Int
ern
at
i
onal
Journ
al of Ele
ctrical
an
d
Co
mput
er
En
gin
eeri
ng
(IJ
E
C
E)
Vo
l.
9
, No
.
3
,
J
un
e
201
9,
pp. 172
0~17
31
IS
S
N: 20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v9
i
3
.
pp.17
20
-
1731
1720
Journ
al h
om
e
page
:
http:
//
ia
es
core
.c
om/
journa
ls
/i
ndex.
ph
p/IJECE
A revie
w of
remotely
sens
ed satelli
te im
age classific
ation
Sa
ks
hi D
hing
ra
,
Dharmi
n
d
er Kum
ar
Depa
rt
m
ent
o
f
C
om
pute
r
Scie
n
ce a
nd
Engi
n
ee
rin
g,
Guru Jambhes
hwar
Univer
sit
y
of
Scie
n
ce a
nd
T
ec
hnolog
y
,
Ind
ia
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
A
pr
24
, 201
8
Re
vised
Dec
3
,
2018
Accepte
d Dec
15
, 201
8
Sate
llite
image
c
la
ss
ifi
c
at
ion
h
as
a
vital
ro
le
fo
r
th
e
ext
r
action
and
ana
l
y
sis
of
the
useful
satellite
image
informati
on
.
Thi
s
paper
comprises
the
stud
y
of
th
e
sate
llit
e
images
cl
assifi
ca
t
ion
and
Remo
te
S
ensing
al
ong
with
a
bri
ef
over
vie
w
of
the
pre
vious
studie
s
tha
t
ar
e
propose
d
in
thi
s
fie
ld
.
In
thi
s
pape
r,
the
exi
st
ing
work
has
bee
n
explaine
d
ut
il
i
zi
ng
t
he
cl
assifi
cation
te
chni
qu
es
on
satellite
images
of
Alwar
re
gion
in
India
th
at
cove
rs
d
ec
en
t
la
nd
cover
fea
tur
es
li
k
e
Ve
get
a
ti
on,
W
at
er
,
Urban,
Ba
rre
n,
a
nd
Rock
y
r
egi
on
s.
The
post
-
implementa
t
ion
of
the
cl
assifi
ca
t
ion
al
gor
it
h
m
s,
the
class
ified
image
is
obta
in
ed
displaying
diffe
ren
t
class
es
tha
t
are
rep
r
e
sente
d
b
y
d
iffe
r
e
nt
col
ours.
Ea
ch
f
ea
tur
e
is
r
e
pre
sent
ed
b
y
a
diffe
ren
t
col
our
and
ca
n
b
e
ea
si
l
y
per
ceive
d
from
the
imag
e
obta
in
ed
af
te
r
class
ifi
cation.
Th
e
foc
us
of
thi
s
st
ud
y
is
on
enha
nc
ing
th
e
class
ifi
cation
accu
racy
b
y
using
pr
oper
c
la
ss
ifi
e
rs
al
ong
wi
th
the
novel
f
eatur
e
ext
r
ac
t
ion
tec
hnique
s
and
pre
-
proc
essing
step
s.
W
ork
of
diffe
ren
t
au
thors
is
bei
ng
dis
cussed
in
a
ta
bul
ar
form
def
ini
ng
th
e
m
et
hods
and
out
comes of
the re
spe
ct
iv
e
st
udie
s
.
Ke
yw
or
d
s
:
Feat
ur
e
ex
tr
act
ion
Rem
ote sen
sin
g
Sate
ll
it
e i
m
age classi
ficat
ion
Super
vised
classi
ficat
ion
Un
s
uper
vise
d cl
assifi
cat
ion
Copyright
©
201
9
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
:
Sakshi D
hi
ngr
a,
Dep
a
rt
m
ent o
f C
om
pu
te
r
Scie
nce a
nd E
ng
i
ne
erin
g,
Guru Jam
bh
es
hw
a
r Un
i
versi
ty
o
f
Science
and Tec
hnol
og
y,
H
issar,
H
a
ryan
a
-
12
5001, In
di
a.
Em
a
il
: saksh
i2
4.dh
i
ngra@
gma
il
.co
m
1.
INTROD
U
CTION
A
sat
el
li
te
i
m
a
ge
ca
n
be
de
fin
ed
as
a
re
pr
ese
ntati
on
of
com
plete
or
eart
h
pa
rt
acq
uire
d
with
arti
fici
al
sat
el
li
te
s;
in
wh
ic
h
i
m
ages
hav
e
num
ber
of
util
iz
at
ion
li
ke
cartog
ra
ph
y,
m
ilit
ary
intel
le
ct
,
and
weathe
r
pr
e
dicti
on
[
1].
These
im
ages
cou
l
d
be
wate
r
vapor
im
ages,
visible
li
gh
t
and
in
fr
a
re
d
i
m
ages.
T
he
visib
le
li
gh
t
i
m
age
com
pr
ise
s
of
im
ages
po
rtrayi
ng
the
s
ky
as
well
as
the
gro
und
de
m
on
strat
ion
f
r
om
ou
te
r
sp
ac
e
at
a
sp
eci
fied
ti
m
e.
W
at
er
va
pour
i
m
ages
m
a
inly
il
lustrate
the
hum
idity
within
the
e
nv
i
ronm
e
nt,
usual
ly
in
c
loud
form
. I
t co
uld
gen
e
rall
y be e
m
plo
ye
d
for
tr
ackin
g
im
m
inent sto
rm
s w
it
h t
he
weat
her p
r
edict
ion
sit
uations.
Sate
ll
it
e
i
m
age
ge
ner
al
ly
enc
om
passes
rich
co
ntents
a
nd
play
s
a
sig
nifi
cant
r
ole
in
de
li
ver
in
g
t
he
env
i
ronm
ental
evide
nce.
Sate
ll
it
e
as
well
as
rem
ote
sensing
i
m
ages
gi
ves
qu
a
ntifia
ble
a
s
well
as
qual
it
at
ive
evide
nce
w
hic
h
m
ini
m
iz
e
the
fiel
d
ef
fort
and
t
he
inter
val
com
plexity
.
The
se
ns
in
g
eq
ui
pm
ent
integra
te
s
the
data
an
d
the
im
ages
at
no
rm
al
intervals.
T
he
data
qua
ntit
y
ob
ta
ine
d
fro
m
the
data
centres
is
enorm
ou
s
with
the
ex
pone
ntial
ly
rising
li
ke
the
te
ch
no
l
og
y
is
dev
el
opin
g
at
a
qu
ic
k
s
pee
d.
