Indonesi
an
Journa
l
of El
ec
t
ri
cal Engineer
ing
an
d
Comp
ut
er
Scie
nce
Vo
l.
9,
No.
3,
Ma
rch
2018,
pp.
72
2
~
7
30
IS
S
N: 25
02
-
4752, DO
I: 10
.11
591/ijeecs
.v9.i
3.pp
72
2
-
7
30
722
Journ
al h
om
e
page
:
http:
//
ia
es
core.c
om/j
ourn
als/i
ndex.
ph
p/ij
eecs
Assessin
g t
h
e Crown
Closu
re of
Nypa on
UAV
Im
ages usin
g
Mean
-
Sh
ift Segmentati
on
Algorith
m
Robert
Pa
r
uli
an
Sil
ala
hi
1
,
I Nen
gah Sura
t
i Jay
a
2
, T
atang Ti
ry
ana
3
, F
ai
rus Mul
ia
4
1,
Bogor
Agric
ul
t
ura
l
Univ
ersity
,
Campus
IPB Dramaga, Bogor, I
ndonesia
2
,3
Depa
rtment
of
Forest
Man
agem
ent
,
Fa
cul
t
y
of
Forestr
y
,
Bogor
Agric
ul
tura
l
Un
ive
rsit
y
,
Campus
IPB Dramaga,
Bogor,
Indon
esia
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Sep
19, 201
7
Re
vised Dec
30, 2
017
Accepte
d
Ja
n 1
7,
2018
Util
izati
on
of
v
er
y
h
igh
-
re
solu
t
ion
images
becom
es
a
new
tre
nd
in
fore
st
m
ana
gement,
pa
rti
cularl
y
in
the
det
e
ct
ion
and
id
ent
ifica
ti
on
of
f
ore
st
stand
var
ia
b
le
s.
Thi
s
p
ape
r
d
esc
ribe
s
th
e
use
of
m
ea
n
-
s
hift
segm
ent
at
io
n
al
gori
thm
on
unm
anne
d
ae
rial
vehi
c
le
s
(UA
V)
i
m
age
s
to
m
ea
sure
cro
w
n
cl
osure
of
n
y
pa
(N
y
pa
fru
ticans)
and
gap
.
T
he
27
combinat
i
ons
of
the
par
am
et
er
val
u
es
such
as
spatial
r
adi
us
(hs),
r
ange
ra
dius
(hr)
,
and
m
ini
m
um
re
gion
size
(M)
.
Gap
det
e
ct
ion
an
d
n
y
pa crown
clos
ure
m
ea
surem
ent
s were
per
for
m
ed
using a
h
y
brid
b
et
wee
n
pixe
l
-
base
d
(
m
axi
m
um
li
kel
ihood
cl
assifie
r)
and
obje
ct
-
base
d
appr
oac
h
es
(segm
ent
at
i
on).
For
eva
luation
of
the
appr
oach
per
form
anc
e
,
th
e
ac
cur
acy
asses
sm
ent
was
done
b
y
compari
ng
obje
c
t
-
b
ase
d
cl
assifi
ca
t
ion
re
s
ult
s
(segm
ent
ati
on)
and
visua
l
in
te
rpre
ta
t
ion
(gro
und
che
ck
).
The
stud
y
found
tha
t the
b
est com
bina
ti
on
of
segm
ent
a
ti
on
p
ara
m
eter
was t
h
e
combinat
ion
of
hs
10,
hr
10
and
M
50,
with
the
over
a
ll
a
cc
ur
a
c
y
of
76
,
6%
and
kapp
a accur
acy
of
55.
7%
.
Ke
yw
or
d
s
:
Nypa
U
nm
ann
e
d
ae
rial
v
ehicl
es
(UAV)
M
ean
-
s
hift
S
egm
entat
ion
Copyright
©
201
8
Instit
ut
e
o
f Ad
vanc
ed
Engi
n
ee
r
ing
and
S
cienc
e
.
Al
l
rights re
serv
ed
.
Corres
pond
in
g
Aut
h
or
:
Rob
e
rt Pa
ru
li
a
n
Sil
al
ahi
,
Bogor A
gr
ic
ultur
al
U
niv
er
sit
y, Ca
m
pu
s IPB
Dr
am
aga,
Bogor,
Ind
on
e
sia
1668
0
Em
a
il
:
ins
-
j
ay
a@ap
ps
.i
pb.ac.i
d
1.
INTROD
U
CTION
Ind
on
esi
a
is
on
e
of
th
e
tro
pica
l
countries
ha
vi
ng
a
wide
var
i
et
y
of
ecosyst
e
m
s
,
sta
rting
from
coastal
forest
/
lowla
nd
(m
ang
rove,
sw
a
m
p
forest,
pe
at
swam
p
fo
res
t
)
,
dry
-
lo
w
la
nd
f
or
e
st,
up
to
m
ou
ntain
ous
f
or
est
s
,
su
b
-
al
pi
ne
a
nd
al
pin
e
)
.
T
o
obta
in
a
s
ound
basis
f
orest
m
anag
em
ent
pla
nn
i
ng,
the
re
is
a
nee
d
to
pr
ov
i
de
accurate
a
nd
t
i
m
ely
su
pp
or
ti
ng
data
relat
ed
to
ecosyst
e
m
ty
pe,
forest
cl
asses,
forest
de
ns
it
y,
biodi
ver
sit
y,
sta
nd
i
ng
stoc
k,
et
c.
On
e
of
the
un
i
qu
e
ec
os
yst
em
ty
pes,
wh
ic
h
ec
onom
ic
al
l
y,
ecolog
ic
al
ly
,
so
ci
al
l
y
an
d
culturall
y
play
s
an
i
m
po
rtant
ro
le
is
the
m
angr
ov
e
ec
os
yst
e
m
.
In
Indon
esi
a,
this
m
ang
r
ove
ecosyst
e
m
is
a
highly
vu
l
ner
a
ble
ecosyst
em
to
co
nv
e
rsion.
On
e
of
the
veget
at
ion
that
go
es
into
a
n
ass
oc
ia
ti
on
of
m
ang
r
ove
ecosyst
em
is
nypa
(
Ny
pa
frut
ic
an
s
).
Nypa
veg
et
at
io
n
is
c
omm
on
in
est
uar
y
a
reas
a
ffec
te
d
by
ti
de
[
1].
I
n
Ind
on
esi
a
,
the
inf
or
m
at
ion
on
the
dynam
ic
g
r
ow
t
h,
sta
tus
and
pote
ntial
of
ny
pa
has
not
bee
n
m
uch
st
ud
ie
d,
since
the
nypa
ecosyst
em
is
le
ss
at
tract
ive
and
e
ve
n
f
requen
tl
y
co
ns
ide
red
as
wastel
a
nd.
S
om
e
research
e
r
s
ind
ic
at
e
that
ny
pa
veg
et
at
io
n
can
pro
vid
e
po
te
ntial
econom
ic
value,
ei
ther
at
sm
a
ll
-
sca
le
or
la
rg
e
-
s
cal
e
bu
si
ness.
Ny
pa
is
on
e
ty
pe
of
ve
getat
ion
t
hat
m
ay
pr
ov
i
de
daily
nee
d
pro
du
ct
f
or
ho
us
e
ho
l
d
co
nsu
m
pt
ion
.
Seve
ral
of
ny
pa
researc
h
fou
nd
that
ny
pa
ha
s
a
po
te
ntial
to
be
a
so
urce
of
foo
d
beca
us
e
of
it
s
have
hig
h
carbo
hydr
at
e
a
nd
protei
n
[
2]
,
e
ve
n
acco
r
din
g
t
o
the
stu
dy
[
3
]
,
nypa
c
onta
ins
a
n
et
ha
no
l
t
o
pr
oduce
fu
el
energy
.
Currentl
y,
the
us
e
of
rem
ote
sensing
te
ch
nolog
y
in
natu
ral
resou
rce
m
anag
em
ent
is
a
m
u
st,
an
d
eve
n
for
a
sm
all
sca
le
fo
rest
m
anag
em
ent
.
Now,
t
her
e
is
a
tren
d
of
usi
ng
ve
ry
h
igh
-
res
olu
ti
on
i
m
ages
fo
r
detai
le
d
forest
m
anag
em
ent
up
to
the
tree
-
le
vel
f
or
es
t
m
anag
em
ent
.
Since
the
la
unch
of
natu
ral
re
so
urces
sat
el
li
te
s
for
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Assessin
g
t
he C
ro
wn Cl
osure
o
f N
y
pa on U
AV Im
ag
es
usi
ng Me
an
-
Shif
t
…
(
Rob
ert
Paruli
an S
il
al
ahi
)
723
ci
vili
an
us
e
i
n
the
19
70
s
,
the
dev
el
op
m
ent
of
rem
ote
sensi
ng
te
c
hnol
og
y
has
grown
rap
i
dly,
sta
rtin
g
f
r
om
a
low
,
m
od
erate
to
ver
y
high
reso
l
utions.
