emo e sensing
A icle
Repea able Seman ic Ree -Mapping h ough Pho og amme y
and Label-Augmen a ion
Ma an Yu al 1,2,*,† , Iñigo Alonso 3, Gal Eyal 4,5 , Dan Tche no 2, Yossi Loya 6,
Ana C. Mu illo 3and Tali T eibi z 1


Ci a ion: Yu al, M.; Alonso, I.; Eyal,
G.; Tche no , D.; Loya, Y.; Mu illo,
A.C.; T eibi z, T. Repea able Seman ic
Ree -Mapping h ough
Pho og amme y and
Label-Augmen a ion. Remo e Sens.
2021,13, 659. h ps://doi.o g/
10.3390/ s13040659
Academic Edi o : John Bu ns
Recei ed: 19 Janua y 2021
Accep ed: 9 Feb ua y 2021
Published: 11 Feb ua y 2021
Publishe ’s No e: MDPI s ays neu-
al wi h ega d o ju isdic ional clai-
ms in published maps and ins i u io-
nal a ilia ions.
Copy igh : © 2021 by he au ho s. Li-
censee MDPI, Basel, Swi ze land.
This a icle is an open access a icle
dis ibu ed unde he e ms and con-
di ions o he C ea i e Commons A -
ibu ion (CC BY) license (h ps://
c ea i ecommons.o g/licenses/by/
4.0/).
1Ha e Depa men o Ma ine Technologies, Cha ney School o Ma ine Sciences, Uni e si y o Hai a,
Hai a 3498838, Is ael; eibi z@uni .hai a.ac.il
2
Depa men o Ma ine Biology, Cha ney School o Ma ine Sciences, Uni e si y o Hai a,
Hai a 3498838
, Is ael;
d che no @uni .hai a.ac.il
3A agón Ins i u e o Enginee ing Resea ch (I3A), Uni e si y o Za agoza, 50009 Za agoza, Spain;
inigo@uniza .es (I.A.); acm@uniza .es (A.C.M.)
4
ARC Cen e o Excellence o Co al Ree S udies, School o Biological Sciences, The Uni e si y o Queensland,
Douglas, QLD 4814, Aus alia; [email p o ec ed]
5The Mina & E e a d Goodman Facul y o Li e Sciences, Ba -Ilan Uni e si y, Rama Gan 5290002, Is ael
6School o Zoology, Tel-A i Uni e si y, Tel A i 6997801, Is ael; [email p o ec ed]
*Co espondence: [email p o ec ed]
† Cu en add ess: Ha e Depa men o Ma ine Technologies & Mo is Kahn Ma ine Resea ch S a ion,
Cha ney School o Ma ine Sciences, Uni e si y o Hai a, Hai a 3498838, Is ael.
Abs ac :
In an endea o o s udy na u al sys ems a mul iple spa ial and axonomic esolu ions,
he e is an u gen need o au oma ed, high- h oughpu amewo ks ha can handle ple ho a o
in o ma ion. The coalescence o emo e-sensing, compu e - ision, and deep-lea ning elici s a new e a
in ecological esea ch. Howe e , in complex sys ems, such as ma ine-ben hic habi a s, key ecological
p ocesses s ill emain enigma ic due o he lack o c oss-scale au oma ed app oaches (mms o kms) o
communi y s uc u e analysis. We add ess his gap by wo king owa ds scalable and comp ehensi e
pho og amme ic su eys, ackling he p o ound challenges o ull seman ic segmen a ion and 3D
g id de ini ion. Full seman ic segmen a ion (whe e e e y pixel is classi ied) is ex emely labou -
in ensi e and di icul o achie e using manual labeling. We p opose using label-augmen a ion, i.e.,
p opaga ion o spa se manual labels, o accele a e he ask o ull segmen a ion o pho omosaics.
Pho omosaics a e syn he ic images gene a ed om a p ojec ed poin -o - iew o a 3D model. In
he lack o na iga ion senso s (e.g., a di e -held came a), i is di icul o epea ably de e mine
he slope-angle o a 3D map. We show his is especially impo an in complex opog aphical
se ings, p e alen in co al- ee s. Speci ically, we e alua e ou app oach on ben hic habi a s, in h ee
di e en en i onmen s in he challenging unde wa e domain. Ou app oach o label-augmen a ion
shows human-le el accu acy in ull segmen a ion o pho omosaics using labeling as spa se as 0.1%,
e alua ed on se e al ecological measu es. Mo eo e , we ound ha g id de ini ion using a le ele
imp o es he consis ency in communi y-me ics ob ained due o occlusions and opology (angle
and dis ance be ween objec s), and ha we we e able o s anda dise he 3D ans o ma ion wi h
wo pe cen e o in size measu emen s. By signi ican ly easing he anno a ion p ocess o ull
segmen a ion and s anda dizing he 3D g id de ini ion we p esen a seman ic mapping me hodology
enabling change-de ec ion, which is p ac ical, swi , and cos -e ec i e. Ou wo k low enables
epea able su eys wi hou pe manen ma ke s and specialized mapping gea , use ul o esea ch
and moni o ing, and ou code is a ailable online. Addi ionally, we elease he Ben hos da a-se , ully
manually labeled pho omosaics om h ee oceanic en i onmen s wi h o e 4500 segmen ed objec s
use ul o esea ch in compu e - ision and ma ine ecology.
Keywo ds:
pho og amme y; o ho ec i ica ion; change-de ec ion; communi y ecology;
label-augmen a ion; co al- ee s; ben hic mapping; compu e - ision; mul i-le el supe pixels
Remo e Sens. 2021,13, 659. h ps://doi.o g/10.3390/ s13040659 h ps://www.mdpi.com/jou nal/ emo esensing
Remo e Sens. 2021,13, 659 2 o 19
1. In oduc ion
Accele a ions in echnologies [
1
] ha e empowe ed ecological s udies by acili a ing
digi al ep esen a ions o na u al sys ems [
2
], hus educing unce ain ies in p edic ing hei
u u e-s a e [
3
]. Ad ances in compu e - ision and emo e-sensing enable c oss-scale e-
sea ch. In he nea u u e, deep neu al ne wo ks will help o deciphe p ocess- om-pa e n
as pa o au oma ed wo k lows; p eceded by da a acquisi ion om obo ic pla o ms and
seman ic segmen a ion o image-based maps [
4
,
5
]. Image-based mapping and seman ic
segmen a ion a e used in an a ay o ecological s udies and applica ions, anging om
s udying ege a ion pa e ns [
6
–
8
] and ci y-scapes [
9
] o a m-managemen [
10
]. Speci -
ically, pho og amme y has become a popula app oach o ben hic esea ch and ee
moni o ing [
11
–
21
]. S uc u e-F om-Mo ion (SFM) pho og amme y es ima es he 3D scene
s uc u e and ela i e mo ion using subsequen images. I is now possible o iew an
ecosys em wi hin a digi al amewo k as a con inuum ac oss spa ial scales, and examine
he indi iduals, popula ions, and communi ies ha comp ise i . Ne e heless, pho og am-
me y is no ye ully ma u e as a epea able me hod o wide scale ecological su eys. Fi s ,
he ou pu 3D models and pho omosaics need o be labeled igo ously o analysis. This is
labo ious and equi es expe knowledge. Thus, he e is an u gen need o au oma ion
in he ull segmen a ion ask (i.e., labeling each pixel) o pho omosaics. Second, a 3D g id
needs o be consis en ly de ined o epea ed su eys. Wi hou p ope da a ex ac ion ha
includes ull, pixel-wise classi ica ion and labeling, he ele an in o ma ion emains con-
cealed in he image. He e, we add ess bo h issues, p o iding a mo e cohe en solu ion o
habi a -mapping and unde wa e pho og amme y. While ou me hods a e applicable o
all domains in which pho og amme y is used, he e we ocus on he ben hic en i onmen .
