Detecting Movement Patterns With Wireless SensorNetworks: Application To Bird Behavior

Erick Stattner, Martine Collard, Philippe Hunel, Nicolas VidotLAMIA, University of Antilles and Guyane, France

e.stattner@gmail.com, {mcollard, phunel, nvidot}@univ-ag.fr

ABSTRACTMore and more animal species are endangered every dayon earth. In order to study their adaptation to world andclimate change and their chances of survival, numerous ini-tiatives have been taken that mostly need human intrusioninto animal communities. Today mobile devices enable re-searchers to go beyond this limit. In this paper, we proposean original solution that consists on a new framework for de-tecting individual songs in a bird population and identifyingremotely by this way their collective behavior in movementswithout human interaction. Movement patterns are elicitedby analyzing data collected via wireless sensors fitted withmicrophone. Whereas similar methods use mobile devicesfitted on some specimens, we rather propose fixed sensors.We demonstrate that this solution provides a good answer totechnical constraints assessed by the context and we discussresults of experimental simulations that allow to define op-timized parameters for the architecture to be set up on theground. Experimental results are provided and show therelative impact of different parameters such as the numberof sensors or the population size on the detection rate.

Categories and Subject DescriptorsC.2.1 [Computer-Communication Networks]: NetworkArchitecture and Design; H.2.8 [Database Management]:Database Applications; I.6.3 [Simulation and Modeling]:Applications; H.5.5 [Information Interfaces and Pre-sentation]: Sound and Music Computing

Keywordswireless sensors, spatio-temporal data, trajectories, move-ment pattern mining, bird behavior, habitat monitoring

1. INTRODUCTIONNew challenges regarding our planet appeared in the last

decades. Environmental troubles threatening our ecosystemhave led scientists to work for meeting alarming issues such

Permission to make digital or hard copies of all or part of this work forpersonal or classroom use is granted without fee provided that copies arenot made or distributed for profit or commercial advantage and that copiesbear this notice and the full citation on the first page. To copy otherwise, torepublish, to post on servers or to redistribute to lists, requires prior specificpermission and/or a fee.MoMM2010 8-10 November, 2010, Paris, FranceCopyright 2010 ACM 978-1-4503-0440-5/10/11 ...$10.00.

as global warming, pollution, exhaustion of energy resourcesor animal species disappearance. This last phenomenon, of-ten a sign of an unbalanced environment, has motivated nu-merous scientific experiences to observe the behavior of ani-mal populations in their original area in order to reintroducethem in similar and protected environments.

Movements act as a fundamental witness of animal speciesbehavior, especially for understanding the way endangeredcommunities are organized and how they live in an area [5,9]. Changes which occur in their habitat and local abioticfactors such as wind, noise or temperature are usually thecause of major modifications in fundamental ecological pro-cesses, such as animal movements. Consequently, studyingmovements provides important information not only on so-cial structure and habits, but on the way these movementsevolve and which elements influence them.

However this task is not easy especially in the case of birdssince environments are generally unsuitable to large-scalehuman studies that are unable to catch real motions andmay only reveal a partial view of them. Moreover, humanobservation [10], which is the current method for studyingbirds behavior may itself cause changes in their movements.

In this context, mobile devices fitted on some individualshave provided new perspectives enabling them to track an-imal motions remotely. Miniaturization and technologicaladvancements in geographic locator devices both at com-putational levels and in terms of memory capacity are in-creasing and have recently allowed researchers to foreseenew perspectives in many fields. Since they are able to pro-vide regular information on their positioning in time andspace, micro-controllers are used to collect spatial-temporaldata for further analyzes of various problems. Moreover,this technique is not conceivable for observing little birdsthat may greatly participate in the ecological balance andneed attention. While they bring a very interesting alterna-tive to physical intrusion in animal communities areas, theynevertheless need at least a minimal human action that issometimes not advisable. We have focused on this particu-lar issue with the objective to provide an original alternativethat combines wireless sensor networks to collect data anddata mining techniques to process them.

In our work, the objective is to analyze and mine spatio-temporal data collected on movements of an endangeredbird population for eliciting patterns among animal motionsand thus improve current knowledge on it. We have beenparticularly interested in the case of the “Moqueur GorgeBlanche”, an endangered species endemic to the island ofMartinique. Since we faced the need to exclude any human

interaction in the bird area, we have proposed a new archi-tecture based on sensors fitted with microphones where sen-sors are located randomly as if they were thrown by plane.Spatio-temporal data deduced from song records are broad-casted through a two layer sensor network and they are an-alyzed for eliciting flock movement patterns.

This work explains how our approach differs from othermethods and its advantages and issues. We also presentour simulation tool, called Lypus, which provides a way tomodelize bird behaviour in their habitat.In this first step,our objective is to simulate various configurations in orderto define optimal parameters for the real architecture. Tothe best of our knowledge, this kind of framework based onbird song record and wireless sensor fitted with microphoneshas not yet been exploited to elicit flock motion patterns.

The paper is organized as follows. Section 2 presents anoverview of main related works. In Section 3, we describethe proposed sensor architecture and we give a more for-mal definition of movement patterns. Section 4 is devotedto the pattern elicitation algorithms involved. Section 5 de-tails experimental results obtained by simulating the sensornetworks with various parameters. Finally, in Section 6, weconclude and present our future research directions.