In
li
ne
with
thei
r
de
vel
op
m
ent,
the
rem
ote
sensing
has
bee
n
al
so
widely
us
ed
f
or
sup
portin
g
ei
ther
reg
i
on
al
le
vel,
nationa
l
le
vel
or
ind
ivid
ual
tree
le
vel
fo
rest
inv
e
nt
or
y
.
To
day,
the
ad
ven
t
of
the
hi
gh
a
nd
ve
ry
hi
gh
re
s
olu
ti
on
had
pro
vid
e
d
a
go
od
oppo
rtu
nity
as
wel
l
as
a
chall
enge
al
l
at
on
ce
.
T
he
hig
he
r
the
spa
ti
al
reso
l
ution,
th
e
m
or
e
detai
le
d
in
form
at
ion
cou
l
d
be
der
iv
ed,
t
he
m
or
e com
plex
and reli
able a
ppr
oac
hes
a
re r
e
qu
i
red.
In
a
ver
y
hi
gh
-
reso
l
ution
im
a
ger
y,
the
us
e
of
ordi
nar
y
pi
xe
l
-
base
d
ap
proa
ch
su
c
h
as
s
uperv
ise
d
an
d
un
s
uper
vise
d
c
la
ssifie
r
is
of
t
en
pro
vid
e
d
le
ss
accu
rate
in
f
or
m
at
ion
,
sinc
e
the
cl
assifi
c
at
ion
is
s
olely
al
on
e
base
d
upon
th
e
bri
ghtness
va
lues
(
re
flect
ance
)
of
t
he
obj
ect
.
T
he
a
dvent
of
obj
ect
-
base
d
cl
assifi
cat
io
n
te
chn
iq
ue
of
te
n
cal
le
d
the
obj
ect
-
based
im
a
ge
analy
sis (
O
BIA), which c
on
si
der
not o
nly t
he
br
ig
htn
e
s
s v
al
ue
s
of
t
he
ob
j
ect
bu
t
al
s
o
the
s
patia
l
aspect
of
the
obj
ect
be
ing
a
naly
zed
has
offer
e
d
a
prom
i
sing
so
lu
ti
on
to
ov
e
rc
om
e
the
dr
a
w
back
of
t
he
existi
ng
pi
xel
-
based
a
nal
ysi
s.
The
o
bje
ct
-
base
d
proc
e
ssing
is
c
on
si
de
red
to
hav
e
a
bette
r
pe
rfor
m
ance
in
processi
ng
hi
gh
-
res
olu
ti
on
di
gital
i
m
age
b
ecause,
in
a
dd
it
ion
to
c
on
si
der
i
ng
the
value o
f
the
p
i
xel it
sel
f,
it
als
o
c
onsider
s the
size (a
rea) an
d sha
pe of
the
obj
ect
withi
n
th
e i
m
age.
In
t
he
la
st
fe
w
deca
des,
t
he
s
patia
l
reso
l
utio
n
of
im
ages
ha
s
bee
n
im
pr
oved
ra
pid
ly
.
In
add
it
io
n
to
sat
el
li
te
i
m
age
ry,
one
of
the
i
m
aging
platfo
r
m
s
that
pr
od
uc
e
ver
y
hi
gh
-
r
es
ol
utio
n
im
ages
is
UAV
(
Un
m
an
n
e
d
Aer
ia
l
Veh
ic
le
).
T
he
U
A
V
i
s
al
so
cal
le
d
t
o
as
dr
on
e
(
dy
nam
ic
rem
ote
ly
op
e
rated
na
vig
at
io
n
e
quip
m
ent).
Now
cam
era
te
chnolo
gy
us
e
d
in
the
U
AV
is
m
ai
nly
convent
ion
al
cam
era
that
m
ay
pr
ovid
e
sp
at
ia
l
res
olut
io
n
up
to
5
cm
.
H
ow
e
ve
r,
it
s
rec
ordin
g
platfo
r
m
is
ver
y
pe
rs
pecti
ve
becaus
e
it
can
be
fl
own
un
der
t
he
cl
oud
y
weathe
r
(
fly
be
low
the
cl
oud)
,
unm
ann
e
d,
l
ow
c
os
t,
fast
a
nd
relat
ively
low
risk.
Th
us,
t
he
U
AV
te
c
hnol
og
y
beco
m
es
a
go
od
al
te
r
native
be
cause
t
he
data
obta
ined
w
oul
d
be
ver
y
detai
l
and
real
ti
m
e,
as
well
as
c
ould
be
ob
ta
ine
d
easi
l
y
with
chea
pe
r
pri
ce
.
T
he
u
ti
li
zat
ion
of
U
AV
te
c
hnolog
y
has
bee
n
widely
us
e
d
to
i
n
m
any
aspects
su
c
h
as
m
app
in
g
act
ivit
y
[4
]
,
quantify
ing
s
patia
l
gap
patte
r
n
[5
]
,
cha
ng
e
detect
ion
[
6],
fo
rest
inv
e
nt
ory
act
ivit
ie
s
[7
]
,
ca
nopy
s
pectral
m
app
in
g
[
8],
m
easur
in
g
sta
nd
va
riables
s
uch
as
cr
own
diam
e
te
r,
canopy
pe
rce
nt
age
an
d
num
ber
of
tree
s
[
9
,
10]
as
well
as
est
i
m
ation
of
the
sta
ndin
g
stock
a
nd
sit
e
ind
e
x
qu
al
it
y o
f
te
ak
sit
es [
11]
.
The
In
c
reasin
g
res
olu
ti
o
n
of
the
di
gital
i
m
a
ges
from
low
reso
l
ution
to
ve
ry
hi
gh
res
ol
ution
ha
ve
encou
rag
e
d
a
dev
el
op
m
ent
of
i
m
age
pr
oce
s
sing.
The
cu
rrent
i
m
age
pr
oc
essing
tre
nds
are
the
us
e
of
obj
ect
-
base
d
cl
assifi
c
at
ion
m
et
ho
ds,
as a co
m
pli
m
e
ntary of th
e p
i
xe
l
-
base
d
a
ppr
oa
ch.
A
pix
el
-
ba
sed
a
ppr
oach
c
an be
us
e
d
as
lo
ng
as
a
pix
el
is
the
s
a
m
e
siz
e
as
a
par
ti
cular
ob
j
ect
[12].
A
n
ob
j
ec
t
-
base
d
ap
proa
ch
is
now
a po
pu
la
r
appr
oach in
hi
gh and
ve
ry h
i
gh
-
re
so
l
ution i
m
age pro
ces
sing.
Ba
sic
al
ly
,
t
he
o
bject
-
cl
assif
ic
at
ion
m
et
ho
d
c
onsist
s
o
f
two
par
ts
nam
el
y
seg
m
e
ntati
on
a
nd
cl
assifi
cat
ion
.
Conve
ntion
al
l
y,
the
se
gm
entat
ion
is
the
init
ia
l
process
of
a
n
obj
e
ct
-
ba
sed
cl
assifi
e
r,
the
n
fo
ll
owe
d
by
cl
assifi
cat
ion
.
I
m
age
segm
ent
at
ion
is
gen
e
ra
ll
y
def
ined
as
a
pr
oc
ess
of
div
idin
g
an
im
a
ge
int
o
gro
up
s
w
hich
are
s
patia
ll
y
or
s
pectrall
y
un
ifor
m
[13].
Object
-
based
a
ppro
ac
hes
ha
ve
been
ap
plied
in
s
om
e
works
s
uc
h
as
cl
assifi
cat
ion
of
a
gr
ic
ult
ur
al
l
and
usi
ng
S
POT
-
5
[
14]
,
e
xp
l
or
at
io
n
of
the
urba
n
in
form
ation
i
n
Ci
anjur
[
15
]
,
m
app
in
g
of
c
or
al
reef
ha
bi
ta
t
[16
]
,
m
app
in
g
la
nd
c
ov
ers
c
hanges
c
hanges
i
n
m
angr
ov
e
ecosyst
em
s
[1
7],
detect
in
g
of
f
or
e
sts
cha
ng
es
cause
d
by
hurr
ic
a
nes
[18
]
,
the
ob
j
ect
-
ba
sed
cl
assifi
ca
ti
on
of
la
nd
us
e
in O
ntario, Ca
na
da h
ad
al
s
o be e
xa
m
ined
by
[
19
]
.
In
t
his
stu
dy,
the
a
uthors
fo
c
us
e
d
on
e
xam
ining
the
ap
propriat
e
obj
ect
-
base
d
c
la
ssific
at
ion
par
am
et
ers
for
analy
zi
ng
t
he
nypa
c
row
n
c
losure
da
n
ga
p.
