From Sensor Data to Animal Behaviour: An Oystercatcher ExampleJudy Shamoun-Baranes1*, Roeland Bom1'2, E. Emiel van Loon1, Bruno J. Ens3, Kees Oosterbeek3, Willem Bouten11 Computational Geo-Ecology, Institute for Biodiversity and Ecosystem Dynamics, University of Amsterdam, Amsterdam, The Netherlands, 2 Department of Marine Ecology and Evolution, Royal Netherlands Institute for Sea Research (NIOZ), AB Den Burg, Texel, The Netherlands, 3S0V0N Dutch Centre for Field Ornithology, Coastal Ecology Team, AB Den Burg, Texel, The Netherlands
AbstractAnimal-borne sensors enable researchers to remotely track animals, the ir physiological state and body movements. Accelerometers, for example, have been used in several studies to measure body movement, posture, and energy expenditure, although predominantly in marine animals. In many studies, behaviour is often inferred from expert interpretation o f sensor data and not validated w ith direct observations o f the animal. The aim o f this study was to derive models that could be used to classify oystercatcher (Haematopus ostralegus) behaviour based on sensor data. We measured the location, speed, and tri-axial acceleration o f three oystercatchers using a flexible GPS tracking system and conducted simultaneous visual observations o f the behaviour o f these birds in the ir natural environment. We then used these data to develop three supervised classification trees o f behaviour and finally applied one o f the models to calculate tim e-activity budgets. The model based on accelerometer data developed to classify three behaviours (fly, terrestrial locomotion, and no movement) was much more accurate (cross-validation error = 0.14) than the model based on GPS-speed alone (cross- validation error = 0.35). The most parsimonious acceleration model designed to classify eight behaviours could distinguish five: fly, forage, body care, stand, and sit (cross-validation error = 0.28); other behaviours that were observed, such as aggression or handling o f prey, could not be distinguished. Model lim itations and potential improvements are discussed. The workflow design presented in this study can facilitate model development, be adapted to a wide range o f species, and together w ith the appropriate measurements, can foster the study o f behaviour and habitat use o f free living animals throughout the ir annual routine.
Citation: Shamoun-Baranes J, Bom R, van Loon EE, Ens BJ, Oosterbeek K, e t al. (2012) From Sensor Data to Animal Behaviour: An Oystercatcher Example. PLoS ONE 7(5): e37997. doi:10.1371/journal.pone.0037997
Editor: Gonzalo G. de Polavieja, Cajal Institute - Consejo Superior de Investigaciones Científicas, Spain
Received January 2, 2012; Accepted April 26, 2012; Published May 31, 2012
Funding: The tracking research infrastructure is supported by LifeWatch and the BiG Grid infrastructure for e-Science (http://www.biggrid.nl). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Com peting Interests: The authors have declared that no competing interests exist.
IntroductionU nderstand ing how anim als in teract w ith their environm ent is
one o f the fundam ental aims o f anim al ecology. In o rder to acquire this knowledge we need inform ation w hich can be used to quantify w hat anim als are doing, w hen, where, how an d for how long. For a b ro ad spectrum of ecological research, from theoretical to applied, quantitative tim e budget inform ation a t the individual level is im portan t [1-3]. A quantitative approach can provide essential inform ation for species an d hab ita t conservation [4-5], understand ing ecosystem dynam ics [2,6], understand ing and m itigating the spread o f anim al borne diseases [7-8], anim al adap tation to clim ate and landuse change [4,9], spread of in troduced and invasive species [3] and the developm ent o f environm ental policy [10], For exam ple, w hen addressing the direct and indirect im pact o f fisheries on seabirds (see question 26 [10]), we w ould like to know w here, w hen and how a species forages [2,11-12].
O u r ability to visually observe the behaviour o f free-ranging anim als is generally quite restricted in space and time. In recent decades, technological advances have enabled researchers to track anim als during local an d m igratory m ovem ents, in the air, on land
and in the sea [13-17]. Similarly, bio-logging features such as body acceleration, heart rate, stom ach tem perature, diving depth enable rem ote m onitoring o f an an im al’s physiological state and its activity in 3 dim ensional space and in tim e [18-20], T h e data collected by these sensors can then be used to infer w hat an anim al is doing. For exam ple, speed m easured directly using G PS (global positioning system) or derived from consecutive tracking locations has been used to infer behaviour, to distinguish betw een travelling and resting during m igration [21-23], and during foraging trips [24-25]. Yet, instantaneous speed m easured with a G PS is p robably too inaccurate for distinguishing small differences in locom otion, especially at low speeds [26], Accelerom eters are a prom ising sensor for studying anim al behaviour rem otely since accelerom eters can m easure the posture and body m ovem ents ([27] and references therein) as well as estim ate the speed and energy expenditure [3,28-30] o f the anim al to w hich it is attached. In the last decade, dynam ic and static body acceleration have been used to study a diverse range o f behaviours including diving [31- 33], swim m ing and flight strategy [34—37], feeding an d b reath ing [38] and m ating behaviour [39]. B ehavioural studies utilizing accelerom eter da ta have focused prim arily on m arine anim als [40]
PLoS ONE I www.plosone.org 1 May 2012 | Volume 7 | Issue 5 | e37997
and very few studies have focused on terrestrial locom otion in wild anim als [3,41-43].
Q uantifying behaviour from bio-logging da ta requires an interm ediate step to translate the m easured sensor da ta into specific behaviours. T hree general approaches to achieve this translation are: (1) non-au tom ated in terpreta tion o f sensor da ta by an expert, w ith [2,44—45]) o r w ithout [35,41] field observations o f the an im al’s behaviour; (2) au tom ated segm entation or clustering o f sensor da ta w ithout field observations o f anim al behaviour, sometimes followed by labelling o f the identified segments by an expert [46]; (3) au tom ated classification o f sensor da ta in com bination w ith observations o f the an im al’s behaviour [3,38,47]. For brevity we will call the first m ethod expert interpretation, the second clustering, and the th ird m ethod classification. In m ost studies o f wild anim als, behaviour has been inferred by expert in terpretation; however, the inferred behaviour canno t be validated via this m ethod. In cases w here behavioural observations are no t available, only the m ethods o f expert in terpreta tion and clustering can be applied to sensor data. T he essential difference betw een the two m ethods is th a t for expert in terpreta tion the behavioural classes have to be specified prio r to the classification task, w hereas in clustering the specification (and meaning) o f behavioural classes follow from the clustering results. H ence, expert in terpreta tion is deductive w hereas the clustering is inductive in nature . D ue to the lack of behavioural observations w hich m atch the sensor data, the uncertain ty o f the results cannot be assessed for either o f these m ethods. In contrast, the classification m ethod can provide inform ation abou t the uncertainty o f the classification result. K now ledge o f classification uncertain ty can be used to answer various kinds o f inferential questions such as w hether a given m odel (using sensor data) is able to pred ic t behaviour bette r th an a null-m odel o r w hether a given num ber o f behavioural classes can be distinguished.
T h e prim ary aim o f this study was to derive, evaluate and com pare m odels to classify m easurem ents from sensors a ttached to individual shorebirds, the oystercatcher (Haematopus ostralegus), into pre-defined behaviours; one m odel w ould only be based on speed m easured by G PS and o ther models w ould include accelerom eter data. T h e use o f accelerom eter da ta to rem otely determ ine the behaviour o f terrestrial wild anim als is still quite new. W e expected th a t classification m odels based on G PS-speed alone w ould im prove if accelerom eter da ta w ould be included. W e describe a m ethodological workflow th a t we used in this study to develop and apply classification m odels o f anim al behaviour. T he aim o f such a workflow is to provide researchers w ith a clear outline o f the diverse processing an d analysis steps needed to quantify behaviour based on sensor data; it can be applied to o ther studies, stream line analysis o f new data and the reanalysis o f existing data. By using sensor da ta to quantify behaviour in com bination with location data, tim e-activity budgets o f anim als can be quantified a t the individual level and in relation to their environm ent [19]. T o show the added value of incorporating behavioural inform ation with location da ta we calculated the tim e-activity budget o f an individual b ird for areas an d times o f day th a t are norm ally difficult to visually observe in the field. M ore specifically, we w anted to determ ine if an individual spends its tim e differently during the day com pared to during the night an d how does it spend its tim e w hen outside the territory.