The e a e inc easing e o s o au oma ic labeling using machine lea ning [
4
,
6
,
9
,
20
,
22
].
Howe e , he commonly used ools [
23
,
24
] s ill p o ide poin classi ica ion and no ull
segmen a ion. Such spa se sampling is o e looking objec /pa ch le el in o ma ion, such
as he mo pho-me ics (shape and size) o indi idual o ganisms ha can be p o ided
by ull seman ic segmen a ion. Se e al me hods o segmen a ion o ben hic images and
pho omosaics ha e been demons a ed [
25
–
29
], including using mul i- iew images [
30
]
and 3D models [
31
]. These wo ks ha a e based on deep lea ning p o ide imp essi e
esul s; howe e , deep neu al ne wo ks ely on a high numbe o lea ning pa ame e s and
because o ha , hey need o be ained wi h a la ge amoun o da a o a oid o e i ing.
Then, he main p oblem o success ul au oma ic iden i ica ion o ma ine species is he lack
o aining da a and ex ensi e a iabili y wi hin axa [
32
,
33
] ha p e en s using labeled
da a om o he loca ions and p edic ing labels ha we e no used in he aining da a. To
o e come his, we p opose p opaga ing spa se labels using ou Mul i-Le el Supe pixel
(MLS) app oach [
25
]. In [
25
] his me hod was sugges ed as a way o quickly gene a e
aining da a o deep lea ning seman ic segmen a ion in se e al e es ial and unde wa e
domains. He e we show ha e en by i sel i enables ob aining as ull segmen a ion
wi h minimal human in e en ion. We es i ex ensi ely on pho omosaics wi h espec o
ecological measu emen s and show ha i p o ides e y high accu acy. Thus, i can be used
as a complimen a y me hod o gene a ing dense aining da a in cases whe e he e a e no
a ailable ained deep ne wo ks as i is gene al and no domain speci ic. Challenges o
deep-lea ning algo i hms in unde wa e imaging include illumina ion and ange, as well
as image deg ada ion caused by e ac ion and wa eleng h-speci ic a enua ion [
34
,
35
].
An o hopho o is gene a ed om a single angle-o - iew on he 3D model h ough he
p ocess o o ho ec i ica ion whe e a planime ically co ec image is c ea ed by emo ing
he e ec s o pe spec i e ( il ) and elie ( e ain). In an o hopho o, he objec s a e scaled
and loca ed in hei ue posi ions ( opology), enabling di ec measu emen s o a eas and
dis ances [
36
]. Howe e , in ansi ion om 3D o 2D (o ho ec i ica ion) he e a e six
deg ees-o - eedom ha need o be se . In opog aphically complex s uc u es, such as
co al ee s, expo ing di e en pe spec i es o he same 3D model a ec s he occlusions
(Figu e 1) and map- opology, as well as a i ac ing and dis o ion on non-plana objec s
wi h limi ed inpu iews. Thus, he dis ance and angle be ween o ganisms may di e
Remo e Sens. 2021,13, 659 3 o 19
wi hou consis ency in o ho ec i ica ion. This can be de imen al, o example, in s udies
ega ding neighbo - ela ions and size-dis ibu ions.
Figu e 1.
The e ec o o ho ec i ica ion wi h a le ele on co al opology, p e alence and size: Th ee pho omosaic eplica es
we e gene a ed subsequen ly. (
a
) G ound- u h (baseline) pho omosaic. (
b
) Replica e which was o ho ec i ied using he
spi i le ele as a e e ence. (
c
) No le ele (Naï e) was o ho ec i ied wi hou in e en ion (3D ans o ma ion). The numbe s
in yellow (le ) a e close-ups on he columns.
Mos solu ions o de ining he plane o p ojec ion y o de ine he Z-axis acco ding
o dep h in he wa e -column. Usually, pe manen ma ke s such as plas ic ubes o s eel
bol s a e used o his pu pose [
15
], and hei dep h and he dis ance be ween hem need o
be measu ed di ec ly o indi ec ly [
37
]. O he means o sol e his p oblem include owed
buoys moun ed wi h GPS senso s [
38
] in shallow wa e su eys, and posi ioning wi h
acous ic da a [
39
]. Ye , hese solu ions a e imp ac ical o deep and emo e ee habi a s
such as Mesopho ic Co al Ecosys ems (MCEs, 30–150 m dep h) [40].
To ackle his p oblem, we de ine he Z-axis as he dep h axis by placing a spi i le ele
wi hin he su ey plo and using i o ans o m he 3D model.
Ben hic habi a mapping using acous ic and op ic senso s encompassess a ange o oci
and scales, om species dis ibu ion models o communi y mapping and abio ic habi a
mapping [
41
]. Op ical imaging can p o ide much g ea e de ail han acous ic senso s,
which ha e wide scalabili y. Howe e , ben hic habi a s a e di icul o map due o he
complex in e ac ions be ween physical, chemical, biological, and beha io al elemen s
ha comp ise hem [
42
]. He e we p esen a mul i-class communi y mapping scheme o
ben hic su eys.
The sessile communi ies ha o m and inhabi he ee a e linked h ough c oss-scale
p ocesses. Fo ins ance, in scle ac inian co als, g ow h- a es and neighbo in e ac ions
occu a e y small spa ial scales, ye hey ope a e wi hin a much mo e expansi e sys em,
whe e dispe sion is enhanced by p eda ion and ex eme wea he e en s [
43
], and ica iance
is e icula e h ough ocean cu en s [
44
]. Acco dingly, bo h he minu e and he eno mous
scales a e signi ican in cha ac e izing he physical and biological ea u es o ee s uc u es.
The composi ion o axa in space and ime has been he ocus o many s udies in ben hic
ecology. Howe e , ee s a e so in ica e (Figu e 2) ha in he lack o adequa e echnology
Remo e Sens. 2021,13, 659 4 o 19
o communi y-le el in es iga ion, he dynamics o sessile o ganisms emain puzzling.
Thus, undamen al ques ions ega ding key ecological p ocesses in he ee ha e emained
la gely he same o o e i e decades [
45
–
49
], as a simpli ied compa men aliza ion o he
ben hos is o en made o handling complex phenomena.