This
st
ud
y
in
te
gr
at
ed
t
he
pi
xel
-
based
ap
proach
with
m
axi
m
u
m
li
kelihoo
d
c
la
ssifie
r
an
d
the
m
ean
-
sh
ift
al
gorithm
in
segm
entat
ion
m
et
hod.
T
he
us
e
of
th
e
m
ean
-
sh
i
ft
seg
m
entat
ion
al
go
rithm
is
widely
us
ed
i
n
earli
er
stu
dies
f
or
be
nth
ic
m
app
in
g
[
20]
,
en
vir
on
m
ental
m
anag
em
ent
m
on
it
or
ing
[
21]
,
la
nd
c
over
c
la
ssific
at
ion
[
22]
,
cha
nge
detect
ion
in
SA
R
i
m
age
[23],
m
edical
diag
nosis
[24].
Although
the
m
ean
-
sh
ift
al
gorithm
had
be
en
su
cces
f
ully
exa
m
ined
in
cl
assify
ing
buil
ding
obj
ect
s
i
n
s
ub
urba
n
ar
ea
[
25
]
,
the
a
pp
li
cat
ion
in
cl
assify
ing
ny
pa
ve
getat
ion
is
sti
ll
chall
eng
in
g.
U
p
t
o
now
,
ver
y
high
-
resol
ution
im
age
ut
il
iz
ation
resea
rch
for
nypa
e
cosyste
m
asse
ssm
ent
is
still
ver
y
ra
re.
The
m
ai
n
obj
ect
ive
of
th
e
stud
y
is
to
dev
el
op
a
cl
assifi
cat
ion
te
ch
nique
for
assessi
ng
the
n
ypa
c
rown
cl
osu
res
an
d
their
gap
by
com
bi
ning
the
m
ean
-
s
hi
ft
segm
ent
at
ion
al
gorith
m
and
m
axi
m
um
l
ikeli
ho
od
cl
assifi
er
of
th
e
pix
el
-
base
d
cl
assifi
c
at
ion
.
2.
RESE
A
R
CH MET
HO
D
2.1
Si
te d
escri
pt
ion
The
resea
rc
h
was
co
nduct
ed
within
t
he
con
cessi
on
area
of
IUP
HHK
PT
Ka
nd
el
ia
Alam
,
geog
raphical
ly
locat
ed
bet
w
een
10
9°34
'
10
,82”
E
&
109°4
1'
14
,
85"
E
;
and
betwee
n
0°
35'
18,21
"
S
&
0°39
'
53,
45
"
S.
A
dm
inist
rati
vely
,
the
st
ud
y
sit
e
is
locat
e
d
within
K
ub
u
Ra
ya
Re
gency
,
W
est
Kali
m
antan
Pr
ovi
nce
(F
ig
ure
1).
Fiel
d
ob
serv
at
io
n
an
d
f
ie
ld
m
easur
em
ents
wer
e
c
ondu
ct
e
d
in
20
16,
especial
ly
in
area
s
cov
e
re
d by U
A
V
(
U
nm
ann
ed
Aer
ia
l
Veh
ic
le
)
im
ager
y.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
9
,
No.
3
,
Ma
rc
h 201
8
:
72
2
–
7
30
724
Figure
1.
Stu
dy
sit
e
m
ap
2.2 Da
ta and
So
f
tware
The
m
ai
n
data
us
e
d
in
this
stu
dy
is
a
UAV
im
age
hav
i
ng
10
cm
sp
at
ia
l
re
so
luti
on,
8
bits
rad
iom
et
ric
value,
4
ba
nds,
nam
el
y
Red
,
G
ree
n,
Bl
ue
,
and
Alpha.
The
pr
e
proc
essing
work
s
uch
as
recti
fi
cat
ion,
rad
i
om
et
ric
co
rr
ect
io
n,
a
nd
ge
om
et
ric
cor
recti
on
was
done
by
data
su
pp
li
er.
The
us
e
d
UAV
im
ages
wh
ic
h
cov
e
r
ap
pro
xi
m
at
ely
40
87
ha
,
wer
e
recor
de
d
in
Fe
bruar
y
2016
.
Ot
her
s
upportin
g
data
us
ed
i
n
this
stud
y
are
the
data
c
ollec
te
d
from
gr
ou
nd
m
e
asur
em
ents
su
c
h
as
an
in
div
id
ual
diam
e
te
r
of
nypa,
le
ng
t
h
a
nd
diam
et
er
of
le
aves,
a
num
ber
of
le
aves,
a
nu
m
ber
of
stu
m
p
and
weig
ht
of
sam
ple
nypa
(100
-
200
gra
m
).
The
p
r
oce
ssing
and
data
analy
sis
wer
e
done
us
in
g
s
om
e
so
ftwar
e
su
c
h
as
QGIS
2.18,
Orfeo
To
olbo
x
/
Mon
te
ve
rd
i
1.24,
ERDA
S I
m
agine 9.
1
a
nd Mi
cro
s
of
t
Ex
cel
20
10.
2.3
Fie
ld
Me
asure
men
t and Data Pr
ocessi
ng
2.3.1
Fie
ld me
as
ureme
nt
Fo
r
c
om
par
iso
n
an
d
accu
rac
y
assess
m
ent
pu
r
poses
as
wel
l
as
to
ob
ta
in
qu
a
ntit
at
ive
an
d
qual
it
at
ive
inf
or
m
at
ion
ab
ou
t
the
a
ct
ual
conditi
on
of
ve
getat
io
n
i
n
th
e
fiel
d
,
the
n
groun
d
obse
rv
at
i
on
a
nd
m
easur
e
m
ent
wer
e
perform
e
d.
T
hese
fiel
d
ob
s
er
vations
and
m
easur
em
e
nts
wer
e
c
onduct
ed
on
f
our
cl
us
te
r
plo
ts
ha
ving
a
siz
e
of
60
m
x
100
m
.
Each
cl
us
te
r
was
div
id
ed
int
o
15
s
ub
cl
us
te
rs
(
cl
us
te
r
el
em
ents),
ha
ving
a
siz
e
of
20
m
x
20
m
each.
The
posit
ion
of
the
cl
us
te
r
was
m
ade
in
su
ch
a
way
so
that
i
ts
po
sit
ion
rela
ti
vely
per
pe
nd
i
cula
r
to the ri
ve
r flo
w.
2.3.2 Pr
e
-
pr
oc
essing
Pr
e
processin
g
of
U
A
Vs
inclu
des
im
age
cro
ppin
g
that
fit
the
sel
ect
ed
cl
us
te
r
locat
ions.
Crop
ping
wa
s
done
t
o
f
a
ci
li
tate
the
pr
ocess
of
segm
entat
ion
a
nd
cl
assifi
c
at
ion
.
The
siz
e
of
the
U
AV
im
age
f
or
eac
h
area
of
interest
is
1100
x
11
00
pi
xels
or
a
ppr
ox
im
at
e
ly
11
0
x
110
m
et
ers
in
siz
e.
T
he
area
of
interest
was
sel
ect
ed
t
o
cov
e
r
three
cl
asses
nam
el
y
m
ang
r
ove,
ny
p
a
and
ga
p.
Ea
ch
cr
opping
pr
ocess
of
the
a
rea
of
inte
rest
(A
O
I)
i
m
age w
as car
r
ie
d
out by fi
rstly
re
m
ov
in
g
th
e
al
ph
a
ba
nd, so that the im
age o
nly h
as
a
Re
d
ba
nd, Gree
n band
,
and Bl
ue ba
nd.
2.3.3 Se
gme
ntati
on
Segm
entat
ion
process
is
the
init
ia
l
ste
p
i
n
ob
j
ect
-
based
i
m
age
cl
assifi
cat
ion
(
OBI
A
),
w
he
re
it
s
cl
assifi
cat
ion
a
lgorit
hm
was
do
ne
by
m
erg
in
g
sm
al
le
r
segm
ents
into
la
r
ge
r
ob
j
ect
s
base
d
on
it
s
hom
ogeneit
y
(i.e.,
a
sim
i
la
rit
y
of
s
pectral
va
lue
an
d
s
patia
l
char
act
erist
ic
s)
of
t
he
im
age.
The
segm
entat
ion
m
e
tho
d
use
d
in
this
resea
rch
is
us
in
g
a
m
ean
-
s
hift
al
go
rithm
of
Orfe
o
T
oo
l
box/
M
on
te
verdi
1.2
4.
I
n
this
st
udy,
the
segm
entat
ion
appr
oach
us
ed
was
base
d
on
sp
ect
ral
an
d
sp
at
ia
l
app
r
oac
hes.