Methods
S tu dy spec ie s a n d s tu d y a reaT h e oystercatcher is a long lived, m onogam ous w ader th a t feeds
on intertidal prey, such as hard-shelled bivalves they can open with
their strong bill an d large m arine worm s. T hey b reed predom inantly in coastal habitats, although inland breeding increased during the second ha lf o f the previous century [48]. O n the D utch W adden Island Schierm onnikoog (53.26°N, 06.10°W , Figure 1) a population o f oystercatchers has been studied and individuals have been colour ringed since the 1983 (e.g. [49-51]). In this population, colour ringed individuals can be easily identified and a range of behaviours can be visually observed in the field from two observation towers (Figure 1).
M e th o d o lo g ica l w orkflowIn the following sections we briefly describe how each of the
following research steps was applied in the curren t study: data collection, da ta processing, m odelling an d m odel application. A m ore detailed description is provided in T ex t S I. T h e steps are also shown in a schematic workflow diagram (Figure 2) and present a general m ethodological approach th a t can be applied to any study w here m easurem ents from sensors a ttached to anim als will be used in com bination with observations o f behaviour to derive and apply a classification m odel o f anim al behaviour.
D ata collectionIn this study, we used the recently developed UvA Bird
T racking System (UvA-BiTS, University o f A m sterdam Bird T racking System) w hich has been used to study several resident and m igratory b ird species (e.g. [24]). T h e tracking device is solar- pow ered an d weighs 13.5 g, an d includes a tri-axial accelerom eter and a G PS receiver w hich m easures geographic position, altitude above m ean sea level, tim e and instantaneous speed. T h e tri-axial accelerom eter m easurem ents were converted to acceleration in g (1 g = 9.8 m s 2) w ith respect to the earth ’s gravitational field in three directions: surge (X), sway (Y) and heave (Z).
D uring the b reeding season o f 2009 (May-July 2009), oystercatchers were observed for several weeks and three colour ringed birds b reeding in high quality territories adjacen t to the m udflats were selected for our tracking study and trapped towards the end o f the b reeding season (Table SI). O bservations in the breeding area p rio r to trapp ing were described in m ore detail in previous studies [50-52]. T he birds were caught on their nest w ith a walk-in trap. After the birds were weighed and m orphological m easurem ents were taken, a tracking device was fitted on their back using a Teflon ribbon harness (weight ~ 2 g). T h e harness was a ttached to the b ird using a figure eight configuration. T h e straps were connected a round the neck and the wings to one weak p o in t a t the sternum . T he weak p o in t was m ade ou t o f cotton thread , w hich is expected to deteriorate in two to three years. T h e harness and tracking device weighed less th an 3% o f the m ean body mass o f the birds (Table SI). Birds were released within 60 m inutes o f capture.
A G PS fix was taken every 10 m inutes from 30 Ju n e 2009 th rough 20 Ju ly 2009 and every 30 m inutes from 21-31 July. D irectly following each G PS fix, acceleration was m easured with a frequency o f 20 H z for 3 seconds. From 30 Ju n e th rough 14 Ju ly 2009 each b ird was observed daily for 30 m inutes with a telescope (20-60 X , Zeiss D iascope 85 T*FL) positioned in one o f the observation towers (Figure 1). D uring visual observations, the tracking device was set to take a G PS fix a t 10 s intervals followed by 3 seconds o f acceleration m easurem ents. W hen a b ird started a new behaviour, it was reported by the observer and recorded by a field assistant in a P S IO N handheld com puter (W orkabout Pro) with O bserver X T software (ww w.noldus.com ). T o accurately link the visual observations with the GPS an d accelerom eter m easurem ents, the handheld com puter was synchronized to G PS tim e using a handheld GPS. T h e recording procedure was
PLoS ONE I www.plosone.org 2 May 2012 | Volume 7 | Issue 5 | e37997
Figure 1. The study area on the island of Schiermonnikoog, the Netherlands (53.29°N , 06.10 E) at different spatial scales. The points represen t GPS fixes of th ree oystercatchers (green - tag 166, red - tag 167, blue - tag 169; Table S I) from 1 July 2009 to 31 July 2009, with consecutive points connected by lines. The black circles are the nests o f these birds. The locations o f the observation tow ers are indicated by a square and the base station by a triangle. Black lines represents creeks, dark grey lines represent urban infrastructure. doi:10.1371/journal.pone.0037997.g001
first practised extensively on non tagged birds. T h e m ain behaviours defined by K ersten [53] were extended with subbehaviours observed in the field (Table 1); the classes to express the behaviours as well as the sub-behaviours are bo th exhaustive and exclusive.
D ata p ro cess ingIn this study da ta processing (Figure 2) included da ta storage,
m erging datasets an d filtering data. All the G PS and accelerom eter da ta that were collected during the study period w ere stored in a dedicated postgreSQ L database (h ttp ://w w w .uva-b its .n l/v irtua l- lab /) and the visual observations were stored in a separate data base. This enables researchers to systematically explore o r re-use (parts of) the data sets if needed. T h e GPS and accelerom eter data were labelled w ith the visually observed behaviours, while accounting for a m axim um o f 10 s recording delay on the handheld com puter (‘m erge’ in Figure 2), see T ext S I for m ore details. Next, all da ta were checked for anom alies (‘filter’ in Figure 2). For exam ple, da ta that could no t be unam biguously linked to a behavioural observation were rem oved from further analysis.
M odellingN ote that the decisions m ade during the m odel building phase
(grey ‘m odel building’ rectangle in Figure 2) regarding the data,
m odel design an d analysis steps are dependent on each o ther and, in general, can also be dealt w ith in a different o rder th an chosen here. O ne o f the first steps in our analysis was defining the m odel aim (‘m odel a im ’ in Figure 2) w hich was to accurately predict behaviour, w hereby all behavioural classes w ere considered equally im portant. In this study, we report three m odelling cycles, each leading to a different m odel. W e first start a simple m odel with three behavioural classes an d increm entally working towards m ore detailed models. T h e aim o f the first m odel (‘m odel S3’, speed m odel o f three behaviour classes) was to predict three behaviour classes. T h e predictor variable specified for this m odel (‘specify predictors’ in Figure 2) was G PS speed. T h e original behaviour classes were grouped (‘reclassify’ in Figure 2) into three behaviour classes (Table 1, colum n 4). W ith the aim o f predicting three behaviour classes, we then w ent th rough feedback loop 1 to develop the second m odel (‘m odel SA3’, speed-acceleration m odel o f three behaviour classes) using pred ic tor variables based on accelerom eter data, described in m ore detail below, as well as GPS speed.
W hile speed was provided by the G PS sensor, the acceleration m easurem ents h ad to be processed to calculate m eaningful p red ic tor variables. W e derived 15 pred ictor variables from the tri-axial acceleration segments (m easurem ent frequency of 20 H z for 3 seconds). All p red ic tor variables used in this study are listed in T able 2 (see T ex t SI for m ore detailed inform ation on how the
PLoS ONE I www.plosone.org 3 May 2012 | Volume 7 | Issue 5 | e37997
Figure 2. A schematic workflow of the d ifferent methodological steps conducted in this study. The workflow is broken dow n into three main categories of activity show n on the upper bar: Data collection, Data processing and Modelling. The objects in the grey rectangle indicate the aspects involved in building classification m odels and th e objects in the dark grey rectangle indicate application of th e classification m odels for diverse analyses such as calculating tim e budgets. Ovals indicate data in various form ats (files from data loggers, w ritten field forms, etc). Cylinders indicate inform ation th a t is stored in a database. White rectangles indicate (com putational) activities and decisions. Solid arrows presen t the workflow to m ove from field data to the establishm ent and application of a m odel. Dashed arrows presen t feed-back loops w here a certain part of the workflow is repeated in response to progressive insights (only th e m ost im portant feed-back loops are shown). Feed-back loops are p resen t from a point after m odel calibration as well as a point after m odel evaluation back to the beginning of th e modelling sequence (2) or later in the m odelling sequence (1). These steps are generalized so th a t they can be applied to o ther studies, for exam ple visual observations may be replaced by video observations or expert interpretation o f sensor data. doi:10.1371/journal.pone.0037997.g002
accelerom eter data were processed). All the predictors, except for the m ean dynam ic body acceleration in single dimensions (odbaX, odbaY an d odbaZ), have been used in o ther studies [27,35,44,47] and are described in [47]. T he overall dynam ic body acceleration (odba) was calculated as the sum of odbaX , odbaY and odbaZ and has been used in o ther studies as a single m easure o f body m ovem ent and a potential proxy for energy expenditure, see [30] for a detailed explanation.