Figu e 2.
The main challenges in ben hic image segmen a ion a e due o plas ici y, i egula shapes,
and elabo a e 3D s uc u es. (
a
) The ben hic communi y s uc u e in Eila , he Red Sea, is composed
mainly o Scle ac inian co als. (
b
) The ee s in he eas e n Ca ibbean a e shi ing owa ds a sponge
and so -co al domina ed communi y. (
c
) In he eas e n Medi e anean, he ocky ee is empo a ily
domina ed by u algae.
In ecological s udies, he scale o in es iga ion depends on he a e o e en s [
50
,
51
].
Ben hic o ganisms ha e g ow h a es on he scale o mms o cms pe yea [
52
]. The e o e,
ou in es iga ion necessi a es cm scale change-de ec ion abili ies. To assess and alida e
he change-de ec ion abili y o ou wo k low, we conduc a epea ed su ey and show
ha such o ho ec i ica ion enables consis en ly examining he g ow h and decay, spa ial
opology, and p esence/absence coun s o sessile ee o ganisms.
Ou me hodology o au oma ed and epea able seman ic mapping can de ec and
eloca e sessile o ganisms on he cm-scale ac oss hund eds o me es. Such a ool can assis
in cons uc ing a mul i-le el, c oss-scale iew o unde wa e and e es ial ecosys ems,
use ul o esea ch and moni o ing e o s. In his pape , we desc ibe i s applica ion on a
new da a-se ha includes manually segmen ed pho omosaics om h ee di e en egions:
a ocky ee in he Eas e n Medi e anean, a co al ee in he No he n Red-Sea, and a co al
communi y in he Eas e n Ca ibbean. We alida e ou app oach h ough compu e - ision
me ics as well as ele an ecological me ics.
Ou speci ic con ibu ions a e:
•
Ex ensi e ecological alida ion o seman ic segmen a ion h ough label-augmen a ion
o spa se anno a ions.
• Valida ion o 3D g id s anda disa ion wi h a consume -g ade spi i -le ele .
•
The Ben hos da a-se ha includes h ee segmen ed pho omosaics om di e en
oceanic en i onmen s.
2. Ma e ials and Me hods
2.1. Imaging Sys em and Pho og amme ic Equipmen
A NIKON D850 came a wi h a 35 mm NIKKOR lens in a Nau icam housing wi h
ou INON Z-240 s obes was used (Figu e 3b). Pho og amme ic a ge s a e objec s wi h
dis inguishable ea u es and o ien a ion. Ou a ge s included measu ing apes, 0.5 m
scale-ba s, unde wa e colou cha s (DGK), a spi i le ele , and di e sla es wi h elec ical
ape ma kings (Figu e 3a).
Remo e Sens. 2021,13, 659 5 o 19
Figu e 3.
Wo k low o seman ic mapping: (
a
) Scale ba s a e loca ed nex o a dis inguishable objec
a he su ey s a ing poin , sla e and colou -ca d a e used as pho og amme ic a ge s, and he
spi i le ele is aligned in he scene and used o 3D g id de ini ion in pos p ocessing. (
b
) Image
acquisi ion is ca ied ou using a di e -held imaging sys em. (
c
) An RGB pho omosaic is p oduced
and (
d
) labeled spa sely. (
e
) Labels a e augmen ed o ull e ain depic ion using Mul i Le el
Supe pixels (MLS). (
) Communi y s a is ics such as class speci ic size- equency dis ibu ion a e
ex ac ed au oma ically.
2.2. Plo Se up and Acquisi ion P o ocol
When eaching he a ge dep h, a dis inguishable na u al o a i icial objec which is
ela i ely simple o na iga e o was de ec ed as a s a ing poin o he su ey. F om ha
poin , we measu ed he equi ed ansec leng h (5–30 m) using a measu ing ape and
ma ked i s su oundings using pho og amme ic a ge s and scale ba s. In he o ho ec i-
ica ion expe imen s, we aligned a spi i le ele in he su ey plo . The spi i le ele has
h ee bubble indica o s (Figu e 3a). When i is placed in such a way ha he bubbles a e
cen ed, he le ele can be used o de ine a plane-o -p ojec ion. Op imally, he le ele was
placed in he cen e o he plo , and pa allel o he ansec . The le ele is used o ans o m
he 3D model, hus i is pa amoun o ob ain a good econs uc ion o i by acqui ing many
(>15) images om di e en angles and dis ances.
Be o e each su ey, se e al es images we e aken o adjus came a se ings: ISO, ape -
u e, shu e speed, and ocus. When eaching he op imal came a se ings he su ey was
ini ia ed, and se ings we e no changed h oughou i . Images we e acqui ed a 1 Hz using
he came a’s in e al ime shoo ing unc ion. The came a was held mainly downwa d-
looking while he di e swam in a lawn-mowe (bous ophodonic) pa e n, pe o ming
close ecip ocal passes o e he su ey plo o ensu e o e lap be ween pa allel legs.

Remo e Sens. 2021,13, 659 6 o 19
2.3. S udy Si es and Da a-Se s
We used image-se s om h ee dis inc oceanic en i onmen s (Figu es 2and 4).
This comes o show he implemen a ion o ou wo k low in di e en ecological zones,
and demons a e he gene ali y o his me hod (Table 1).
Figu e 4.
[Top] The global dis ibu ion o he h ee s udy zones, and he main pho omosaics used in his s udy; he ben hos
da a-se [bo om], (Table 1). (
A
)Spa an Ree o sho e Hai a on he eas e n Medi e anean coas line (
MD
da a-se ). (
B
) IUI o
Eila ee , Gul o Aqaba, no he n Red-Sea (
RS
and
RS20
da a-se s). (
C
)Double-W eck ee , island o S . Eus a ius in he
eas e n Ca ibbean (CR da a-se ).
Table 1.
The di e en da a-se s used in his s udy a e om h ee oceanic egions. Some o he da a-se s a e labeled coa sely
(no all pixels ha e a label) and some a e manually segmen ed ( ull manual labeling; e e y pixel has a label). Classi ica ion
is di ided be ween a genus-speci ic scheme and a lowe le el habi a -mapping scheme (Te ain) wi h eigh classess ha
ep esen he e ain ype.
Region Name Dep h (m) Size in m2Labeling Classi ica ion Map Replica es
Red Sea RS20 20 10 ×1 Coa se Genus 3
Red Sea RS 24–28 5 ×5 Full Te ain 2
Medi e anean MD 20 5 ×4 Full Te ain 2
Ca ibbean CR 20 12 ×2 Full Te ain 1
Remo e Sens. 2021,13, 659 7 o 19
Labeling and Classi ica ion
We used wo manual labeling schemes: coa se labeling (a polygon inside he objec
co e ing i s cen e bu no all o i s pixels) and ull segmen a ion, and wo classi ica ion
schemes: genus-speci ic (57 classes), and habi a mapping (eigh classes) (Table 1). We
used labelbox, a dedica ed ool o compu e - ision applica ions, because o i s lexibili y,
academic p icing bene i s, and simple in e ace. Images we e uploaded and labeled wi h
a polygon p ojec se up. In da a-se Red Sea 20 (
RS20
) we used genus-speci ic classes o
scle ac inian co als, and o he sessile g oups a lowe axonomic esolu ions. In da a-se s
Red-Sea (
RS
), Ca ibbean (
CR
), and Medi e anean (
MD
) we used eigh classes in ull
manual labeling, by he e ain ype. This equi es less expe ise and can be dis ibu ed
among non-expe labele s such as unde -g adua e o high-school s uden s, and e en
ex e nal wo k o ces.