T
her
e
is
a
n
extrem
e
req
uir
e
m
ent
of
a
r
obus
t
te
c
hn
i
qu
e
f
or
the
ex
tract
io
n
of
us
ef
ul
in
form
a
ti
on
f
r
om
the
i
m
ages
[2
]
.
T
he
cl
assifi
cat
io
n
has
e
m
erg
ed
as in
f
luentia
l an
d robu
st
tech
nique
s that f
ulfil t
he
needs f
or the
e
xtracti
on
of
im
portant i
nfor
m
at
ion
.
The
cl
assifi
cat
ion
has
a
sig
ni
ficant
r
ole
in
the
rem
ov
al
of
i
m
po
rtan
t
in
form
ation
f
ro
m
t
he
sat
el
li
te
i
m
ages.
The
c
la
ssific
at
ion
is
usual
ly
def
in
ed
as
a
proce
dure
of
i
ntegrat
ing
the
pix
el
s
with
the
sig
nifican
t
cl
asses. T
his cl
assifi
cat
ion
pro
cess is
gen
e
rall
y per
form
ed
w
it
h
w
orkf
l
ow in m
ulti
steps [3].
Its classi
ficat
ion
c
an
al
s
o
be m
entioned
as t
he
proce
dure
of m
ining
valua
ble d
at
a from
l
arg
e
vo
l
um
es
of
t
he
sat
el
li
te
i
m
ages
and
it
s
cl
assifi
cat
ion
is
not
com
po
un
d;
howe
ve
r,
th
e
exam
iner
has
to
exec
ute
m
ulti
ple
op
ti
ons
wh
ic
h
range
in
the
process
of
sat
el
li
te
i
m
age
cl
assif
ic
at
ion
[
4].
T
his
inclu
des
a
na
ly
sis
of
sp
at
ia
l
data
m
ining
,
rem
ote
sensi
ng
im
ages,
a
naly
zi
ng
va
ried
ve
getat
io
n
kinds
li
ke
fa
rm
ing
,
f
or
e
stry
et
c.
a
nd
e
xam
ining
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
A revie
w of r
e
mo
te
ly
se
ns
e
d satel
li
te
ima
ge
cl
as
sif
ic
ation
(
Sa
k
sh
i
D
hingr
a
)
1721
urba
n
re
gions
for
determ
ining
se
ver
al
la
nd
us
a
ges
of
the
reg
i
on
.
In
la
te
1972,
Lan
dsa
t
-
I
was
the
init
ia
l
sat
el
li
te
to
int
egr
at
e
Eart
h
r
eflect
ance
on
the
60
m
reso
luti
on.
O
n
that
tim
e
per
iod,
the
su
pe
r
vised
an
d
un
s
uper
vise
d
cl
assifi
cat
ion
s
wer
e
ta
ke
n
as
two
m
et
ho
ds
a
nd
it
was
en
ou
gh
in
case
of
s
patia
l
reso
luti
on
[
5].
W
it
h
the
pass
age
of
ti
m
e,
t
he
dem
and
of
rem
otely
sense
has
been
inc
r
eased.
N
um
bers
of
rem
ote
se
ns
in
g
app
li
cat
io
n
are
there
relat
ed
to
en
vir
on
m
ental
,
public
safety
and
foo
d
sec
ur
it
y.
To
f
ulfil
these
dem
and
s,
the
sat
el
li
te
i
m
age
s
aim
ed
at
an
enh
a
nce
d
s
patia
l
reso
luti
on
at
higher
fr
e
que
ncy
ra
ng
e
,
t
ho
ugh,
t
his
does
nass
ur
e
i
m
pr
oved
la
nd
cov
er
.
The
us
age
of
im
age
cl
assifi
cat
ion
te
chn
i
qu
e
s
is
con
si
der
e
d
as
a
sign
ific
ant
fa
ct
or
f
or
i
m
pr
oved
accu
racy [
6].
Fi
gur
e 1
show
s a s
at
el
li
te
i
m
age which
c
onta
ins
va
rios segm
ents li
ke
cr
op, bar
r
en
et
c.
Figure
1
.
Sate
ll
it
e i
m
age in
cl
udin
g
trees
, low
er
veg
et
at
io
n,
i
m
per
vio
us s
urf
aces an
d hyd
rol
og
ic
al
f
eat
ur
es
Fr
om
the
histor
ic
al
po
int
of
vi
ew,
the
cl
assif
ic
at
ion
accurac
y
ca
m
e
ou
t
to
be
le
ss
du
e
to
the
la
ck
of
appr
opriat
e
featur
e
extracti
on
an
d
sel
ect
ion
al
gorithm
.
In
sat
el
li
te
i
m
a
ge,
feat
ur
e
ext
racti
on
a
nd
se
le
ct
ion
process
play
a
vital
ro
le
to
a
chieve
bette
r
a
ccur
acy
.
I
n
th
e
tradit
ion
al
w
ork,
t
her
e
is
no
existe
nce
of
su
c
h
com
bin
at
ion
(
f
eat
ur
e
e
xtracti
on
a
nd
sel
ect
io
n)
[7
]
.
T
he
e
xisti
ng
w
ork
is
de
fici
ent
in
fin
di
n
g
plane
acc
ordi
ng
to
dif
fer
e
nt
f
re
qu
e
ncies
from
sat
el
li
te
i
m
ages.
S
o,
it
is
m
and
at
ory
to
a
pp
ly
pr
e
-
processin
g
to
fin
d
the
pl
ane
as
per f
reque
ncies.
The
ab
ove
pro
blem
can
be
so
lve
d
by
us
i
ng
the
basic
pr
e
-
processi
ng
ste
ps
su
c
h
as
pla
ne
extracti
on,
ROI
e
xtra
ct
io
n,
plan
e
filt
erin
g
usi
ng
m
or
ph
ologica
l
operat
ion
s
et
c.
I
n
the
post
-
pr
ocessin
g
ste
p,
hy
br
i
d
f
eat
ur
e
sel
ect
ion
al
gor
it
h
m
li
ke
the
com
bin
at
ion
of
GA
(
Gen
et
ic
A
lgorit
hm
)
with
ABC
(
Ar
ti
fici
al
bee
col
on
y
)
c
an
be
us
e
d.
