RESEARCH ARTICLE
Amazonian mana ee c i ical habi a e ealed by a i icial
in elligence-based passi e acous ic echniques
Flo ence E bs
1
, Mike an de Schaa
1
, Mi iam Ma mon el
2
, Ma ina Gaona
1,2
, Emiliano Ramalho
2
& Michel And ´
e
1
1
Labo a o y o Applied Bioacous ics (LAB), Uni e si a Poli `
ecnica de Ca alunya-Ba celonaTech (UPC), Ba celona, Spain
2
Ins i u o de Desen ol imen o Sus en ´
a el Mami au´
a, Te ´
e, B azil
Keywo ds
bioacous ics, deep lea ning, habi a ,
Si enians, Sou h Ame ica, ocal epe oi e
Co espondence
Michel And ´
e, Labo a o y o Applied
Bioacous ics (LAB), Uni e si a Poli `
ecnica de
Ca alunya-Ba celonaTech (UPC), Ba celona,
Spain. Tel: (34) 938 967 227; E-mail: michel.
[email p o ec ed]
Funding In o ma ion
This esea ch is pa o P ojec P o idence
(h p://www.p ojec p o idence.o g/) unded
h ough he Sense o Silence Founda ion by
he Go don and Be y Moo e Founda ion.
Addi ional unding was p o ided by he Rolex
Ins i u e and he Fonda ion P ince Albe II de
Monaco.
Edi o : D . Vincen Lecou s
Associa e Edi o : D . Da id Cu nick
Recei ed: 26 No embe 2023; Re ised: 30
Ap il 2024; Accep ed: 4 July 2024
doi: 10.1002/ se2.418
Abs ac
Fo many species a isk, moni o ing challenges ela ed o low isual de ec abili y
and elusi e beha io limi he use o adi ional isual su eys o collec c i ical
in o ma ion, hinde ing he de elopmen o sound conse a ion s a egies. Passi e
acous ics can cos -e ec i ely acqui e e es ial and unde wa e long- e m da a.
Howe e , o ex ac aluable in o ma ion om la ge da ase s, au oma ic me hods
need o be de eloped, es ed and applied. Combining passi e acous ics wi h deep
lea ning models, we de eloped a me hod o moni o he sec e i e Amazonian
mana ee o e wo consecu i e looded seasons in he B azilian Amazon lood-
plains. Subsequen ly, we in es iga ed he ocal beha io pa ame e s based on
ocaliza ion equencies and empo al cha ac e is ics in he con ex o habi a use.
A Con olu ional Neu al Ne wo k model success ully de ec ed Amazonian mana ee
ocaliza ions wi h a 0.98 a e age p ecision on aining da a. Simila classi ica ion
pe o mance in e ms o p ecision ( ange: 0.83–1.00) and ecall ( ange: 0.97–1.00)
was achie ed o each yea . Using his model, we e alua ed mana ee acous ic p es-
ence o e a o al o 226 days comp ising eco ding pe iods in 2021 and 2022.
Mana ee ocaliza ions we e consis en ly de ec ed du ing bo h yea s, eaching 94%
daily empo al occu ence in 2021, and up o 11 h a day wi h de ec ions du ing
peak p esence. Mana ee calls we e cha ac e ized by a high emphasized equency
and high epe i ion a e, being mos ly p oduced in apid sequences. This ocal
beha io s ongly indica es an exchange be ween emales and hei cal es. Com-
bining passi e acous ic moni o ing wi h deep lea ning models, and ex ending em-
po al moni o ing and inc easing species de ec abili y, we demons a ed ha he
app oach can be used o iden i y mana ee co e habi a s acco ding o seasonali y.
The combined me hod ep esen s a eliable, cos -e ec i e, scalable ecological mon-
i o ing echnique ha can be in eg a ed in o long- e m, s anda dized su ey p o o-
cols o aqua ic species. I can conside ably bene i he moni o ing o inaccessible
egions, such as he Amazonian eshwa e sys ems, which a e acing immedia e
h ea s om inc eased hyd opowe cons uc ion.
In oduc ion
F eshwa e en i onmen s a e amongs he mos endange ed
ecosys ems (Dudgeon e al., 2006). In pa icula , eshwa e
mega auna a e disappea ing wice as as as e es ial and
ma ine mega auna (He e al., 2019; Jenkins, 2003). In he
Amazon basin, human ac i i ies, including dam cons uc-
ion and mining, a e inducing apid and la ge-scale deg a-
da ion o eshwa e sys ems (Cas ello & Macedo, 2016)
wi h di ec and p o ound consequences o he Amazonian
mana ee (T ichechus inunguis) popula ions (A au
e al., 2017; A au & Ma mon el, 2016).
Wi h hei la ge body, hei la e sexual ma u i y, and
hei low unipa ous ep oduc i e a e (one cal e e y 2 o
3 yea s, Ma mon el, 2009; Rod igues e al., 2008) si enians
a e especially a isk o ex inc ion (Deu sch e al., 2008;
Kei h Diagne, 2015; Ma mon el e al., 2016). His o ically
hun ed h oughou hei dis ibu ion ange (Domning,
ª2024 The Au ho (s). Remo e Sensing in Ecology and Conse a ion published by John Wiley & Sons L d on behal o Zoological Socie y o London.
