Computational Intelligence Applied to Monitor Bird Behaviour

Compu a ional In elligence Applied o Moni o Bi d Beha iou D. F. La ios1, C. Rod íguez², J. Ba bancho1, M. Baena3, F. Simón¹, J. Ma ín2, C. León1 and J. Bus aman e2 1Depa men o Elec onic Technology, Uni e si y o Se ille, Se ille, Spain 2Depa men o We land Ecology, Doñana Biological S a ion (EDB-CSIC), Se ille, Spain 3ICTS, Singula Scien i ic and Technological In as uc u e, Doñana Biological S a ion, Se ille, Spain Keywo ds: Neu onal Ne wo k, Compu a ional In elligence, Da a Fusion, En i onmen al Moni o ing, Senso Ne wo ks. Abs ac : The bes way o ob ain ele an in o ma ion abou he beha iou o animals is di ec obse a ion (o indi iduals). Howe e , adi ional close-up obse a ions can in e e e on he beha iou , and aking biome ic measu emen s equi es he cap u e o indi iduals, which also causes s ess. This pape desc ibes an au oma ic mo o ing sys em o bi ds b eeding in nes boxes. The main goal is o signi ican ly inc ease he amoun and quali y o da a acqui ed on bi d beha iou wi hou s essing he indi iduals o in e e ing. This sys em is based in an in e connec ed embedded senso ne wo k, which pe mi s sha ing his aluable in o ma ion wi h esea che s all o e he wo ld h ough he in e ne . Each de ice o he ne wo k is a sma nes -box ha allows a c oss- alida ion o senso in o ma ion and da a quali y. This sys em has been e alua ed o he speci ic case o a lesse kes el b eeding colony in Sou he n Spain. The lesse kes el is an insec i o ous mig a o y alcon ha eadily accep s nes -boxes. The sys em has been named HORUS and he esul s ob ained om a yea expe imen demons a e he e iciency o his app oach. 1 INTRODUCTION Fo zoologis s, one o he mos impo an pe iods o a bi d is he b eeding pe iod, being his pe iod one o he mos equen ly s udied. Ma e acquisi ion, nes de ence, ma e eeding, incuba ion, and chick ea ing (including p o isioning ligh s) a e s udied in he su oundings o he nes . Many o hese aspec s help us unde s and key opics in ecology, such as wha ac o s in luencing li e ime ep oduc i e success (LRS: New on, 1992), he pa en -o sp ing con lic (T i e s, 1974; Schlome e al., 2010), o e olu i e s able s a egies (ESS: Mayna d-Smi h and P ice, 1973) ega ding sex oles in ep oduc ion (Kokko and Wong, 2007). Classic moni o ing me hods ha equi e he cap u ing o indi iduals o close-up obse a ions limi he amoun and quali y o da a ha can be ob ained. The e o e, in his pape a emo e moni o ing sys em based on sma nes -boxes is p oposed. These sma nes -boxes allow acqui ing high amoun o da a wi hou s essing he animals, ga he ing long- e m and highly eliable in o ma ion on he species. The p oposed sys em, called HORUS, pe mi s ga he ing basic in o ma ion on he iden i y o indi iduals, s udying i s beha iou and he empo al changes in indi idual body mass. All his in o ma ion can be made accessible h ough he in e ne o scien is all o e he wo ld. Some o he in o ma ion eco ded by he sys em can be used o s udy, wi hou in e e ence, he beha iou o species du ing he b eeding pe iod. One o he mos impo an biome ic pa ame e s in bi ds is body mass. I allows us o measu e he impac o pa en al ca e on b eeding indi iduals. Manually, i is impossible o de elop a con inuous moni o ing o his pa ame e . Cap u ing causes oo much s ess on he indi idual in i s mos sensi i e pe iod. Ano he impo an pa ame e is he measu emen o he amoun o ood b ough o he nes by indi iduals o eed hei o sp ing. The p oposed sys em allows us o pe o m a con inuous moni o ing o he ep oduc ion wi hou s essing he indi iduals, e.g. ob aining eliable body mass measu emen s e e y ime a bi d en e s o lea es he