With t
his co
m
bin
at
ion, i
f
the
classi
fier
can
be use
d
tha
n
the
outcom
e w
ould
b
e
b
et
te
r.
The
m
anu
scrip
t
cou
ld b
e
orga
nized
in
the b
e
low
a
ppro
ac
h: Th
e
fir
st
segm
ent
deals
with
i
ntr
oductio
n
of
sat
el
li
te
i
m
a
ge
cl
assifi
cat
io
n
that
def
i
nes
t
he
ba
ck
gro
und
of
t
he
c
on
ce
pt
with
the
na
rr
a
ti
on
of
t
he
pr
oble
m
fo
ll
owin
g
t
he
pro
po
se
d
w
ork
that
ca
n
be
a
dopted
to
s
olve
the
pro
blem
.
The
c
on
ce
pt
of
rem
ote
sens
ing
i
s
exp
la
ine
d
la
te
r
.
Af
te
r
that,
the
fo
c
us
ha
s
been
giv
e
n
to
diff
ere
nt
stu
di
es
via
wh
ic
h
the
cl
assifi
cation
of
sat
el
li
te
i
m
ages
can
be
do
ne
.
Lat
er,
t
he
f
oc
us
has
been
giv
e
n
to
dif
fere
nt
cl
assifi
ers
thr
ough
w
hich
the
cl
assifi
cat
ion
will
be
done
on
the
basis
of
se
ver
al
t
ypes
of
ba
nds
tha
t
exist
i
n
the
sat
el
li
te
i
m
age.
A
com
par
is
on
has
been
i
n
the
on
the
basis
of
sat
el
li
te
i
m
age
and
cl
assifi
ed
sat
el
li
te
i
m
age
f
or
to
detect
in
g
the
pr
ese
nce
of
ve
getat
ion
base
d
on
the
dif
fe
ren
t
sce
nar
i
os
.
At
the
en
d
of
this
pa
per,
a
glance
of
e
xisti
ng
te
chn
iq
ues
is provide
d
i
n
the
tabu
la
r
f
orm
.
2.
RE
M
OTE
SE
NS
I
NG
The
data
bein
g
ac
qu
i
red
in
the
obj
ect
s
wit
hin
t
he
ea
rth
s
urface
la
c
king
in
a
ppr
oach
i
ng
in
c
on
ta
ct
with
th
em
is
re
m
ote
sensi
ng.
Senso
rs
a
re
de
vices
us
e
d
f
or
c
onduct
ing
inv
e
sti
gations.
It
cl
aim
ed
that
it
is
feasible
f
or
the
accum
ulati
on
of
data
on
risky
or
isolat
ed
re
gions.
Its
ap
plica
ti
on
s
kee
ps
a
n
ey
e
on
reg
i
ons
of
defor
e
sta
t
ion
l
ike
Am
azon
Ba
sin,
cl
im
a
te
var
ia
ti
on
sig
nifi
cances
in
glaci
ers
as
well
as
the
Ar
ct
ic
wit
h
the
An
ta
rcti
c
reg
i
ons
with
the
pitch
coa
sta
l
sou
nd
i
ng
with
t
he
ocean
pitches
[8
]
.
Its
data
ha
s
sp
at
ia
l
as
w
el
l
as
sp
ect
ral reso
l
ut
ion
s: Sp
at
ia
l
re
so
luti
on
is de
fi
ned
as
a
com
pu
ta
ti
on
of
s
hort
est
pointed
or
l
inear p
arti
ng
a
m
on
g
two
obj
ect
s
w
hich
c
ou
l
d
be
determ
ined
via
a
sens
or.
I
f
th
e
reso
l
ution
of
sens
or
is
m
ore
then
t
he
cap
aci
ty
of
data
is
gr
eat
er
bu
t
sm
aller
are
a
is
encl
os
ed
.
The
c
overa
ge
of
fiel
d
with
t
he
reso
l
ution
is
interc
onnecte
d
with
the
exte
nt
f
or
the
re
presentat
ion
.
S
pectral
r
esolutio
n
is
c
on
si
der
e
d
a
s
t
he
m
agn
it
ude
with
the
sta
ti
sti
c
of
pr
eci
se
i
nter
va
ls
of
wa
velen
gt
h
withi
n
the
el
ect
ro
m
agn
et
ic
sp
ect
r
um
fo
r
w
hich
a
se
ns
or
i
s
vulne
rab
le
.
Du
e
to
restrict
ed
ba
nd
widths
in
se
ve
ral
dom
ai
ns
of
the
el
ect
r
om
a
gn
et
ic
s
pectr
um
,
nu
m
ero
us
featur
e
s
a
re
re
adily
sp
eci
fied
.
Im
a
ge
s
pectr
os
c
opy
or
hy
per
-
s
pe
ct
ral
inv
e
sti
gation
is
the
proc
ess
of
sca
nnin
g
nu
m
ber
s
of
cl
os
el
y
arr
a
ng
e
d
a
nd
quit
e li
near
s
pec
tral
b
an
ds
[9
-
10
].
T
he
w
ord h
yper
-
sp
ect
ral is
u
se
d
to
re
fer
t
he
s
pectra t
hat
dw
el
l
a
la
rg
e
f
racti
on
of
na
rro
w,
s
hort
s
pace
d
s
pe
ct
ral
ba
nds.
T
o
ta
c
kle
with
t
he
hype
rs
pectr
al
i
m
age
cl
assifi
cat
ion
qu
e
ry, m
any super
vised
and
unsupe
rv
ise
d
a
na
ly
sis has been
appr
oac
hed
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
9
, N
o.
3
,
June
2019
:
17
10
-
1731
1722
2.1.
Rem
ot
e
sensin
g
im
ag
es
The
cl
assifi
cat
i
on of
rem
ote sen
sin
g
im
ages is into
t
hr
ee
p
a
rts as s
how
n
i
n
T
able
1 [11].
Table
1
.
Var
i
ous
Re
m
ote Sensi
ng I
m
ages
I
m
ag
es
Descripti
o
n
Multi
-
sp
ectra
l
I
m
a
g
e
A
m
u
ltisp
ectr
al i
m
ag
e is co
m
p
rised
o
f
sev
eral
b
an
d
s o
f
d
ata.
The b
an
d
s in
t
h
e
m
u
lti
-
sp
ectra
l i
m
ag
e
ar
e l
ess
than
30
.