2. RELATED WORKRecent development of technologies for real-time tracking

of moving entities (GPS, mobile phones, RFID chips, etc.)has allowed the collection and exploitation of a new type ofdata so called spatio-temporal data. The exploitation of thiskind of data to track spatio-temporal movements of groupsor communities has motivated early works. However, mostmethods on moving entities (humans, animals or objects) as-sume either the observation by a human observer, or mobiledevices placed on the entities.

A first study initiated in movement patterns focused onflock movement [23, 19]. A flock pattern in a time intervalT can be defined informally by the movement of m entities,such as for each point taken in the time interval T , we canfind the same set of m entities at a different location. Forexample, this problem was studied in [3] by Benkert et al.and assumed human interaction and manual data collection.Subsequently, it was established that collective movementsof entities are limited to a finite number of patterns. Thus,Laube et al. [22] defined a set of spatio-temporal modelsbased on the characteristic of movements such as flock pat-tern and periodic pattern. With a different point of viewGudmundsson and al. [16] distinguished new movement pat-terns. Other kinds of movement such as flock patterns [3],leadership patterns, periodic patterns [24] and meeting pat-terns or frequent patterns were identified too.

Various methods are proposed for detecting these patterns[21, 17] and most of them employ data mining techniques,including clustering [23, 19] and association rules [27, 6].Verhein [27] showed how it is possible to describe the entitiesmovements by association rules. Similarly, Gudmundsson etal. [16] provide a very complete overview of data miningcontributions in the development of algorithms for researchpatterns in the movement.

The low cost of micro-devices that provide spatial-temporalinformation about moving entities are now involved in theethology domain. Various studies using GPS collars to col-lect the animals motions were conducted. For example, GPScollars were experimented by Rumble and al. [26] on elks to

obtain their positions every twenty minutes and by Dumonand al. [11] on sheep for studying their collective motions.Associated with environmental data such as temperature,light or noise, spatial-temporal data can provide valuableinformation about the movement of endangered species andfacilitate their reintegration in similar environments.As proved by a lot of experiments, GPS systems gather sev-eral advantages. But in our context, it was not conceivableto consider such a solution for several reasons listed below.A GPS system only works effectively in open areas withoutany object which may obstruct the field of view of the re-ceiver [12]. Thus, operation under a dense foliage is difficult[4]. It is particularly expensive and for large populations ofbirds, it would be difficult to fit each bird with a GPS de-vice. It consumes energy enormously. It may have a sizeableweight, and often cannot be placed on a small bird becauseof risks it cannot fly anymore, or it is tired very rapidly. Itmay be harmfulness since it needs to operate not only in theanimal area but on selected specimens too.

Unlike GPS collars, sensors are not grafted onto a singleindividual, but simply set up on the target area. By fittingeach sensor with a microphone, it is possible to detect thepresence of a bird when it sings. Obviously, although thesensors do not have the same problems as the GPS devices,they also have some disadvantages that have to be consid-ered. First of all, as explained just above, sensors detect thepresence of a bird only if and when it sings. And we knowthat birds do not sing all the time. In addition, signalsrecorded by sensors do not allow to identify one individual.It only ensures to recognize the species. Since we are onlyinterested in collective patterns, this is not a real drawbackas we show in further sections.

Other technologies like mobile phones supply different me-dia to collect spatio-temporal data too. For instance, trian-gulation schemes are proposed [2] to locate, track and countmobile phones. For obvious reasons, this cannot be a solu-tion for this experiment.

3. DETECTING PATTERNIn the work presented here, the experimental context set

strong constraints that contribute to the originality of theapproach. Indeed we propose a quite novel solution with anetwork of fixed sensors to collect audio signals (bird songs)that are interpreted in terms of individual positions in spaceand in time. Two main correlated questions arise:(1) how to ensure a relevant individual localization with afixed sensor network? (2) how patterns can be elicited fromthese spatio-temporal data?

3.1 Our architectureIn order to configure the wireless sensors network fitted

with microphone, we assume that the target area is dividedinto sub-areas, that we call “regions” and that are supposedto be delimited by experts. Typically, a region may delimitan area around a water supply point. Another region maydefine an area of dense vegetation, or freshness, etc. Theadvantage of such a definition is that it eventually allows toobtain semantic information on the movements of individu-als based on characteristics of regions. Unfortunately, thisform of delimitation is not always possible. In a first stage,we define regions in an arbitrary manner.

Thus, each region has a unique identifier A, B, C, etc.Each sensor is attached to one region only. We will see that

A

B

C

D

E

F

G

I

H

J

K

L

Detection radiusIntra-region com.Inter-region com.

0

200

400

600

800

1000

250 500 750 1000

Figure 1: The sensor architecture involved in Lypusto collect data on birds

depending on the sensor placement method, a sensor can bein a region boundary and may have its detection area thatoverlap several regions. All sensors of a same region cancommunicate with each other through intra-region commu-nications. For example, these communications are used todetermine the number of birds in a region. The detail ofmethods that we defined for counting birds with a similarsensor network [14, 15] are out of scope in this paper. Theyuse triangulation techniques to locate songs and signal com-parisons. Thus in this paper, we are able to consider thatbirds in a region may be counted on the basis of their songs.