T
he
m
ean
-
sh
i
ft
al
go
rit
hm
was
first
intr
oduce
d
by
[
26
]
a
nd
has
bee
n
pro
ven
that
this
m
et
ho
d
is
a
ve
rsati
le
,
non
-
pa
ram
et
ric
m
e
t
hod
for
est
i
m
ating
gr
a
dients
i
n
the
cl
us
te
rin
g
pro
cess.
S
om
e
research
suc
cess
fu
ll
y
im
ple
m
e
nted
t
he
m
et
ho
d
f
or
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Assessin
g
t
he C
ro
wn Cl
osure
o
f N
y
pa on U
AV Im
ag
es
usi
ng Me
an
-
Shif
t
…
(
Rob
ert
Paruli
an S
il
al
ahi
)
725
so
lvi
ng
prob
le
m
s
at
low
-
le
ve
l
vision
[
27]
.
The
pa
ram
et
e
rs
us
e
d
in
se
gm
enta
ti
on
pro
cess
us
in
g
m
e
an
-
s
hift
al
gorithm
are
the
s
patia
l
ra
diu
s,
ra
nge
ra
di
us,
a
nd
m
ini
m
u
m
reg
ion
siz
e.
I
n
this
stu
dy
,
t
he
a
uthor
s
e
xa
m
ined
27
c
om
bin
at
io
ns
of
se
gm
entat
ion
pa
ram
et
ers
to
obta
in
se
gm
entat
ion
inf
orm
ation
on
ca
nopy
a
nd
ga
p
of
nypa
veg
et
at
io
n
a
t
r
esearch
locati
on
(Tab
le
1)
.
Table
1.
C
om
bin
at
ion
of the
param
et
er ex
am
ined o
n
the
se
gm
entat
ion
No
Settin
g
hr
hs
M
Co
m
b
in
atio
n
No
Settin
g
hr
hs
M
Co
m
b
in
atio
n
1
2
3
4
5
6
7
8
9
10
11
12
13
14
Settin
g
01
Settin
g
02
Settin
g
03
Settin
g
04
Settin
g
05
Settin
g
06
Settin
g
07
Settin
g
08
Settin
g
09
Settin
g
10
Settin
g
11
Settin
g
12
Settin
g
13
Settin
g
14
5
5
5
5
5
5
5
5
5
10
10
10
10
10
10
10
10
20
20
20
30
30
30
10
10
10
20
20
50
100
150
50
100
150
50
100
150
50
100
150
50
100
5
-
10
-
50
5
-
10
-
100
5
-
10
-
150
5
-
20
-
50
5
-
20
-
100
5
-
20
-
150
5
-
30
-
50
5
-
30
-
100
5
-
30
-
150
10
-
10
-
50
10
-
10
-
1
0
0
10
-
10
-
1
5
0
10
-
20
-
50
10
-
20
-
1
0
0
15
16
17
18
19
20
21
22
23
24
25
26
27
Settin
g
15
Settin
g
16
Settin
g
17
Settin
g
18
Settin
g
19
Settin
g
20
Settin
g
21
Settin
g
22
Settin
g
23
Settin
g
24
Settin
g
25
Settin
g
26
Settin
g
27
10
10
10
10
15
15
15
15
15
15
15
15
15
20
30
30
30
10
10
10
20
20
20
30
30
30
150
50
100
150
50
100
150
50
100
150
50
100
150
10
-
20
-
1
5
0
10
-
30
-
50
10
-
30
-
1
0
0
10
-
30
-
1
5
0
15
-
10
-
50
15
-
10
-
1
0
0
15
-
10
-
1
5
0
15
-
20
-
50
15
-
20
-
1
0
0
15
-
20
-
1
5
0
15
-
30
-
50
15
-
30
-
1
0
0
15
-
30
-
1
5
0
Remarks:
hs
:
Spatial
r
adi
us (p
i
x
el)
,
hr
:
R
ange radius/spec
tr
al
v
alue
(DN
)
and
M
:
Minim
um
re
gion
siz
e
(p
ix
el
)
Sp
at
ia
l
rad
ius
is
a
par
am
e
te
r
that
has
a
fu
nc
ti
on
to
co
ntro
l
the
distance,
m
easur
ed
by
a
nu
m
ber
of
pix
el
s.
T
he
s
pa
ti
al
rad
ius
wil
l
gr
ou
p
a
num
ber
of
pix
el
s
i
nto
on
e
se
gm
e
nt
or
one
ob
j
e
ct
,
wh
e
reas
th
e
range
rad
i
us
is
a
se
gm
entat
ion
pa
ra
m
et
er
that
correspo
nd
s
to
t
he
sp
ect
ral
valu
e
of
each
pi
xe
l.
The
ra
ng
e
r
adius
ref
e
rs
to
sp
ect
ral
var
ia
bili
ty
(d
ist
anc
e
in
n
-
dim
ension
al
sp
at
ia
l
sp
ace)
to
gro
up
a
nu
m
ber
of
pix
el
s
into
a
sing
le
se
gm
ent.
Furthe
rm
or
e,
it
is also d
efin
ed
that t
he
m
in
i
m
u
m
reg
ion
siz
e (M)
is a p
ar
a
m
et
er r
el
at
ed
to the
m
ini
m
u
m
s
iz
e
of
the
nu
m
ber
of
pix
el
s that f
or
m
a sing
le
obj
ect
. O
bj
ect
s t
ha
t ha
ve
the nu
m
ber
o
f
pix
el
s
belo
w
the
pa
ram
et
er
value
will
be
com
bin
ed
with
t
he
nea
rest
ob
j
e
ct
.
Af
te
r
t
he
pa
ram
et
er
values
are
determ
ined,
t
he
segm
entat
ion
process
is
perf
or
m
ed
on
al
l
fo
ur
res
earc
h
cl
us
te
rs
.
At
the
end
proces
s,
si
nce
the
se
gm
e
ntati
on
was
proces
sed
on
the
ra
ste
r
-
f
or
m
at
te
d
data,
an
outp
ut
of
th
e
segm
entat
ion
would
al
so
ra
ste
r
form
at
.
Fi
nal
l
y,
it
is n
eede
d
t
o conve
rt r
ast
er
data into
a
vect
or form
at
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
9
,
No.
3
,
Ma
rc
h 201
8
:
72
2
–
7
30
726
Figure
2. Flo
w
diag
ram
o
f
t
he
stu
dy pro
ce
ss
2.3.4
Clas
si
ficat
i
on
In
this
stu
dy,
the
auth
or
a
ppli
ed
the
hybr
i
d
betwee
n
the
obj
ect
-
base
d
(seg
m
entat
ion
,
OBIA)
a
nd
pix
el
-
base
d
m
axim
u
m
l
ikelih
oo
d
cl
assifi
er
(MLC).
T
he
resu
lt
s
of
se
gm
entat
ion
cl
assifi
cat
ion
re
sul
ts
of
segm
entat
ion
r
un
us
i
ng
O
rf
e
o
T
oo
l
box/Mo
nteve
rd
i
was
t
hen
i
nc
orporat
ed
with
th
e
re
su
lt
s
of
s
up
e
r
vised
cl
assifi
cat
ion
r
un
us
in
g
ERD
AS
im
agine.
Du
ri
ng
the
cl
assifi
cat
ion
us
i
ng
MLC
,
sever
al
cl
asses
wer
e
de
fine
d,
nam
ely
nypa
canopy,
m
angrove
ca
nopy,
bar
e
la
nd,
ga
p,
a
nd
water
body
.
T
he
r
esults
of
s
upe
rv
ise
d
cl
assifi
cat
ion
ha
ve
the
n
us
e
d
an
at
tribu
te
of
each
poly
gon
ob
ta
ine
d
du
ring
se
gm
entat
i
on
.
T
he
cl
assif
ic
at
io
n
ste
p ou
tl
ine
d
i
n
this
stu
dy is
dep
ic
te
d i
n Fi
gure
2.
2.3.5
Accur
ac
y Assessme
nt
On
e
of
t
he
im
portant
issue
s
in
determ
ining
the
optim
al
se
gm
entat
ion
is
i
n
the
acc
ur
acy
assessm
en
t.
Althou
gh the
re
are
s
om
e
m
et
h
od to
ass
ess th
e p
er
f
or
m
ance o
f segm
entat
ion
s
uch as
fr
a
gm
entat
ion
in
de
x
[
28]
,
the
area
fit
ind
ex
[
29]
and
there
is
al
so
ano
t
her
m
et
ho
d
to
evaluate
the
segm
entat
i
on
re
su
lt
us
in
g
area,
per
im
et
er
and
sh
a
pe
ind
e
x
(SI)
[
30
]
.
T
his
stud
y
us
e
d
a
co
m
par
ison
betw
een
the
segm
e
ntati
on
res
ults
with
ref
e
ren
ce
area
.