In the th ird m odelling cycle (‘m odel SA8’, speed-acceleration m odel o f eight behaviour classes) we went th rough feedback loop 2; the m odel aim is to classify the eight m ain behavioural classes (Table 1, colum n 1) using all available predictors. In o rder to ensure a sufficient sample size pe r behaviour (Table 1, colum n 5) to train the models, the observations o f the three individuals were pooled (‘pool d a ta ’ in Figure 2), treating each individual observation an d each individual equally. Even after pooling the data, the sample size was very small for several o f the subbehaviours.
W e selected classification trees (‘select m odel type’ in Figure 2) as our m odelling approach [54—55], using the im plem entation in
the rp a rt R package [56-57]. O verall cross-validation erro r was used as a single criterion to m easure the degree o f success (‘error criteria’ in Figure 2). W e did not split the da ta into sub-sets for m odel calibration and evaluation because the dataset was already lim ited in size (with only a few observations for some behavioural classes) and also because da ta splitting is an integral p a rt o f the m odel calibration procedure for regression trees as described below.
Classification trees were derived (‘calibrate m odel’ in Figure 2) by initially growing a m axim um (over-fitted) tree, w hich was subsequently p ru n ed to an optim al size. T o determ ine the optim al tree size, we applied the ‘one standard deviation ru le’: select the smallest tree whose cross validation erro r is less th an the m inim um cross validation erro r +1 standard deviation [54], M odel perform ance was evaluated (‘evaluate m odel’ in Figure 2) by 10- fold cross-validation in w hich the dataset is split into 10 partitions, 9 o f w hich are used to calibrate the m odel and 1 is used to evaluate the m odel; the calibration and evaluation is then repeated 10 times using a new da ta partition [58], O nce a m odel could not
PLoS ONE I www.plosone.org 4 May 2012 | Volume 7 | Issue 5 | e37997
T a b le 1. Different behaviours and sub-behaviours visually observed during the study and linked to GPS and accelerometer measurements.
Behaviour Sub behaviour Description 3-class m odel behaviours n
Aggression Bobbing Bird is standing and moves its body up and down No locomotion 4
Chasing Bird is chasing conspecifics Terrestrial locomotion 3
Stand Solitary piping Bird is calling loudly while standing, conspecifics are nearby
No locomotion 18
Piping ceremony Bird is calling loudly together with other birds, while walking
Terrestrial locomotion 12
Walk Solitary piping Bird is calling loudly while walking, conspecifics are nearby
Terrestrial locomotion 12
Body care Preen Bird is preening its feathers No locomotion 82
Wash Bird is bathing No locomotion 3
Fly normal flight Bird is flying Fly 13
Forage By sight Bird is searching for prey by sight while walking Terrestrial locomotion 249
By touch Bird is searching for prey by touch while walking Terrestrial locomotion 5
Handle Handling a t surface Bird is handling the prey a t the surface No locomotion 15
Handling in situ Bird is handling the prey beneath the surface No locomotion 29
Walking with prey Bird is walking with the prey Terrestrial locomotion 7
Sit Bird is sitting No locomotion 100
Stand Bird is standing No locomotion 125
Walk Bird is walking Terrestrial locomotion 25
The column '3-class model behaviours' shows the behavioural classes reclassified a priori and used to calibrate the 3-class models (S3 and SA3). The behavioural classes in the first column were used as the predicted variable in the 8-class model (SA8). The num ber of visual observations (n) is provided per behaviour. doi:10.1371 /journal.pone.0037997.t001
T a b le 2. Predictive parameters used in this study, derived from the GPS (speed) and the accelerometer sensors.
predictor direction label explanation
body pitch (c pitchX angle of the body along the surge axissurge
body roll (c rollY angle of the body along the sway axissway
mdbaY maximum dynamic body acceleration along the sway axissway
overall dynamic body acceleration (g) odbaX Mean dynamic body acceleration along the surge axissurge
heave odbaZ Mean dynamic body acceleration along the heave axis
odba overall dynamic body acceleration (odbaX+odbaY+odbaZ)
heave pitchZ angle of the body along the heave axis
heave mdbaZ maximum dynamic body acceleration along the heave axis
maximum dynamic body acceleration (g) surge mdbaX maximum dynamic body acceleration along the surge axis
sway odbaY Mean dynamic body acceleration along the sway axis
GPS speed (m s speed 3D speed
dominant power spectrum (g2Hz 1) surge dpsX maximum power spectral density (psd) of dynamic acceleration along the surgeaxis
sway dpsY maximum psd along the sway axis
heave dpsZ maximum psd along the heave axis
frequency at the dominant power surge fdpsX frequency at the maximum psd along the surge axisspectrum (Hz)
sway fdpsY frequency at the maximum psd along the sway axis
heave fdpsZ frequency at the maximum psd along the heave axis
The dominant power spectrum measures the relative am ount of kinetic energy that is spent a t the dominant periodicity in a signal (see Text S1 for more details).The integration interval of the measurement for the accelerometer sensor is 3 seconds with 20 Hz. The direction in which each variable is defined is given in Cartesian coordinates relative to the ground surface: surge represents the x-axis, sway the y-axis and heave the z-axis. The m easurement units (SI) are provided in parentheses. doi:10.1371 /journal.pone.0037997.t002
PLoS ONE I w w w .plosone.org 5 May 2012 | Volume 7 | Issue 5 | e37997
accurately p red ict behaviour, the m odelling cycle w ould no t be repeated to pred ict behaviour in m ore detail.
M odel ap p lica t ionT o exemplify m odel application we calculated tim e budgets per
individual, including areas and times o f day th a t are difficult to visually observe in the field. W e applied m odel SA8 to the sensor da ta collected in Ju ly 2009, a period for w hich visual observations were no t collected, to classify each da ta po in t into discrete behaviours associated to the geographic position provided by the GPS (‘pred ic t’ in Figure 2). W e then used the pred ic ted behaviours to calculate the tim e budget during the day and at night for three different habitats (territory, m udflats and salt marsh). For m ore details see T ex t S I.
T h e percentage of tim e devoted to each o f the classified behaviours was calculated by dividing the num ber o f observations pe r behaviour by the total num ber o f observations during the day or during the night. W e associated each G PS fix to one o f the following habitats: territory, m udflats and salt m arsh using the geographic database o f global adm inistrative areas (GADM , h ttp ://w w w .g ad m .o rg ). A b ird was considered to be in its territory w hen it was within 150 m from its nest. T his distance was chosen after visually inspecting the locations o f each individual in relation to their nest. All terrestrial areas outside the territory, w hich are p redom inantly salt m arsh in the study area, were labelled salt m arsh. T h e inter-tidal areas were labelled mudflats. D ay was defined as the hours betw een sunrise an d sunset a t Schierm onnikoog (53.47 N, 6.23 E).
S oftw are im p le m e n ta t io n o f th e various analysis s t e p sT h e da ta processing steps, the definition o f prediction variables
and subsequent m odelling were conducted using the R language for statistical com puting [57]. W e provide a m odelling package with the scripts developed for da ta analysis (model building and m odel application) and the data presented in this study (Dataset S I ) .
Results
Behavioura l m e a s u r e m e n t sT able SI provides an overview o f the num ber o f G PS and
acceleration segments (60 m easurem ents pe r segment) collected for each bird. D uring visual observations 16 behaviours were observed and 702 G PS fixes and acceleration segments could be linked to the visual observations (Table 1). Forage by sight was the m ost frequently observed behaviour. T h e m ean value o f each pred ic to r variable is provided pe r observed behaviour in Tables S2A -C . M ean speeds did no t differ significantly betw een behaviours w ithin the no locom otion and terrestrial locom otion behaviours (P> 0 .05 , T ukey F1SD test). This justified the reclassification o f behaviours into three categories, fly, terrestrial locom otion and no locom otion, before fitting the speed m odel (Table 1, C olum n 4).