This
dep
en
d
s o
n
th
e app
licatio
n
and
it
is alwa
y
s less
th
an
30
ban
d
s
Su
p
er
-
sp
ectra
l
I
m
a
g
e
Co
n
te
m
p
o
rary
sat
ellite
sen
so
rs
are
sk
illed
,
co
n
cernin
g
th
e
capt
u
ring
o
f
i
m
ag
es
at
s
ev
eral
h
ig
h
er
wav
elen
g
th
ban
d
s.
Fo
r
i
n
stan
ce,
n
u
m
erou
s
satellites
are
co
m
p
rised
o
f
th
irty
six
sp
ectra
l
b
an
d
s
with
th
e
m
an
ag
e
m
en
t
o
f
th
e
regio
n
s wav
elen
g
th
r
an
g
es f
ro
m
the
n
ear
-
in
f
rar
ed
,
v
isib
le and
sh
o
rtwave i
n
f
rar
ed
to th
e ther
m
a
l inf
rared
.
The b
an
d
ban
d
wid
th
is r
estrict
ed
,
allo
win
g
exact
sp
ectra
l asp
ects o
f
th
e targ
ets to
be acqu
ired v
ia
th
e sen
so
r.
Hy
p
er
-
sp
ectra
l
I
m
ag
e
The
h
y
p
ersp
ectra
l
i
m
ag
e
h
as
h
u
n
d
red
o
r
m
o
re,
ad
j
acent
sp
ectra
l
b
an
d
s
h
av
i
n
g
a
th
r
ee
-
d
i
m
en
sio
n
al
i
m
ag
e cub
e.
The
co
m
p
ressio
n
o
f
d
i
m
en
sio
n
s
is
u
tilized
in
h
y
p
ersp
ect
ral
i
m
ag
e
ex
a
m
in
a
tio
n
f
o
r
d
ecre
asin
g
d
ata
v
o
lu
m
e
an
d
r
ed
u
n
d
an
cy
.
2.2.
Taxon
om
y of
im
ag
e
classific
at
i
on
me
thods
Im
age
cl
assifi
cat
ion
is
cat
eg
ori
zed
i
nto
diff
e
ren
t
cat
e
gories
on
the
basis
of
di
ff
e
ren
t
c
rite
ria’s
[12].
Table
2
is
giv
i
ng an o
ve
rv
ie
w of
m
et
hods
on the
basis
of
char
act
e
risti
cs w
it
h
cl
assifi
e
rs
ex
am
ple.
Table
2
.
Im
age
Cl
assifi
cat
ion
Me
thod Ta
xonom
y
Criteria
Categ
o
ries
Ch
arac
teristics
Clas
sif
iers exa
m
p
l
es
Usag
e of
trainin
g
sa
m
p
l
es
Su
p
ervis
ed
Prov
id
in
g
land
cover
classes
.
More likelih
o
o
d
Ap
p
rop
riate
r
ef
ere
n
ce data is giv
en
a
n
d
can be u
tilized
as
trainin
g
sa
m
p
les.
Less d
istan
ce
ANN
(Ar
tif
icial
N
eu
ral
N
et
wo
rk)
The sig
n
atu
res prod
u
ced are
u
tilized
f
o
r
trainin
g
the class
if
iers
f
o
r
class
if
y
in
g
the sp
ectra
l data in
the
th
e
m
atic
m
ap
.
The classi
f
ier
o
f
th
e decisi
o
n
tr
ee
Un
-
su
p
ervis
ed
The alg
o
rith
m
of
c
lu
sterin
g
is utilized
f
o
r
p
o
rtion
in
g
the
sp
ectra
l i
m
ag
e into
dif
f
erent sp
ectra
l
class
es o
n
the b
asis
of
statis
tical data f
ro
m
th
e i
m
ag
e.
ISODA
TA
K
-
m
e
an
clu
sterin
g
alg
o
rith
m
There
is n
o
us
ag
e
o
f
pri
o
r
class
es d
ef
in
itio
n
.
It
is th
e
resp
o
n
sib
il
ity
of
the an
aly
st
t
o
label and
m
e
rge t
h
e
sp
ectra
l classes
in
d
esired class
es.
Util
iza
ti
o
n
o
f
diff
erent
p
a
ram
e
ters
lik
e M
ean
Vector/C
o
v
ariance
m
a
trix
Para
m
etric
class
if
iers
The ass
u
m
p
tio
n
of the Gaus
sian
d
istr
ib
u
tio
n
.
More
Likelih
o
o
d
The g
en
eration
of
para
m
et
ers fro
m
tr
ain
in
g
sa
m
p
les.
LD
A
(L
in
ea
r
Disc
r
i
m
in
an
t
An
aly
sis
)
The n
o
isy
resu
lts a
re
u
sed
when
the l
an
d
scap
e is co
m
p
l
ex
.
It
is to
u
g
h
to co
m
b
in
e the sp
atial,
co
n
tex
tu
al and
ancillar
y
data
in
to
th
e classif
icatio
n
pro
cess
.
No
n
-
Para
m
etric
class
if
iers
The d
ata is no
t ass
u
m
ed
f
o
r
th
e r
eq
u
i
re
m
en
t.
ANN
(Ar
tif
icial
N
eu
ral
Net
wo
rk)
It
d
o
esn
’t
n
eed s
tat
istical p
ara
m
ete
rs f
o
r
calculatin
g
the c
lass
sep
aration
and
is a
p
p
rop
riate
f
o
r
in
teg
ration
of
non
-
re
m
o
te
sen
sin
g
data in
the class
if
icatio
n
pro
cess
.
Evid
en
tial r
easo
n
in
g
SVM
Decisio
n
tr
ee
class
if
ier
Exp
ert
S
y
ste
m
Pix
el
in
f
rm
a
ti
o
n
ty
p
e
Per
-
p
ix
el
class
if
iers
Sig
n
atu
re
is d
ev
elo
p
ed
by
traditio
n
al
class
if
iers by
in
teg
rating
th
e tr
ain
in
g
set sp
ec
tra
o
f
the p
ix
els v
ia f
eatu
re.
Clas
sif
iers like
m
o
re
lik
elih
o
o
d
Less d
istan
ce
The sig
n
atu
re
h
as t
h
e con
tribu
tio
n
o
f
eac
h
m
at
erial
withi
n
the
p
ix
els o
f
the trainin
g
sets with
o
u
t com
b
in
ed
pix
el pro
b
l
e
m
s.