To
ide
ntify
th
e
m
os
t
accurat
e
segm
entat
ion
pa
ram
et
er
fo
r
asse
ssin
g
the
ny
pa
cr
own
cl
osure
a
nd
gaps,
the
c
onve
ntion
al
c
onf
usi
on
m
at
rix
an
al
ysi
s
was
perf
or
m
ed.
The
da
ta
ref
ere
nce
use
d
f
or
e
xpressi
ng
t
he
act
ual
crown
c
losure
an
d
ga
ps
der
i
ved
is
th
e
data
der
i
ved
from
visu
al
interp
retat
ion
a
nd
groun
d
obse
rvat
ion.
The
c
om
par
ison
betwee
n
th
e
aut
om
at
ed
cl
assifi
cat
ion
(
OBI
A
a
nd
M
LC)
was
t
hen
us
e
d
to
cal
cu
la
te
the
ov
e
rall
accu
rac
y (OA)
and
ka
pp
a
accu
racy
a
s the
fo
ll
ows
[
31
]
.
1
100
(1)
1
-
1
2
-
100
(2)
Rem
ark
s
:
OA
: Ov
e
rall
accu
r
acy
(
%)
K
: Kappa acc
ura
cy
(
%)
X
ii
: C
oin
ci
de
d val
ue
(num
ber
of
pix
el
)
N
: Total
pi
xel
K
:
Kappa acc
ura
cy
(%)
X
i+
:
The
s
um
o
f
c
olu
m
n
j
X
+i
:
The
s
um
o
f r
ow
i
2.3.6
Cro
w
n
Clos
ur
e
The
resu
lt
s
of
the
accu
racy
a
ssessm
ent
fo
r
al
l
com
bin
at
ion
s
of
t
he
pa
ra
m
et
er
set
ti
ng
wer
e
us
e
d
t
o
identify
the
optim
al
seg
m
entat
ion
param
eter
s.
T
he
best
com
bin
at
ion
of
segm
entat
ion
par
am
et
er
was
then
us
e
d
to
cal
cul
at
e
the
value
of
cr
own
cl
osu
r
e
(Cc)
of
n
ypa
in
eac
h
cl
us
te
r.
Pr
io
r
t
o
a
ny
f
ur
the
r
proces
s,
the
cl
us
te
r
was
si
m
ula
te
d
into se
ver
al
p
lot form
s
acco
rd
i
ng
to the
sha
pe
an
d
s
iz
e
of
the f
ie
ld p
lot
. S
el
ect
ion
o
f
t
he
m
os
t
op
tim
a
l
plo
t
siz
e
si
m
ul
at
ion
was
done
by
con
si
der
i
ng
the
c
oeffici
ent
of
c
o
va
riance
(CV
)
val
ue
of
the
bio
m
ass
,
volu
m
e
and
nypa
densi
ty
var
ia
ti
on.
T
he
pe
rc
entage
of
c
rown
cl
osure
is
the
rati
o
bet
w
een
total
crow
n
c
ov
e
rage (m
2
)
and p
l
o
t si
ze (m
2
).
3.
RESU
L
TS
A
ND AN
A
LYSIS
The
sel
ect
io
n
of
t
he
best
se
gm
entat
ion
par
a
m
et
ers
is
done
by
eval
uatin
g
t
he
a
ver
a
ge
val
ue
of
overall
accuracy
(OA
)
an
d
k
ap
pa
acc
ur
acy
(KA)
f
or
al
l
f
our
cl
us
te
rs.
Re
ca
pitula
ti
on
of
accu
rac
y
values
of
O
A
a
nd
KA
are
su
m
m
arized
i
n
Fi
gur
e
3
.
T
he
acc
uracy
value
was
der
i
ved
by
c
om
par
ing
the
se
gm
entat
ion
an
d
visu
a
l
interp
retat
ion
a
s w
el
l as
fiel
d ob
s
er
vation as
ref
e
ren
ce
d
at
a.
Of
the
27
com
bin
at
io
ns
of
th
e
segm
entat
ion
par
am
et
er
exam
ined
,
we
fou
nd
t
hat
the
rel
at
ively
hig
h
accuracy
in
pr
edict
ing
the
c
r
own
cl
osure
of
nypa
we
re
pr
ov
i
ded
by
the
set
ti
ng
-
10
(10
-
10
-
50),
set
ti
ng
-
01
(
5
-
10
-
50),
set
ti
ng
-
04
(
5
-
20
-
50)
and
set
ti
ng
-
19
(1
5
-
10
-
50
)
ha
ving
overall
accuracy
bet
we
en
76.
5
%
an
d
76.
6
%
and
k
a
pp
a
acc
ur
acy
b
et
wee
n
55.
6
a
nd
55.
7%
(F
ig
ure
3)
.
The
cl
assifi
cat
ion
acc
uracy
ob
ta
ined
by
a
pp
l
yi
ng
the
com
bin
at
ion
of
m
ean
-
sh
i
ft alg
or
it
hm
an
d
pi
xe
l
-
base
d
m
axi
m
u
m
li
kelihood
cl
assifi
er is s
li
gh
tl
y l
ow
er t
han
t
he
accuracy
of
t
he
com
bin
at
io
n
m
ean
-
sh
ift
al
gorithm
and
support
vector
m
a
chine
(
SV
M)
cl
assi
ficat
io
n
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Assessin
g
t
he C
ro
wn Cl
osure
o
f N
y
pa on U
AV Im
ag
es
usi
ng Me
an
-
Shif
t
…
(
Rob
ert
Paruli
an S
il
al
ahi
)
727
exam
ined
by
[
25
]
.
T
his
is
m
is
du
e
to
t
he
siz
e,
sh
a
pe
an
d
bri
ghtness
va
lue
of
the
ny
pa
ve
getat
ion
m
or
e
com
plica
te
d
than
siz
e,
sh
a
pe a
nd brig
htn
es
s
value o
f
the
bu
il
din
g o
bj
e
ct
s.
0
.
0
1
0
.
0
2
0
.
0
3
0
.
0
4
0
.
0
5
0
.
0
6
0
.
0
7
0
.
0
8
0
.
0
9
0
.
0
1
0
0
.
0
01
02
03
04
05
06
07
08
09
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
A
c
c
u
r
a
c
y
(%
)
Pa
r
a
m
e
te
r
Se
tti
n
g
K
a
p
p
a
a
n
d
O
v
e
r
a
l
l
A
c
c
u
r
a
c
i
e
s
Ka
p
p
a
OA
Figure
3
.
O
veral
l and
Ka
pp
a
Accuracy
fo
r
a
ll
co
m
bin
at
ion
of seg
m
entat
ion
par
am
et
er
Fr
om
these
accuracy
val
ues,
it
is
no
te
d
tha
t
the
var
ia
ti
on
in
the
value
s
of
sp
at
ia
l
rad
i
us
and
r
an
ge
rad
i
us
do
e
s
no
t
pro
vid
e
a
si
gnific
ant
di
ff
e
ren
ce
of
acc
ur
acy
(F
i
gure
4).
No
sig
nificant
dif
fer
e
nce
i
n
acc
uracy
was
obta
ined
wh
e
n
t
he
s
pat
ia
l
rad
iu
s
was
increas
ed
or
decr
ease
d.
Of
al
l
set
ti
ng
,
it
seem
s
that
the
m
os
t
accurate
sp
at
ia
l
rad
ius
pa
ram
et
er
fo
r
pr
e
dic
ti
ng
the
crow
n
cl
os
ur
e
is
10
pix
el
s.
Wh
e
n
the
sp
at
ia
l
rad
i
us
is
change
d
to
1
5
then
the
accu
ra
cy
decr
eased
.
Fr
om
the
ran
ge
rad
ius
po
i
nt
of
vie
w,
the
in
crease
of
range
rad
iu
s
from
10
to
20
causin
g
t
he
de
cl
ine
in
OA
f
r
om
76
.
6%
to
76.
3%,
ve
ry
ti
ny
change
(s
ee
set
ti
ng
-
10
(
10
-
10
-
50)
and
s
et
ti
ng
-
13
(10
-
20
-
50)
)
(Fi
gure
4
b
)
.
T
he
range
ra
di
us
di
ff
ere
d
sig
nific
antly
wh
e
n
it
r
ai
sed
f
ro
m
20
to
30
pix
el
s
(F
ig
ure
4
b
), causi
ng a
decr
ease
of
ov
erall
accu
racy
2.8% a
nd
kapp
a accu
racy o
f
a
bout
5.1%.
7
6
.
5
5
5
.
6
7
6
.
6
5
5
.
7
7
6
.
5
5
5
.
6
0
.
0
2
0
.
0
4
0
.
0
6
0
.
0
8
0
.
0
1
0
0
.
0
OA
Ka
p
p
a
Sp
a
ti
a
l
R
a
d
i
u
s
(h
s
)
Se
t
t
i
n
g
0
1
Se
t
t
i
n
g
1
0
Se
t
t
i
n
g
1
9
a)
7
6
.