2.4. Label-Augmen a ion
This expe imen e lec s he amoun o labeling e o equi ed in o de o ob ain he
highes quali y o label-augmen a ion.
Augmen a ion om Spa se Anno a ions
Label-augmen a ion consis s o expanding spa se labels o ull segmen a ion by aug-
men ing he numbe o labeled samples. We use he me hod p e iously de eloped by
us [
53
] ha was since alida ed ex ensi ely on di e en ypes o da a including ci y-scape
images o au onomous d i ing, e es ial o hopho os, and luo escen and RGB co al
images [
25
,
54
] (code a ailable online h ps://gi hub.com/Sha he/ML-Supe pixels (accessed
on 19 Janua y 2021)). He e, we examine his me hod wi h espec o meaning ul ecological
measu es. We apply label-augmen a ion on pho omosaics (Figu e 5), whe e he inpu is
spa se anno a ions, and he ou pu is a ully segmen ed map. A supe pixel is a low-le el
g ouping o neighbo ing pixels. The MLS app oach uses supe pixels o p opaga e he
spa se labels. I compu es se e al supe pixel le els o di e en sizes and uses he spa se
anno a ions as o es. I consis s o applying he supe pixel image segmen a ion i e a i ely,
p og essi ely dec easing he numbe o supe pixels gene a ed in each i e a ion. In he i s
i e a ion, he numbe o supe pixels is e y high, leading o e y small-sized supe pixels
o cap u ing small de ails o he images. The ollowing i e a ions dec ease he numbe o
supe pixels, leading o la ge supe pixels co e ing unlabeled pixels. Successi e i e a ions
do no o e w i e in o ma ion; hey only add new labeling in o ma ion un il all pixels
a e co e ed.
Figu e 5.
Augmen a ion alida ion wo k low. (
a
) A 5
×
5 m
2
pho omosaic om he
RS
da a-se .
(
b
) The mosaic was ully labeled manually acco ding o he eigh classes in he colou code ( igh ).
(
c
) The ull labels we e spa si ied ( he example depic s he emaining 0.1% pixels and a magni ica ion
o he op le co ne ). (
d
) The spa se labels we e augmen ed using ou me hod, and (
e
) e alua ed
agains he ull manual labels.
To e alua e he me hod, we conduc ed an expe imen o es ima e how many ini ial
seeds a e equi ed o achie e an accu a e ull segmen a ion and how di e en spa si ies
a ec he augmen a ion pe o mance. As ou pho omosaics we e manually labeled densely,
i.e., all he pixels we e labeled, we simula e he spa se labeling by andomly sampling
Remo e Sens. 2021,13, 659 8 o 19
ini ial seeds in se e al spa si y le els (10%, 1%, 0.1%, 0.01%, 0.001%) o he o iginal dense
labels, and augmen ing i using he same me hod. These spa se labels simula e he way
ben hic da a-se s a e usually labeled o educing he labeling cos .
2.5. O ho ec i ica ion
The pu pose o his expe imen was o simula e epea ed su eys wi hou pe manen
ma ke s o na iga ion senso s. In his manne , epea ed su eys can ake place wi h he
aid o na u al and a i icial e e ences such as dis inc i e ee ea u es o moo ing sinke s.
These objec s se e as a s a ing poin o he su ey, and o hopho os can be egis e ed in
pos -p ocessing as long as hey a e consis en ly o ho ec i ied.
2.5.1. 3D G id De ini ion and O ho ec i ica ion
We used Agiso Me ashape 1.5 o cons uc ing he 3D map models and o hopho-
os (Agiso Me ashape P o essional Ve sion 1.5, Agiso LLC, S . Pe e sbu g, Russia,
2016). In da a-se s
RS20, RS
, and
MD
, a 0.5
×
0.05 m
2
spi i le ele was used o de-
ine he 3D g id. The models we e scaled using he known size o he scale ba s. We
ma ked he poin s o known dis ance on 5–10 images un il he scale e o was lowe
han 0.0005 m. Expo ing he o hopho o has se e al deg ees-o - eedom ha ha e o
be se o epea abili y. The loca ions o h ee co ne s o he spi i le ele we e ma ked
as
(X,Y,Z) = (0, 0, Z),(0, 0.05 m, Z),(0.5 m, 0, Z)
whe e Z is he known dep h measu ed
in si u. This was done wi hin he e e ence pane o Agiso , ma king 15–20 images om
di e en angles and dis ances. The model was hen o a ed and ansla ed acco dingly,
and a pho omosaic was expo ed. O ho ec i ied pho omosaics we e expo ed as .png
image iles a a esolu ion o 0.5 mm pe pixel. These we e hen c opped o he a ea o
in e es and adjus ed o con as in Ma lab using he imadjus unc ion [MATLAB R2019].
2.5.2. Repea ed-Su ey Simula ion
To es ima e he abili y o ou pipeline o o ho ec i ying using a le ele and i s change-
de ec ion sensi i i y we epea ed image acquisi ion wo o h ee imes du ing he same
di e, esul ing in image se s ha cons i u e echnical eplica es. Be ween epea s, he spi i
le ele was mo ed a ound he scene. The g ound- u h pho omosaic ep esen s a i s
empo al epea o baseline su ey, and i was ully manually labeled. The labels we e
hen spa si ied (subsampled), and augmen ed on all eplica es esul ing in ully segmen ed
pho omosaic eplica es. These we e compa ed o he g ound- u h mosaic o e alua ion
(Figu e 6e).
In his expe imen ou eplica es a e expec ed o be iden ical and he nega i e-con ol
(naï e) is expec ed o show he highes a iance om he g ound- u h. To gene a e he
naï e (
no le ele
) pho omosaic, one o he image-se s was expo ed wice, be o e and a e
ans o ma ion. We egis e ed he eplica e o hopho os using Ma lab’s manual image
egis a ion ool cpselec and 15–20 egis a ion poin s (Figu e 6c).
2.6. E alua ion Me ics
In all augmen a ion expe imen s, he augmen ed labels we e e alua ed agains he
o iginal manual dense anno a ions. Se e al me ics we e used o assess he pe o mance o
he augmen a ion including ecall, accu acy (pe pixel) and he In e sec ion o e Union
(IoU, pe class):
Recall =T ue Posi i es
T ue Posi i es +False Nega i es (1)
Accu acy =T ue Posi i es +T ue Nega i es
T ue Posi i es +T ue Nega i es +False Posi i es +False Nega i es (2)
IoU =T ue Posi i es
T ue Posi i es +False Posi i es +False Nega i es (3)
Remo e Sens. 2021,13, 659 9 o 19
Figu e 6.