This is an open access a icle unde he e ms o he C ea i e Commons A ibu ion-NonComme cial-NoDe i s License, which pe mi s use and
dis ibu ion in any medium, p o ided he o iginal wo k is p ope ly ci ed, he use is non-comme cial and no modi ica ions o adap a ions a e made.
1
1982), Amazonian mana ee popula ions ha e declined sig-
ni ican ly and he species is cu en ly classi ied as ulne a-
ble (Ma mon el e al., 2016). Al hough mana ee hun ing
was banned in B azil in 1967, popula ions ha e no eco -
e ed and hun ing o subsis ence o illegal ade s ill occu s
(Cal imon es & Ma mon el, 2022). The di icul y o de ec -
ing mana ees h ough isual me hods complica es hei
conse a ion s a us de e mina ion and he implemen a ion
o adequa e conse a ion s a egies. Visual moni o ing o
mana ees is possible unde ce ain condi ions, such as
when hey a e es ing nea he su ace in clea shallow
wa e s (e.g., Flo ida mana ees agg ega ing in sp ings du ing
win e ). Usually, mana ees spend mos o hei ime deepe
unde wa e , su acing o e y b ie momen s wi h only
he ip o he snou usually isible. Due o his inconspicu-
ous beha io , isual obse a ions a e di icul , especially in
u bid whi e wa e s o he Amazon wi h ex emely poo
isibili y (less han 1 m, Hende son & C amp on, 1997).
Amazonian mana ees a o ma s o loa ing meadows
(Gu e es-Pazin, 2014), and a e o en concealed by he
aqua ic co e o he aqua ic plan s, e en when su acing o
b ea he.
Visual su eys (boa , ae ial o sho e-based) ypically
achie e low de ec ion a es (Fac heu e al., 2023). Unde -
wa e imaging echniques such as a side scan sona ha e
been used ecen ly wi h some success (Gonzalez-Socoloske
& Oli e a-G´omez, 2023) and can be pa icula ly use ul in
mu ky o u bid wa e s wi h limi ed isibili y. A limi a ion
o side scan sona me hods is ha animals may ac i ely
a oid he esea ch boa , mo ing ou side o he sona beam
de ec ion ange. Addi ionally, his me hod has a sho e
de ec ion ange in sedimen - illed wa e wi h aqua ic ege-
a ion. This can in oduce noise in he sona da a which
makes i di icul o iden i y he shape o a mana ee.
All he ex an si enian species, including he h ee mana-
ee species and he dugong species, p oduce ocaliza ions
o communica ion pu poses (O’Shea e al., 2022), making
hem well-sui ed o s udying h ough passi e acous ic
moni o ing (PAM). A c i ical ad an age o ixed PAM is
he abili y o con inuously su ey mana ee p esence a a
loca ion o e mon hs o yea s wi hou in e e ence (i.e., no
in oduc ion o addi ional dis u bance and noise associa ed
wi h boa s o isual o sona su eys). This moni o ing
echnique can e eal aluable in o ma ion abou a species’
empo al occupancy o a habi a , including seasonal a ia-
ions, a he cos o a lowe spa ial esolu ion. To ap in o
PAM esea ch po en ial, bioacous icians collec inc easing
amoun s o da a ha canno be e icien ly p ocessed manu-
ally. Technological ad ances in da a acquisi ion ha e inau-
gu a ed a “big da a” e a o bioacous ics, demanding
au oma ed me hods o deal wi h he analysis o la ge da a-
se s. In ecen yea s, deep lea ning has been inc easingly
used o classi ica ion o bioacous ics signals and has been
success ully applied o au oma ic de ec ion o a wide ange
o soni e ous axa (S owell, 2022), becoming a majo asse
o da a p ocessing and analysis. P e ious wo k has dem-
ons a ed he e iciency o hese me hods in de ec ing man-
a ee ocaliza ions (Fac heu e al., 2023).
The Amazonian mana ee epe oi e cha ac e iza ion
comes mainly om s udies conduc ed in cap i i y (E ans
& He ald, 1970; Land au-Gio anne i e al., 2014; Sonoda
& Takemu a, 1973; Sousa-Lima e al., 2002) wi h only
one s udy om wild animals desc ibing simila signals
(based on a limi ed sample size; Sousa-Lima e al., 2013).
The majo i y o ocaliza ions desc ibed a e sho , onal
ha monic complexes ha con ain one o ou no es.