6
5
5
.
7
7
6
.
3
5
5
.
2
7
3
.
5
5
0
.
1
0
.
0
2
0
.
0
4
0
.
0
6
0
.
0
8
0
.
0
1
0
0
.
0
OA
K
a
p
p
a
R
a
n
g
e
R
a
d
i
u
s
(h
r
)
S
e
t
t
i
n
g
1
0
Se
t
t
i
n
g
1
3
S
e
t
t
i
n
g
1
6
b)
7
6
.
6
5
5
.
7
7
5
.
0
5
2
.
7
7
4
.
0
5
0
.
6
0
.
0
2
0
.
0
4
0
.
0
6
0
.
0
8
0
.
0
1
0
0
.
0
OA
Ka
p
p
a
M
i
n
i
m
u
m
R
e
g
i
o
n
Si
z
e
(M
)
Se
t
t
i
n
g
1
0
Se
t
t
i
n
g
1
1
Se
t
t
i
n
g
1
2
c)
Figure
4
.
O
veral
l and
Ka
pp
a
accuracy
for
al
l par
am
et
er
The
thir
d
par
a
m
et
er
we
ob
se
rv
e
d
was
the
m
ini
m
u
m
reg
ion
siz
e
(M).
T
his
M
par
am
eter
determ
ine
s
the
num
ber
of
pix
el
s
that
m
a
y
com
po
se
a
sing
le
se
gm
ent
(o
bject
)
.
W
e
se
t
up
the
M
siz
e
on
the
basis
of
the
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
9
,
No.
3
,
Ma
rc
h 201
8
:
72
2
–
7
30
728
ind
ivi
du
al
diam
et
er
of
nypa
tree,
that
rangin
g
from
0.
5
m
2
~
1.
5
m
2
.
These
M
valu
es
are
from
50
to
15
0
pix
el
s.
Th
e
stu
dy
f
ound
t
hat
the
high
acc
ur
a
cy
set
ti
ng
s
a
re
obta
ine
d
f
r
om
the
value
M
of
50
pi
xels.
T
he
M
hav
i
ng
siz
e
m
o
re
tha
n
50
pix
e
ls
reduce
the
overall
accu
racy
.
The
m
ini
m
um
reg
io
n
siz
e
diff
e
re
d
sig
nifi
cantl
y
wh
e
n
it
raised
to
100
pi
xels
(
Figure
4c
),
an
d
causin
g
a
decre
ase
of
ove
rall
accuracy
2.6
%
and
ka
ppa
accurac
y
5.1%.
Am
on
g
the
t
hree
segm
entat
ion
pa
ram
et
ers
evaluated
,
it
is
sh
ow
n
that
the
M
is
the
m
os
t
aff
ect
in
g
par
am
et
ers,
la
r
ger
tha
n
t
he
s
pa
ti
al
rad
ius
a
nd
range
ra
diu
s
.
This
is
in
li
ne
with
the
stu
dy
car
ried
out
by
[
2
7],
expressi
ng
t
hat
sp
at
ia
l
rad
i
us
was
le
ss
se
ns
it
ive
than
t
he
oth
er
se
gm
entat
i
on
pa
ram
et
er.
The
M
is
ex
pressi
ng
the
ny
pa
siz
e
(
crow
n
c
ov
e
ra
ge
).
T
he
ra
diu
s
range
w
hich
e
xpress
es
the
ra
ng
e
of
sp
ect
r
al
is
le
ss
se
ns
it
ive
in
disti
nguish
i
ng the
ny
pa
le
a
ve
s
an
d
ga
p
. Thu
s
it
caus
e
s
le
ss sign
ific
a
nt
in
s
egm
entat
ion
pr
ocesses
of
t
he
crow
n
cl
os
ure a
naly
sis o
f
the
ny
pa.
(1)
(2)
(3)
a)
Clust
er
1
(1)
(2)
(3)
b)
Cluste
r
2
(1)
(2)
(3)
c)
Clust
er
3
(1)
(2)
(3)
d)
Cluste
r
4
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Assessin
g
t
he C
ro
wn Cl
osure
o
f N
y
pa on U
AV Im
ag
es
usi
ng Me
an
-
Shif
t
…
(
Rob
ert
Paruli
an S
il
al
ahi
)
729
Figure
5
.
Com
par
is
on b
et
wee
n
sam
ple o
f
ori
gin
al
im
age o
f
the UA
V
(le
ft
colum
n
: 1
) wit
h
se
gm
entat
ion
resu
lt
s
us
in
g se
tt
ing
-
10 (
m
idd
le
co
lum
n
:
2
)
a
nd the
set
ti
ng
-
04 (rig
ht co
l
um
n
: 3
)
at
clu
ste
r 1 (
a
)
~ cl
us
t
er
4
(d).
To
inc
rease
the
accu
racy
of
crow
n
cl
osu
re
de
li
neati
on,
the
auth
or
s
a
lso
ap
plied
a
pix
el
-
base
d
cl
assifi
er
that
reli
es
on
it
s
sp
ect
ral
sig
na
ture.
T
he
se
gm
entat
ion
re
su
lt
was
the
n
com
bin
ed
wi
th
the
cl
assifi
cat
ion
r
esult
ob
ta
ine
d
fr
om
the
su
pe
rv
ise
d
cl
assifi
cat
ion
.
T
his
is
in
li
ne
with
t
he
stud
y
of
[
20
]
fo
r
ident
ific
at
ion
of
be
nth
ic
habi
ta
t.
How
eve
r,
his
fi
nd
i
ng
pro
vid
e
d
a
ver
y
l
ow
overall
acc
ur
acy
of
only
27.7%
.
His
stu
dy
note
d
that
lo
w
acc
ur
acy
m
ay
be
du
e
to
the
ga
p
from
seg
m
entat
ion
process
w
hen
integ
rati
ng
phot
o
-
transect
an
d
cl
us
te
r
.
Be
si
des,
the
low
acc
uracy
m
ay
be
due
to
the
sam
pl
e
locat
ion
that
do
es
not
m
a
tch
wit
h
the ob
j
ect
in
th
e m
ap.
The
ex
am
inatio
n
on
the
use
of
a
pix
el
-
base
d
m
et
ho
d
to
cl
assify
crow
n
cl
os
ure
a
nd
t
he
gap
was
al
s
o
perform
ed.
Howev
e
r,
the
st
udy
resu
lt
s
s
how
that
it
s
abili
ty
to
identify
the
obj
ect
is
sti
ll
low.
O
ne
of
it
s
dr
a
w
back
s
is
i
n
th
e
la
belin
g
of
ob
j
ect
s,
w
hich
is
only
bas
ed
on
the
bri
ghtness
val
ue
or
c
olor
of
the
obj
ect
,
without
co
ns
i
de
rin
g
the
s
patia
l
aspects
su
c
h
as
locat
io
n,
siz
e,
sh
a
pe,
te
xture,
et
c.
T
he
su
ccess
or
fai
lure
of
pix
el
-
base
d
cl
assifi
cat
ion
is
so
le
ly
deter
m
ined
by
it
s
br
i
gh
t
ness
val
ue.
Th
us
,
the
qual
it
y
of
t
he
obj
ect
cl
assifi
cat
ion
r
esults
dep
e
nds
on
the
qu
al
it
y
of
pix
el
-
base
d
qu
al
ific
at
io
n
re
su
lt
s.
The
diff
i
culty
in
determ
ining
the
sp
ect
r
al
separ
at
io
n
of
a
c
la
ss
al
so
bec
om
es
a
chall
enge
in
pixe
l
-
ba
se
d
cl
assifi
cat
ion
par
ti
cula
rly
in
us
i
ng
the
high
-
res
olut
ion
data.
T
he
low
accu
racy
of
usi
ng
sp
ect
r
al
ran
ge
(
ra
diom
et
ric)
m
igh
t
be
due
to
the
im
age
qu
al
it
y
(b
lu
rr
i
ng
/i
m
age
m
oti
on
a
nd
lo
w
s
pe
ct
ral
reso
luti
on)
[
32
,
3
3
]
.
T
he
i
m
a
ge
blu
r
r
ing
due
to
the
hig
h
i
m
age
m
otion
aff
ect
the
a
bili
ty
of
U
AV
im
age
in
ide
ntify
ing
the
obj
e
ct
of
i
nterest
[
32
,
33
]
.
T
he
pr
es
ence
of
no
ise
al
so
re
duces
the
acc
urac
y
of
segm
entat
ion
.
I
n
t
he
very
high
-
res
olu
ti
on
im
ager
y,
it
is
quit
e
com
m
on
that
the
sm
a
ll
siz
e
of
gap
s
am
ong
trees,
br
a
nc
h
e
s,
tw gs,
a
nd
l
eaves
caus
ng
“sal
t
and
pe
ppe ”
no se
.