Tes ing o ho ec i ica ion h ough label-augmen a ion. Two pho omosaics a e gene a ed
subsequen ly, he g ound- u h ( op) ep esen s he baseline su ey, and he eplica e (bo om)
ep esen s he empo al epea . (
a
) The g ound- u h (baseline) pho omosaic is labeled. (
b
) Labels
a e spa si ied (
c
) Manual egis a ion o pho omosaics is done using dis inc i e ea u es in he
scene. (
d
) Spa se labels om (
b
) a e augmen ed on he baseline and ans o med mosaic eplica e.
(e) augmen ed maps a e used o e alua ion.
These me ics a e no mally used o assess he pe o mance o CNNs in segmen a-
ion asks.
2.7. Communi y-Me ics Compa isons
We de eloped a Ma lab code o communi y da a ex ac ion. All he objec s below size
0.0002 m
2
we e excluded om analysis because hey come om noise in he segmen a ions.
•Class-speci ic size- equency dis ibu ions.
We di ided he classes in nine bins,
s a ing om 0.0002 m
2
o 0.045 m
2
wi h a s ep size o 0.005 m
2
. We used
χ2
dis ance
o assess he simila i y o class size dis ibu ion be ween maps. Low alues indica e
high simila i y be ween se s o da a whe e ze o is he maximal simila i y.
χ2Dis ance =
n
∑
i=1
(Obse edi−Expec edi)2
Obse edi+Expec edi
(4)
•Rela i e amoun o indi iduals pe class.
The numbe o objec s om each class
di ided by he o al numbe o objec s in he map.
•Rela i e a ea by class.
The size in m
2
pe class di ided by he o al size o he map.
The pho omosaics a e expo ed a 0.5 mm pe pixel, and o ans e o m
2
we use he
ollowing equa ion:
Size in m2=∑Pixels
4×10−6(5)
3. Resul s
3.1. Label-Augmen a ion
In his expe imen we used ully manually labeled images o es he MLS augmen-
a ion app oach om spa se seeds on wide scale da a; pho omosaics om di e en ee
en i onmen s. We used da a-se s
RS, MD
, and
CR
. The label-augmen a ion expe imen
shows ha augmen ed labeling and dense manual anno a ions p o ide e y simila ecolog-
ical ou pu s. Figu e 7depic s bo h he pe -pixel accu acy and he IoU o all spa si y le els.
The pe -pixel me ic (accu acy) is highe because o he backg ound classes (Sand, Rock)
Remo e Sens. 2021,13, 659 16 o 19
Howe e , compa ing he simila i y o o hopho os is no s aigh o wa d, due o di e ences
such as colo and a i ac s (blu /holes) caused o example by sligh di e ences in he
dis ance and angle o image. We compa ed he maps indi ec ly h ough label-augmen a ion.
Compa ing he segmen ed maps gene a ed om augmen ing a single se o spa se labels
( om he o iginal image) es s all s eps o he wo k low in ac , and includes noise om
he pho og amme ic (di e en inpu images) and o ho ec i ica ion p ocesses. Thus, i
simula es an obse e e ec and a noisy eal-wo ld si ua ion.
When applying his wo k low in any se ing, he mos impo an ac o s o conside
a e he classi ica ion le el ( axonomic/ unc ional speci ici y), as i implies on he le el o
expe knowledge equi ed as well as he accu acy in au oma ic iden i ica ion, and he
expec ed change-de ec ion abili y which is go e ned by he e ec i e esolu ion and signal-
o-noise a io. Fu he mo e, i is impo an o conside he e ec o he slope on he ee ,
in he sense ha he op down iew is no always pe pendicula o he ee - able.
A he momen , he e a e se e al ools o image segmen a ion wi h weak human-
in e e ence [
67
]. Mo eo e , new ools will soon be eleased wi h p omising ou look on he
ben hic pho omosaic segmen a ion asks [68].
Au ho Con ibu ions:
Concep ualiza ion, All Au ho s; me hodology, M.Y., T.T., I.A., and A.C.M.
alida ion, M.Y. and I.A.; o mal analysis, M.Y., I.A., T.T., and A.C.M.; in es iga ion, M.Y., I.A.;
esou ces, M.Y., T.T., I.A., and A.C.M.; da a cu a ion, M.Y., T.T.; w i ing—o iginal d a p epa a ion,
M.Y., I.A., T.T.; w i ing— e iew and edi ing, All au ho s; isualiza ion, M.Y., I.A.; supe ision, T.T.,
D.T., A.C.M.; p ojec adminis a ion; M.Y.; unding acquisi ion, M.Y., T.T. All au ho s ha e ead and
ag eed o he published e sion o he manusc ip .
Funding:
T.T. was suppo ed by he The Leona M. and Ha y B. Helmsley Cha i able T us , The
Mau ice Ha e Founda ion, he Is ael Minis y o Na ional In as uc u es, Ene gy and Wa e
Resou ces G an 218-17-008, he Is ael Minis y o Science, Technology and Space g an 3-12487,
and he Technion Ollendo Mine a Cen e o Vision and Image Sciences. I.A. and A.C.M. we e
suppo ed by p ojec PGC2018-098817-A-I00 MCIU/AEI/FEDER, UE. M.Y. was suppo ed by he
PADI Founda ion (applica ion #32618), he Mu ay Founda ion o s uden esea ch, ASSEMBLE+
Eu opean Ho izon 2020 ( ansna ional access #216), and Mic oso AI o Ea h; AI o Co al Ree
Mapping. Y.L. was unded by he Is ael Science Founda ion (ISF) g an No. 1191/16. GE was
suppo ed by he Eu opean Union’s Ho izon 2020 esea ch and inno a ion p og am unde he Ma ie
Skłodowska-Cu ie g an ag eemen #796025.
Da a A ailabili y S a emen :
The Ben hos da a-se and Ma lab code is a ailable in he D yad Digi al
Reposi o y: h ps://doi.o g/10.5061/d yad.8cz8w9gm3 The code o ML Supe pixels can be ound
a : h ps://gi hub.com/Sha he/ML-Supe pixels.
Acknowledgmen s:
We hank he Mo is-Kahn Ma ine Resea ch S a ion, he In e uni e si y Ins i u e
o Ma ine Sciences o Eila , and he Ca ibbean Ne he land Science Ins i u e o making hei acili ies
a ailable o us, and mos impo an ly he s uden s, s a , and di ing eams o hese acili ies o
ieldwo k and echnical assis ance; A iad A ni, Debo ah Le i, Assa Le i, Ophe Ba -Na han, Leonid
Deh e , Yu al Gold ach , Sha on Fa be , Ilan Ma dix, Inbal Ayalon, Li az Le y, Lindsay Boni o, Pim
Bongae s, Basha Elnasha , De ya Akkaynak, and A i Ba -Massada o aluable in ellec ual and
echnical con ibu ions; The image labeling eam: Shai Zilbe man, Gal E ia a , Adi Zwei le , Osh a
Yosse , Ma Dohe y; Guilhem Banc-P andi o he pic u e in Figu e 3b; Ne a Kashe o d awing
he g aphical abs ac ; NVIDIA Co po a ion o he dona ion o he Ti an Xp GPU used in his wo k.