A onal ocaliza ions ha e also been no ed. In con as ,
he acous ic epe oi e o Wes Indian mana ees is ai ly
well desc ibed (e.g., B ady e al., 2020; Ramos e al., 2020,
Ri e a Cha a ia e al., 2015; Sousa-Lima e al., 2008)
and A ican mana ee ocaliza ions we e ecen ly cha ac-
e ized (Rycyk e al., 2021). The li e a u e demons a es
ha mana ee ocal epe oi e is simila ac oss species
(O’Shea e al., 2022; Rycyk e al., 2021). Va iabili y in
ocaliza ion equency and du a ion ha e been linked o
beha io and g oup composi ion (e.g., B ady e al., 2021;
Miksis-Olds & Tyack, 2009; O’Shea & Poch´e J , 2006).
Mana ees p oduce calls ac oss di e en beha io al con-
ex s (B ady e al., 2021; O’Shea & Poch´e J , 2006), mak-
ing his species a good candida e o PAM s udies.
Addi ionally, calling a es may a y widely depending on
beha io al s a es, po en ially p o iding new insigh s on
mana ee beha io in he moni o ed habi a . The highes
calling a es a e p oduced du ing cow-cal in e ac ions
and se e al s udies sugges ed ha one o he main unc-
ions o acous ic communica ion in mana ees is main ain-
ing con ac be ween a mo he and he cal (Beng son &
Fi zge ald, 1985; Bullock e al., 1980; Ha man, 1979;
O’Shea & Poch´e J , 2006; Reynolds, 1981).
He e, we p esen he esul s o ou ongoing e o s o
moni o Amazonian mana ees in he Mami au´aSus ainable
De elopmen Rese e. The i s objec i e was cha ac e izing
mana ee p esence o e wo consecu i e yea s a Mami au´a
Lake, a his o ically impo an loodplain lake o he popu-
la ion. The second objec i e was desc ibing he ocal epe -
oi e o he wild Amazonian mana ee and analyzing ocal
beha io o in e hei daily ac i i y pa e ns.
Ma e ials and Me hods
S udy a ea
The s udy a ea is loca ed in he Mami au´a Sus ainable De el-
opmen Rese e (MSDR), in he s a e o Amazonas, B azil
(Fig. 1). Si ua ed a he con luence o he Solim˜
oes (uppe
Amazon Ri e ) and he Japu ´a i e s, wi h ci ca 11 000 km
2
2ª2024 The Au ho (s). Remo e Sensing in Ecology and Conse a ion published by John Wiley & Sons L d on behal o Zoological Socie y o London.
AI-based acous ic moni o ing o Amazonian mana ees F. E bs e al.
20563485, 0, Downloaded om h ps://zslpublica ions.onlinelib a y.wiley.com/doi/10.1002/ se2.418 by Readcube (Lab i a Inc.), Wiley Online Lib a y on [04/12/2024]. See he Te ms and Condi ions (h ps://onlinelib a y.wiley.com/ e ms-and-condi ions) on Wiley Online Lib a y o ules o use; OA a icles a e go e ned by he applicable C ea i e Commons License
o loodplains (whi ewa e i e loodplain o ´a zea) habi a s,
he MSDR is he la ges B azilian p o ec ed a ea dedica ed o
he conse a ion o looded ain o es s. Whi ewa e i e
loodplains, seasonally inunda ed by he sedimen - and
nu ien -laden Amazon Ri e wa e s and ibu a ies, a e
highly p oduc i e ecosys ems. Du ing high wa e le els,
(semi-)aqua ic plan s a e abundan and cons i u e he main
ood sou ce o he Amazonian mana ee (Gu e es-
Pazin, 2014). B eeding and cal ing pe iods seem synch onized
wi h he seasonal a ailabili y o ood (Bes , 1982).
Da a collec ion
An au onomous SM4 eco de (Wildli e Acous ics, USA)
wi h an HTI-96-Min hyd ophone (High Tech Inc., USA,
sensi i i y: 165 dB e 1 V/uPa) was ins alled in he
Mami au´a Lake (2°59055.500S, 64°54018.14400W), ixed
abo e he wa e line o a ee on he lake edge. The
hyd ophone was placed h ee me e s deep a mid- lood in
Janua y wi h a maximum dep h o ca. 12 m a he highes
wa e le els in June. Acous ic eco dings we e collec ed
on a du y cycle o 5 min on/55 min o , while sampling a
96 kHz. In 2021, 53 days o da a we e collec ed om he
end o May o mid-July. In 2022, da a we e collec ed
om mid-Janua y o he end o May and om he begin-
ning o July o mid-Augus p o iding 173 eco ded days.
CNN classi ica ion p ocedu e
The CNN a chi ec u e was ini ially de eloped unde P o-
jec P o idence (Zaugg e al., 2023) and was p e iously
applied o unde wa e da a o he classi ica ion o
Figu e 1. Map o he Mami au´
a Sus ainable De elopmen Rese e showing he eco ding loca ion a Mami au´
a Lake ( ed s a ). The habi a ypes
a e modi ied om Fe ei a-Fe ei a e al., 2015.
