment since Amp M is the dominant power spectrum of sometagdesignsnowincludeGPSsensorsthatworksyner-
the magnitude and DFZ is the dominant frequency of gistically with accelerometers (i.e. accelerometer-informed
the Z-axis, which measures the heave (up-down motion) GPS), allowing for more flexible and intelligent GPS sam-
of the animal. For terrestrial quadrupeds, movement be- pling intervals [9,17]. Such accelerometer-informed GPS
haviors result in cyclic accelerometer patterns that are tags also reduce battery consumption (thereby prolonging
dominated by one frequency because the accelerations field deployment) by only recording GPS data points when
are primarily produced by footfalls and body movements the animal is actively moving [17]. Accelerometers could
(Figure 7) [14]. Thesedominantfrequenciescorrespondto also be programmed to trigger onboard cameras like Kitty-
footfallpatternsandcanbeusedforbiomechanicalanden- cam, which was used to document domestic cat predation
ergetics analyses [15,16]. The third parameter identified is on wildlife [18], to more efficiently quantify hunting at-
the standard deviation of the magnitude, which is higher tempts and kills of small prey that are difficult to identify
during mobile than non-mobile behaviors. Taken together, usingGPSdataalone.Camerasonwildanimalscanalsobe
theseparameterscanclearlydistinguishmovementbehavior, used to verify behaviors (e.g. footfall counts for energetics)
Wangetal.MovementEcology (2015) 3:2 Page6of12
Figure5Plotsoftwopredationeventsbypuma2F.Thetoppanelforeachplotillustratesthenumber(N)ofhighaccelerationmovements
perminuteoveraperiodoftwodays.Thedarkgreyrectanglehighlightstheperiodoftimeassociatedwiththepredationeventasverified
independentlyfromfieldvisitstoclustersofGPSlocations.Thebottompanelshowstherawaccelerometermeasurementsinunitsofgravitygforthe
Z-axis.Thebottominsertsmagnifyaone-minuteperiodofaccelerometermeasurementsfromselectedlargeclusterstoshowthemagnitudeand
durationoftheaccelerationduringthosehighaccelerationevents.Thearrowindicateswhenwehypothesizethekilleventtohaveoccurred.
predicted by models generated using accelerometer In our study, accelerometer sensors provided new in-
data[19]. sights on the variability and duration of predation events
We noted that puma predation events were associated identified by GPS locations and confirmed by subsequent
with high acceleration movements, which were predicted fieldvalidations. Whilepredationbyfemales onadult prey
by the behavior model. Although clusters of high acceler- occurred over several minutes and included some of the
ation movement occurred throughout the day, the events highest magnitude ranges, events involving larger-bodied
that we classified as kills had raw accelerometer readings malepumasandsmallerjuvenilepreywerenotasextreme.
withmagnitudesthatweregenerallylongerandlargercom- More broadly, we may be able to characterize predation
paredtoother events(Figures5and6).Additionally,there andattemptedpredationeventsofothermediumandlarge
were several other prolonged high acceleration movement bodiedmammalsusingpairedaccelerometerandGPSdata
clusters of similar or even larger magnitudes that were andcorroboratingthesepredictionswithfieldvisitationsof
not associated with identified kills. These events may kill site GPS clusters [3]. Combining such detailed loca-
have been unidentified kills or unsuccessful hunting tional and acceleration information can reveal the dur-
attempts. ation, energetic expenditure, and chase sequence of
Wangetal.MovementEcology (2015) 3:2 Page7of12
Figure6Histogramsshowinghowthesize(top)andmagnitude(bottom)ofclustersofpredictedhighaccelerationmovements
associatedwithpredationeventscomparetothosenotassociatedwithpredationeventsforpumas2F,5M,7Fand16M.Binscontaining
valuescorrespondingtoverifiedpredationeventsarehighlightedinredandaccentedbyanarrow.Killsby2Fand7Fatthefarrightofthe
histogramwereofadultdeerwhereasthoseof16Mand5Minthecenterofthehistogramwereoffawns.