Each region has at least one specific sensor called“gateway-sensor”. It is able to communicate with other gateway-sensors of neighbor regions through inter-regions commu-nications. These communications allow regions to exchangeinformation and communicate with the central base respon-sible for collecting the whole records for further analysis.Figure 1 presents this architecture.

These two-layer communication system, intra-region andinter-region, not only enables sensors to locally determinethe detected birds, but prevents that redundant informationwas sent to the the central base. This has the effect ofseverely limiting useless exchanges in the network.

We assume that all sensors have the same properties: bothsame detection radius r and same energy level. Each sensorknows the region to which it is affiliated and its positionin this region. Sensors are able to detect an individual ofthe given species by analyzing its song [15], but they arenot able currently to differentiate individuals between them.They can communicate with the sensors in the same regionto send information to gateway-sensors. The gateway-sensordetermines the number of birds currently in its region andsends this information to the base station. The central baseknows all regions and receives the information on detectedbirds count in each region from gateway-sensors. In thiswork, we are not interested in the routing information tothe base station. We focus only on the study of movements.

Before presenting the problem more formally, it is impor-tant to detail sensor placement, i.e. how to locate sensorsagainst each other. Indeed, while only one region is assignedto a sensor, the detection radius of a sensor can cover severalregions and thus strongly influences the results. There aredifferent kinds of configuration:

Uniform configuration: this configuration consists onlaying out the sensors uniformly so that they cover the entirezone. Unfortunately, this scheme is not realistic, in the caseof dropping sensors by plane, as it is often the case.

Dense configuration: this configuration consists on plac-ing a maximum of sensors to cover the maximum area. This

configuration naturally raises problems of cost and interfer-ence between signals.

Incremental configuration: this method consists onoptimizing the overall calculation, with an approximate op-timal place for each new sensor integrated into the network.This mechanism is good for adjustment after installation.

Random configuration: this method consists on plac-ing the sensors randomly on the area, usually by droppingof the sensors by plane. This method is the most realistic,but it produces unbalanced sensor distribution.

Self-deployment: the objective of this method is tomaximize the surface covered by the network, and ensure theinclusion of obstacles. The disposition by self-deploymentshould be done independently by the nodes.

3.2 Problem definitionThe technical solution we adopted due to domain con-

straints provides only a limited set of collected information:a sound, the region where it has been recorded, an esti-mate of its position (using triangulation scheme) and therecording time. In this configuration, we assume that suchinformation is collected over a period long enough to be ex-ploited. Thus, whenever a desired sound is detected, forexample the specific song of a bird species, communicationsintra and inter regions are done to send to the central basethe information which will be used to study the movements.Then, the base station stores a dataset which is analyzedand investigated to discover patterns in movements.

More formally, let denote R = {r1, ..., rn} the set of all theregions which divide the area, and C = {c1, ..., cm} the setof all the sensors which cover the regions. Each sensor ci isassigned to a region r with r ∈ R. We denote I = 〈t1, ..., tk〉,the time sequence during which data are collected. We knowthat each gateway-sensor is able to count locally the numberof birds at a time t on its region [14, 15]. We denote Ft thecounting function which returns at each time t of I, thenumber of entities detected into a region.

∀t ∈ I, Ft : R→ N

Thus, the detection vector−→Dt, defined by (1), represents

the number of entities detected into all regions at a time t.

∀t ∈ I,−→Dt = [Ft(r1), ..., Ft(rn)] (1)

The movement analysis is done over long time periods.Thus, for a given time sequence I = 〈t1, ..., tk〉, the centralbase stores a n ∗ k matrix M , called “detection matrix”, inwhich each element Mij = Ftj (ri) with ri ∈ R and tj ∈ I.Each row i represents the detection into the region ri at eachtime in I. Similarly, a column j represents the number ofentities detected by each region in R at time tj . So each

column j is equivalent to−→Dj

t.

M =

Ft1(r1) Ft2(r1) ... Ftk (r1)Ft1(r2) Ft2(r2) ... Ftk (r2)Ft1(r3) Ft2(r3) ... Ftk (r3)... ... ... ...

Ft1(rn) Ft2(rn) ... Ftk (rn)

Figure 2(a) shows an example of the Flock movement of

three birds following globally a same path through regionsover a time sequence I = 〈t1, ..., t6〉. Data collected by sen-sors allow us to represent the detection in the form of thematrix presented in Figure 2(b).

t1 t2 t3 t4 t5 t6

A 3 3

B

C 2 3

D

E 1

F

G 1 2

H 3

(a) Movements of three birds

(b) Detection matrix

0

200

400

600

800

1000

250 500 750 1000

A

BC

D

E

F

G

I

H

J

K

L

Bird 1Bird 2Bird 3

Figure 2: An example of detection matrix for amovement of three birds

Different patterns may be elicited from lines and columnsof M : flock and periodic patterns. These data aggregatedby regions, enables to handle the “noise” which may existwhen considering individual position. Indeed, the entitiessurveyed have never exactly the same coordinates while mov-ing. Moreover, sensors that we use, can only provide anestimate of the position by triangulation techniques. How-ever, this region oriented approach is highly dependent onthe size of regions. If an area is too large, the movementmay be detected only in this region. On the other hand, ifregions are too narrow, movements may occur in differentregions. Anyway in the following subsections, we show howit is possible to find the two kinds of pattern with this sensorarchitecture.