ª2024 The Au ho (s). Remo e Sensing in Ecology and Conse a ion published by John Wiley & Sons L d on behal o Zoological Socie y o London. 3
F. E bs e al. AI-based acous ic moni o ing o Amazonian mana ees
20563485, 0, Downloaded om h ps://zslpublica ions.onlinelib a y.wiley.com/doi/10.1002/ se2.418 by Readcube (Lab i a Inc.), Wiley Online Lib a y on [04/12/2024]. See he Te ms and Condi ions (h ps://onlinelib a y.wiley.com/ e ms-and-condi ions) on Wiley Online Lib a y o ules o use; OA a icles a e go e ned by he applicable C ea i e Commons License
Amazon Ri e dolphins (E bs e al., 2023). The ame-
wo k was o iginally c ea ed using Tenso low 1.14 and
Py hon 3.6.8 o backwa ds compa ibili y, al hough o
model execu ion, mo e ecen e sions a e gene ally used.
The classi ica ion p ocedu e is pe o med in i e s eps
(c. . E bs e al., 2023): labeling, augmen a ion, aining,
alida ion and inal es e alua ion. In he ollowing, label
da ase e e s o all he da a ha con ains labels used o
building he classi ica ion model. The aining da ase is
he po ion o he label da ase used o ain he CNN
model. The model lea ns pa e ns and ela ionships om
his da a. The alida ion da ase is ano he po ion o he
label da ase used o assess he model’s pe o mance a e
aining, ha is, p o ide pe o mance me ics. The es
da ase e e s o he da a used o a inal e alua ion o he
model’s pe o mance. In ou case, i is he o ali y o he
da a eco ded in 2021 and 2022, o which we sepa a ely
assess pe o mance pe eco ding pe iod. Fou eco ding
pe iods we e de ined based on looding s ages.
Labeling
The da ase s om ou sou ces om he Mami au´a and
Aman˜
a Sus ainable De elopmen Rese es (label da ase )
we e anno a ed o c ea e a da abase o mana ee call labels.
To ensu e p ope soundscape ep esen a ion in he label
da ase , wo da ase s ha did no con ain mana ee ocali-
za ions, bu did con ain o he sound sou ces such as dol-
phin whis les and echoloca ion clicks, ain, and boa
passages, we e added. Adding hese addi ional sounds-
capes esul ed in 12 classes in o al which allowed o
imp o e he pe o mance o he single mana ee class o
in e es . An o e iew o he label da ase is p esen ed in
Table 1. A cus om Py hon GUI p esen ed 5-s
spec og ams o sound labeling. The in e ace allowed o
de ining he equency and ime bounda ies o he a ge
sounds by d awing boxes on he spec og am. These
labels we e hen au oma ically impo ed in o a dedica ed
da abase. A o al o 5063 segmen s we e labeled, including
726 segmen s wi h mana ee ocaliza ions. The label da a-
se was andomly spli in o aining and es ing da ase s,
wi h 50% o he labels de o ed o aining he classi ie
and 50% o es ing.
Da a augmen a ion
As pa o he aining p ocedu e, he aining da ase was
a i icially expanded using a da a augmen a ion p ocedu e
ha included da a duplica ion and ans o ma ion
h ough equency and ime shi s, con as adjus men s,
and ime-wa ping. All new samples we e augmen ed, he
augmen a ion unc ions we e always applied o he ain-
ing da a a each ound; ha is each da a sample unde -
wen each augmen a ion. All pa ame e s ela ed o da a
augmen a ion we e uni o mly selec ed om a de ined
ange. These anges a e gene ally kep iden ical once
de ined o a speci ic model (speci ic sound ypes ha
need o be classi ied). The anges a e p o ided in
Table A1. The aining se , a e da a augmen a ion, com-
p ised 6567 segmen s, o which 1014 con ained mana ee
ocaliza ions.
Model aining
A classi ie model was ained and alida ed using a con-
olu ional neu al ne wo k (CNN). De ails o he CNN
a chi ec u e can be ound in E bs e al. (2023). Following
he same p ocedu e, he log-Mel spec og am was used as
Table 1. O e iew o he labeling da ase used o aining and es ing he CNN model. All au onomous eco dings we e collec ed a 96 kHz. The
boa -based eco dings we e collec ed a 576 kHz and esampled o 96 kHz o he analysis.
Da ase
loca ion Yea Habi a
Reco ding
equipmen
Reco ding
me hod
Du a ion
(h)
No. o
segmen s wi h
mana ee calls
No. o segmen s
wi hou mana ee
calls
ASDR
Aman˜
a
Lake
2018 En ance o he lake SM4 eco de and
HTI-96-Min
hyd ophone
Au onomous 0.67 103 471
MSDR
Mami au´
a
Lake
2021 Floodplain lake SM4 eco de and
HTI-96-Min
hyd ophone
Au onomous 2.42 623 1740
MSDR 2019
2020
Di e en loodplain habi a s (bay,
channel, loodplain lakes,
looded o es )
SoundT ap 300 HF
eco de
Boa -based 2.40 0 1728
MSDR 2019 Flooded o es SM4 eco de and
HTI-96-Min
hyd ophone
Au onomous 1.56 0 1124
4ª2024 The Au ho (s). Remo e Sensing in Ecology and Conse a ion published by John Wiley & Sons L d on behal o Zoological Socie y o London.