nes . The main p oblem ob aining he measu emen s is he mo emen o he animal, which p oduces uns able alues. These alues ha e been calib a ed using a neu onal ne wo k p ocessing, Figu e 1: Pic u e au oma ically aken by he ou side came a o he moni o ed colony wi h he HORUS ex e nal senso s. F om op o down i shows a pigeon, a lesse kes el male and a emale. ob aining high accu a e measu emen s. The es o he pape is o ganized as ollows: Sec ion 2 ocuses on he lesse kes el beha iou s udy, especially compa ing a adi ional app oach e sus au oma ed da a eading. Sec ion 3 b ie ly desc ibes he HORUS sys em in as uc u e. A de ailed desc ip ion o he in o ma ion ea men de eloped in his p ojec can be ound in sec ion 4. The esul s ob ained wi h ou sys em a e shown in sec ion 5. Finally, sec ion 6 sum-up conclusions and p o ides ema ks. 2 LESSER KESTREL BREEDING BEHAVIOR STUDY habi a s (Pee and Gallo-O si, 2000). The species is also sensi i e o clima e wa ming (Rod iguez and Bus aman e, 2003). So i makes o a good model species o s udy he impac o global change on an endange ed species. 2.1 T adi ional Moni o ing The “Es ación Biológica de Doñana” (EDB-CSIC) has been moni o ing lesse kes el colonies since 1988. I has been eco ding colony occupancy and b eeding success in e ms o numbe o ledglings and p opo ion o success ul nes s. Rega ding indi idual moni o ing, bi ds ha e been ma ked wi h me al and PVC ings wi h a unique alphanume ic code ha allows iden i ying indi iduals by using elescopes. Biome ic measu es we e aken spo adically o all ma ked indi iduals when cap u ed. Due o e hical easons, howe e , he numbe o cap u es in he nes is limi ed ( he cap u e al e s b eeding beha iou and may jeopa dize he su i al o he o sp ing) and he majo i y o esigh ings we e made wi h elescopes. This causes high di e ences in he equency o ecap u es among indi iduals mainly due o di e ences in de ec abili y. In a classic da a base moni o ing, 2,135 bi ds igu ed as ecap u ed (including esigh ings wi h elescope). On a e age hey we e cap u ed 3 imes on he same b eeding season ( ange: 1-70). In app oxima ely 45% o cases, body mass was measu ed and maximum numbe o measu emen s pe bi d and yea was 4. Because o ha , he pa e n o body mass a ia ion o b eeding adul s om a i al o he colony in mid-Feb ua y o he end o he nes ling pe iod in mid-July is no well known. Acco dingly, we ha e no in o ma ion on he pa e n o in aday a ia ion in body mass. The p oposed emo e moni o ing sys em aims o b idge he abo e de ailed logis ic and e hic gaps, hus allowing us o ge enough in o ma ion o documen bo h pa e ns. 2.2 Au oma ed Da a Reading Habi a moni o ing has e ol ed g ea ly e olu ion due o he boom o senso ne wo ks echnology. Se e al consequences ha e been caused due o he inc ease o senso s: Fi s ly he quali y o in o ma ion g ows in ime and on he spa ial domain; secondly he possibili y o ansmi ing he measu ed da a h ough he ne wo k inc eases he need o ha ing high bandwid h communica ions; and hi dly, i o he educ ion o he cos o he da a The lesse kes el (Falco naumanni, igu e 1) is a small (body mass a ound 150 g ams) mig a o y alcon inhabi ing open landscapes (C amp and Simmons, 1980). I is a colonial species ha b eeds in old buildings, such as chu ches o cas les wi hin u ban a eas in Wes e n Eu ope. The species expe ienced a ma ked decline in i s Wes e n Palea c ic b eeding ange in he middle o he 20 h cen u y (C amp and Simmons, 1980; Bibe , 1990). Conside ed p e iously one o he mos abundan ap o s in Eu ope (Bijle eld, 1974) he lesse kes el became ex inc in se e al coun ies (e.g. Aus ia, Hunga y, Poland) and p ac ically disappea