3.3 Flock patternSeveral definitions have been given for the expression of

Flock pattern. Let us consider a sequence of successive timesI = 〈t1, ..., tk〉. Benkert et al. [3], define a Flock pattern asbeing a movement of at least m entities such as for each timein I, we can find a disk of radius v which contains m entities.However, in the context of this paper, we have to take aless precise definition for two reasons. First, the currentnetwork does not supply the birds position with precision.We cannot use disks of a given radius to seek such groups.In addition, the detection is highly dependent on bird songand we know that at a time t, all individuals in the groupdoes not necessarily sing. Thus, we are unlikely to find agroup of the same size exactly at each point in I. That iswhy, in our context, we give another definition.

Definition 1: Let α ∈ N be the detection margin ofa group, I a time sequence 〈t1, t2, ..., tk〉 with ∀j ∈ [1..k],tj < tj+1, R the set of given regions and E is a set ofm entities. We call α−Flock pattern over E, R and I, thesequence 〈rt1 , rt2 , ..., rtk 〉 of regions in R such as at each timetj of I, a maximal set of g entities among E occur in regionrtj with |m− g| ≤ α.

Thus, we consider a α−Flock pattern as a motion of al-most the entire group observed. If the whole area is divided

in regions so that they are not too large, we can considerthat the birds appearing in same region at same time allalong the I time sequence represent a flock movement. Theexample of the matrix M1 below, shows the flock movementof six birds whose songs have been partially recorded sincethey are not singing at every time in I. Indeed, it is expectedthat the number of birds detected is not always equal to thegroup size. This is the result of two phenomena inherentin our solution based on sound detection. First, sensors areonly fitted for sound recordings, but at a time t, birds inthe group are unlikely singing together. Secondly, the noisethat can occur on the signal may prevent the detection.

M1 =

6 6 0 0 0 0 0 0 0 00 0 0 0 0 0 0 2 5 60 0 5 5 0 5 5 4 0 00 0 1 0 6 0 1 0 0 0

According to the definition, an α−Flock pattern may be

seen as the sequence S = 〈rt1 , rt2 , ..., rtk 〉 of regions, if itexists, that maximizes the number of entities detected. Thus∀rtj ∈ S, rtj is the only region r of R, if it exists, such that:

Ftj (r) =n

maxl=1

Mlj and |m−Mij | ≤ α (2)

For now, our architecture only allows us to study the col-lective movements of the entire group. According to theabove definition, we consider that if at the same time t, sev-eral regions satisfy this definition, birds do not adopt a flockmovement. Thus, for this example if we take α = 2, wedetect a Flock pattern of the six birds group across regions〈A,A,C,C,D,C,C,C,B,B〉.

3.4 Periodic patternA periodic pattern is a movement which repeats itself after

a certain time called the period.Definition 2: Let us consider S as a sequence of re-

gions. We call periodic segment of period T , a sub-sequence〈rta , ..., rtb〉 of S, where (tb − ta) mod T = 0 and for each(tp, tq) ∈ N × N, a ≤ p ≤ b, a ≤ q ≤ b we have rtp = rtq iftp mod T = tq mod T . s = 〈ra, ra+1, ..., ra+T 〉 is repeatedin S that is called a periodic pattern of period T .

The detection matrix M2 below shows for instance a Pe-riodic pattern observed for six birds even if the number ofbirds detected is not always equal to the size of the group(cf. section 3.3).

M2 =

4 6 0 0 0 6 6 0 0 01 0 5 1 0 0 0 6 1 00 0 0 5 0 0 0 0 4 00 0 0 0 2 0 0 0 0 6

The analysis of M2 is done in two steps. As described

for Flock pattern, we begin by extracting from M2 the se-quence S which maximizes the number of detected entitiesas follows: ∀rtj ∈ S, rtj is the region r of R such that:

Ftj (r) =n

maxl=1

Mlj (3)

Then, we extract from S the periodic segment s whichrepeats itself in S. In this example, S matches the periodicsegment exactly. A periodic pattern is not a flock pattern,it only reveals tendencies to adopt periodic paths throughregions.

For instance, the sequence 〈A,A,B,C,D,A,A,B,C,D〉is deduced from M2. The objective is to extract from Sthe periodic segment 〈A,A,B,C,D〉. How can we find thissegment? Some algorithms had already been developed tomining periodic pattern in spatio temporal data. We detailin the next section the reason why these algorithms cannotbe applied directly and we present a simple algorithm todiscover periodic segment in S.

4. PERIODIC PATTERN ELICITATIONThis issue can be considered as apparently similar to ques-

tions studied in data stream or spatio-temporal data mining.The search for sequential patterns among large data streamshas been conducted since earliest data mining works as anextension of classical association mining [1, 20]. These pat-terns provide useful information on sequential relationshipsbetween objects in a dataset in multiple areas like web de-sign, DNA sequences, intrusion detection or sensor networkanalysis for instance. In on line data streams, due to the con-tinuous and high speed data flow it is not possible to storethe whole data and to perform multi-scan of the dataset likein traditional data mining solutions. It is thus necessary tokeep only the least possible data and to mine approximatepatterns in order to avoid memory overflow [25]. Patternmining in temporal and spatio-temporal data [17, 28, 24,16, 18] is mainly based on time-series analysis where dataare collection of time-series of each object over time. Indyket al.[18] present this problem as follows. Given a long se-quence S and a period T , the aim is to discover the most rep-resentative trend that repeat itself in S every T timestamps.Wang et al. [28] introduced flow patterns which describethe changes of events over space and time. They considerevents occurring in regions and dependencies among changesin neighboring regions. The work conducted by Mamouliset al. [24] on the search for periodic patterns focused on themovement of entities by using an area divided into regionstoo. They partition the space into a set of regions which al-lows them to define a pattern P as a sequence [r0, r1, ..., rn]of given length n, where ri is a spatial region or the specialcharacter *, indicating the whole spatial universe.