ANN
SVM
Decisio
n
tr
ee
Su
b
-
p
ix
el
cla
ss
if
iers
The sp
ectra
l
valu
e
f
o
r
th
e pix
el is taken
as linear/ no
n
-
li
n
ear
with
the in
teg
ratio
n
of
m
en
tio
n
ed
pu
re
m
at
erials
with
p
rop
o
rtion
al
m
e
m
b
ersh
ip
.
Fu
zzy
set
classif
ier
s
Sp
ectra
l
m
ix
tu
re
a
n
aly
sis
Su
b
pix
el classif
iers
The o
u
tp
u
t
is sp
ecif
ic
f
o
r
lan
d
c
o
v
er
class
Hard
class
if
icatio
n
Every p
ix
el sh
o
u
ld
be allo
cated to
the
un
iq
u
e class.
More likelih
o
o
d
Less d
istan
ce
The esti
m
a
tio
n
of
the are
a
m
a
y
resu
lt
s in
er
rors
f
ro
m
dat
a of
co
arse spatial
reso
lu
tio
n
becau
se o
f
th
e pro
b
le
m
of
m
ix
e
d
p
ix
els.
ANN
SVM
Decisio
n
-
tree
So
f
t
class
if
icatio
n
A si
m
il
arit
y
deg
re
e is pro
v
id
ed
f
o
r
ev
ery
cl
ass
.
The classi
f
iers of
f
u
zzy
set
The analysis
of
sp
ectra
l
m
ix
tu
re
It
g
iv
es
m
o
re
inf
o
rm
a
tio
n
and
m
o
re
precise
resu
lts.
Su
b
-
p
ix
el classif
iers
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
A revie
w of r
e
mo
te
ly
se
ns
e
d satel
li
te
ima
ge
cl
as
sif
ic
ation
(
Sa
k
sh
i
D
hingr
a
)
1723
Table
2
.
Im
age Cla
ssific
at
ion
Me
thod Ta
xonom
y
(
con
ti
nue
)
Criteria
Categ
o
ries
Ch
arac
teristics
Clas
sif
iers exa
m
p
l
es
Util
iza
ti
o
n
o
f
sp
atial
Inf
o
rm
-
atio
n
Sp
ectra
l
class
if
iers
Usag
e of
sp
ectra
l i
n
f
o
r
m
atio
n
in th
e c
lass
if
icatio
n
of
i
m
ag
es.
More likelih
o
o
d
Less d
istan
ce
Becau
se o
f
hig
h
va
riation
s in
the sp
ectra
l dis
tribu
tio
n
of
si
m
ilar
class, no
isy
classif
icatio
n
m
a
y
r
esu
lt
ANN
Co
n
tex
tu
al
class
if
iers
Usag
e of
sp
atially
n
eig
h
b
o
u
ring
class
if
icatio
n
info
r
m
ati
o
n
Co
rr
ectio
n
of
a
p
o
in
t to p
o
in
t
co
n
tex
t
Clas
sif
ier
o
f
f
requency
dep
en
d
en
t
co
n
tex
tu
als
Sp
ectra
l
-
c
o
n
tex
tu
al
cl
ass
if
iers
Utilizatio
n
of
sp
ectral
as
wella s
sp
atial d
ata.
Integ
ration
of
param
e
tri
c as
well as
non
-
p
ara
m
etric
an
d
con
tex
tu
al
alg
o
rith
m
s
ECH
O.
3.
FEATU
RE E
X
T
R
AC
TI
ON
Feat
ur
es
a
re
t
ho
s
e
thin
gs
w
h
ic
h
disti
nctiv
el
y
def
ine
a
n
obj
ect
,
s
uc
h
a
s
siz
e,
sh
a
pe,
com
po
sit
ion
,
locat
ion
,
et
c.
[
13
]
. W
it
h
the
effor
tl
ess
avail
a
bili
ty
and
adv
a
ncem
ent
in
the
reso
luti
on
of
s
at
el
li
te
i
m
ages,
m
or
e
stud
y
ef
f
or
ts
hav
e
ha
d
bee
n
la
id
in
t
he
recent
ti
m
es
on
e
xtracti
ng
fea
tu
res
fro
m
sat
el
l
it
e
i
m
ages.
These
featu
res
pri
m
arily
com
pr
ise
structu
r
es
an
d
facil
it
ie
s
(i.e
.
bu
il
di
ngs,
fly
ove
rs,
ai
r
ports,
et
c.),
la
nd
use
,
la
nd
c
ov
e
r,
ve
getat
ion
cl
as
sific
at
ion
(i.e
.
agr
ic
ult
ur
e,
f
or
est
ry,
en
vir
on
m
ental
,
m
ineral),
water
bodies
,
well
loca
ti
ons, t
op
s
oil va
riet
ie
s,
c
onduit
s,
pi
pe
li
nes
an
d
m
any
m
or
e [
13]
.
Feat
ur
e
e
xtract
ion
ca
n
be
de
fi
ned
pr
im
aril
y
as
an
im
age
pr
ocessin
g
te
ch
ni
qu
e
w
hic
h
is
us
e
d
to
spot
and
cl
assi
fy
m
utu
al
relat
io
nships
or
m
utu
al
m
eaning
c
on
c
ern
i
ng
t
he
im
a
ge
areas
.
Se
gme
ntati
o
n
m
et
hods
are
e
m
plo
ye
d
to
separ
at
e
the
pr
efer
red
entit
y
from
the
i
m
ag
e
fo
r
de
velo
pi
ng
the
com
puta
ti
on
.
The
fe
at
ur
e
extracti
on
sta
ge
is
intende
d
t
o
acq
uire
a
c
om
pact,
no
n
-
re
dunda
nt
an
d
sig
nificant
dep
ic
ti
on
of
inter
pr
et
at
ion
s
.
It
is
accom
plis
hed
by
el
i
m
ina
t
ing
re
dundant
and
ina
ppr
opri
at
e
m
at
erial
fr
om
the
data.
Th
ese
featur
e
s
ar
e
then
util
iz
ed
by
the
cl
assifi
er
in
or
der
to
cl
assi
fy
the
data.
T
he
c
la
ssifie
r
that
use
s
con
ci
se
a
nd
appr
opriat
e
fea
tures
will
d
el
iver
m
o
re p
recisi
on
a
nd also invo
l
ves
less
m
e
m
or
y,
w
hic
h
is t
he
m
os
t anti
ci
pated a
sp
ect
o
f
a real
-
tim
e
syst
e
m
.