ed in o he s (e.g. F ance, Po ugal, Bulga ia). Medi e anean Spain cons i u es i s s onghold in he Wes e n Palea c ic (Bibe , 1990). Howe e , he Spanish popula ion also su e ed a p ecipi ous decline, as i d opped om an es ima ed 20,000– 50,000 pai s in he 1970s (Ga zón, 1977) o 4,000– 5,000 b eeding pai s in 1988 (González and Me ino, 1990). This decline has been a ibu ed o he educ ion in bo h he ex en and quali y o o aging Figu e 2: HORUS Ne wo k scheme. s o age makes possible o sa e huge amoun s o da a. All hese consequences imply some nega i e e ec s: an inc ease o da a a ic and inc ease o powe consump ion. Some au ho s ha e aken hese e ec s in o accoun (Cook, 2007; S idha , 2007) and ha e exp essed he need o employ p ocessing echniques in o de o educe hese handicaps. The e a e di e en app oaches o habi a moni o ing. Some o hem use wi eless senso ne wo k echnology in o de o acqui e and p ocess he physical in o ma ion (Ga cía-Sánchez e al, 2010; Handcock e al, 2010; Valen e e al, 2011; Ca ullo e al, 2009). O he s ocus on he needed middlewa e ha allows access o he physical in o ma ion (Hwang e al., 2010; Fa shchi e al., 2007). In ou app oach, bo h aspec s a e conside ed. 3 HORUS INFRAESTRUCTURE The p oposed in as uc u e is a dis ibu ed sys em as igu e 2 depic s. This igu e shows he mos impo an de ices o he p oposed a chi ec u e. These de ices will now beb ie ly desc ibed. 3.1 Ne wo k In as uc u e The HORUS in as uc u e is made up o di e en subsys ems in e connec ed h ough a low da a a e communica ion ne wo k. This ne wo k has been designed conside ing he ollowing es ic ions: The de ices ha p o ided in o ma ion o he ne wo k a e deployed in a sp ead way wi hou any p e ious planning. The da a a e associa ed wi h he da a sou ces is low (< 250 kbps) The sys em could be easily scalable. The ne wo k used in HORUS can be accessible h ough di e en physical media (wi eless o wi ed based). Robus ness o he ne wo k is e y impo an , o he p oposed applica ion: du ing he b eeding pe iod, i is no possible o ealize main enance asks, because i can dis u b and s ess he colony. All de ec ed ailu es would be epai ed in win e , a e he bi ds ha e le he colony. 3.2 Base S a ion The p ocess se e is a sys em ha o e s he ollowing se ices: Da abase se e . Moni o ing and con ol sys em. Remo e con ol access. The da abase s o es all he his o ic senso s in o ma ion ga he ed om he sys em. The moni o ing con ol sys em is a p og am esponsible o adding addi ional aluable in o ma ion o he senso measu emen s, such as in o ma ion abou he nes sende , a ime s amp egis e o a con ol sequence, ha pe mi s Figu e 3: Nes cabine . de e mining he numbe o loss packe s. This sys em s o es he in o ma ion in he da abase. Remo e access con ol o e s he cloud se ices o emo e use s, such as biologis s. These se ices pe mi emo e access o he senso s da abase. 3.3 Sma Nes -Boxes The sma nes -boxes a e he main componen s o he moni o ing sys ems. I consis s o he nex wo blocks: The nes cabine . The elec onic sys em The nes cabine (Figu e 3) is di ided in o wo pa s: a co ido and he incuba ion chambe . This nes cabine has a sma design o ensu e ha he bi ds pass he co ido each ime hey en e o lea e he nes . The ad an age o his is o allow he dis ibu ion he senso s in a small a ea ( he co ido ) whe e he animal is o ced o pass and, he e o e, i ensu es ob aining he senso in o ma ion. The elec onic sys em (Figu e 4) o each sma nes -box is accomplished wi h he nex subsys ems: 3.3.1 Mic ocon olle Boa d This boa d is based on he ATmega2560, an economic, low powe and obus mic ocon olle . I con ols and p ocesses he nes ’s senso in o ma ion. This