The
o
the
r
so
urce
of
no
is
e
m
ay
co
m
e
fr
om
the
m
isregistrat
ion
a
nd
geo
m
et
ric
corr
ect
ion
durin
g
pr
e
-
pr
ocessin
g
sta
ge.
Ba
sed
on
the
be
st
-
sel
ect
ed
pa
ram
et
er
seg
m
e
ntati
on
set
t
ing
-
10,
the
cr
own
cl
os
ure
of
ny
pa
within
the
sa
m
pl
in
g
plo
t a
re r
a
ngin
g from
3
8.4% t
o 61.6%
.
4.
CONCL
US
IO
N
Fr
om
the
fo
re
goin
g
res
ults
an
d
disc
us
sio
ns
,
t
his
stud
y
co
ncl
ud
e
s
that
the
m
os
t
op
tim
a
l
s
egm
entat
ion
par
am
et
ers
in
m
easur
in
g
the
crow
n
cl
osure
of
n
ypa
an
d
ga
p
is
prov
i
de
d
by
the
set
ti
ng
-
10
,
wh
ic
h
ha
s
10
f
or
sp
at
ia
l
rad
ius
,
10
f
or
range
rad
ius
a
nd
50
f
or
m
ini
m
u
m
reg
ion
siz
e
.
The
accu
racy
of
this
segm
entat
ion
par
am
et
er
co
m
bin
at
io
n
prov
i
des
a
ppr
ox
im
a
te
ly
7
6.6
%
of
ov
e
rall
accu
ra
cy
and
5
5.7
%
for
kappa
acc
ur
acy
.
This
stu
dy
al
so
co
nclu
de
d
that
var
ia
ti
on
of
m
in
i
m
u
m
re
gion
siz
e
(M)
con
t
rib
ute
a
sign
i
ficant
va
riat
ion
in
accuracy
asse
ssm
ent,
gr
eat
er
tha
n
an
ot
he
r
pa
ram
et
er
sp
at
ia
l
r
adi
us
(h
s
)
a
nd
ra
nge
rad
i
us
(hr)
.
The
cl
assifi
cat
ion
te
chn
i
qu
e
by
com
bin
ing
of
pix
el
-
base
d
an
d
obj
ect
-
base
d
has
giv
e
n
a
prom
isi
ng
resu
lt
in
delineat
in
g
a
ve
ry sm
all f
eat
ure
within t
he
n
ypa v
e
getat
io
n.
REFERE
NCE
S
[1]
Duke
NC.
Man
grove
Floristi
cs
and
Biog
eogr
a
ph
y
.
In:
Roberts
on
AI,
Alongi
DM
.
Edi
tors
.
T
ropic
a
l
Mangro
v
e
Ec
os
y
st
ems
.
W
a
shington
DC:
A
m
eri
ca
n
G
eoph
ysica
l
Union.
199
2.
63
-
100
.
[2]
Heri
y
ant
o
NM
,
Subiandono
E
,
Karl
ina
E.
Pote
nsi
dan
seb
ara
n
Nipah
(
Ny
fa
fru
ct
i
cans
(Thunb.
)
W
urm
b)
Sebaga
i
Sum
ber
da
y
a
Pan
gan.
Jurnal Pe
n
didi
kan
Hutan
d
an
Konservasi
A
lam
.
2011;
8(4):
327
-
335.
[3]
Ta
m
unai
du
S,
Matsui
N,
Oki
m
ori
Y,
Saka
S
.
Nipa
(
N
ypa
fr
uti
cans
)
sap
as
a
pote
n
ti
a
l
f
ee
d
stock
for
e
tha
n
ol
produc
ti
on
.
Journal
B
iomass
and
Bi
o
ene
rgy
.
201
3;
52:
96
-
102.
[4]
Eve
ra
ert
s
J.
Th
e
Us
e
Of
Unm
an
ned
Aeri
a
l
Vehi
cl
es
(UA
VS
)
for
Remote
Sensing
And
Mapping.
The
Inte
rnation
al
Archi
v
es
of
The
Phot
ogram
metry
,
R
emote Se
nsin
g
and
Spat
ial Inf
orm
ati
on
Scienc
es
.
2008;
38:
11
87
-
1192.
[5]
Getz
in
S,
Nus
ke
RS
,
W
ie
gand
K.
Us
ing
Un
m
anne
d
Aeri
al
Ve
hic
l
e
(UA
V)
to
Qu
ant
if
y
Spatial
Gap
Patt
ern
s
In
Forests.
Re
mot
e Sens
.
2014;
6(8):
6988
-
7004.
[6]
W
al
la
c
e
L
,
Lucie
er
A,
W
at
son
C.
Assesing
th
e
Fe
asibi
li
t
y
of
UAV
-
based
LiDAR
for
High
Re
soluti
on
Fore
st
Change
Detect
i
on
.
Th
e
12th
C
ongre
ss
of
the
Inte
rna
ti
ona
l
Societ
y
for
Photo
gra
m
m
et
r
y
and
Remote
Sensing
.
Melbourne
.
201
2;
39:
499
-
504.
[7]
W
al
la
c
e
L
,
Lu
ciee
r
A,
Turne
r
D
,
W
at
son
C.
Error
Assess
ment
And
Mit
iga
ti
on
F
or
Hype
r
-
Tempor
al
UAV
-
Borne
Lidar
Surve
ys
Of
Forest
In
ve
n
tory
.
The
1
1
th
Int
er
nat
ion
al
Conf
erence
on
Li
DA
R
Applic
a
ti
ons
for
As
sess
ing
Forest
Ec
os
y
st
ems
.
Hobart
.
2012;
1
-
13.
[8]
Dandois
JP
,
El
li
s
EC.
High
Spatial
Resolution
Thre
e
-
Dim
ensi
onal
Mapping
Of
Vege
ta
ti
on
Spect
ra
l
D
y
nam
ic
s
Us
ing
Com
pute
r
Vision.
Journal Re
mote
Sensing
of
En
vi
ronm
ent
.
2013;
136:
259
-
276.
[9]
Ja
y
a
INS
,
Cah
y
ono
AB.
Kajian
Te
kn
is
Pem
anf
aata
n
Potr
et
Udara
Non
-
Metri
k
Form
at
K
ec
i
l
Pada
B
idang
Kehuta
nan
.
Jurnal
Manaj
eme
n
H
utan
Tr
opik
a.
20
01;
7(1):
55
-
64.
[10]
Ja
y
a
INS
,
Kle
in
n
C,
Mela
ti
D,
F
ehr
m
ann
L,
Pére
z
-
Cruza
do
C
,
Sept
y
awa
rda
n
i
E,
Dhani
FA
R,
W
ac
hjuni
S.
Utiliz
i
ng
Mult
i
-
Sourc
e
Data
For
Sustainabl
e
Forest
Ma
nageme
nts
In
Indone
sia.
Bridgi
n
g
The
Gap
Be
twe
en
Informatio
n
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
9
,
No.
3
,
Ma
rc
h 201
8
:
72
2
–
7
30
730
Nee
ds
And
For
est
Inv
en
tory
C
apaci
t
y
.
Proce
e
dings
of
the
5
th
I
nte
rna
ti
ona
l
DA
AD
W
orksh
o
p.
Durban
da
n
Piet
ermari
tz
burg
.
2015:
163
-
181.
[11]
Kus
nadi
,
Ja
y
a
I
NS
,
Pus
pani
ngsih
N,
Basuki
M,
Hakim
L.
Model
Penduga
Kua
li
t
as
Te
m
pat
Tum
buh
Jati
(
Tect
on
a
gra
ndis
)
Mengg
unaka
n
Ci
tra
Re
solusi
Sangat
Tinggi
Pesawat
Tidak
Ber
awa
k
Di
KP
H
Nganjuk.
Jurnal
Pe
n
eliti
a
n
Kehut
anan
Wa
llace
.
2016;
5(2):
185
-
194.
[12]
Bla
schke
T.
Obj
ec
t
B
ase
d
Im
age
Anal
y
s
is
for
Remote
Sensing.
ISPR
S
Journal
of
Phot
ogram
metry
and
R
emot
e
Sensing
.
2010
;
6
5(1):
2
-
16.
[13]
Pal
NR,
Pa
l
NK
.
A Re
vi
ew
on
I
m
age
Segm
ent
a
t
ion
T
ec
hniqu
es.
Pat
te
rn
Recogni
t
ion
.
1993
;
26(9)
:
1277
-
1294
.
[14]
Duro
DC,
Frankli
n
SE,
Dube
MG
.