Be
side
s
the
increasin
g
preci
sion,
th
e
featur
e
e
xtra
ct
ion
proce
dur
e
al
so
enh
a
nce
s
the
com
pu
ta
ti
on
al
sp
ee
d
of
the
cl
assifi
er.
T
her
e
are
certai
n
ste
ps
that
are
requi
red
to
be
f
ollo
wed
i
n
orde
r
to
pe
rf
or
m
the
feature
extracti
on
fro
m
an
o
bject
as
def
i
ned b
el
ow
[14].
a.
Im
age p
re
-
proc
essing
b.
Im
age S
egm
entat
ion
Wh
e
n
pre
-
pro
cessi
ng
a
nd
th
e
favoured
se
gm
entat
ion
le
vel
is
create
d,
then
the
capt
ur
e
strat
e
gy
featur
e
will
be
i
m
ple
m
ented
in
se
gm
ents
to
get
feat
ur
es
,
fol
lowed
by
the
i
m
ple
m
entat
ion
of
cl
assifi
cat
io
n
a
nd
po
st
-
pr
ocessin
g
m
et
ho
ds.
In
pre
-
processi
ng
the
im
age,
the
se
ns
ors
re
cord
t
he
im
age
data
of
t
he
s
at
el
li
te
wh
ic
h
per
m
it
s
the
bugs
t
o
ha
ve
pix
el
a
nd
ge
om
et
ry
values.
The
c
orrecti
on
of
the
bu
gs
i
s
us
i
ng
ex
act
m
at
he
m
at
ic
a
l
m
od
el
s
wh
ic
h
can
be
a
sta
ti
st
ic
al
or
sp
eci
fic
m
od
el
.
The
en
ha
nce
m
ent
of
i
m
ages
is
the
a
m
end
m
ent
of
an
im
age
by
al
te
ring
the
pi
xel
value
f
or
t
he
i
m
pr
ovem
e
nt
of
an
im
age
eff
ect
.
It
usua
ll
y
has
a
nu
m
ber
of
te
ch
niqu
es
that
enh
a
nce
the
i
m
age’
s
vis
ual
app
ea
ra
nce
tha
t
would
la
te
r
be
con
si
der
e
d
a
s
bette
r
f
or
the
m
achine
an
d
hu
m
an
analy
sis [
15
]
.
Occasio
nally
i
m
ager
y
ta
ken
from
sat
e
ll
i
te
s
is
le
ss
in
con
trast
and
li
ght
du
e
to
li
m
it
s
of
im
aging
su
bsy
ste
m
s
wit
h
the
li
gh
ti
ng
ci
rcu
m
sta
nces
with
the
capt
ur
ing
of
the
im
a
ge.
The
im
age
s
ha
ve
var
ie
d
no
ise
s
.
W
it
h
the
e
nh
a
ncem
ent
of
som
e
i
m
age,
the
aim
is
to
cap
ture
the
featu
r
es
of
s
om
e
i
mages
f
or
su
cce
ssive
rev
ie
w.
For
i
ns
ta
nce;
e
dge
an
d
c
ontrast
en
han
c
em
ent,
noise
filt
ering
,
ps
e
udoc
olorin
g
a
nd
ex
pa
ns
io
n.
It
is
ne
cessary
f
or
ca
pt
ur
in
g
featur
e
s
li
ke
i
m
age
dis
play
and
im
age
ana
ly
sis.
The
pro
cess
of
e
nhanc
e
m
ent
do
e
s
no
t e
nhan
ce the
data
being in
her
it
ed
. It
r
el
ie
s on som
e i
m
age f
eat
ures
[16]
.
Segm
entat
ion
i
s
kn
own
as
the
m
ai
n
issue
in
i
m
age
pr
ocessi
ng.
It
is
the
pr
oce
du
re
that
gen
erates
the
i
m
age
and
the
obj
ect
s.
T
he
degree
acc
ordi
ng
to
wh
ic
h
t
he
par
t
is
bas
ed
s
olv
es
t
he
issue
m
eans
that
the
segm
entat
ion
has
to
be
sto
pped
w
hile
the
obj
ect
s
of
inte
rest
in
so
m
e
app
li
cat
io
n
are
separ
at
e
f
or
in
sta
nce,
W
it
h
in
de
pe
ndent
use
of
ta
r
geted
ai
r
-
to
-
groun
d
ac
qu
isi
ti
on,
ass
um
ing
our
i
nterest
li
es
in
Determ
ini
ng
th
e
veh
ic
le
s
on
a
r
oad,
the
fir
st
ste
p
is
t
o
se
gm
e
nt
the
r
oad
f
rom
the
i
m
age
a
nd
the
n
se
gm
e
nt
the
c
onte
nts
of
th
e
ro
a
d
t
o po
s
sibl
e v
e
hicle
s.
Im
age thres
ho
l
ding tech
niques
ar
e
util
iz
ed
f
or i
m
age seg
m
entat
ion
.
It
beco
m
es
i
mp
ort
ant
to
f
oc
us
on
the
sta
ge
of
ta
king
th
is
featur
e
with
a
striking
ef
f
ect
on
the
eff
ect
ive
ness
of
the
rec
ogniti
on
syst
em
[1
7]
.
The
sel
ect
ion
of
featu
res
i
s
a
sign
ific
a
nt
factor
for
ach
ie
ving
recog
niti
on
of
hi
gh
pe
rfor
m
ance.
T
he
extra
ct
ion
of
featu
r
es
is
“extract
ing
from
the
raw
data
inform
at
i
on
that
is
m
os
t
su
it
able
fo
r
cl
assi
ficat
ion
pu
rpos
es
wh
il
e
m
ini
m
iz
ing
the
wi
thin
cl
ass
patt
ern
var
ia
bili
ty
an
d
enh
a
ncin
g
the
betwee
n
-
cl
ass
patte
rn
va
ria
bili
ty
”.
Ther
ef
or
e
,
ch
oo
si
ng
an
ap
pro
pr
ia
te
appro
ac
h
to
getti
ng
featur
e
s
acco
r
ding
to
the
in
put
pro
vid
e
d.
F
or
the
se
facto
r
s,
it
wo
ul
d
be
i
m
po
rtant
to
c
heck
var
ie
d
ac
cessi
ble
featur
e
s fo
r gett
ing
feature
s in
a s
pecified
do
m
ai
n
by cove
ring the
vast
pro
bab
il
it
ie
s of the
cases.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
9
, N
o.