AI-based acous ic moni o ing o Amazonian mana ees F. E bs e al.
20563485, 0, Downloaded om h ps://zslpublica ions.onlinelib a y.wiley.com/doi/10.1002/ se2.418 by Readcube (Lab i a Inc.), Wiley Online Lib a y on [04/12/2024]. See he Te ms and Condi ions (h ps://onlinelib a y.wiley.com/ e ms-and-condi ions) on Wiley Online Lib a y o ules o use; OA a icles a e go e ned by he applicable C ea i e Commons License
inpu in o he classi ie he e. The inal model was selec ed
wi h andom weigh ini ializa ion, 300 epochs, 32 ba ch
size, 0.001 lea ning a e, and 0.005 decay a e. Da a we e
eshu led o each epoch (see Table A2 o he model
dimensions and Figu e A1 o he ne wo k layou ).
Model pe o mance on he alida ion da ase
The pe o mance o he inal model was e alua ed by
compu ing he p ecision and he ecall. The p ecision (o
posi i e p edic i e alue, PPV) ep esen s he a io o co -
ec posi i e p edic ions o he o al posi i e p edic ions
made by he model. The ecall (o ue posi i e a e,
TPR) ep esen s he a io o co ec posi i e p edic ions
ou o all ac ual posi i e ins ances in he da ase . These
wo me ics o m he p ecision- ecall cu e, wi h each
poin o he cu e associa ed o a decision h eshold
(Fig. 2). The a ea unde he p ecision- ecall cu e, also
called he a e age p ecision, was calcula ed o he mana-
ee ocaliza ion class. A decision h eshold o 0.28, co e-
sponding o a ecall o 0.9, was selec ed om his cu e
as a s a ing poin o ine uning he h esholds on he
es da ase .
P ecision PPVðÞ=
T ue posi i e
T ue posi i e þFalse posi i eðÞ
Recall TPRðÞ=
T ue Posi i e
T ue Posi i e þFalse nega i e
A e age p ecision APðÞ¼∑ ¼T1
¼0TPR ðÞTPR þ1ðÞ½
PPV ðÞ
wi h T he numbe o h esholds and TPR(T) =0,
PPV(T) =1.
The model pe o mance achie ed a 0.98 a e age p eci-
sion on he alida ion da ase . This alue co esponds o
he a ea unde he p ecision- ecall cu e o his model
(Fig. 2).
Model pe o mance on he es da ase
The classi ica ion model ob ained a he aining s age
was un o e each 5-min ile eco ded a he Mami au´
a
Lake in 2021 and 2022. A andom selec ion o 100 seg-
men s pe eco ding pe iod was hen e alua ed by a
human expe (FE) o assess whe he he segmen p edic-
ion was co ec a he decision h eshold selec ed a he
aining s age. This p ocedu e was epea ed o ine- une
he decision h eshold in o de o achie e simila ecall
and p ecision alues o each eco ding pe iod. The deci-
sion h esholds and associa ed p ecision and ecall alues
a e lis ed in Table A3. The model could con iden ly iden-
i y mana ee ocaliza ions wi h a ecall anging om 0.97
o 1 and a p ecision be ween 0.83 and 1. The lowe p eci-
sion ob ained in Janua y–Ma ch 2022, when wa e le els
we e s ill low and he hyd ophone close o he su ace,
was ela ed o he p esence o ai bo ne sounds (absen
om he aining da ase ).
Mana ee occu ence based on CNN
classi ica ion o ocaliza ions
I he pe -segmen ou pu o he classi ie sco ed abo e
he decision h eshold, he segmen was assigned posi i e
o mana ee p esence. Segmen s assigned as posi i e we e
used o compu e he daily numbe o eco ding hou s
wi h mana ee de ec ions.
Mana ee diel pa e n o acous ic ac i i y
The numbe o posi i e segmen s (i.e., pe -segmen ou -
pu alue abo e he decision h eshold) was coun ed by
day and by hou o he day. Time pe iods be ween 06:00
and 17:59 (AMT, Amazon ime) we e assigned o he
“day ime” ca ego ical alue, and pe iods be ween 18:00
and 05:59 we e assigned o “nigh ime” based on sun ise
and sunse o he eco ding pe iods. To es i he e was
mo e ac i i y du ing he day han he nigh we es ed he
p opo ion o day ime ac i i y o he p opo ion o nigh -
ime ac i i y by pe o ming a p opo ions z- es o each
sepa a e yea .
Cha ac e iza ion o he Amazonian mana ee
ocal epe oi e
To a oid sampling he same indi iduals ha migh be
p esen in he a ea o e hou s du ing he same day, only
Figu e 2. P ecision- ecall cu e o he Amazonian mana ee CNN
classi ie .