boa d communica es wi h senso s and o he componen s, and p ocesses he collec ed in o ma ion ha is sen o he p ocess se e o e he communica ion in e ace. The p og am implemen ed in he mic ocon olle pe o ms he ollowing asks: Communica es wi h he p ocess se e o e a communica ion in e ace, and synch onize clock ime wi h his. Figu e 4: A chi ec u e o he elec onic sys em. Checks in a- ed ba ie s. Each nes -box has wo in a- ed ba ie s a bo h ex emes o he co ido . The sequence in which hey a e ac i a ed indica es whe he bi ds en e o lea e he nes -box. Checks i he RFID eade has ead a code om inged kes els. Ob ains he body mass measu emen om a digi al balance. Reads he empe a u e and humidi y o he nes . Con ols he RFID eade o iden i y indi iduals. 3.3.2 Senso s Boa d A senso boa d adap s he logic le els om he nes senso s o he mic ocon olle boa d’s equi emen s. All he nes ’s senso s a e sp ead on o he co ido o he nes . Posi ions o senso s a e designed o ensu e ha e e y ime he bi ds pass he co ido he sys em egis e s a leas one eco d pe senso . The deployed senso s a e: A digi al balance. I allows a maximum weigh o 600 g . and an accu acy o 0.01 g , o e ing 16 measu es pe second. I pe mi s ge ing an es ima e o he body mass o he indi iduals in mo emen . Al hough he pan is ound, i has been modi ied o be ec angula in o de o i he shape o he co ido . (a) (b) (c) (d) Figu e 5: Di e en weigh pa e n Y-axis, weigh in g ams. X-axis, samples. An in eg a ed empe a u e senso loca ed in he window. I is calib a ed o ope a e in en i onmen al empe a u e ange. I is used o measu e he nes empe a u e. An in eg a ed humidi y senso . I is used o measu e he nes humidi y. Two in a- ed ba ie s, used o ace he di ec ion o bi ds’ mo emen s. A RFID eade . I communica es ia RS-232C and o e s a eading on he unique ID o a agged bi d, when i is passing h ough he nes en ance. This sys em has mechanisms o a oid collisions, pe mi ing ope a ion e en when he e a e se e al bi ds a ound he en ance. A Se omechanism. I is used o emo ely cap u e bi ds when hey en e in he nes -box. 4 TREATMENT OF THE INFORMATION As desc ibed be o e, e e y nes -box p o ides he ollowing in o ma ion: Measu emen s o body mass: The digi al balance used o e s 16 measu emen s pe seconds wi hou calib a ion and classi ies measu emen s as s able o uns able. IR in o ma ion: These senso s pe mi de e mining i he bi ds go in o o go ou o he nes . RFID in o ma ion: I pe mi s a ibu ing he in o ma ion o o he senso s o an indi idual bi d. All his in o ma ion is ob ained om he senso s deployed in e e y nes -box. The senso s o e ele an in o ma ion on he indi idual b eeding a he colony o he biologis s ha s udy hem. This in o ma ion, excep he body mass, canno be added, as hey in o m abou disc e e e en s. The e o e, his in o ma ion is sen di ec ly o he da abase wi hou any local p ocessing o ea men . On he o he hand, he digi al balance o e s a high amoun o in o ma ion. I s equency o measu emen is much highe han he body mass e olu ion o he animal. i.e., he animal body mass e olu ion has mo e ine ia han he weigh p o ided by he balance. Due o his, i is possible o pe o m a da a p e-p ocessing abou weigh in o ma ion, educing wi h ha he amoun o in o ma ion send o he cen al p ocessing. 