3-class s p e e d m o d e lT h e best m odel for three behavioural classes, based on speed
alone (model S3), classified 470 ou t o f 695 observations correctly (7 ou t o f the 702 observations in our dataset did no t have a speed m easurem ent) resulting in an absolute cross-validation erro r o f 0.35. Speeds below 0.18 m s-1 w ere classified as no locom otion, speeds h igher or equal to 3.4 m s_1 as fly an d interm ediate speeds as terrestrial locom otion (Figure 3A). Fly was classified incorrectly as terrestrial locom otion in 8% (1 ou t o f 13) o f the cases. O bserved behaviours belonging to the terrestrial locom otion group were
incorrectly classified as no locom otion in 44% o f the cases and behaviours belonging to the no locom otion group w ere incorrectly classified as terrestrial locom otion in 24% o f the cases.
3-class acce le ra t io n m o d e lT h e best m odel for three behavioural classes, using speed and
acceleration data (model SA3), classified 609 o f the 702 observations correctly (absolute cross-validation erro r = 0.14). O bserved behaviours belonging to the terrestrial locom otion group were incorrectly classified as no locom otion in 9% o f the cases, and no locom otion observations were incorrectly classified as terrestrial locom otion in 18% of the cases. T he predictors that were included in the m odel were the m ean dynam ic acceleration in the surge axis (odbaX) an d m axim um pow er spectral density (psd) o f dynam ic acceleration along the heave axis (dpsZ) (Figure 3B, T able 2). I f odb aX was less th an 0.09 g, equivalent to no dynam ic acceleration in the surge axis, then behaviour was classified as no m ovem ent, if odb aX was h igher and dpsZ was greater th an or equal to 5.1 W Flz , then behaviour was classified as fly an d if odb aX was greater th an or equal to 0.09 g and dpsZ was less th an 5.1 W F lz— 1, then behaviour was classified as terrestrial locom otion (Figure 3B). Speed was not re ta ined as a p red ictor variable.
8-class acce le ra t io n m o d e lT h e best m odel for the eight behavioural classes, using speed
and acceleration data (model SA8), classified 517 o f the 702 observations correctly (absolute cross-validation erro r = 0.282). O nly five o f the eight behaviours were classified: ‘fly’, ‘forage’, ‘body care ’, ‘stand’ and ‘sit’ (Figure 4). W alk was generally misclassified as forage w hich is no t surprising as forage included, by definition, walking m ovem ent (see T able 1). Aggression was generally misclassified as body care o r forage, and handle was predom inantly misclassified as forage. F rom the 15 different explanatory variables, only four variables were re ta ined in the classification m odel: od b aX an d dpsZ, bo th also included in m odel SA3, as well as overall dynam ic body acceleration (odba) an d the p itch angle m easured in the surge (pitchX). As w ith m odel SA3, odb aX can be used to distinguish betw een forw ard locom otion (fly and forage) and no locom otion (body care, stand and sit). Similarly, as in m odel SA3, dpsZ greater th an or equal to 5.1 g2 Flz 1 could be used to distinguish fly from forage.
T h e decision rules in this m odel (Figure 4) are easy to in terpret w ithin the context o f the field observations and locom otion. D ynam ic acceleration or deceleration in the surge axis (odbaX) describes active forw ard m ovem ent, an d the b ird is either flying or walking, depending on the am o u n t o f energy invested a t the dom inan t periodicity in the heave axis signal (dpsZ). W hen there is no t m uch m ovem ent in the heave or surge, the b ird is either standing or sitting, depending on the p itch angle o f the body; a zero or slightly positive angle m eans th a t the logger is horizontal and the b ird is sitting an d a negative angle m eans th a t the anterior o f the b ird is tilted upw ards and the b ird is standing (see T ex t SI for m ore details). W hile standing, the b ird m ay p reen its feathers m oving its bill and body, resulting in h igher overall dynam ic body acceleration (odba) th an w hen standing still. Figure 5 shows characteristic exam ples o f dynam ic and static acceleration signals for fly, forage, body care, stand an d sit behaviours, w hich were correctly classified by the SA8 m odel. See V ideo SI for an exam ple o f foraging behaviour coupled w ith accelerom eter data.
T h e SA8 m odel effectively reduced the num ber o f behavioural classes from 8 to 5. This result and the cross validation error indicated th a t m ore detailed behavioural classes o r trying to
PLoS ONE I www.plosone.org 6 May 2012 | Volume 7 | Issue 5 | e37997
Figure 3. Decision tree and confusion m atrix for models S3 and SA3. For m odel S3 (A) and m odel SA3 (B), th e num ber of observations correctly classified per behaviour is show n in bold. See Table 2 for a description o f the predictor variables. Out of the 702 observations, there w ere no speed m easurem ents in 7 cases, hence the sam ple size o f 695 for m odel S3. doi:10.1371/journal.pone.0037997.g003
classify the 16 m utually exclusive classes o f behaviour observed (column sub-behaviour in T able 1) was no t feasible.
Tim e b u d g e t analysisT h e SA8 m odel was applied to classify behaviour per
accelerom eter segm ent an d then associated to the respective GPS fix (time an d location). Subsequently, the tim e spent on five
different behavioural activities (‘fly’, ‘forage’, ‘body care ’, ‘stand’ and ‘sit’) in several habitats was calculated for each bird. T h e time budget analysis results for the b ird fitted with logger 169 (a female) are presented in Figure 6 an d for birds fitted with loggers 166 and 167 (both males) in Figures SI an d S2. T he tim e budgets differed betw een individuals p redom inantly in w here an d w hen they spent their tim e on different activities ra th e r than the total p roportion of tim e spent on any one activity. All three birds spent a similar
PLoS ONE I www.plosone.org 7 May 2012 | Volume 7 | Issue 5 | e37997
Figure 4. Decision tree and confusion m atrix for model SA8. The num ber of observations correctly classified per behaviour is show n in bold. See Table 2 for a description o f th e predictor variables. doi:10.1371/journal.pone.0037997.g004
proportion o f tim e during the day foraging as during the night (166: 39% an d 45% ; 167: 38% and 40% ; 169: 37% and 38% respectively). Individual 169 spent m ost o f this tim e foraging on the mudflats. W hen on the salt m arsh, w hich functions p redom inantly as a roosting site outside the b reeding season, the three birds spent m ost o f their tim e on other activities, such as standing, sitting (barely at night) and body care (Figure 6). All three birds spent relatively little tim e in flight during the day and at night (< 2 % o f total time) and spent relatively m ore tim e during the day sitting than a t night. W hile the p roportion o f tim e spent foraging barely differed betw een day and night, the spatial distributions o f the classified behaviours clearly differed (Figure 6).
Discussion
Classification m o d e l sT h e prim ary aim o f this p ap er was to develop and assess
classification m odels to convert sensor da ta into specific behaviours observed in the field. As we expected, variables derived from body acceleration are clearly bette r predictors o f behaviour th an speed alone. Thus, w hen tracking anim als, collecting acceleration has a great added value if inform ation about behaviour is desired. How ever, since m any GPS tracking studies only provide inform ation on speed and location it is useful to note that ground speed m easured by the G PS can, in some cases, be used to distinguish flight from non-flight quite reliably. Yet the threshold will differ pe r species, flight strategy used (e.g. soaring or flapping
flight [59]) and environm ental conditions such as w ind speed and direction. In this study, 3.4 m s an d higher is associated with flight (Figure 3a and T able S2), however this threshold is based on a very small sample o f 13 observations m ade close to the nest. In a study on M anx Shearw aters (Puffinus puffinus), using a different m ethodology, a g round speed of 2.5 m s-1 was found as the optim al threshold betw een sitting and flying [60]. Since terrestrial locom otion in oystercatchers is quite slow (Table S2) and GPS speed is no t accurate enough [26], distinguishing betw een terrestrial locom otion and no locom otion is m ore difficult.