Fieldwo k in Eila was ca ied ou unde pe mi #42-128 om he Is aeli Na u e and Pa ks Au ho i y.
Con lic s o In e es : The au ho s decla e no con lic o in e es .
Re e ences
1.
Ku zweil, R. The law o accele a ing e u ns. In Alan Tu ing: Li e and Legacy o a G ea Thinke ; Sp inge : Be lin/Heidelbe g,
Ge many, 2004; pp. 381–416.
2.
Da ies, N.; Field, D.; Ga aghan, D.; Holb ook, S.J.; Planes, S.; T oye , M.; Bonsall, M.; Claude , J.; Rode ick, G.; Schmi , R.J.; e al.
Simula ing social-ecological sys ems: The Island Digi al Ecosys em A a a s (IDEA) conso ium. GigaScience
2016
,5, s13742-016.
[C ossRe ] [PubMed]

Remo e Sens. 2021,13, 659 17 o 19
3.
Laplace, P.S. A Philosophical Essay on P obabili ies, Six h F ench ed.; T usco , F.W., Emo y, F.W., T ansla o s; Chapman & Hall,
Limi ed: London, UK, 1902.
4.
B od ick, P.G.; Da ies, A.B.; Asne , G.P. Unco e ing ecological pa e ns wi h con olu ional neu al ne wo ks. T ends Ecol. E ol.
2019,34, 734–745. [C ossRe ]
5.
De Kock, M.; Gallache , D. F om d one da a o decision: Tu ning images in o ecological answe s. In P oceedings o he Con e ence
Pape : Inno a ion A abia, Dubai, Uni ed A ab Emi a es, 7–9 Ma ch 2016; Volume 9.
6.
Ka enbo n, T.; Eichel, J.; Fassnach , F.E. Con olu ional Neu al Ne wo ks enable e icien , accu a e and ine-g ained segmen a ion
o plan species and communi ies om high- esolu ion UAV image y. Sci. Rep. 2019,9, 1–9. [C ossRe ]
7.
Sil e , M.; Tiwa i, A.; Ka nieli, A. Iden i ying ege a ion in a id egions using objec -based image analysis wi h RGB-only ae ial
image y. Remo e Sens. 2019,11, 2308. [C ossRe ]
8.
Zimudzi, E.; Sande s, I.; Rollings, N.; Omlin, C. Segmen ing mang o e ecosys ems d one images using SLIC supe pixels. Geoca o
In . 2019,34, 1648–1662. [C ossRe ]
9.
Maggio i, E.; Ta abalka, Y.; Cha pia , G.; Alliez, P. High- esolu ion ae ial image labeling wi h con olu ional neu al ne wo ks.
IEEE T ans. Geosci. Remo e Sens. 2017,55, 7092–7103. [C ossRe ]
10.
Tsuichiha a, S.; Aki a, S.; Ike, R.; Shige a, M.; Takemu a, H.; Na o i, T.; Aikawa, N.; Shindo, K.; Ide, Y.; Tejima, S. D one and
GPS senso s-based g assland managemen using deep-lea ning image segmen a ion. In P oceedings o he 2019 Thi d IEEE
In e na ional Con e ence on Robo ic Compu ing (IRC), Naples, I aly, 25–27 Feb ua y 2019; IEEE: Pisca away, NJ, USA, 2019;
pp. 608–611.
11.
Johnson-Robe son, M.; Piza o, O.; Williams, S.B.; Mahon, I. Gene a ion and isualiza ion o la ge-scale h ee-dimensional
econs uc ions om unde wa e obo ic su eys. J. Field Robo . 2010,27, 21–51. [C ossRe ]
12.
B yson, M.; Fe a i, R.; Figuei a, W.; Piza o, O.; Madin, J.; Williams, S.; By ne, M. Cha ac e iza ion o measu emen e o s using
s uc u e- om-mo ion and pho og amme y o measu e ma ine habi a s uc u al complexi y. Ecol. E ol.
2017
,7, 5669–5681.
[C ossRe ]
13.
Bu ns, J.; Delpa e, D.; Ga es, R.; Takabayashi, M. In eg a ing s uc u e- om-mo ion pho og amme y wi h geospa ial so wa e
as a no el echnique o quan i ying 3D ecological cha ac e is ics o co al ee s. Pee J 2015,3, e1077. [C ossRe ]
14.
Calde s, K.; Phinn, S.; Fe a i, R.; Leon, J.; A ms on, J.; Asne , G.P.; Disney, M. 3D Imaging Insigh s in o Fo es s and Co al Ree s.
T ends Ecol. E ol. 2020,35, 6–9. [C ossRe ]
15.
Edwa ds, C.B.; Eynaud, Y.; Williams, G.J.; Pede sen, N.E.; Zgliczynski, B.J.; Gleason, A.C.; Smi h, J.E.; Sandin, S.A. La ge-a ea
imaging e eals biologically d i en non- andom spa ial pa e ns o co als a a emo e ee . Co al Ree s
2017
,36, 1291–1305.
[C ossRe ]
16.
Fe a i, R.; Figuei a, W.F.; P a che , M.S.; Boube, T.; Adam, A.; Kobelkowsky-Vid io, T.; Doo, S.S.; A wood, T.B.; By ne, M. 3D
pho og amme y quan i ies g ow h and ex e nal e osion o indi idual co al colonies and skele ons. Sci. Rep.
2017
,7, 16737.
[C ossRe ]
17.
González-Ri e o, M.; Beijbom, O.; Rod iguez-Rami ez, A.; Hol op, T.; González-Ma e o, Y.; Ganase, A.; Roel sema, C.; Phinn, S.;
Hoegh-Guldbe g, O. Scaling up ecological measu emen s o co al ee s using semi-au oma ed ield image collec ion and analysis.
Remo e Sens. 2016,8, 30. [C ossRe ]
18.
He nández-Landa, R.C.; Ba e a-Falcon, E.; Rioja-Nie o, R. Size- equency dis ibu ion o co al assemblages in insula shallow
ee s o he Mexican Ca ibbean using unde wa e pho og amme y. Pee J 2020,8, e8957. [C ossRe ]
19.
Lange, I.; Pe y, C. A quick, easy and non-in asi e me hod o quan i y co al g ow h a es using pho og amme y and 3D model
compa isons. Me hods Ecol. E ol. 2020,11,714-726. [C ossRe ]
20.
Mohamed, H.; Nadaoka, K.; Nakamu a, T. Towa ds Ben hic Habi a 3D Mapping Using Machine Lea ning Algo i hms and
S uc u es om Mo ion Pho og amme y. Remo e Sens. 2020,12, 127. [C ossRe ]
21.