4.1 Weigh P e-p ocessing The algo i hm desc ibed in his pape , is ocused on locally p e-p ocessing he weigh in o ma ion, o educe he amoun o unnecessa y in o ma ion and inc ease i s accu acy. I is designed o be execu ed in each nes -box, in he mic ocon olle boa d. I has been designed o ul il he nex goals: To educe he amoun o useless in o ma ion in he da abase using local p e-p ocessing. To inc ease he accu acy o he measu emen s, calib a ing he esul s ob ained. To inc ease accu acy o he communica ion ne wo k, educing he amoun o packe loss, he delays and he collisions. To inc ease he amoun o use ul in o ma ion in he da abase, es ima ing a body mass om each pa e n wi h non-s able measu emen s. To pe mi i s execu ion on de ices wi h low esou ces. To inc ease he accu acy a a e calib a ion is necessa y. The balance used o e s measu emen s wi hou a a e calib a ion. This calib a ion would be ob ained consul ing he body mass measu ed by he balance, when he e is no animal on he pan i.e., when he measu ed weigh is below a ce ain es ima ed h ough a compu a ional in elligence algo i hm. In bo h cases, only one selec ed weigh pe pa e n is sen o he da abase. These selec ed weigh s a e calib a ed wi h he a e, be o e sending hem. The p oposed algo i hm is summed-up in he nex pseudo-code: while 1: wai new(meas_weigh ); i meas_weigh >= h eshold weigh [i]:=meas_weigh – a e; inc ease I; i s able(meas_weigh )==1 s able :=1; end i else i i!=0 i s able==1 es _weigh :=a e age( s able_weigh ); s able:=0; else es _weigh :=model(weig h); end i send_se e (es _weigh ); i:=0; else new_ a e:=meas_weigh ; a e:=I e _RMS(p e _ a es, new_ a e); end i end i end while 4.1.1 Applying Machine Lea ning o Weigh Recogni ion Fo his applica ion, an algo i hm has been e alua ed. Ini ially, an algo i hm wi hou machine lea ning based on he di e ences be ween consecu i es measu emen s has been conside ed. This algo i hm conside s a weigh s able i he e a e mo e han a ce ain numbe o measu emen s o he same weigh . This is simila o he in e nal algo i hm o he balance o agging measu emen s as s able o uns able, bu i is less es ic i e: he balance equi es a high numbe o measu emen s wi h he same alue o conside a measu emen s able. I pe mi s he e ie al o some weigh s om he uns able pa e ns, bu i ails wi h complex pa e ns. Ou p oposal o using compu a ional in elligence (machine lea ning) inc eases he pe cen age o success. Machine lea ning is widely used in pa e n ecogni ion, bu i s use in animal moni o ing is less widesp ead. O he supe ised lea ning echniques apa o he neu onal ne wo k ha e been conside ed. h eshold. This h eshold can be ob ained as a unc ion o he body mass o he animals o moni o . In ou deploymen o he Lesse Kes el (wi h a body mass ange o 100-190 g) a h eshold o 100 g ams o has been used. In he eal deploymen we ha e p o en ha he a e does no change signi ican ly du ing a yea . The e o e, measu ing he a e only once pe day o e s enough accu acy o he p oposed sys em. On he o he hand, as desc ibed be o e, he balance o e s 16 weigh measu emen s agging hem e e y second by i sel as s able (i.e., measu emen s ha emain a same alue du ing a long pe iod o ime) o uns able. Bu bi ds usually do no pass o e he balance slow enough o ob ain s able measu emen s. This causes he da abase o ha e a high amoun o he in o ma ion as uns able measu es. In he eal p o o ype only abou 15.25% o he measu ed pa e ns had a s able measu emen , conside ing a pa e n as he collec ion o measu emen s ob ained om he ime he bi d ge s on he balance (i.e., when he balance acqui es a weigh o e he h eshold) un il he animal ge s ou o balance (i.e., when he balance acqui es du ing 5 seconds weigh s below he h eshold). Fig. 5 shows di e en examples o weigh pa e ns ob ained in he eal deploymen wi h hese condi ions. This igu e shows di e en eal weigh pa e ns ob ained om he same animal in di e en days. Only pa e n (a) has some s able measu emen . These s able measu emen s ha e been compa ed wi h measu emen s o he animal done manually cap u ing he bi