puma predation events and help scientists gain more oystercatchers, cheetahs, and other species have used ac-
insight about the hunting behavior of pumas [15]. celerometers to identify feeding bouts in terrestrial ani-
Our behavior model was weak at predicting feeding mals with some success [5,10]. However, both studies
events and was not able to predict grooming behaviors. were able to observe the collared individual engaging in
This is most likely due to the complexity of identifying these behaviors and use that information to classify add-
non-locomotive motion [7] and to the relatively few in- itional events. While we are unable to directly observe
stances of these behaviors observed in the captive pumas in the wild, we may be able to record their feed-
pumas. While grooming is a relatively unimportant be- ing behavior using cameras placed at bait sites or fresh
havior to identify, accurately predicting feeding events is carcasses [20]. Using this information, we could then
crucial in understanding behavior and energetics. We identify accelerometer sequences that correspond with
believe these issues can be overcome in the future feeding in the field and then use that to identify future
through a variety of innovative strategies. Studies on feeding events. Additionally, we could perform more
Wangetal.MovementEcology (2015) 3:2 Page8of12
Figure7Two-secondwindowsofZ-axisaccelerationforfourbehaviorswithassociateddominantfrequenciesanddominant
powerspectrums.
feeding trials with captive pumas using whole deer and location, and energy expenditure has the potential to pro-
raccooncarcassessoastomorecloselymimictheirdietary videnovelinsightsintohowlandscapestructureinfluences
habitsinourstudyarea. the allocation of energy to different behaviors [25]. Such
Integrating accelerometer technology into dataloggers information would be valuable for conservation and man-
has great potential for animal behavioral research and is agement issues by revealing the detailed responses of
being increasingly adopted in ecological and conserva- individual animals to their surrounding landscape.
tion studies [4,8]. Data derived from accelerometers can
also be used to assess how animals respond behaviorally
and energetically to anthropogenic influences [15,21]. Methods
For example, pumas and other large carnivores may Studyspeciesandarea
change their overall activity levels and hunting patterns, Pumas are territorial, apex predators that live in diverse
thus impacting their caloric demands, when moving habitatsthroughouttheAmericas[26].Individualsarepri-
through human dominated habitats [3,22,23]. Addition- marily nocturnal and solitary, although females will typic-
ally, caloric expenditure by pumas can be calculated ally raise and accompany cubs for 15-21 months after
more accurately using our accelerometer derived activity birth. In our study area in the Santa Cruz Mountains of
budgetsandfootfallfrequenciesthanfromGPSinforma- California (37° 10.00’ N, 122° 3.00’ W), pumas primarily
tion alone [14,15,24]. feed on black-tailed deer(Odocoileushemionuscolumbia-
Our research demonstrates that accelerometers can suc- nus) but occasionally on other species, including wild
cessfully predict movement behaviors in animals that are boars (Sus scrofa), raccoons (Procyon lotor) and domestic
difficult to observe in the wild. However, more complex cats[3].
behaviors,suchasfeeding,mightonlybeaccuratelyidenti- Our 1,700 km2 study area encompasses a diverse land-
fiedwithadditionalobservationsfromcaptiveandwildan- scape ranging from dense, urban development to large
imals. Accelerometer sensors can be used with any tracts of intact and relatively undisturbed native vegeta-
terrestrial mammal to create a complete activity budget, tion primarily comprised of redwood and Douglas fir,
catalogue behaviors, including predatory ones, and poten- oak woodland, or coastal scrub communities. It is
tiallymeasureenergeticexpenditureaswellasforagingef- bisected by a large freeway and further crisscrossed by
ficiency. We believe this ability to link behavior, spatial numerous smaller roads providing access to rural houses
Wangetal.MovementEcology (2015) 3:2 Page9of12
and developments. The climate is Mediterranean, with tochangesinvelocityresulting from patternsoflocomo-
precipitation concentrated between November and April. tion and generally reflects the movement of the animal.
Elevationrangesfromsealevelto1155m. We subtracted static acceleration from collar measure-
ments using 2-second windows [29], and then extracted
Captiveanimaldatacollectionandanalysis 16 predictor variables from the three-accelerometer axes
We used custom-built collars [15,27] equipped with a tri- and the magnitude from each 2-second segment of dy-
axial accelerometer sampling continuously at 64Hz to namic acceleration (see Table 1). We also applied a Fast
monitorbehaviorincaptiveandwildpumas.Thetri-axial Fourier Transform to accelerometer measurements to
accelerometer was mounted such that the x-, y-, and z- extract the dominant frequency and dominant power
axeswereparalleltotheanterior-posterior,thetransverse, spectrumvaluesofeachbehavior.
and the dorsal-ventral planes of the animal, respectively. We selected Random Forests (RF) [30] as our modeling
Captive pumas were housed and trained by the Colorado tooltopredictunobservedbehaviorsinwildanimalsbased
Parks and Wildlife (Foothills Wildlife Research Facility) onmeasurementsofobservedbehaviorsincaptiveanimals.