While this topic seems quite similar to the issues addressedin our case, the solutions are not suited to it for two reasons.The first one is that we do not have any information aboutthe period of repetition of the pattern like in [24] or [7]. Itis in fact what we want to determine. The second is thatunlike the proposed methods, we do not have personal dataon the movements of each individual, but we have only oneestimate sequence of collective movements of birds. Indeed,the detection matrix M introduced in Section 3 only allowsus to obtain the most representative sequence in movements.That’s why we do not seek to obtain a frequent periodic seg-ment for each individual, but rather a fixed segment, whichis repeated in a collective sequence S. Indeed, if the group ofbirds really adopt a movement of periodic nature, we expectto identify a corresponding sub-segment into S that shouldbe exactly the same at every period T .

This specific problem seems to have been thoroughly stud-ied in bioinformatics [13, 7]. The search for repeated pat-terns in DNA and protein sequences is proved to be impor-tant in sequence analysis, in particular in large-scale genomesequencing projects. Although these methods are efficient,they often aim to seek the repetition of segments known inadvance or unequally spaced in the sequence.

We propose the simple algorithm Periodic Mine verymuch inspired by these methods, to dynamically generatea candidate segment from a sequence S and automaticallycheck if it is periodic in S. This algorithm extracts a periodicsegment Scand from a sequence S. The rehearsal period Tof the segment Scand in S is equal to the size of the segment,i.e. T = |Scand|.

Input : A sequence SOutput : A periodic sequence Scand extracted from S

Function PeriodicMine( S : Sequence) : SequenceisFrequent : booleanisFrequent← falseSCand : SequenceSCand← ∅While (S 6= ∅ and isFrequent = false) do

add S[0] to ScandremoveElement 0 from SisFrequent← is(Scand periodic in S)

doneIf (isFrequent) then

return Scandend Ifreturn ∅

End

Algorithm 1: PeriodicMine for Mining periodic sequence

5. EXPERIMENTAL EVALUATIONIn order to understand the collective behavior of the Mo-

queurs Gorge Blanche which motivated this work, we searchfor useful information on their movements. This knowledgemay be significant to reintroduce them in environments thathave similar characteristics to their original habitat. How-ever, before deploying our solution on the ground in a realsituation, a first necessary step is to design a prototype andthus measure the impact of all factors implicated on re-sults. Indeed, to the best of our knowledge, the solution wepropose has no equivalent in the literature. Consequently,simulations are expected to reveal relevant information foroptimizing the deployment of wireless sensors. Simulationshave indeed become essential steps in the study and compre-hension of complex phenomena. They will allow us to testdifferent configurations, but also to analyze parameters thatare most prominent for the detection of movements. There-after, we plan to use the results obtained in simulation toadapt the architecture in real environment. This section isdevoted to the simulation tool Lypus and the results.

In order to evaluate the efficiency and the role of variousparameters in the discovery of patterns, we built the 2Dsimulation environment Lypus, which models the behavior ofbirds in their habitat. This simulation aims at representing anatural environment, that virtually reproduces bird habitatin which virtual sensor network is set up.

In this tool, the area is first divided into uniform regions.We do not seek to obtain information of correlation betweenmovements and region characteristics. We are just inter-ested in determining whether the proposed approach can beused to detect movement patterns.

Once the area divided into regions, the sensors are then

(a) (b)

Figure 3: random(a) and uniform(b) sensor place-ment in Lypus

placed and are each associated to the region to which itbelongs. All sensors have the same capabilities. Their de-tection radius and battery level are equivalent. Sensors ofa same region communicate with each other, and sensor-gateways communicate with the sensors-gateways of neigh-boring regions to route information to the central database.In this tool, we are interested nor in optimizing exchanges,neither in reducing loss or corruption of information fre-quently observed on sensor networks. We assume that interand intra regions communications are always successful.

Lypus is fully customizable. Thus, it is possible to definethe size of the area, the number of regions and their size,the number of sensors, the size of their detection radius, thekind of sensor configuration, the size of the bird population,the probability of bird song, the type of movements, etc.Figure 3(a) shows the Lypus 2D interface visualization withan area of 200m × 200m. The largest rectangle representsthe global area and each small rectangle represents a region.Sensors are figured by a point and their detection radius bycircles. Birds are represented by small red filled squares.

The example in Figure 3(a) illustrates sensors which havebeen placed randomly, a kind of layout obtained by droppingof the sensors by plane. In this case some sensors overlap sev-eral regions, so that they produce distortion on results. Weobserve the impact of sensors layout on results. In our sim-ulation, sensors are able to correctly identify birds species.