d. The s able a e measu emen s a e co ec , bu no equen enough o ob ain a long e m sequence o body mass empo al change o he bi ds a he colony. To sol e his, a compu a ional in elligence algo i hm o es ima e he body mass o animals om he pa e ns wi h non-s able weigh s has been de eloped, inc easing he amoun o use ul in o ma ion. This neu onal ne wo k algo i hm is desc ibed below. Ini ially, he sys em has been designed o s o e, in he cen al se e , all weigh measu emen s o he pa e n acqui ed by he balance, s able o uns able, bu i causes high bandwid h consump ion in he communica ions in e ace. To educe he amoun o useless in o ma ion, he p oposed algo i hm only sends one es ima ed weigh o he da abase o each measu ed pa e ns. I he pa e n has some s able measu emen s, he es ima ed weigh sen o he cen al se e will be he a e age o he ob ained s able measu emen s. I no measu emen s o he pa e n a e s able, he weigh is Non supe ised echniques, such as Sel -O ganized maps (SOM, Kohonen, 1990) o Suppo Vec o Machine (SVM, Co es 1995) we e disca ded, because we ha e some s able measu emen s ha pe mi pe o ming aining. One example o he conside ed supe ised machine lea ning echniques is he use o A i icial Neu o-Fuzzy In e ence Sys ems (ANFIS; Jang, 1993). ANFIS has many applica ions in he e alua ion o complex sys ems, bu i equi es a p e ious knowledge o he sys em o design he ules and he ini ial sys em. This sys em was disca ded; due o he complex o ms o he pa e ns ha do no easily pe mi acqui e his ini ial sys em. Expe sys ems o case based expe s sys em we e no conside ed, due o he amoun o p e ious in o ma ion ga he ed om he sma nes -box was no su icien o hese kinds o sys ems. Fo hese easons, a neu onal ne wo k model was inally chosen. The a iables used as inpu s o he model a e as ollows: Max_1: The mos epea ed weigh in a pa e n ( he la ges i mul iple). N_1: Numbe o epe i ions o he p e ious a iable in a pa e n. Max_2: The second mos epea ed weigh in a pa e n ( he la ges i mul iple). N_2: Numbe o epe i ions, in a pa e n, o he p e ious a iable. Max_C1: The mos consecu i ely epea ed weigh in a pa e n. NC_1: Numbe o epe i ions o he p e ious a iable in a pa e n, Max_C2: The second mos epea ed weigh , consecu i ely, in a pa e n. NC_2: Numbe o epe i ions, in a pa e n, o he p e ious a iable. N_EL: To al numbe o weigh measu es in a pa e n. In o de o ob ain hese pa ame e s, a pa e n wi h a leas 5 weigh measu emen s is needed. As Table 1: Analysis o he da abase. Cap ion Value Measu emen weigh 2583565 Numbe o pa e n 51517 Pa e ns wi h s able weigh s 7856 A e age pa e n ime 23,18 seconds Days o es 399 days an alue ou pu , he neu onal ne wo k model o e s a alue, called “Ou pu weigh ”. This ou pu e lec s he es ima ed weigh o he neu on model and i is he in o ma ion sen h ough he ne wo k o he se e da abase. The s eps execu ion o his neu onal ne wo k model is summed-up in he nex pseudo-code, whe e he neu onal ne wo k is he execu ion o a h ee laye ne wo k. Neu onal ne wo k needs a se o pa ame e s o i s aining. These se s ha e been ob ained o each pa e n wi h s able measu es, by execu ing he ollowing s eps: S ep 1: A a iable name “Ta ge weigh ” was de ined o e e y pa e n. This a iable s o es he a e age alue o all s able weigh s. This is he a ge esul o he aining o he neu onal ne wo k. S ep 2: Fo e e y pa e n, a new pa e n has been c ea ed, elimina ing all s able measu emen s. S ep 3: The inpu s ha e been ob ained om his new pa e n wi hou s able alues. S ep 4: The inpu alues o each pa e n we e s o ed, oge he wi h hei espec i e Ta ge weigh in o a able, named “T aining in o ma ion” Wi h hese ables wo se s o in o ma ion we e ob ained, one o aining and he o he o e alua ing he accu acy o he sys em. In o al, he aining in o ma ion able has 1163 se s o alues. 