T h e m odels we developed can be applied to autom atically classify additional sensor da ta from the same individuals and potentially the same species. How ever, as w ith any m odel, if the dataset used to fit the m odel is very lim ited, for exam ple in the num ber o f m easurem ents per behaviour o r the environm ental conditions experienced, the chance o f misclassification m ay increase. In general, once behaviours can be reliably classified, the locom otion param eters such as flight speed, wing beat frequency, gait rates, odba can then be used for com parative analysis betw een species, individuals, environm ental conditions or for com parison w ith theoretical estim ates [37,40,61-63]. O ne aspect w hich deserves m ore a ttention in the future, especially w hen samples are large enough, is the extent to w hich predictor variables differ w ithin and betw een individuals. I f predictor variables differ significantly betw een individuals and enough data is available, then building an d applying m odels pe r individual m ay result in lower classification errors th an w hen using m odels
PLoS ONE I www.plosone.org 8 May 2012 | Volume 7 | Issue 5 | e37997
Figure 5. Examples of behaviours with characteristic signals from dynamic acceleration and static acceleration. Characteristic signals from dynam ic acceleration (A-D) and static acceleration (E-F) are show n. In all panels, acceleration in the surge (X) axis is show n with a continuous grey line, in the sway (Y) axis with a dashed line and in th e heave (Z) axis with a continuous black line. Fly and forage (A, B) are especially characterized by high-am plitudes of all dynam ic acceleration com ponents , b u t th e frequency of the signals is higher for fly than it is for forage (especially in th e Z direction, see dpsZ in Figure 4). Many of th e accelerom eter signals for foraging are characterized by the alternation betw een
PLoS ONE I www.plosone.org May 2012 | Volume 7 | Issue 5 | e37997
relatively sm ooth lateral m ovem ent (changes in acceleration predom inantly in the surge axis) and short bursts of high frequency changes in acceleration in all th ree axes (e.g. catching prey, a t 2.2 s panel B, see also Video S I). The changes in dynam ic acceleration for body care (C) are much smaller than for fly and forage, b u t still considerably higher than for stand and sit (D, see also odbaX and odba in Figure 4). The static acceleration can be used to distinguish sit (E) and stand (F) due to differences in body posture (see pitchX in Figure 4). doi:10.1371/journal.pone.0037997.g005
calibrated on pooled data. How ever, if predictive variables are robust enough, they could encom pass individual variability.
U nfortunately, in the curren t study we could no t derive a reliable m odel th a t could classify the 16 sub-behaviours observed in the field, and could only classify five o f the eight m ain behaviours. Nevertheless, classification m odels could potentially be im proved in several ways in the future. W e strongly believe that video observations w ould be extrem ely useful for classifying behaviour, developing pred ic to r variables, and re-evaluating models [44—45]. As video can be observed again after the activity has taken place, synchronization betw een observations and m easurem ents can be im proved, observations and interpretations can be cross-validated an d the im portance o f context (for exam ple presence o f o ther individuals, o r past events) can also be considered w hen classifying behaviour. As hum ans we are not always conscious o f all the inform ation we are visually processing
to reach a certain conclusion an d yet w hen only using pa rt o f this inform ation for au tom ated classification we expect the same conclusions to em erge. By re-exam ining videos carefully we m ay be able to identify these gaps an d fill them . For exam ple, studying posture, properties o f m ovem ent and the m easurem ents sim ultaneously (see video SI), m ay provide a better understanding of how they are related and enable researcher to derive m ore suitable p red ic tor variables. T h e pred ic tor variables included in this study are all aggregate m easures which, for exam ple, do no t p a ram eterize dependencies w ithin the 3 s observation period, and are hence crude in some respects. A good exam ple is given in Figure 5B, showing alternating patterns (with regard to total energy as well as frequency) o f acceleration w ithin the 3-second observation period. Thus, predictor variables w hich account for dependencies w ithin an acceleration segm ent m ay also result in m odel im provem ent.
Day b e h a v io u r July
Territory45%
Saltmarsh15%
Mudflats39% m
m fly□ forage□ body care□ stand□ sit
i 1--------1--------1--------i--------10 20 40 60 80 100
% Behaviour
Night behaviour July
Territory55%
Saltmarsh11%
Mudflats34% 1
■ fly□ forages body care□ stand□ sit
i i i i I
20 40 80 80 100
% Behaviour
Figure 6. Diurnal and nocturnal tim e budget of one oystercatcher during July 2009, using model SA8 to classify behaviours. Diurnal (top) and nocturnal (bottom ) tim e b udgets for one oystercatcher (logger 169, Table S I) during July 2009, using m odel SA8 (Figure 4) to classify behaviours. The locations of each behaviour (fly, forage, body care, stand and sit) are presen ted on the m ap; th e colours of the icons on the m ap correspond to those in the tim e b u d g e t bar graphs. doi:10.1371/journal.pone.0037997.g006
PLoS ONE I www.plosone.org 10 May 2012 | Volume 7 | Issue 5 | e37997
Q uan tify ing b e h a v io u r in s p a c e a n d t im eBy com bing inform ation on the location o f the b ird an d the time
from the GPS, behaviour from the accelerom eter an d inform ation abou t the environm ent we can calculate spatio-tem poral activity budgets for com parative analysis. T h e strength o f this approach is th a t once a classification m odel is built, it can be applied to data w here additional observations (visual o r video) are no t available or no t possible. In the cu rren t study we apply the classification m odel to da ta from three individuals for w hich sim ultaneous observations were no t always available. O n e aspect we were interested in was a com parison o f d iurnal and nocturnal tim e budgets as oystercatch- ers are know n to forage a t night in tidal areas. In a GPS tracking study o f oystercatchers in the W adden Sea [64], the authors showed th a t oystercatchers travel farther a t night than during the day, suggesting th a t they foraged extensively a t night, although inform ation on behaviour was no t available. In our study, we showed th a t a lthough individuals visited different locations during the day and night, all three individuals spent similar proportions o f tim e foraging during the night as during the day (Figures 6, S I, S2). O u r study also showed th a t the three individuals spent very little tim e in flight (>2% ) bo th inside an d outside o f the territory, w hich is similar to findings from a tim e budget analyses based on visual observations w ithin the territory and im m ediate surroundings [53,65]. Furtherm ore, our study supports previous suggestions th a t oystercatchers forage predom inantly in their territory an d in the m udflats close by [53,65]. W hile we canno t generalize these results on the basis o f the small sam ple used in this case study, it shows how these m ethods can be used to com pare tim e budgets w ithin and betw een individuals. In the future, we will apply the classification m odel in the future to a longer tim e series and m ore individuals to study inter-seasonal carry-over effects o f hab ita t selection and tim e-activity budgets. In this context, the type of tracking system is very relevant, the UvA -BiTS enables the retrieval o f da ta or re-program m ing the sensors rem otely while with m ost o f the com m ercially available tracking equipm ent an individual m ust be recap tured to retrieve the da ta (e.g. [27,64]).
M e th o d o lo g ica l w orkflowT h e m ethodological workflow presented here can be used for
similar studies regardless o f the study species o r the environm ent in w hich the study is conducted (e.g. terrestrial or m arine). By im plem enting such a workflow in a p rogram m ing language with a connection to a database w here the da ta is stored, the researcher greatly facilitates the reproducibility o f results, re-analysis, m odel im provem ent, knowledge transfer and collaboration, especially for researchers first entering the field o f bio-logging. T o facilitate the transfer o f knowledge, we have provided a m odelling package (Dataset SI) w hich includes a database an d R-scripts w ritten to ru n the analysis in this study. As shown in this study, several processes m ay be iterative, such as the specification of m odel predictors o r m odel design; each iteration m ay im prove our understand ing an d in terpreta tion o f the da ta as well as our models and a m ethodological workflow can stream line this process.
C o n c lu d in g rem arksT h e application of accelerom eters in behavioural research has
greatly increased in the last few years. Similarly, new developm ents a t the interface o f ecology and com puter science m ay greatly facilitate the analysis, visualization and exploration o f such da ta [46,66]. R ecent studies have also shown th a t m easures o f dynam ic body acceleration can be used to estim ate energy expenditure in a num ber o f species during active locom otion as well as m ore sedentary behaviour [3,28,30,67-68]. T hus, the potential for using accelerom eters to quantify behaviour and
energy expenditure makes it a very powerful tool in ecological research. O nce different characteristics o f behaviour an d body locom otion are quantified they can be com pared betw een studies, individuals, species, environm ental conditions, seasons or even different life history stages such as m igratory com pared to foraging m ovem ents. C om parative studies m ay also help increase our understand ing o f biom echanics an d evolution o f locom otion [37]. Perhaps m ost exciting is the possibility to link behaviour and energy expenditure to space use and tim e a t the individual level to gain new insight into the ability o f anim als to ad ap t to an ever changing world. In this study we provided a b lueprin t for the developm ent an d application o f classification m odels for this purpose.