To evaluate the efficiency of the architecture in patterndiscovery, we give predefined behaviors to birds. We haveimplemented the two types of movements formalized in sec-tion 3: Flock and Periodic movements. The objective is tocheck and evaluate how the method we propose can retrievethese behaviors. Results which are proposed have been madewith a calibration that we consider close to the real environ-ment of birds. The dimensions of the area is 1000m × 1000mand it is divided into 100 regions of 100m × 100m. The sen-sors have detection radius of 25m. Furthermore, we assumethat when several birds are singing in a sensor neighbor-ing, it generates noise that cancels the recognition process.These birds are not detected by this sensor.

The results were obtained on a Core 2 Duo P8600 2.4 Ghz,with 4Go RAM, Ubuntu 9.04 and JDK 1.6. Each test hasbeen conducted 100 times. For each generation, we changedthe sensors location and the routes taken by birds. Then, webased pattern discovery efficiency on results average. Thediscovery of the α−Flock pattern is done with α = 2.

However, it is clear that the discovery of patterns withsuch an architecture is confronted with two main issues. In-deed, the pattern discovery rate is strongly linked to the factthat the birds are singing or not. The song is currently theonly data we can collect. It is obvious that if birds do notsing, we will not detect any pattern. Moreover, according to

NbS NbReal NbDet Periodic Flock

100 4,15 1,11 40,15 14,99

200 4,14 1,86 54,27 28,18

300 4,06 2,39 62,95 43,63

400 4,13 2,92 67,81 47,72

500 4,15 3,19 71,56 57,27

600 4,12 3,43 69,48 55,45

700 4,14 3,59 72,57 61,36

800 4,18 3,8 73,9 59,093

900 4,15 3,9 74,79 58,63

1000 4,11 3,94 74,88 59,09

0

10

20

30

40

50

60

70

80

100 200 300 400 500 600 700 800 900 1000

Det

ectio

nra

te

Number of sensors

Periodic

Flock

Figure 4: Detection rates observed with differentnetwork sizes for a population of 20 individuals

the way sensors are set up, either the area may be partiallycovered or sensor detection regions may overlap. This situ-ation may create a bias on the process of pattern discovery.That is why we have studied the impact of three parametersthat are the song occurrence probability, the sensor numberand the kind of sensor layout.

First, performances were evaluated with different networksizes, as shown in Figure 4, then we analyzed different birdpopulations sizes (Figure 5), finally we studied the impactof a uniform disposition of sensors (Figure 8). In the resultspresented NbS is the number of sensors, NbB is the size ofthe population, NbReal is the real number of birds singing,NbDet is the number of birds detected, Periodic is therate of discovered Periodic patterns for 100 tests and Flockis the rate of discovered α−Flock patterns for 100 tests

Figure 4 shows the detection rate of the two patterns whenthe sensors number varies with a population of 20 individu-als. We notice that when the number of sensors is increasing,the detection rate is improving too. This is because the areais rather wide and when the sensors number is too low, thearea is hardly entirely covered. Thus, in that case numer-ous birds that are singing in not covered regions are notdetected. When the sensors number increases, the surfacecoverage becomes more important, the detection of the twopatterns is improved and the number of detected birds isclose to the number of real birds. The detection rate be-comes insignificant after 500 sensors.

We are aware that a major issue related to such an ar-chitecture is the probability of song occurrence. Indeed, thesong and its location are the only data that sensors can cur-rently collect. However, we know that birds do not sing allthe time. To study the impact of this phenomenon, we havechosen to make the population size vary, rather than theprobability of singing. Thus if we increase the size of thepopulation, we increase the probability that several individ-uals sing. The results are shown in Figure 5 for a area of1000m × 1000m with a random configuration of 1000 sensorson 100 regions of 100m × 100m.

On Figure 5, we can observe that when the bird popula-tion increases, the detection rate decreases. Indeed, whenthe number of birds is increasing, the probability that sev-eral birds sing at the same time on the same region is moreimportant. This phenomenon generates noise in the sensors

NbB NbReal NbDet Periodic Flock

10 2,073 1,98 64,99 77,72

20 4,25 4,04 72,09 60,45

30 6,34 5,96 69,36 53,63

40 8,36 7,77 64,45 44,08

50 10,61 10,08 60,13 46,81

60 12,12 11,49 56,51 34,99

70 14,38 13,61 49,74 26,81

80 16,67 15,71 43,76 35,45

90 18,66 17,55 39,24 29,99

100 21,19 20,08 33,68 29,08

0

10

20

30

40

50

60

70

80

90

10 20 30 40 50 60 70 80 90 100

Det

ectio

nra

te

Population size

Periodic

Flock

Figure 5: The impact of population size on the de-tection rate for 1000 randomly placed sensors

neighboring. However, this problem could be partly solvedby adding sensors if they are not affected by this noise.

These results highlight the dependency of the proposedmethod efficiency on these two factors: the area coverageand the noise in the sensors neighboring. It is easy to un-derstand that our model is very dependent on the number ofsensors located in the area. Indeed, when sensors detectionregions cover the area, the chances of detecting birds aremore important. That is highlighted in Figure 4. When thenumber of sensors increases, the number of detected birdsincreases and the pattern discovery rate also increases.