Naugh on, P.; Edwa ds, C.; Pe o ic, V.; Kas ne , R.; Kues e , F.; Sandin, S. Scaling he anno a ion o sub idal ma ine habi a s.
In P oceedings o he 10 h In e na ional Con e ence on Unde wa e Ne wo ks & Sys ems; ACM: New Yo k, NY, USA, 2015; pp. 1–5.
22.
Williams, I.D.; Couch, C.; Beijbom, O.; Oli e , T.; Va gas-Angel, B.; Schumache , B.; B aina d, R. Le e aging au oma ed image
analysis ools o ans o m ou capaci y o assess s a us and ends on co al ee s. F on . Ma . Sci. 2019,6, 222. [C ossRe ]
23.
Beijbom, O.; Edmunds, P.J.; Kline, D.I.; Mi chell, B.G.; K iegman, D. Au oma ed anno a ion o co al ee su ey images.
In P oceedings o he 2012 IEEE Con e ence on Compu e Vision and Pa e n Recogni ion (CVPR), P o idence, RI, USA, 16–21
June 2012; IEEE: Pisca away, NJ, USA, 2012; pp. 1170–1177.
24.
Beijbom, O.; Edmunds, P.J.; Roel sema, C.; Smi h, J.; Kline, D.I.; Neal, B.P.; Dunlap, M.J.; Mo ia y, V.; Fan, T.Y.; Tan, C.J.; e al.
Towa ds au oma ed anno a ion o ben hic su ey images: Va iabili y o human expe s and ope a ional modes o au oma ion.
PLoS ONE 2015,10, e0130312. [C ossRe ]
25.
Alonso, I.; Yu al, M.; Eyal, G.; T eibi z, T.; Mu illo, A.C. Co alSeg: Lea ning co al segmen a ion om spa se anno a ions. J. Field
Robo . 2019,36, 1456–1477. [C ossRe ]
26.
F iedman, A.L. Au oma ed In e p e a ion o Ben hic S e eo Image y. Ph.D. Thesis, Uni e si y o Sydney, Sydney, Aus alia, 2013.
27.
Pa oni, G.; Co sini, M.; Callie i, M.; Palma, M.; Scopigno, R. Seman ic segmen a ion o ben hic communi ies om o ho-mosaic
maps. In P oceedings o he In e na ional A chi es o he Pho og amme y, Remo e Sensing & Spa ial In o ma ion Sciences,
Limassol, Cyp us, 2–3 May 2019.
Remo e Sens. 2021,13, 659 18 o 19
28.
Rashid, A.R.; Chennu, A. A T illion Co al Ree Colo s: Deeply Anno a ed Unde wa e Hype spec al Images o Au oma ed
Classi ica ion and Habi a Mapping. Da a 2020,5, 19. [C ossRe ]
29.
Teixidó, N.; Albajes-Eizagi e, A.; Bolbo, D.; Le Hi , E.; Demes e, M.; Ga abou, J.; Guigues, L.; Gili, J.M.; Pie a, J.; P elo , T.; e al.
Hie a chical segmen a ion-based so wa e o co e classi ica ion analyses o seabed images (Seascape). Ma . Ecol. P og. Se .
2011
,
431, 45–53. [C ossRe ]
30.
King, A.; M Bhanda ka , S.; Hopkinson, B.M. Deep Lea ning o Seman ic Segmen a ion o Co al Ree Images Using Mul i-View
In o ma ion. In P oceedings o he IEEE Con e ence on Compu e Vision and Pa e n Recogni ion Wo kshops, Long Beach, CA,
USA, 16–20 June 2019; pp. 1–10.
31.
Hopkinson, B.M.; King, A.C.; Owen, D.P.; Johnson-Robe son, M.; Long, M.H.; Bhanda ka , S.M. Au oma ed classi ica ion o
h ee-dimensional econs uc ions o co al ee s using con olu ional neu al ne wo ks. PLoS ONE
2020
,15, e0230671. [C ossRe ]
32. Todd, P.A. Mo phological plas ici y in scle ac inian co als. Biol. Re . 2008,83, 315–337. [C ossRe ] [PubMed]
33.
Schlich ing, C.D.; Pigliucci, M. Pheno ypic E olu ion: A Reac ion No m Pe spec i e; Sinaue Associa es Inco po a ed: Sunde land,
MA, USA, 1998.
34.
Be man, D.; T eibi z, T.; A idan, S. Di ing in o hazelines: Colo es o a ion o unde wa e images. In P oceedings o he B i ish
Machine Vision Con e ence, London, UK, 4–7 Sep embe 2017; Volume 1.
35.
Akkaynak, D.; T eibi z, T. A e ised unde wa e image o ma ion model. In P oceedings o he IEEE Con e ence on Compu e
Vision and Pa e n Recogni ion, Sal Lake Ci y, UT, USA, 18–22 June 2018; pp. 6723–6732.
36.
Deng, F.; Kang, J.; Li, P.; Wan, F. Au oma ic ue o hopho o gene a ion based on h ee-dimensional building model using
mul i iew u ban ae ial images. J. Appl. Remo e Sens. 2015,9, 095087. [C ossRe ]
37.
Rossi, P.; Cas agne i, C.; Cap a, A.; B ooks, A.; Mancini, F. De ec ing change in co al ee 3D s uc u e using unde wa e
pho og amme y: C i ical issues and pe o mance me ics. Appl. Geoma . 2020,12, 3–17. [C ossRe ]
38.
Piza o, O.; F iedman, A.; B yson, M.; Williams, S.B.; Madin, J. A simple, as , and epea able su ey me hod o unde wa e
isual 3D ben hic mapping and moni o ing. Ecol. E ol. 2017,7, 1770–1782. [C ossRe ]
39.
Abadie, A.; Boisse y, P.; Viala, C. Geo e e enced unde wa e pho og amme y o map ma ine habi a s and subme ged a i icial
s uc u es. Pho og amm. Rec. 2018,33, 448–469. [C ossRe ]
40.
Pyle, R.L.; Copus, J.M. Mesopho ic co al ecosys ems: In oduc ion and o e iew. In Mesopho ic Co al Ecosys ems; Sp inge :
Be lin/Heidelbe g, Ge many, 2019; pp. 3–27.
41.
B own, C.J.; Smi h, S.J.; Law on, P.; Ande son, J.T. Ben hic habi a mapping: A e iew o p og ess owa ds imp o ed unde s and-
ing o he spa ial ecology o he sea loo using acous ic echniques. Es ua . Coas . Shel Sci. 2011,92, 502–520.
42.
Lecou s, V.; De ille s, R.; Schneide , D.C.; Luciee , V.L.; B own, C.J.; Edinge , E.N. Spa ial scale and geog aphic con ex in ben hic
habi a mapping: Re iew and u u e di ec ions. Ma . Ecol. P og. Se . 2015,535, 259–284. [C ossRe ]
43. McKinney, F.K.; Jackson, J.B. B yozoan E olu ion; Uni e si y o Chicago P ess: Chicago, IL, USA, 1991.
44.