50% o hese alues ( andomly selec ed) we e used o aining, and he o he 50 % we e used o alida ion. 5 SIMULATION, TESTS AND RESULTS The esul s ob ained wi h his sys em can be classi ied in wo ypes: analysis o he ne wo k pe o mance and weigh es ima ion accu acy ob ained wi h he eal deploymen . This sec ion summa izes hese wo ypes o esul s. 5.1 Ne wo k Pe o mance Du ing he i s yea o he deploymen (2010), he p o o ype was sending in o ma ion om all senso s, e en he 16 eco ds pe second o he balance, o he da abase o he cen al se e . The main cha ac e is ic o he ga he ed in o ma ion in he da abase is summa ized in able 1. A e a yea o deploymen , he analysis o da a allowed us o de ec some ne wo k con lic s. Fo example, i di e en nes -boxes a e acqui ing weigh s om indi iduals a he same ime, hey a e compe ing o con ol o he bus, causing da a collisions and delays in ansmi ing in o ma ion. Table 2: Cos pe message wi h CC2420 Radio anscei e . Cap ion Ene gy (J) Wi hou da a usion 255.3 Wi h da a usion 0.608 The p oposed sys em allows a oiding hese con lic s, using he p oposed da a usion. In his sec ion we a e going o quan i y he ad an age o da a usion agains he classical cen alized sys ems. Due o ha , in his kind o applica ions i is impo an o educe he use o bandwid h as much as possible. The analysis o da abase in o ma ion has been summed-up in he able 1. I shows ha only a 15.25% o he acqui ed pa e ns ha e any s able measu emen . Knowing ha he balance o e s 16 Samples Pe Second (SPS), he a e age payload o he applica ion laye pe pa e n o he sys em wi hou da a usion can be ob ained wi h he equa ion 1. ,16 · · · T aw SPS T By es msg N P N N (1) Whe e ,T aw N is he numbe o by es o send pe day a applica ion laye ; T P is he leng h o he pa e n in seconds; By es N is he numbe o by es o send. 16 by es in his case and msg N is he numbe o messages pe day. On he o he hand, wi h he p oposed algo i hm, only one message pe pa e n is sen . In his case, he payload pe pa e n can be ob ained acco ding o equa ion 2. ,· T aw By es msg N N N (2) This shows ha he amoun o in o ma ion sen o he da abase a ies in unc ion o he numbe o pa e ns and he leng h (in ime) o he pa e n. Figu e 6 depic s hese esul s. Concluding, he local p ocessing pe mi s one o d as ically educe he used h oughpu o he ne wo k, especially in days wi h a high numbe o pa e ns. This da a usion and agg ega ion scheme is especially impo an o i s use in low bandwid h sys ems, due o i pe mi ing one o sa e ene gy. Wi h he p oposed sys em, only one message pe pa e n is sen , ins ead o 16 measu emen s pe second du ing he cap u e o he pa e n. These esul s a e summed-up in able 2. They conside he a e age pa e n leng h o 23.18 seconds, i.e. he a e age ime while he bi d is on he balance. Wi h hese condi ions and wi h he CC2420 adio anscei e , widely used in wi eless senso ne wo k, pe mi sa ing 99.76 % o he ene gy used in da a ansmissions, conside ing a powe consump ion o 38mW in ansmission mode (Polas e e al, 2005. Using all weigh pa e ns ob ained in he yea 2010, he p oposed body mass es ima ion algo i hm pe mi s he e ie al o a ound 56.21% o he pa e ns wi hou s able measu emen s. This is a good esul ha pe mi s us o ob ain an a e age o 4 body mass es ima ions pe day and nes , which is 4 imes highe han using only pa e ns wi h s able measu emen s. I pe mi s o ha e a con inuous acing o body mass in indi iduals. As a conclusion, he local p ocessing pe mi s us o d as ically educe he used h oughpu o he ne wo k, especially