However, we also notice that the size of the sensor networkwas not alone responsible for errors. Indeed, for large pop-ulations of birds, we observed on Figure 5, that the perfor-mance of the pattern discovery decreases, even when the net-work size is high. That is due to the co-occurrence of birdssinging in same sensors neighboring that generates noises.When the bird population size is high, the probability thatindividuals were singing in close proximity to each other isobviously important. It is interesting to see how the methodbehaves according to these two parameters. We have repre-sented the rate of detected patterns according to these twoparameters, using a colored 3D representation in which thecolor intensity represents the detection rate level: blue forlow level, whereas red indicates a very high rate. Figure6 presents the evolution of the Periodic pattern detectionand Figure 7 presents the evolution of the α−Flock patterndetection. These figures highlight the variation of the detec-tion and allow to observe the detection rate optima.

100

0

8060 40 20

0

500

1000

10

20

30

40

50

60

70

80

Population size

Dete

ctio

nra

te

Figure 6: Detection rate of Periodic pattern depend-ing on the number of sensors and population size

Dete

ctio

nra

te

Number of sensors

20

40

60

80

0

1000

0

800 600400 200

50

100

0

Figure 7: Detection rate of Flock pattern dependingon the number of sensors and population size

Figure 6 shows that with about 400 sensors, the archi-tecture is able to efficiently identify the Periodic patternfor bird populations below 60 individuals. Figure 7 reflectsthe difficulty to identify the Flock pattern. We can observethat for populations size below 20, the architecture is ableto detect correctly a Flock pattern with about 200 sensors.However, we conducted these tests with α = 2. It is obviousthat the detection of Flock pattern would be better with ahigher margin α.

Although the assumption of random placement of sensorsis the most realistic, this method produces unbalanced sen-sor distribution over regions. It cannot ensure to cover en-tirely the area and a region may be covered by several sensordetection areas. For example, a bird can sing into a regionA, but be detected by a sensor of region B. Such a placementis therefore a source of errors in the detection. Figure 3(b)shows what could be the same area of 200m × 200m with auniform set of sensors.

Although a uniform placement is not conceivable in prac-tice, we have studied this configuration on the test area of1000m × 1000m, divided into 100 regions of 100m × 100m.We placed nine sensors, uniformly distributed on each re-gion. This represents 900 sensors placed on the area. Theresults are shown on the Figure 8.

NbB NbReal NbDet Periodic Flock

10 2,01 1,83 85,45 76,82

20 4,2 3,3 95 60

30 6,2 5,7 97,18 49,09

40 8,33 7,56 96,4 43,18

50 10,01 8,99 93,29 40,9

60 12,32 11,1 84,03 34,99

70 14,07 12,64 77,16 29,54

80 16,81 15,51 63,86 27,72

90 18,83 16,69 51,13 31,36

100 20,61 18,66 43,4 26,81

0

20

40

60

80

100

120

10 20 30 40 50 60 70 80 90 100

Det

ectio

nra

te

Population size

Periodic

Flock

Figure 8: Detection rate with a uniform placementof 900 sensors

We can compare these results with those obtained in thesame conditions for the random placement in Figure 5. Wecan notice that the Flock pattern detection is still difficult.We even notice the results for this pattern seem slightly less

accurate than those obtained with a random configuration.This result may be explained by the fact that when the sen-sors are placed uniformly, situations avoiding noise are lessfrequent. However, Periodic Pattern detection is improved.Indeed, we obtain rates higher than 90 % for populationsunder 50 individuals.

We realize that several factors may influence the detectionrate: the region size, the probability of singing, the popula-tion size and the number of sensors. However, these resultsseem to show that an optimal configuration must be studiedas a function of the population size. The population size ap-pears to be the most relevant factor that influences results.So it would be possible, if the population size is known orapproximated, to adjust the placement of sensors in orderto achieve optimal detection of bird movements.

6. CONCLUSION AND FUTURE WORKSensor networks offer real opportunities for observing col-

lective and individual behaviors in situations where othermobile devices like GPS equipments are undesirable. Thework presented in this paper addresses issues we met in sucha situation for studying, monitoring and modeling the habi-tat of an endangered bird population.

In this paper, we show how a wireless sensor network, inwhich each sensor is fitted with a microphone, can be usedto study collective birds movements. We have shown thatthe proposed architecture that currently only records birdsongs enables to elicit relevant information on bird move-ment patterns. We have studied the impact of different net-work configurations on the pattern discovery rate throughsimulations. Thus, we are able to conclude that the maindrawback in this framework is the noise in sensor environ-ment when the size of the population becomes too large.

In future works, in the short term we will focus on theimprovement of the periodic pattern elicitation process, inorder to detect the slight variations that can occur into theperiodic segment. Moreover, in the longer term we will alsofocus on the ability to differentiate individuals by their song,as suggested by [8] in order to adapt the pattern discoveryprocess. These first results confirm the opportunity to setup the network on the ground in the real environment ofthe Moqueur Gorge Blanche. We will use sensors that willbe able to collect, at regular time intervals, data such astemperature, light or noise. We will particularly target in-formation on correlation between the region characteristicsand birds movements.

7. REFERENCES[1] R. Agrawal and R. Srikant. Mining sequential patterns.

Proceeding of the 11th International Conference on DataEngineering, Taipei, Taiwan, pages 3–14, 1995.

[2] R.E. Bell. A sound triangulation method for counting barredowls. The Wilson Bulletin. The Wilson OrnithologicalSociety, 1964.

[3] M. Benkert, J. Gudmundsson, F. Hubner, and T. Wolle.Reporting flock pattern. Computational Geometry: Theoryand Applications, pages 111–125, 2008.