3
,
June
2019
:
17
10
-
1731
1724
4.
CLASSIFI
C
A
TION
METH
ODS
Ther
e
are
se
ve
ral
ap
proac
he
s
an
d
m
et
ho
ds
that
are
ass
oc
ia
te
d
with
sat
e
ll
it
e
i
m
age
cl
assifi
cat
ion
.
But
m
os
t
widely
sat
el
l
it
e
i
m
a
ges
are
cl
assi
fied
into
t
wo
m
ai
n
cat
eg
or
ie
s
as
show
n
i
n
t
he
Figure
2
i.e.
Pixel
-
base
d
as
w
el
l
as
O
bj
ect
-
ba
sed.
T
he
pix
e
l
-
base
d
te
ch
ni
qu
e
s
are
furth
er
di
vid
e
d
int
o
uns
uper
vise
d
an
d
su
pe
r
vised
techn
i
qu
e
s [1
8].
A br
ie
f
abo
ut vario
us
classi
fic
at
ion
m
et
ho
ds
are tab
ulate
d
i
n
the
Ta
ble 3
.
Figure
2
.
Sate
ll
it
e i
m
age classi
ficat
ion
tech
ni
qu
e
s
Table
3
.
Cl
assi
ficat
i
on
M
et
hods
Metho
d
s
Descripti
o
n
Pix
el
-
b
ased
class
if
icatio
n
Pix
el
-
b
ased
class
if
icatio
n
s
are
b
ased
o
n
th
e
g
rey
v
alu
e
o
f
p
ix
els
an
d
f
o
r
th
e
class
if
icatio
n
p
u
rpo
se
o
n
ly
sp
ectra
l
in
f
o
r
m
atio
n
is u
sed
.
Thes
e ar
e
con
sid
ered as the least u
n
it
th
at dep
icts so
m
e i
m
a
g
e.
This
techn
iq
u
e utilizes
the statis
tics o
f
r
ef
lectance f
o
r
p
articular
p
ix
els.
It
ass
e
m
b
les
pix
els to
exp
ress lan
d
cov
er
f
eatu
res.
The lan
d
cov
erage
h
as to
be f
o
rested, m
e
trop
o
litan
,
ag
ric
u
ltu
ral
an
d
ano
th
er
f
eatu
res variet
y
.
The classi
f
icat
io
n
of
the p
ix
el is f
u
rth
er
catego
riz
ed
into
un
su
p
ervis
ed
as well
as su
p
ervis
ed
cl
ass
if
icatio
n
.
Un
su
p
ervis
ed
Clas
sif
icatio
n
In th
is, pix
els are
i
n
teg
rated accord
in
g
to th
e r
ef
lectance
pro
p
erties.
Thes
e gro
u
p
s are
i
n
d
iv
id
u
ally
called
as ‘Clu
sters’
.
The analyst
classif
ies th
e varied clu
st
ers to p
rod
u
ce a
n
d
the b
an
d
s to
utilize.
The analyst
r
ecog
n
izes th
e clus
ters
wi
th
the class
es o
f
land
cov
erage classes
.
Later, anal
y
st allo
c
ates sig
n
if
ican
t labels
to th
e clus
ters a
n
d
deliv
ers prop
erly
satellit
e i
m
ag
e.
It
is f
requ
en
tly
the
case in wh
ich
sev
e
ral
clu
sters
stan
d
f
o
r
a dis
tin
ct land
c
o
v
er
class
.
The analyst
integ
ra
tes th
e clus
ters int
o
a
lan
d
cov
er
cate
g
o
ry
.
This
classif
icatio
n
m
e
th
o
d
is us
u
ally
e
m
p
lo
y
ed
while
no
sa
m
p
le locatio
n
s e
x
ist.
K
-
m
e
an
s an
d
I
SO
DATA
ar
e
the tech
n
iq
u
es u
sed
f
o
r
u
n
su
p
ervis
ed
classif
i
catio
n
of
the satellite i
m
ag
es
Su
p
ervis
ed
Clas
sif
icatio
n
In th
e Sup
ervis
ed
classif
icatio
n
,
an
inp
u
t is r
eq
u
ired f
ro
m
the an
al
y
st.
The in
p
u
t of
the an
aly
sts
is t
er
m
ed
as
‘tr
ain
in
g
set’.
The
sa
m
p
le
o
f
trai
n
in
g
is
co
n
si
d
ered
to
b
e
an
i
m
p
o
rt
an
t
asp
ect
o
f
th
e
m
eth
o
d
s
o
f
su
p
ervis
ed
class
if
icatio
n
an
d
t
h
e
accurac
y
o
f
thes
e
m
eth
o
d
s v
astly
o
n
the sa
m
p
l
es e
m
p
lo
y
ed
f
o
r
th
e pu
rpo
se o
f
tr
ain
in
g
.
The classi
f
icatio
n
is bas
ed
on
the sp
ectra
l sig
n
atu
res in t
h
e tr
ain
in
g
set.
Each class is
de
m
o
n
strated
on
the b
asis
of
what it is si
m
i
lar
to
the
m
o
st in
t
h
e sets o
f
tr
ain
in
g
.
Fo
re
m
o
st class
if
ier
s f
o
r
su
p
ervis
ed
classif
icatio
n
ar
e M
i
n
i
m
u
m
-
d
istan
ce,
parallelepip
ed
and
m
a
x
i
m
u
m
likelih
o
o
d
.
The alg
o
rit
h
m
pri
m
a
ril
y
sep
a
rates th
e pix
els f
ro
m
eac
h
o
th
er
b
ased
on
the trainin
g
sa
m
p
les th
at den
o
te a
site o
n
th
e gro
u
n
d
.
Ob
ject
-
Bas
ed
Clas
sif
icatio
n
It
is d
if
f
erent f
ro
m
p
ix
el
-
b
ased
classif
icatio
n
app
roach
as it
wo
rks
on
the g
rou
p
of
pi
x
els in
stead
of
direct pi
x
els.
In
Ob
ject
-
Bas
ed
i
m
a
g
e
class
if
icatio
n
,
so
m
e
i
m
ag
e
is
in
terpreted
n
o
t
o
n
l
y
f
o
r
sin
g
le
p
ix
el
b
u
t
is
also
v
alid
i
n
sig
n
i
f
ican
t
i
m
ag
e ob
jects an
d
th
eir
co
m
m
o
n
r
elat
io
n
sh
ip
s.