ª2024 The Au ho (s). Remo e Sensing in Ecology and Conse a ion published by John Wiley & Sons L d on behal o Zoological Socie y o London. 5
F. E bs e al. AI-based acous ic moni o ing o Amazonian mana ees
20563485, 0, Downloaded om h ps://zslpublica ions.onlinelib a y.wiley.com/doi/10.1002/ se2.418 by Readcube (Lab i a Inc.), Wiley Online Lib a y on [04/12/2024]. See he Te ms and Condi ions (h ps://onlinelib a y.wiley.com/ e ms-and-condi ions) on Wiley Online Lib a y o ules o use; OA a icles a e go e ned by he applicable C ea i e Commons License
one ile pe day was andomly selec ed om all he iles
wi h classi ica ion alues abo e he decision h eshold. A
o al o 85 iles we e selec ed, wi h 48 iles om 2021 and
37 om 2022. Each ile was manually inspec ed in Adobe
Audi ion .5 (Adobe Sys ems Inc., USA) o con i m he
p esence o mana ee calls. Up o i e calls pe ile we e
selec ed wi h an SNR o e 10 dB and measu ed in Ra en
P o 1.5 (Bioacous ics Resea ch P og am, 2011) using a
2048-poin FFT (Fas Fou ie T ans o m), Hann window
wi h 50% o e lap, co esponding o a spec og am e-
quency and ime esolu ion o 46.9 Hz and 21.3 ms,
espec i ely. Va iables we e selec ed based on p e ious li -
e a u e on he mana ee acous ic epe oi e (B ady
e al., 2020; Rycyk e al., 2021; Sousa-Lima e al., 2008).
Mana ee ocaliza ions ypically p esen se e al ha monics.
Fo each mana ee call, he ollowing pa ame e s we e
measu ed on he emphasized equency: s a equency,
end equency, s a ime, end ime, du a ion, peak e-
quency and numbe o he emphasized equency ha -
monic. The emphasized peak equency is de ined as he
peak equency o he ha monic displaying he highes
acous ic ene gy; he undamen al peak equency as he
peak equency o he i s ha monic. Mos o he ime,
he undamen al equency was no measu able on he
spec og am due o a lowe SNR and/o masking by o he
biological sounds (amphibians and ish calls). The unda-
men al peak equency was compu ed by di iding he
alue o he emphasized peak equency by he numbe o
he co esponding ha monic ( ollowing Sousa-Lima
e al., 2002). Call equency con ou ca ego ies (unmodu-
la ed, hill-shaped, mul iple in lec ions and downsweep)
we e assigned based on simila i y wi h call ypes desc ibed
in li e a u e (B ady e al., 2020,2022; O’Shea & Poch´
e
J , 2006). The numbe o no es (as desc ibed in
Sousa-Lima e al., 2002) was coun ed. To in es iga e
po en ial changes in ocal beha io be ween yea s, a - es
was pe o med on emphasized equency and du a ion.
Compa ison o Mana ee ocaliza ion
pa ame e s and acous ic beha io
be ween yea s
Using 64 days o da a ( andomly selec ing 34 iles om
2021 and 30 om 2022 o ob ain a balanced da a se ), we
compa ed he emphasized peak equency and du a ion
be ween he 2 yea s (using a - es ). The in e -call in e al
(ICI) was compa ed ( - es ) be ween he yea s based on
10 andomly selec ed iles o each yea , on call sequences
wi h simila equency con ou shape. The same iles ha
we e used o he ICI we e used o compu e he ocaliza-
ion a e. Calls we e manually coun ed and his coun
was di ided by he du a ion o he ile o ob ain a alue
in no. he numbe o calls pe minu e. The ocaliza ion
a e was compu ed o e he whole ile leng h, including
he pe iods o silence be ween epe i i e call sequences.
Mana ee indi idual coun pe ile
The 85 iles selec ed o ocaliza ion measu emen we e
manually inspec ed in Adobe Audi ion .5. The numbe
o call con ou shapes was coun ed based on spec al
cha ac e is ics (especially equency modula ion). As calls
o indi iduals p esen less a ia ion han in e indi idual
ones (Sousa-Lima e al., 2002), and because mos calls
occu ed in sequences (Figu e A2), he numbe o con-
ou ypes was used as a p oxy o he numbe o indi id-
uals p esen in one ile.
Resul s
Mana ee occu ence in 2021 and 2022
Mana ee ocaliza ions we e de ec ed du ing he wo con-
secu i e yea s a a ime pe iod co esponding o he high-
es wa e le els a Mami au´a Lake (Fig. 3). In 2021, da a
we e collec ed exclusi ely du ing high wa e s, om he
end o May o mid-July. On he 53 eco ding days, mana-
ees we e acous ically p esen du ing 50 consecu i e days,
wi h a daily p esence anging om one o 11 h. The peak
p esence occu ed du ing he i s hal o June. In 2022,
he da a collec ion was ex ended o include he s a o
he looded pe iod (Janua y o Ma ch) and he beginning
o he eceding wa e pe iod (Augus ). Due o echnical
issues wi h he eco ding equipmen , da a we e no col-
lec ed be ween mid-May and he beginning o July. Man-
a ees we e de ec ed spo adically du ing he i s imes e ,
wi h 5 days o p esence. De ec ions inc eased om Ap il
on, and du ing May mana ees we e de ec ed on 17 o he
23 eco ding days, wi h daily p esence anging om 1 o
5 h, simila o he p e ious yea o he same ime pe iod.