[4] M. Bosse, N. Nourani-Vatan, and J. Roberts. Coveragealgorithms for an under-actuated car-like vehicle in anuncertain environment. International Conference on Roboticsand Automation, pages 698 – 703, 2007.

[5] L. Brotons and S. Herrando. Effect of increased food abundancenear forest edges on flocking patterns of coal tit parus aterwinter groups in mountain coniferous forests: Individualsconcentrated near forest edges in bigger social groups than inforest interiors and foraged more on pine cones which weremore abundant there. Bird Study, 50:106–111, 2003.

[6] H. Chok and L. Gruenwald. Spatio-temporal association rulemining framework for real-time sensor network applications.Conference on Information and knowledge management,pages 1761–1764, 2009.

[7] E. Coward and F. Drablos. Detecting periodic patterns inbiological sequences. Bioinformatics, 14:498–507, 1998.

[8] L. Dabouineau. Pourquoi les oiseaux chantent-ils ? Le raled’eau, 119:10–14, 2004.

[9] A. Desrochers and M.J. Fortin. Understanding avian responsesto forest boundaries: a case study with chickadee winter flocks.Oikos, 91:376–384, 2000.

[10] A. Dolby. Movement patterns of male common voles (microtusarvalis) in a network of y junctions: role of distant visual cuesand scent marks. National Research Council of Canada,Ottawa, 79(12):2228–2238, 2001.

[11] B. Dumont, A. Boissy, C. Achard, A. Silbbald, and H. W.Erhard. Consistency of animal order in spontaneous groupmovements allows the measurement of leadreship in a group ofgrazing heifers. Applied animal behavior science,95(1-2):55–66, 2005.

[12] A. Fakih, S. Ng, and J. Zelek. Improving gps localization withvision and inertial sensing. Geomativa, 63-4:383–396, 2010.

[13] E. F. Glynn, J. Chen, and A. R. Mushegian. Detecting periodicpatterns in unevenly spaced gene expression time series usinglomb scargle periodograms. Bioinformatics, 22(3):310–6, 2006.

[14] H. Gros-desormeaux, P. Hunel, and N. Vidot. Counting birdswith wireless sensors networks. IWCMC 2009 Wireless SensorNetworks Symposium, pages 1163–1167, 2009.

[15] H. Gros-desormeaux, P. Hunel, N. Vidot, and E.Stattner.Acoustic counting algorithms for wireless sensors networks.Proceedings of the 6th ACM symposium on Performanceevaluation of wireless ad hoc, sensor, and ubiquitousnetworks, pages 79–84, 2009.

[16] J. Gudmundsson, P. Laube, and T. Wolle. Movement patternsin spatio-temporal data. Encyclopedia of GIS, pages 726–732,2008.

[17] J. Han, G. Dong, and Y. Yin. Efficient mining of partialperiodic patterns in time series database. InternationalConference on Data Engineering, pages 106–115, 199.

[18] P. Indyk, N. Koudas, and S. Muthukrishan. Identifyingrepresentative trends in massive time series data sets usingsketches. Proc. of Very Large Data Base, pages 363–372, 2000.

[19] C. S. Jensen, D. Lin, and B. C. Ooi. Continuous clustering ofmoving objects. Knowledge and Data Engineering, IEEETransactions on, 19:1161–1174, 2007.

[20] N. Jiang and L. Gruenwald. Research issues in data streamassociation rule mining. SIGMOD Record, 35:14–19, 2006.

[21] P. Laube and S. Imfeld. Analyzing relative motion withingroups of trackable moving point objects. Proceedings of theSecond International Conference on Geographic InformationSciences, pages 132–144, 2002.

[22] P. Laube, M. van Kreveld, and S. Imfeld. Detecting relativemotion patterns in geospatial lifelines. InternationalSymposium on Spatial Data Handling, pages 201–214, 2004.

[23] Y. Li, J. Han, and J. Yang. Clustering moving objects.International Conference on Knowledge Discovery and DataMining, pages 617–622, 2004.

[24] N. Mamoulis and M. Hadjieleftheriou. Mining, indexing, andquerying historical spatiotemporal data. Proceedings of thetenth ACM SIGKDD international conference on Knowledgediscovery and data mining KDD ’04, pages 236–245, 2004.

[25] A. Marascu and F. Masseglia. Mining sequential patterns fromtemporal streaming data. Proceedings of the firstECML/PKDD Workshop on Mining Spatio-Temporal Data(MSTD ’05), held in conjunction with the 9th EuropeanConference on Principles and Practice of KnowledgeDiscovery in Databases (PKDD ’05, 2005.

[26] M. A. Rumble, L. Benkobi, F. Lindzey, and R. Scott Gamo.Evaluating elk habitat interactions with gps collars.International conference held at the Macaulay Land UseResearch Institute, pages 11–17, 2001.

[27] F. Verhein and S. Chawla. Mining spatio-temporal associationrules, sources, sinks, stationary regions and thoroughfares inobject mobility databases. Proceedings of the 11thInternational Conference on Database Systems for AdvancedApplications, 3882:187–201, 2006.

[28] H. Wang, W. Fan, P. S. Yu, and J. Han. Mining conceptdriftingdata streams using ensemble classifiers. InternationalConference on Knowledge Discovery and Data Mining, pages226–235, 2003.