Ve on, J.E.N. Co als in Space and Time: The Biogeog aphy and E olu ion o he Scle ac inia; Co nell Uni e si y P ess: I haca, NY,
USA, 1995.
45. Hughes, T.P. Communi y s uc u e and di e si y o co al ee s: The ole o his o y. Ecology 1989,70, 275–279. [C ossRe ]
46. Hus on, M. Pa e ns o species di e si y on co al ee s. Annu. Re . Ecol. Sys . 1985,16, 149–177. [C ossRe ]
47.
Loya, Y. Communi y s uc u e and species di e si y o he ma ypic co als a Eila , Red Sea. Ma . Biol.
1972
,13, 100–123. [C ossRe ]
48.
Plaisance, L.; Caley, M.J.; B aina d, R.E.; Knowl on, N. The di e si y o co al ee s: Wha a e we missing? PLoS ONE
2011
,
6, e25026. [C ossRe ] [PubMed]
49.
Shlesinge , T.; Loya, Y. Sexual ep oduc ion o scle ac inian co als in mesopho ic co al ecosys ems s. shallow ee s. In Mesopho ic
Co al Ecosys ems; Sp inge : Be lin/Heidelbe g, Ge many, 2019; pp. 653–666.
50.
O’Neill, R.V.; Deangelis, D.L.; Waide, J.B.; Allen, T.F.; Allen, G.E. A Hie a chical Concep o Ecosys ems; Numbe 23; P ince on
Uni e si y P ess: P ince on, NJ, USA, 1986.
51. Mo in, P.J. Communi y Ecology; John Wiley & Sons: Hoboken, NJ, USA, 2009.
52. Ruppe , E.E.; Ba nes, R.D. In e eb a e Zoology, 5 h ed.; WB Saunde s Company: Philadelphia, PA, USA, 1987.
53.
Alonso, I.; Camb a, A.; Munoz, A.; T eibi z, T.; Mu illo, A.C. Co al-segmen a ion: T aining dense labeling models wi h spa se
g ound u h. In P oceedings o he IEEE In e na ional Con e ence on Compu e Vision Wo kshops, Venice, I aly, 22–29 Oc obe
2017; pp. 2874–2882.
54.
Alonso, I.; Mu illo, A.C. Seman ic segmen a ion om spa se labeling using mul i-le el supe pixels. In P oceedings o he 2018
IEEE/RSJ In e na ional Con e ence on In elligen Robo s and Sys ems (IROS), Mad id, Spain, 1–5 Oc obe 2018; IEEE: Pisca away,
NJ, USA, 2018; pp. 5785–5792.
55. Blaschke, T. Objec based image analysis o emo e sensing. ISPRS J. Pho og amm. Remo e Sens. 2010,65, 2–16. [C ossRe ]
56.
Ben-Romdhane, H.; Ma pu, P.R.; Oua da, T.B.; Ghedi a, H. Co als & ben hic habi a mapping using DubaiSa -2: A spec al-spa ial
app oach applied o Dalma Island, UAE (A abian Gul ). Remo e Sens. Le . 2016,7, 781–789.
57.
Luciee , V. Objec -o ien ed classi ica ion o sidescan sona da a o mapping ben hic ma ine habi a s. In . J. Remo e Sens.
2008
,
29, 905–921. [C ossRe ]
58.
Micalle , A.; Le Bas, T.P.; Hu enne, V.A.; Blondel, P.; Hühne bach, V.; Deidun, A. A mul i-me hod app oach o ben hic habi a
mapping o shallow coas al a eas wi h high- esolu ion mul ibeam da a. Con . Shel Res. 2012,39, 14–26. [C ossRe ]
Remo e Sens. 2021,13, 659 19 o 19
59.
Wahidin, N.; Si ega , V.P.; Nababan, B.; Jaya, I.; Wou huyzen, S. Objec -based image analysis o co al ee ben hic habi a
mapping wi h se e al classi ica ion algo i hms. P ocedia En i on. Sci. 2015,24, 222–227. [C ossRe ]
60.
Hess, M.; Pe o ic, V.; Kues e , F. In e ac i e classi ica ion o cons uc ion ma e ials: Feedback d i en amewo k o anno a ion
and analysis o 3D poin clouds. In . A ch. Pho og amm. Remo e Sens. Spa . In . Sci. 2017,42, 343–347. [C ossRe ]
61.
Rossi, P.; Pon i, M.; Righi, S.; Cas agne i, C.; Simonini, R.; Mancini, F.; Ag a io is, P.; Bassani, L.; B uno, F.; Ce ano, C.; e al.
Needs and gaps in op ical unde wa e echnologies and me hods o he in es iga ion o ma ine animal o es 3D-s uc u al
complexi y F on . Ma . Sci. 2021, in p ess.
62.
Akkaynak, D.; T eibi z, T. Sea-Th u: A Me hod o Remo ing Wa e F om Unde wa e Images. In P oceedings o he IEEE
Con e ence on Compu e Vision and Pa e n Recogni ion, Long Beach, CA, USA, 15–20 June 2019; pp. 1682–1691.
63.
Neye , F.; Noce ino, E.; G uen, A. Moni o ing co al g ow h– he dicho omy be ween unde wa e pho og amme y and geode ic
con ol ne wo k. ISPRS-In . A ch. Pho og amm. Remo e Sens. Spa . In . Sci. 2018,42, 2. . [C ossRe ]
64. Con i, L.A.; Lim, A.; Wheele , A.J. High esolu ion mapping o a cold wa e co al mound. Sci. Rep. 2019,9, 1–15. [C ossRe ]
65.
Misiuk, B.; B own, C.J.; Robe , K.; Lacha i é, M. Ha monizing mul i-sou ce sona backsca e da ase s o seabed mapping using
bulk shi app oaches. Remo e Sens. 2020,12, 601. [C ossRe ]
66.
T zcinska, K.; Janowski, L.; Nowak, J.; Rucinska-Zjadacz, M.; K uss, A.; on Deimling, J.S.; Pocwia dowski, P.; Tegowski, J.
Spec al ea u es o dual- equency mul ibeam echosounde da a o ben hic habi a mapping. Ma . Geol.
2020
,427, 106239.
[C ossRe ]
67.
Acuna, D.; Ling, H.; Ka , A.; Fidle , S. E icien in e ac i e anno a ion o segmen a ion da ase s wi h polygon- nn++. In P o-
ceedings o he IEEE con e ence on Compu e Vision and Pa e n Recogni ion, Sal Lake Ci y, UT, USA, 18–22 June 2018;
pp. 859–868.
68.
Maninis, K.K.; Caelles, S.; Pon -Tuse , J.; Van Gool, L. Deep ex eme cu : F om ex eme poin s o objec segmen a ion. In P o-
ceedings o he IEEE Con e ence on Compu e Vision and Pa e n Recogni ion, Sal Lake Ci y, UT, USA, 18–22 June 2018;
pp. 616–625.