F om mid-July o Augus , when wa e le els s a ed o
lowe , he e we e only 2 days wi h one posi i e hou .
Mana ee diel acous ic pa e n
In 2021, mana ees appea ed o be p esen and ocally
ac i e be ween 19:00 and 06:00 (all imes a e local ime
and in 24 h clock o ma ), ha is, du ing nigh ime
(Fig. 4A). Mana ees we e de ec ed in 88% o he nigh -
ime and 44% o he day ime pe iods. The di e ence in
he p opo ion o day ime ac i i y o he p opo ion o
nigh ime ac i i y was s a is ically signi ican (z- alue:
4.78; P- alue 0.001). The peak de ec ion hou in 2021
occu ed a 02:00 and ep esen ed 12.8% o he nigh ime
de ec ion and 11.8% o he o al de ec ions. This pa e n
was no isible in 2022 whe e mana ee de ec ions we e
6ª2024 The Au ho (s). Remo e Sensing in Ecology and Conse a ion published by John Wiley & Sons L d on behal o Zoological Socie y o London.
AI-based acous ic moni o ing o Amazonian mana ees F. E bs e al.
20563485, 0, Downloaded om h ps://zslpublica ions.onlinelib a y.wiley.com/doi/10.1002/ se2.418 by Readcube (Lab i a Inc.), Wiley Online Lib a y on [04/12/2024]. See he Te ms and Condi ions (h ps://onlinelib a y.wiley.com/ e ms-and-condi ions) on Wiley Online Lib a y o ules o use; OA a icles a e go e ned by he applicable C ea i e Commons License
simila be ween day ime and nigh ime (40.3 and 38.4%,
espec i ely; z- alue: 0.2; P- alue =0.84). The highes
numbe o de ec ions occu ed du ing he day, a 06:00
(44.9% o he day ime de ec ions, Fig. 4B).
Cha ac e iza ion o Amazonian mana ee
ocal epe oi e
F om he 85 iles selec ed, a o al o 422 calls we e
selec ed o measu emen s. All calls we e onal (wi hou
a onal componen s), and ou main call ca ego ies we e
obse ed based on hei equency con ou s: unmodu-
la ed, hill-shaped, mul iple in lec ions and downsweep.
Figu e 5p o ides an example spec og am o each call
ca ego y. Nonlinea elemen s such as subha monics we e
obse ed (Fig. 5A). Mul iple in lec ions o he equency
con ou we e p esen in 77.8% o he calls, 11.8% p e-
sen ed a hill-shaped con ou , 9.4% had a p onounced
downsweep componen and less han 1% had a la e-
quency con ou . The majo i y o he calls we e composed
o one (69.0%) o wo (27.2%) no es, wi h only 3.8%
con aining h ee no es. In 98% o he calls, he empha-
sized equency was cen e ed on he second o highe ha -
monic. The mean emphasized peak equency was 10.2
(2.7 SD) kHz. The mean s a and end equencies (mea-
su ed om he emphasized equency) we e ela i ely
simila (9.98 s. 9.63 kHz). The undamen al peak e-
quency was 3.6 (0.4 SD) kHz. The mean du a ion o a
call was 0.294 (0.072 SD) s (Table 2).
Compa ison o ocaliza ion pa ame e s and
acous ic beha io be ween yea s
F equency (emphasized equency) and empo al (du a-
ion, in e -call in e al) pa ame e s we e compa ed
be ween 2021 and 2022 (Table A4). The mean empha-
sized equency o 2021 was highe in 2021 compa ed o
2022 ( - es , (327) =2.52, P=0.01) a α=0.05. Du a ion
alues ( - es , (327) =0.69, P=0.49) we e no di e en
be ween yea s (P[0.05).
In e -call in e als (ICIs) we e compu ed o e he
whole ile (5 min), which o en included calls emi ed in
sequences and silen pe iods (Figu e A2). On a e age,
calls we e emi ed 2.91 s apa in 2021 and 3.42 s apa
in 2022 ( - es , (1273) =1.78, P=0.08). The maxi-
mum ICI alues we e abo e 1 min o bo h yea s,
indica i e o calls no emi ed in sequences. The a e age
calling a e was 13.9 (2021) and 12.5 (2022) calls
pe minu e, wi h a maximum o 27 calls pe minu e
(Table A4).
Coun e calling
Files manually analyzed o ocaliza ion measu emen s
we e also e iewed o iden i ying he numbe o ocaliza-
ion ypes in a ile. Only 3 ou o 64 iles p esen ed wo
di e en equency con ou shapes (Fig. 6). Two iles we e
eco ded du ing 2021, on June 5 and June 13. The hi d
ile was eco ded on Ap il 17, 2022.