[15]. We outfitted one adult male and one adult female RFisarelativelynovelandpowerfulmachinelearningtool
pumawithatestaccelerometercollarduringtrainingses- thatworkswellfornon-linearandcomplexecologicaldata
sions. We observed the collared animals and recorded not easily fitted by traditional methods such as generalized
their behaviors both manually and with a video camera. linearmodels[31].RFalsomakesitpossibletomakeaccur-
We conducted 1-2 trials per animal during two different ate predictions from datasets with correlated variables and
visits,totaling2hoursand40minutesofrecordedbehav- to compare conditional variable importance measures,
ioral observations for collar-outfitted captive pumas. At which identify the extent to which specific predictor
the end of each visit, we retrieved the collar and down- variables influence classification accuracy [31]. A higher
loadedthedata. measureof variableimportanceindicatesthatthe variable
We reviewed all recordings of captive pumas and cate- exerts greater influence on the response relative to other
gorized behaviors into mobile (e.g., slow movement, fast predictorswithlowervalues[32].
movement) or non-mobile (e.g., resting, feeding, groom- Our first model (mobility model) segregated mobile
ing) activities. We divided observations into 2-second from non-mobile behavior. To build our model using
segments encompassing only one behavior type and ex- RF, we fit 500 classification trees to a randomly selected
tracted the corresponding accelerometer data (Figure 7). subsample (n=1000, without replacement) of captive
Animalactivitieswereconstrainedintotheabove5classi- puma data using a random subset of 5 predictor vari-
fications,andwedidnotcataloguetransitionsbetweenbe- ables for each split in the tree [32,33]. Predictions made
haviors since they occurred very quickly (≤1s). We chose by all trees for each observation were then tallied, with
2-second windows because this duration accommodated classification assigned by the majority result and ties
at least 2 full strides of the animal within the 20 m test decided randomly. Model prediction accuracies were
trackusedforcaptivepumafilmingandcalibrationtrials. calculated by comparing predicted and actual classifi-
To compareacross collars we converted allaccelerom- cations. We then obtained unbiased variable importance
eter data into units of g (1g=9.8 m s−2) using tag spe- estimates using a permutation procedure described in
cific calibration values derived prior to deployment. The Strobl [32]. We built our second model (a more compre-
processofcalibratingtheaccelerometersconsistsofgen- hensive behavior model) using the same methodology to
tly tumbling the collar and measuring the body-frame predict five classes of behaviors: low acceleration move-
output of the accelerometer triad [28]. In the case of a ment(e.g.,walking),highaccelerationmovements(e.g.trot-
perfectly calibrated accelerometer, the locus of points ting, running), resting, eating, and grooming. We fit all of
would all be attached to a sphere centered at the origin our models and calculated variable importance estimates
with a radius of 1g. Due to null shift errors, the sphere is using the Party package [34] in the R statistical program
centeredofftheorigin,andscalefactorerrorstransformthe (vers.2.15.1)[35].
sphereintoanellipsoid.WedevelopedacustomMATLAB Wecreateddown-sampleddatasetsofaccelerometerdata
8.0 (The MathWorks, Inc., Natick, Massachusetts, United at 32, 16, 8, 4, and 2 Hz to explore the influence of sam-
States)scripttoextractthenullshiftandscalefactorerrors pling frequency on model prediction accuracy. Using the
byfittinganellipsetothedatatwoaxesatatime,andcom- down-sampleddatasets,webuiltadditionalmobilitymodels
biningtheresultantparameters. andassessedmodelaccuracyasdescribedpreviously.
Accelerometer measurements were deconstructed into
static and dynamic components [7]. Static acceleration Wildanimaldatacollectionandanalysis
relates to the inclination of the accelerometer with re- We captured wild pumas (5 males, 7 females) from
specttotheearth’sgravitationalfieldandtherebyreflects 2010-2012 using trailing hounds, cage traps, or leg hold
the posture of the animal. Dynamic acceleration relates snares as described in Wilmers et al. [3]. Each animal
Wangetal.MovementEcology (2015) 3:2 Page10of12
was tranquilized using Telazol and outfitted with an off- points to be generally larger than if we mostly predicted
the-shelf GPS/VHF collar (Vectronics Aerospace GPS non-movement.