Supporting InformationT ext SI E xtended m ethods.(PDF)
Figure SI D iu r n a l a n d n o c tu r n a l t im e b u d g e t o f o n e o y s te r c a tc h e r d u r in g J u ly 2 0 0 9 , u s in g m o d e l SA8 to c la s s i fy b e h a v io u r s . D iurnal (top) and nocturnal (bottom) time budgets for one oystercatcher (logger 166, T ab le SI) during Ju ly 2009, using m odel SA8 (Figure 4) to classify behaviours. T he locations o f each behaviour (fly, forage, body care, stand and sit) are presented on the m ap; the colours o f the icons on the m ap correspond to those in the tim e budget graph.(PDF)
Figure S2 D iu r n a l a n d n o c tu r n a l t im e b u d g e t o f o n e o y s te r c a tc h e r d u r in g J u ly 2 0 0 9 , u s in g m o d e l SA8 to c la s s i fy b e h a v io u r s . D iurnal (top) and nocturnal (bottom) time budgets for one oystercatcher (logger 167, T ab le SI) during Ju ly 2009, using m odel SA8 (Figure 4) to classify behaviours. T he locations o f each behaviour (fly, forage, body care, stand and sit) are presented on the m ap; the colours o f the icons on the m ap correspond to those in the tim e budget graph.(PDF)
T ab le SI T h e total num ber o f G PS fixes and accelerom eter segments (3 s intervals) obtained from the date o f deploym ent th rough 31 Ju ly 2009 for each o f the three oystercatchers in this study. Individual ring code, logger num ber, sex and body mass (g) on date o f deploym ent are also provided.(PDF)
T ab le S2 List o f behaviours observed in the field an d the m ean and standard deviation o f the p red ictor variables pe r behaviour according to T able 1 as follows: T ab le S2-A, 3-class m odel (S3 and SA3) behaviours (Table 1 colum n 4); T ab le S2-B, behaviours for SA8 m odel (Table 1 colum n 1); T ab le S2-C, 16 sub behaviours (Table 1, colum n 2). T h e pred ic to r variables are described in T ab le 2.(PDF)
D a ta se t SI A d a ta s e t a n d so ftw a r e p a c k a g e . T h e R-scripts and dataset for this study can be found in this self-contained archive w hich also includes a readm e-file th a t explains its contents. (ZIP)
V id eo SI A short video of an oystercatcher foraging by sight (Table 1) shown sim ultaneously w ith corresponding dynam ic and static acceleration in the heave axis (green), surge axis (red) and sway axis (blue) in units o f g (1 g = 9.8 m /s 2). T he m easurem ent d uration is 10 s, the film is shown at a slower rate. This record was no t included in this study.(WMV)
PLoS ONE I www.plosone.org 11 May 2012 | Volume 7 | Issue 5 | e37997
AcknowledgmentsM any people contributed to the Oystercatcher population study, especially J . Hulscher, M. Kersten, D. Heg, L. Bruinzeel and S. Verhuist. Vital support for the population study was initially provided by the late R.H. D rent and more recently by J.M . Tinbergen, including many stimulating scientific discussions. We are very grateful to Natuurm onum enten for allowing us to work in the N ational Park Schiermonnikoog. The D utch ethics committee on animal experiments (DEC) of the Royal Netherlands Academy of Arts and Sciences (KNAW) approved our research plans on fitting GPS-loggers to Oystercatchers on 27 M arch 2008. Hedwig Ens and Jeroen O nrust provided assistance with the field work. Edwin Baaij
References1. N o rb e rg R A (1977) A n ecological th eo ry o n fo rag in g tim e a n d en erg e tics a n d
cho ice o f o p tim a l fo o d -sea rch in g m eth o d . J A n im E co l 46: 511—529.2. R o p e r t-C o u d e r t Y , G rém ille t D , K a to A , R y a n P G , N a ito Y, e t al. (2004) A fm e-
scale tim e b u d g e t o f C a p e g an n e ts p ro v id es insights in to th e fo rag in g stra teg ies o f coasta l seabirds. A n im B ehav 67: 98 5 —992.
3. H alsey L G , W h ite C R (2010) M e asu rin g E nergetics a n d B eh av io u r U sing A cce le ro m etry in C a n e T o a d s Bufo marinus. PL oS O N E 5: e l0 1 7 0 .
4. Aviles J M (2004) C o m m o n c ran es G ru s grus a n d h a b ita t m a n a g e m e n t in ho lm oak d ehesas o f S pain . B iodivers C o n se rv 13: 2015—2025.
5. B o g rad SJ, B lock BA, C o sta D P , G od ley BJ (2010) B io logging technologies: new tools fo r conserv a tio n . In tro d u c tio n . E n d a n g e re d Species R e se a rc h 10: 1—7.
6. P ich eg ru L , R y a n P, d e r L in g en C , C o e tzee J , R o p e r t-C o u d e r t Y , e t al. (2007) F o rag in g b eh a v io u r a n d energe tics o f C a p e g an n e ts M orus capensis feed in g o n live p rey a n d fishery d iscards in th e B enguela u pw elling system . M a r E co l-P ro g Ser 350: 1 27-136 .
7. G a id e t N , C ap p e lle J , T a k ek aw a J Y , P rosser D J, Iv erso n SA, e t al. (2010) P o ten tia l sp re a d o f h igh ly p a th o g e n ic av ian in flu en za H 5 N 1 b y wildfowl: d ispersal ran g es a n d rates d e te rm in e d f ro m large-scale satellite telem etry . J A pp l E co l 47: 1 1 4 7 -1 1 5 7 .
8. C u m m in g G S , H o ck ey P A R , B ru inzee l L W , D u Piessis M A (2008) W ild b ird m o v em en ts a n d a v ian in fluenza risk m a p p in g in S o u th e rn A frica. E co l Soc 13:26.
9. M cL ellan B N , S h ack le to n D M (1988) G rizz ly b e a rs a n d reso u rce -ex trac tio n industries: effects o f ro ad s o n b eh a v io u r, h a b ita t use a n d d e m o g rap h y . J A pp l E co l 25: 4 5 1 M 6 0 .
10. S u th e rla n d W J, A rm stro n g -B ro w n S, A rm sw o rth P R , T o m B, B r ick la n d J , e t al. (2006) T h e id en tificatio n o f 100 eco log ical questio n s o f h ig h po licy re lev an ce in th e U K . J A p p l E co l 43: 6 1 7 -6 2 7 .
11. W e im ersk irch H , A nce l A , C a lo in M , Z ah a rie v A , S pag iari J , e t al. (2003) F o rag in g efficiency a n d a d ju s tm e n t o f en erg y ex p e n d itu re in a p e lag ic seab ird p ro v is io n in g its chick. J A n im E co l 72: 500—508.
12. C a m p h u y sen C J, S h am o u n -B aran es J , B o u ten W , G a r th e S (2012) Iden tify ing ecologically im p o r ta n t m ar in e a reas fo r seab irds using b e h a v io u ra l in fo rm a tio n in c o m b in a tio n w ith d is tr ib u tio n p a tte rn s . B io l C o n se rv doi: 1 0 .1 0 1 6 / j .b io c o n .2 0 1 1.12.024.
13. R o b in so n W D , Bow lin M S , Bisson I, S h am o u n -B aran es J , T h o ru p K , e t al. (2010) In te g ra tin g con cep ts a n d techno log ies to a d v a n ce th e study o f b ird m ig ra tio n . F ro n t E co l E n v iro n 8: 354—361.
14. M illsp au g h J J , M a rz lu ff J M , eds. R a d io track in g a n d a n im a l p o p u la tio n s. S an D iego , C A : A cad em ic Press. 4 7 4 p.
15. K u h n G E , J o h n s o n D S , R e a m R R , G e la tt T S (2009) A d v an ces in th e track in g o f m ar in e species: using G P S loca tions to eva lua te sa tellite track d a ta a n d a co n tin u o u s-tim e m o v em e n t m odel. M a r E co l-P ro g S er 393: 97—109.