in he days wi h a high numbe o pa e ns. 5.2 Body Mass Es ima ion Accu acy Based on he aining and e i ica ion se desc ibed in sec ion 4.1.1, some analysis has been done o he p oposed algo i hm o body mass es ima ion. Wi h he e alua ion se , he sys em o e s an accu acy o 98.7%, i.e., an e o in he o de o 2 (a) (b) Figu e 6: By es pe day send, a applica ion laye , o he cen al se e . (a) Wi hou da a usion. B) Wi h da a usion. Figu e 7: Neu al ne wo k model: impo ance o he a iables in he calcula ion o he es ima ed weigh s. Figu e 8: G ain ele a o used o he p o o ype ins alla ion. g ams, which is qui e small conside ing he ypical body mass o hese animals (150 g ams). This accu acy pe mi s analysis o a long se ies o he e alua ion o empo al changes in body mass, and some imes o de e mine he body mass o p ey, when bi ds b ing medium-sized animals o he nes o eed he nes lings. F om he aining p ocedu e, an analysis o he impo ance o he inpu pa ame e s in ela ionship wi h he a ge body mass can be ob ained. Figu e 7 shows hese esul s. This analysis concludes ha he selec ed pa ame e s a e alid o e ec i ely es ima e he body mass o animals. 5.3 Real Deploymen A p o o ype, o a eal alida ion o he p oposed sys em, has been deployed in he g ain ele a o o “La Palma del Condado (Huel a P o ince, SW Spain” ( igu e 8). A his si e, esea che s o he Es ación Biológica de Doñana ha e been s udying he lesse kes el colony since 1994. A his colony, kes els nes ed on he windowsills o he g ain ele a o ha a e shel e ed and su icien ly enclosed o make a sui able nes ing si e. Fo he p o o ype ins alla ion we selec he windows on he 6 h loo o he building whe e sma nes -boxes we e ins alled, and eadily accep ed, by kes els du ing p elimina y checking (3 and 4 nes -boxes du ing 2008 and 2009, espec i ely) and also when he de ini i e p o o ype ins alla ion was made in 2010. Nes -boxes a e placed in all he windows along he six h loo . They a e named “6XY”, whe e X e e s o he ca dinal poin and Y is an o dinal numbe . Each box has wo sepa a e en ances and wo incuba ion chambe s (I. le and D: igh ) in a symme ical dis ibu ion. En ances a e placed a he ex emes o he box o a oid po en ial agg essions be ween neighbou s, hus maximizing he numbe o po en ial b eeding pai s. None heless, he igh pa has no been opened ye . The esul s p o ided by he sys em a e s ill being analysed by biologis s. Howe e , in i s cu en s a e, i is possible o ob ain some conclusions: 18 o he 20 ins alled nes s-boxes we e used by b eeding kes els. This leads o he conclusion ha he p oposed sys em e ec i ely allows one o ga he a high amoun o in o ma ion abou he beha iou o b eeding indi iduals wi hou s essing hem. I he nes and i s senso s we e hos ile, i would no ha e been chosen by lesse kes els b eeding pai s. A p elimina y esul o he con inuous weigh ing o indi iduals will allow he esea che s o es ima e he cos o b eeding in e ms o body mass. This cos is di ec ly associa ed wi h he o aging ips o eed he nes lings. The lesse kes el mainly eeds on insec s, bu some imes can ca ch sligh ly bigge p ey, such as small oden s, bi ds o liza ds (wi h a ound a dozen o g ams). The p oposed sys em would pe mi an analysis o he equency o big p ey cap u es. 6 CONCLUSIONS The main goal o he p oposed sys em is o use cu en echnological ad ances in a eal-wo ld applica ion in he a ea o Biodi e si y Conse a ion o s udy how global clima e change could a ec a colonial and endange ed bi d species. The esul s ob ained conclude ha he p oposed sys em would pe mi i s use in a sys em wi h low esou ces and wi h a low bandwid h usage. The p o o ype deployed in Spain o e alua ion