Figu e 3. Amazonian Mana ee acous ic p esence a Mami au´
a Lake (B azil), based on CNN classi ica ion. The ligh g ay a eas indica e an absence
o eco dings. The da k g ay a ea indica es ime pe iods no moni o ed due o low wa e le els.
ª2024 The Au ho (s). Remo e Sensing in Ecology and Conse a ion published by John Wiley & Sons L d on behal o Zoological Socie y o London. 7
F. E bs e al. AI-based acous ic moni o ing o Amazonian mana ees
20563485, 0, Downloaded om h ps://zslpublica ions.onlinelib a y.wiley.com/doi/10.1002/ se2.418 by Readcube (Lab i a Inc.), Wiley Online Lib a y on [04/12/2024]. See he Te ms and Condi ions (h ps://onlinelib a y.wiley.com/ e ms-and-condi ions) on Wiley Online Lib a y o ules o use; OA a icles a e go e ned by he applicable C ea i e Commons License
Figu e 4. (A) Radial hea maps o mana ee daily acous ic ac i i y o he yea s 2021 and 2022. The colo scale is based on he p esence coun ,
ha is, he numbe o segmen s wi h de ec ions g ouped by hou (see Me hods sec ion 2.5). Values on he y-axis a e da es in o ma MM-DD.
G ay a eas ep esen absence o da a. (B) His og ams o he mana ee acous ic de ec ions ep esen ing he cumula i e numbe o segmen s wi h
de ec ions g ouped by hou o he day o he yea s 2021 and 2022.
Figu e 5. Spec og am examples o he ou ca ego ies o Amazonian mana ee calls, based on equency con ou . (A) Unmodula ed. (B) Hill-
shaped. (C) Mul iple in lec ions. (D) Downsweep. No e he p esence o subha monics on ype A.
8ª2024 The Au ho (s). Remo e Sensing in Ecology and Conse a ion published by John Wiley & Sons L d on behal o Zoological Socie y o London.
AI-based acous ic moni o ing o Amazonian mana ees F. E bs e al.
20563485, 0, Downloaded om h ps://zslpublica ions.onlinelib a y.wiley.com/doi/10.1002/ se2.418 by Readcube (Lab i a Inc.), Wiley Online Lib a y on [04/12/2024]. See he Te ms and Condi ions (h ps://onlinelib a y.wiley.com/ e ms-and-condi ions) on Wiley Online Lib a y o ules o use; OA a icles a e go e ned by he applicable C ea i e Commons License
Discussion
We combined passi e acous ic moni o ing and deep
lea ning algo i hms o moni o he seasonal p esence o
Amazonian mana ees a a loodplain lake whe e mana ees
we e his o ically abundan . Ou analysis was conduc ed
o e wo looded seasons in 2021 and 2022 and show ha
mana ees we e p esen a Mami au´a Lake du ing mos o
he days a peak p esence, some imes up o 11 h a day,
demons a ing ha his loca ion is likely an impo an
habi a o he local mana ee popula ion.
Classi ica ion pe o mance
The CNN classi ie de eloped he e p esen ed e y high
pe o mance o he au oma ic iden i ica ion o mana ee
ocaliza ions wi h a 0.98 a e age p ecision on aining
da a. On he ull Mami au´a lake da ase , he model
Table 2. Summa y o he desc ip i e s a is ics o he a iables measu ed on wild Amazon mana ee ocaliza ions eco ded a Mami au´
a Lake,
Amazonas, B azil, in 2021 and 2022.
Va iable Fundamen al equency (Hz) Emphasized equency (Hz) S a equency (Hz) End equency (Hz) Du a ion (s)
N265 423 423 423 423
Mean 3676 10 197 9979 9629 0.294
S anda d de ia ion 442 2728 2631 2611 0.072
Min 2625 2766 2766 3000 0.107
Max 4617 18 000 17 016 17 203 0.499
Median 3656 9891 9656 9281 0.286
Figu e 6. Spec og ams o Amazonian mana ee “due s” eco ded a Mami au´
a Lake (B azil) on June 5, 2021 ( op), June 13, 2021 (middle), and
Ap il 17, 2022 (bo om). The blue and pu ple ma ks on each spec og am indica e he empo al loca ion o wo di e en ypes o signals (no e
ha he ypes o signals a e no he same be ween spec og ams).
ª2024 The Au ho (s). Remo e Sensing in Ecology and Conse a ion published by John Wiley & Sons L d on behal o Zoological Socie y o London. 9
F. E bs e al. AI-based acous ic moni o ing o Amazonian mana ees
20563485, 0, Downloaded om h ps://zslpublica ions.onlinelib a y.wiley.com/doi/10.1002/ se2.418 by Readcube (Lab i a Inc.), Wiley Online Lib a y on [04/12/2024]. See he Te ms and Condi ions (h ps://onlinelib a y.wiley.com/ e ms-and-condi ions) on Wiley Online Lib a y o ules o use; OA a icles a e go e ned by he applicable C ea i e Commons License