16. H a r t K M , H y re n b a c h K D (2009) Satellite te lem e try o f m arin e m eg av erteb ra tes : th e co m in g o f age o f a n ex p e rim e n ta l science. E n d a n g e re d Species R esea rch 10: 9 -2 0 .
17. B ridge E S , T h o ru p K , B ow lin M S , C h ilso n PB , D ieh l R H , e t al. (2011) T ech n o lo g y o n th e M ove: R e c e n t a n d F o rth c o m in g In n o v atio n s fo r T ra ck in g M ig ra to ry Birds. B ioS cience 61: 6 8 9 -6 9 8 .
18. R u tz C , H ay s G C (2009) N ew fro n tie rs in b io logg ing science. B iol L e tt 5:2 8 9 -2 9 2 .
19. R o p e r t-C o u d e r t Y , W ilson R P (2005) T re n d s a n d perspec tives in an im al- a tta c h e d rem o te sensing. F ro n t E co l E n v iro n 3: 437—444.
20. C o o k e SJ, H in c h SG , W ikelski M , A ndrew s R D , K u c h e l LJ, e t al. (2004) B io telem etry : a m ech an istic a p p ro a c h to ecology. T re n d s E co l E vol 19:3 3 4 -3 4 3 .
21. K laassen R , S tra n d b e rg R , H a k e M , A le rs tam T (2008) F lexibility in daily trave l ro u tin es causes reg io n a l v a ria tio n in b ird m ig ra tio n speed. B ehav E co l Sociobiol 62: 1 4 2 7 -1 4 3 2 .
22. S h am o u n -B aran es J , B a h a ra d A, A lp e rt P , B e rth o ld P , Y o m -T o v Y , e t al. (2003) T h e effect o f w ind , season a n d la titu d e o n th e m ig ra tio n sp eed o f w h ite storks Ciconia ciconia, a lo n g th e ea ste rn m ig ra tio n ro u te . J A v ian Biol 34: 97—104.
23. T h o ru p K , A le rs tam T , H a k e M , K je llen N (2006) T ra v e lin g o r sto p p in g o f m ig ra tin g b ird s in re la tio n to w ind: a n illu stra tio n for th e osprey. B ehav E co l 17: 4 9 7 -5 0 2 .
provided assistance with the tracking system. W e thank Yaiza Dronkers Londono and Merijn de Bakker for the data and software needed for the video used in this paper. We thank Ja n Ropert-Coudert and two anonymous reviewers for their constructive feedback on an earlier version of this manuscript.
Author ContributionsConceived and designed the experiments: JSB RB EEvL BE K O WB. Performed the experiments: RB K O . Analyzed the data: RB EEvLJSB. Contributed reagents/m aterials/analysis tools: EEvLJSB RB WB. W rote the paper: JSB RB EEvL BE WB.
24. S h am o u n -B aran es J , B o u ten W , C a m p h u y sen C J, Baaij E (2011) R id in g the tide: in tr ig u in g o b serva tions o f gulls res tin g a t sea d u r in g b reed in g . Ib is 153: 4 1 1 M 4 5 .
25. G u ilfo rd T , M e ad e J , W illis J , P h illips R A , Boyle D , e t al. (2009) M ig ra d o n a n d s to p o v er in a sm all p e lag ic seab ird , th e M a n x sh e arw ate r Puffinus puffinus: insights f ro m m ac h in e lea rn in g . P R o y Soc B-B iol Sei 276: 1 2 1 5 -1 2 2 3 .
26. W itte T H , W ilson A M (2004) A ccu racy o f n o n -d iffe ren d a l G P S for the d e te rm in a d o n o f sp eed o v er g ro u n d . J B iom ech 37: 1891—1898.
27. W ilson R P , S h e p a rd E L G , L iebsch N (2008) P ry in g in to th e in tim a te details o f a n im a l lives: use o f a daily d iary o n anim als. E n d a n g e re d Species R e se a rc h 4: 12 3 -1 3 7 .
28. G re e n J A , H a lsey L G , W ilso n R P , F rap p e ll PB (2009) E stim a tin g energy ex p e n d itu re o f an im als using th e acce le ro m etry techn ique : activity , inactiv ity a n d c o m p ariso n w ith th e h e a rt-ra te tech n iq u e . J E x p B iol 212: 47 1 —4-82.
29. B id d er O R , S oresina M , S h e p a rd E L G , H alsey L G , Q u in ta n a F , e t al. (2012) T h e n e e d for speed: tes tin g a c ce le ra tio n fo r e s tim a tin g a n im a l trav e l ra tes in te rres tria l d ead -reck o n in g system s. Z oo logy 115: 58—64.
30. G leiss A G , W ilson R P , S h e p a rd E L G (2011) M a k in g overall d y n a m ic b o d y a c ce le ra tio n w ork: o n th e th eo ry o f a c ce le ra tio n as a p ro x y fo r energy e x p en d itu re . M e th o d s in Ecology a n d E vo lu tio n 2: 23—33.
31. R o p e r t-C o u d e r t Y , D a u n t F, K a to A , R y a n P G , Lewis S, e t al. (2009) U n d e rw a te r w in g b ea ts e x te n d d e p th a n d d u ra tio n o f p lu n g e dives in n o r th e rn M orus bassanus. J A v ian Biol 40: 380—387.
32. G ó m ez L a ich A , W ilson R P , Q u in ta n a F , S h e p a rd E L G (2008) Id en tifica tio n o f im p e r ia l c o rm o ra n t Phalacrocorax atriceps b e h a v io u r u s in g acce le ro m e te rs . E n d a n g e re d Species R e se a rc h 10: 29—37.
33. M ita n i Y , A ndrew s R D , Sato K , K a to A , N a ito Y , e t al. (2010) T h re e - d im en sio n a l res tin g b eh a v io u r o f n o r th e rn e le p h a n t seals: d riftin g like a falling leaf. B iol L e tt 6: 1 6 3 -1 6 6 .
34. W e im ersk irch H , Le C o r re M , R o p e r t-C o u d e r t Y , K a to A, M arsac F (2005) T h e th ree -d im en s io n a l fligh t o f red -fo o ted boobies: a d a p ta tio n s to fo rag in g in a tro p ica l e n v iro n m en t? P R o y Soc B-Biol Sei 272: 53—61.
35. Y o d a K , K o h n o H , N a ito Y (2004) D e v e lo p m e n t o f flig h t p e rfo rm a n c e in the b ro w n bo o b y . P R o y Soc B-Biol Sei 271: S 2 4 0 -S 2 4 2 .
36. M ille r I J O , J o h n s o n M P , T y ack PL, T e r ra y E A (2004) S w im m ing gaits, passived ra g a n d b u o y a n c y o f d iv ing sp e rm w hales Physeter macrocephalus. J E x p B iol 207: 1 9 5 3 -1 9 6 7 .
37. G leiss A G , J o rg e n se n SJ, L iebsch N , Sala J E , N o rm a n B, e t al. (2011) C o n v e rg e n t ev o lu tion in lo co m o to ry p a tte rn s o f flying a n d sw im m ing anim als. N a t C o m m u n 2: 352.
38. O k u y a m a J , K a w a b a ta Y, N a ito Y , A ra i N , K o b ay ash i M (2009) M o n ito rin gb eak m o v em en ts w ith a n acce le ra tio n datalogger: a useful tec h n iq u e forassessing th e feed in g a n d b re a th in g b eh av io rs o f sea tu rdes. E n d a n g e re d Species R e se a rc h 10: 3 9 -4 5 .
39. W h itn e y N , P ra tt H J , P ra tt T , C a rrie r J (2010) Iden tify in g sh ark m atin g b eh a v io u r using th ree -d im en s io n a l acce le ra tio n loggers. E n d a n g e re d Species R e se a rc h 10: 7 1 -8 2 .
40. R o p e r t-C o u d e r t Y , B eau lieu M , H a n u ise N , K a to A (2009) D iv ing in to the w o rld o f b io logging . E n d a n g e re d Species R e se a rc h 10: 21—27.