A
Com
par
ison
of
Pixel
-
B
ase
d
and
Obje
ct
-
Bas
ed
Im
age
Anal
ysis
W
it
h
Sele
c
ted
Mac
hine
Le
arn
i
ng
Algorit
hm
s
for
The
Cla
ss
ific
at
ion
of
Agric
u
ltural
L
andsc
ape
s
Us
ing
Spot
HR
G
Im
age
r
y
.
201
2;
118:
259
-
272.
[15]
Sari
NM
,
Kus
ha
rdono
D.
Objec
t
Segm
ent
at
ion
on
UA
V
Photo
Dat
a
to
Support
the
Provision
of
Rur
al
Area
Spatia
l
Inform
at
ion.
For
um Geografi
.
20
15;
29(1):
49
-
58.
[16]
W
ahi
din
N,
Sire
gar
VP
,
Naba
ba
n
B,
Ja
y
a
I
,
W
outhu
y
ze
n
S.
Obj
ec
t
-
b
ase
d
Im
age
Anal
y
sis
for
Co
ra
l
Re
ef
Bent
h
i
c
Habit
a
t
Mapping
with
Sev
eral
Cl
assific
a
ti
on
Algo
rit
hm
s.
Proc
edi
a
Environm
ent
al
Sci
en
ce
s
.
2015;
24:
222
-
227.
[17]
Conche
dda
G
,
Durie
ux
L,
Ma
yaux
P.
An
Obje
ct
-
base
d
Me
thod
for
Mapping
an
d
Change
Anal
ysis
in
Mangrove
Ec
os
y
st
em.
In
te
r
nati
onal Journal
of Re
mot
e
Se
nsing
.
2008;
63(5):
578
-
589.
[18]
Cheha
t
a
N,
Orny
C,
Boukir
S,
Gu
y
on
D,
W
igne
ron
JP
.
Objec
t
-
base
d
Change
Detect
ion
in
W
ind
Storm
-
Da
m
age
d
Forest
Us
ing
High
R
esolut
ion
Multi
spec
tral
Im
age
s.
Inte
rnational
Journal
of
Re
mote
Sensing
.
201
4;
35(13):
4758
-
4777.
[19]
La
ckn
er
M,
Co
nwa
y
TM.
Det
ermining
La
nd
-
use
Inform
at
ion
from
La
nd
Co
ver
Through
an
Objec
t
-
or
ie
n
te
d
Cla
ss
ifi
c
at
ion
of
IKO
NO
S Im
age
r
y
.
Canadian
Jo
urnal
of
R
emot
e
Sensin
g.
2008;
34(2):
77
-
92
.
[20]
W
ic
aksono
P,
Farda
NM
.
Apl
i
k
asi
Al
goritma
Klasif
ik
asi
M
ean
Shif
t
untu
k
Pe
m
et
aan
Habitat
B
ent
i
k
Studi
Kasus
Kepul
auan
Kari
munjawa
.
Pros
id
ing
Pert
emuan
Il
m
ia
h
Ta
hun
an
X
X.
Bogor.
2015:
663
-
672.
[21]
Bal
lari
D,
Orel
l
ana
D,
Acosta
E,
Espinoza
A,
Morocho
V.
UAV
Monit
oring
fo
r
Env
iromental
Manage
ment
in
Galapagos
Islan
d
.
The
Inte
rn
ati
onal
Archi
ves
o
f
the
Photogramm
et
r
y
,
Remote
Sensing
and
Sp
at
i
al
Inform
at
io
n
Scie
nc
es
.
ISP
RS
Congress XXIII.
Prague
.
20
16: V
ol
XLI
-
B1.
11
05
-
1111.
[22]
Lin
H.
Me
thod
of
Image
Segme
ntat
ion
on
High
-
Re
soluti
on
Ima
ge
and
Classifi
cat
ion
for
Land
Cove
rs
.
Fourth
Inte
rna
ti
ona
l
Co
nfe
re
nc
e
on
Nat
ura
l
Com
putatio
n
.
Jinan
.
2008
:
5
63
-
566
.
[23]
Aiaz
z
i
B
,
Alpar
one
L
,
B
aro
nti
S
,
Gar
ze
l
li
A
,
Zo
ppet
ti
C.
Nonpar
ametr
ic
Chang
e
Dete
c
ti
on
In
Mu
lt
item
pora
l
SA
R
Im
age
s
Based
On
Mea
n
-
Shift
Cluste
ring
.
IE
E
E
Tr
ansacti
ons
on
Geosci
en
ce
and
Re
mote
Sen
sing
.
2013;
51
(
4
):
2022
-
2031
.
[24]
Ti
ng
Y,
Mingxi
ng
G,
Yanm
ing
W
.
Ultra
sound
Im
age
Segm
ent
at
ion
b
ase
d
on
the
Mea
n
-
Shif
t
and
Graph
Cut
s
The
or
y
.
TEL
KOMNIKA
.
2013
;
1
1(10):
5600
-
560
8.
[25]
Bena
rc
h
id
O,
Ra
issouni
N.
Support
Vec
tor
Ma
ch
ine
for
Obje
ct
B
ase
d
Buil
d
ing
E
xtra
c
ti
on
in
Sub
urba
n
Area
Us
in
g
Ve
r
y
High
R
esolut
ion
Sat
el
l
ite
Im
age
s,
a
Ca
se
Stud
y
:
Te
tu
an,
Morocc
o
.
I
ndonesian
Journal
of
Artificia
l
Inte
lligen
ce
.
201
3;
2(1):
43
-
50.
[26]
Fukunaga
K,
Hos
te
tl
e
r
LD.
Th
e
Esti
m
at
ion
of
the
Gradi
ent
of
a
Densit
y
Func
ti
on,
with
Appli
ca
t
ion
in
Patt
er
n
Rec
ognition.
I
E
EE
Tr
ansacti
ons
on
Information
Theory
.
1975;
2
1(1):
32
-
40.
[27]
Com
ani
ci
u
D,
Mee
r
P.
Me
an
Shift:
A
Robust
Approac
h
Towa
rd
Fea
ture
Spa
ce
Ana
l
y
s
is.
IE
EE
Tr
ansacti
on
s
of
Pat
te
rn
Analysis and
Mac
h
ine In
te
lligen
ce
.
2002;
24(5):
603
-
619.
[28]
Straste
rs
KC,
Gerbra
nds
JJ
.
Thr
ee
-
dimensiona
l
I
m
age
Segm
ent
aion
Us
ing
A
Spl
it
,
Merg
e
and
G
roup
Approac
h.
Pat
te
rn
Recogni
t
ion
Letter
.
1991;
12(5):
307
-
325.
[29]
Luc
i
ee
r
A.
Unc
e
rta
inties
in
Segm
ent
a
ti
on
and Their
Visual
isat
ion
.
PhD
The
sis.
Utr
ec
ht
:
Utre
cht Unive
r
sit
y
;
2004.
[30]
Neube
rt
M,
Her
old
H,
Meine
l
G
.
As
sess
ing
Im
a
ge
Segm
ent
at
ion
Quali
t
y
-
Conc
e
pts,
Methods
an
d
Applic
ation.
I
n:
Bla
schke
T,
Ha
y
G,
L
ang
S.
E
dit
ors
.
Obje
ct
-
B
ase
d
Im
age
Ana
l
y
sis
-
Spatial
Conce
pts
For
K
nowledge
-
Drive
n
Remote
Sensing
Applic
ations.
L
ec
tur
e
Notes
in
Geoinf
orm
at
ion
&
Cart
ogra
ph
y
(
LNG&C).
Berl
i
n:
Springer;
200
8:
769
-
784.
[31]
Ja
y
a
INS
.
Anal
i
sis
Cit
ra
Digital
,
Perspekti
f
Pen
ginde
ra
an
Jauh
untuk
Pengel
ol
a
an
Sum
ber
da
y
a
Alam.
Bogor:
Instit
ut
Pe
rta
n
i
a
n
Bogor.
2015:
1
85
-
232.
[32]
W
ah
y
uni
S,
Ja
y
a
INS
,
Pus
pani
ng
sih
N.
Model
for
Esti
m
ating
Abo
ve
Ground
Biom
ass
of
Recla
m
ati
on
Forest
Us
ing
Unm
anne
d
Aeri
al
Vehi
cles.
Ind
onesian
Journal
of
Elec
tri
cal
En
gine
ering
and
Computer
Sci
en
ce
.
2016;
4(3)
:
58
6
-
593.
[33]
Zhe
ng
L
,
Z
h
an
g
J,
W
ang
Q.
Mea
n
-
Shift
Bas
ed
Color
Segm
ent
a
ti
on
o
f
Im
a
ges
Conta
in
ing
Gree
n
Vege
ta
t
io
n.
Computers and Electronic
s
in A
gricul
ture
.
2008;
65
(1)
:
93
-
98
.
Evaluation Warning : The document was created with Spire.PDF for Python.