PLUS model) combined with the custom-built archival We constructed 24-hour movement budgets for all
3-axis accelerometer tag [27], which was incorporated pumas to document the proportion of time pumas spent
into the battery casing (total collar weight=480 g) Data movingthroughoutthe day.To determine the proportion
collected by the accelerometer were recorded in an on- oftimespentmovingforeachone-hourperiod(e.g.,1AM
board 8GB microSD card, which is capable of storing to 2AM), we calculated the number of increments during
more than 200days ofaccelerometermeasurements. whichwepredictedmobileactivityanddividedthatbythe
We programmed each collar to acquire a GPS fix every total number of predictions we recorded between those
4 hours and had a mean fix rate of 86% (±1%) from 4 or time periods. From our behavior model, we generated
moresatellites.ForanimalscapturedpriortoApril2011,we 24-hourbehavioralbudgetsforourfivebehaviorclasses.
programmed accelerometers to record at a duty-cycle of Using the results from our behavior model, we tested
2 weeks on, 4 weeks off commencing immediately upon whether predation events were associated with periods
capture. While accelerometers were recording, the collars of high acceleration movement. We used six feeding
wereprogramedtoacquireadditionalGPSfixesat5-minute events by four pumas on five deer and one unknown
intervals between 8PM and 9PM local time for one week species to examine the corresponding high accelerom-
(GPS intensive sampling period). After April 2011, we eter movements. When feeding, pumas generally remain
programmed accelerometers to operate at cycles of two with the carcass over several GPS acquisitions. We used
consecutive days every week beginning 5 days after the this information to estimate the duration of the feeding
animal was captured to extend battery life while optimiz- event as the interval of time between the first and last
ing data collection. When accelerometers were recording, GPS location at the kill site. We also added the four-
collarsrecordedadditionalGPSlocationsevery15minutes hour interval prior to the first GPS location associated
during a 24-hour period from noon to noon (GPS inten- with the site to the feeding duration since it is possible
sive sampling period). We retrieved all collars either dur- thatthepuma madeakillduringthis periodoftime.
ingarecaptureoftheanimal(n=8)orfollowingitsdeath Five of these feeding events were visited and verified by
(n=4;2depredations,2unknowncauses).TheAnimalCare field personnel, and one was classified as a kill with 80%
andUseCommitteeatUCSantaCruzapprovedallanimal- probabilityusingapredationmodeldevelopedbyWilmers
handlingprocedures(IACUCProtocol#Wilmc1101). etal.[3].Fromobservationsofcaptivepumabehavior,we
We downloaded all available accelerometer data and know that faster and more intense movements, such as
removed the first 24 hours of data following anesthesia. runningandjumping,leadtoaccelerometerreadingswith
We then converted accelerometer data into units of g as ranges spanning 3-5 g or more. We expected that preda-
describedintheprevioussection. tion events would be associated with clusters of high ac-
Weusedourcaptivepuma-derivedmobilityand behav- celeration movements as pumas attack and wrestle with
iormodelstopredictfree-rangingpumabehaviorfromac- their prey. To screen for possible predation events, we
celerometer data obtained from wild puma collars that identified clusters of high acceleration movements with a
collected at least one day of both accelerometer and GPS magnitude range exceeding 3.4g, or two standard devia-
data. Because we could not observe behavior in wild ani- tions above the average and calculated the duration and
mals, we tested the accuracy of our mobility model’s pre- themaximummagnitudeoftheevent.Wedefinedaclus-
dictions by fitting a linear mixed effects model to GPS ter to be two or more successive high acceleration move-
data using the lme4 package [36]. Specifically, we tested ments separated by no more than three minutes between
whether our mobility model predictions were positively consecutive behaviors. We expected that if high acceler-
correlated with the distance traveled by pumas between ation movements were associated with predation events,
GPS points. We used the distance between successive wewouldseeaclusterof highaccelerationmovementsin
15-minute GPS points as our response variable and the first quartile of each predation event. Compared to
treated our model-predicted percentage of time spent non-predation clusters of high acceleration movements,
movingasafixedeffectwithpumaIDasarandomeffect. wealsoexpectedthesizeandmaximummagnitudeofthe
We expected that longer-distanceGPS movementswould cluster representing a potential predation event to be in
becorrelatedwithahigherpercentageofmodel-predicted thetop10%ofthosemeasurementsacrossallclusters.
movement activity, and that this relationship might vary
by individual pumas. For example, during a 15-minute Additional file
gap between successive GPS points, we used the mobility
modeltopredictthepercentageoftimethepumawasac- Additionalfile1:FigureS1.Plotsofpredationeventsbypumas5M,
tively moving. If we mostly predicted movement, we 7F,and16M,analogoustoFigure2.Thetoppanelforeachplot
illustratesthenumber(N)ofhighaccelerationmovementsperminute
would expect that the distance between the two GPS