41. H o llan d R A , W ikelski M , K ü m m e th F, B osque G (2009) T h e S ecre t Life o f O ilb irds: N ew Insights in to th e M o v e m e n t Ecology o f a U n iq u e A vian Frug ivore . PL oS O N E 4: e8264.
42. S h e p a rd E L G , W ilson R P , Q u in ta n a F , G ó m e z L a ich A, L iebsch N , e t al. (2008) Id en tifica tio n o f a n im a l m o v em e n t p a tte rn s using tri-ax ia l acce le ro m etry . E n d a n g e re d Species R e se a rc h 10: 47—60.
43. S h e p a rd E L G , L am b ertu cc i SA, V a llm itjan a D , W ilson R P (2011) E n erg y B ey o n d Food: F o rag in g T h e o ry In fo rm s T im e S p e n t in T h e rm a ls b y a L arge S o a rin g Bird. PL oS O N E 6: e27375.
44. Y o d a K , N a ito Y , Sato K , T a k a h a sh i A , N ish ikaw a J , e t al. (2001) A new tec h n iq u e for m o n ito rin g th e b eh a v io u r o f free -ran g in g A delie p engu ins. J E xp Biol 204: 6 8 5 -6 9 0 .
45. H alsey L G , P o r tu g a l SJ, S m ith J A , M u rn G P , W ilson R P (2009) R e c o rd in g r a p to r b eh a v io r o n th e w in g v ia acce le ro m etry . J F ield O rn ith o l 80: 171—177.
46. S ak am o to K Q , S ato K , Ish izuka M , W a ta n u k i Y , T a k a h a sh i A , e t al. (2009) G a n E th o g ra m s Be A uto m atica lly G e n e ra te d U sin g B ody A cce le ra tio n D a ta fro m F ree -R a n g in g B irds? P L oS O N E 4: e5379.
PLoS ONE I www.plosone.org 12 May 2012 | Volume 7 | Issue 5 | e37997
47. W a ta n a b e S, Izaw a M , K a to A, R o p e r t-C o u d e r t Y , N a ito Y (2005) A newtec h n iq u e fo r m o n ito rin g th e d e ta iled b e h a v io u r o f te rres tria l an im als: A casestu d y w ith th e dom estic cat. A p p l A n im B ehav Sei 94: 117—131.
48. H a g e m eije r W JM . B lair M J, eds. T h e E B C C A d as o f E u ro p e a n B reed in g Birds: th e ir D is tr ib u tio n a n d A b u n d an ce . L o n d o n : T & A D Poyser.
49. E n s BJ, Safrie l U N , H a r r is M P (1993) D iv o rce in th e lo n g -liv ed a n dm o n o g am o u s o y ste rca tch e r, Haematopus ostralegus: in co m p a tib ility o r ch oosing th e b e tte r op tio n ? A n im B eh av 45: 1 1 9 9 -1 2 1 7 .
50. H e g D , E n s BJ, V a n D e r J e u g d H P , B ru inzee l L W (2000) L oca l d o m in a n c e a n d te rr ito ria l s e td e m en t o f n o n b ree d in g oyste rcatchers. B eh av io u r 137: 473—530.
51. E n s BJ, K e rs te n M , B ren n in k m eije r A , H u lsc h e r J B (1992) T e rrito ry quality , p a re n ta l effo rt a n d rep ro d u c tiv e success o f O y s te rc a tc h e rs (.Haematopus ostralegus). J A n im E co l 61: 7 0 3 -7 1 5 .
52. P o lM v d , B ro u w er L, E ns BJ, O o s te rb e e k K , T in b e rg e n J M (2009) F lu c tu a tin g selection a n d th e m a in ten a n c e o f ind iv id u a l a n d sex-specific d ie t specia liza tion in free-liv ing oyste rcatchers. E vo lu tio n 64: 8 3 6 —851.
53. K e rs te n M (1996) T im e a n d en erg y b u d g e ts o f o y ste rcatch ers H a e m ato p u s ostra legus o ccu p y in g terr ito rie s o f d iffe ren t quality . A rd e a 84: 291—310.
54. B re im an L , F r ie d m a n J H , O lsh e n R A , S to n e C J (1984) C lassification a n d reg ression trees. M o n te rey , CA : W ad sw o rth a n d B ro o k s /C o le .
55. P ra sa d A , Iv erso n L, L iaw A (2006) N ew er C lassification a n d R eg ress io n T re e T ech n iq u es: B agging a n d R a n d o m F orests fo r E cological P red ic tio n . Ecosystem s 9: 181-199 .
56. T h e m e a u T M , A tk inson E J (1997) A n in tro d u c tio n to recursive p a rtitio n in g us in g th e R P A R T ro u tin e s . D e p a r tm e n t o f H e a l th S c ien ces R e se a rc h , R o ch es te r , M N : M ay o C linic.
57. R D e v e lo p m e n t C o re T e a m (2011) R : A lan g u ag e a n d e n v iro n m en t for sta tistical c o m p u tin g . R F o u n d a tio n fo r S tatistical C o m p u tin g , V ie n n a , A ustria .
. PLoS ONE I www.plosone.org
IS B N 3 -9 0 0 0 5 1 -0 7 -0 {R w eb site . A vailab le : h t tp : / /w w w .R - p r o j e c t .o r g / . A ccessed 2012 M a y 2.
58. H a s tie T , T ib sh ira n i R , F r ie d m a n J (2009) T h e E lem en ts o f S tatistical L earn ing : D a ta M in in g , In fe ren ce a n d P red ic tio n . Berlin: S pringer-V erlag .
59. S h am o u n -B aran es J , v a n L o o n E (2006) E n e rg e tic in fluence o n guii fligh t s tra tegy selection . J E x p Biol 209: 3 4 8 9 -3 4 9 8 .
60. G u ilfo rd T C , M e ad e J , F re em a n R , B iro D , E vans T , e t al. (2008) G P S track in g o f th e fo ra g in g m o v em en ts o f M a n x S hea rw a te rs Puffinus puffinus b ree d in g o n S k o m er Island , W ales. Ib is 150: 4 6 2 -4 7 3 .
61. B ru d e re r B, B o ld t A (2001) F lig h t ch arac te ristics o f b irds: I. r a d a r m easu rem en ts o f speeds. Ib is 143: 17 8 -2 0 4 .
62. B ru d e re r B, P e te r D , B o ld t A , L iech ti F {2010) W in g -b e a t ch arac te ristics o f b irds rec o rd e d w ith track in g r a d a r a n d cine cam era . Ib is 152: 272—291.
63. Z ijlstra W , H o f A L (2003) A ssessm en t o f sp a tio -te m p o ra l g a it p a ra m e te rs fro m tru n k acce lera tio n s d u r in g h u m a n w alking. J B iom ech 18: 1—10.
64. S ch w em m er P, G a r th e S {2011) S p a tia l a n d te m p o ra l p a tte rn s o f h a b ita t use by E u ra sia n o y ste rcatch ers {.Haematopus ostralegus) in th e ea ste rn W a d d e n Sea rev ea led using G P S d a ta loggers. M a r B iol 158: 5 4 1 -5 5 0 .
65. K e rs te n M , V isser W {1996) F o o d in tak e b y O y s te rc a tc h e rs Haematopus ostralegus b y d ay a n d b y n ig h t m ea su red w ith a n e lec tro n ic n es t b a la n c e . A rd e a 84: 57—72.
66. S h am o u n -B aran es J , v a n L o o n E E , P urves R S , S p eck m an n B, W eisk o p f D , e t al. {2012) A nalysis a n d v isu a liza tio n o f a n im a l m o v em en t. B iol L e tt 8: 6—9.
67. Fossette S, Schofie ld G , Tilley M K S , G leiss A G , H ay s G G {2012) A cce le ra tio n d a ta rev ea l th e en erg y m a n a g e m e n t s tra teg y o f a m ar in e ec to th e rm d u r in g rep ro d u c tio n . F u n c t E co l 26: 3 2 4 -3 3 3 .
68. W ilson R P , W h ite G R , Q u in ta n a F , H alsey L G , L iebsch N , e t al. {2006) M o v in g to w ard s a c ce le ra tio n fo r estim ates o f activity-specific m etab o lic ra te in free-liv ing anim als: th e case o f th e c o rm o ran t. J A n im E co l 75: 1081—1090.