Ob
ject
-
b
ased
inf
o
rm
a
tio
n
extraction
i
s relies up
o
n
sp
ectru
m
cha
racte
r
,
g
eo
m
e
tr
y
as
well
as st
ructu
re
ev
id
en
ce.
This
m
eth
o
d
of
f
ers ac
tu
ally
inn
o
v
ativ
e data and
can si
m
p
ly
acc
ess
ib
le.
Ob
ject
-
Bas
ed
Clas
sif
icatio
n
has
two m
a
in
stag
es:
i.
I
m
ag
e
Seg
m
en
tati
o
n
to g
en
erate
a
se
g
m
en
ted
i
m
ag
e
ii.
Clas
sif
icatio
n
of
th
e seg
m
en
ted
i
m
ag
e
.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
A revie
w of r
e
mo
te
ly
se
ns
e
d satel
li
te
ima
ge
cl
as
sif
ic
ation
(
Sa
k
sh
i
D
hingr
a
)
1725
4.1.
Super
vise
d classi
ficat
i
on al
go
ri
th
ms
The
f
ollo
wing
al
gorithm
s
Ta
ble
4
are
t
he
ty
pes
of
super
vi
sed
cl
assifi
cat
ion
te
ch
niques
that
can
be
util
iz
ed
in t
he
r
esearch
wo
rk for
the classi
fic
at
ion
of the
re
m
ote
ly
sen
sed sat
el
li
te
i
m
ages.
Table
4
.
Algori
thm
s o
f
S
up
e
r
vi
sed
Cl
assifi
cat
ion
Alg
o
rith
m
s
Descripti
o
n
Su
p
p
o
rt
Vector
Machin
e (
S
VM
)
It
was f
irst
intro
d
u
ced b
y
Vapn
ik
in 1
9
7
9
,
b
ased
on
the
p
rincip
les o
f
statistical lea
rnin
g
theo
ry
.
SVMs
are a
d
ecent
an
d
straig
h
t
f
o
rwar
d
ex
a
m
p
le
o
f
a
su
p
ervis
ed
class
if
icatio
n
tech
n
iq
u
e
for
rem
o
t
e
sen
so
r
d
ata cla
ss
if
icatio
n
.
It
is
th
e
m
ajo
r
cla
ss
if
icatio
n
m
eth
o
d
th
at
f
o
ll
o
ws
class
if
icatio
n
task
s
wit
h
th
e
erec
tio
n
o
f
h
y
p
erplan
es
in
th
e
m
u
ltid
i
m
en
sio
n
al sp
ace
with
the
iso
late scenario
s o
f
varied class
l
ab
els.
Reg
ressio
n
as
well
as
class
if
icatio
n
task
s
an
d
u
p
h
o
ld
sev
eral
p
ersisten
t
as
well
as
u
n
co
n
d
itio
n
al
SVM
v
ariables.
This
tech
n
iq
u
e
is
su
itab
le
f
o
r
d
istin
g
u
ish
in
g
th
e
p
attern
s
as
well
as
th
e
o
b
jects
th
at
ar
e
u
tiliz
ed
f
o
r
p
ix
el
-
b
ased
as well
as
ob
ject
-
b
ased
classif
ic
atio
n
.
It
can b
e initiat
ed
b
y
des
cribin
g
the sa
m
p
les o
f
tr
ain
in
g
.
The
ai
m
o
f
th
e
trainin
g
is
th
e
d
isco
v
ery
o
f
so
m
e
f
u
n
cti
o
n
wh
ich
d
ef
i
n
es
t
h
e
relative
b
a
m
o
n
g
o
u
tp
u
t
as
well as inp
u
t.
The
su
b
seq
u
en
t
is
SVM
m
eth
o
d
to
f
in
d
th
e
b
est
ex
trica
tio
n
h
y
p
er
-
p
lan
e
b
etween
class
es
b
y
f
o
cu
ss
i
n
g
o
n
th
e r
eady
data p
lac
ed
at
th
e end
of
the class
des
cripto
rs.
The
o
rgan
ized
d
ata
is
k
n
o
wn
as
Su
p
p
o
rt
Vector
wh
ic
h
h
as
a
v
ital
role
in
th
e
learnin
g
m
a
ch
in
e
role
b
ecaus
e they
a
re
th
e tr
ain
in
g
data ele
m
e
n
ts wh
ich
m
a
y
v
ary
the
h
y
p
erplan
e po
sitio
n
if
di
m
in
ish
sed
.
In
th
is
p
rocess
,
th
e
o
p
ti
m
al
h
y
p
erplan
e
is
f
ix
e
d
an
d
effectiv
ely
u
se
f
ewer
p
reparation
d
ata,
so
t
h
e
accuracy
o
f
the class
if
icatio
n
increases with
th
e prepared
data
set
[
1
6
].
This
m
eth
o
d
is
m
o
st
ad
v
an
tag
eo
u
s
f
o
r
re
m
o
t
e
sen
sin
g
i
m
ag
es
an
d
is
m
o
re
a
ccurate
th
an
o
th
er
p
o
p
u
lar
tech
n
iq
u
es su
ch
as
neu
ral
n
etwo
rk cl
ass
if
iers.
K
-
n
eare
st
n
eig
h
b
o
r
(k
-
NN
)
It
is a
wa
y
f
o
r
clas
sif
icatio
n
of
entitie
s o
n
the b
asis
of
th
e adjo
in
in
g
tr
ain
in
g
sp
eci
m
en
s in
f
ea
tu
re
sp
ace.
In
th
is
cla
ss
if
icatio
n
,
o
u
tco
m
e
is
a
c
o
n
sid
ered
as
class
-
m
e
m
b
e
rsh
ip
.
The
o
b
ject
h
as
b
een
class
if
ied
f
ro
m
h
u
g
e
v
o
te
f
ro
m
n
e
arby
n
eig
h
b
o
rs
h
a
v
in
g
th
e
en
tity
wh
i
ch
is
allo
cated
to
class
m
u
tu
al
b
etw
ee
n
th
e
en
tities
k
-
ad
jo
in
in
g
neighbors.
The
p
ri
m
e
in
ten
tio
n
o
f
th
e
k
Ne
arest
Neig
h
b
o
u
rs
(KNN)
p
rocess
is
to
u
se
th
e
d
atab
ase
wh
erein
th
e
d
ata
are
d
iv
id
ed
in
to
a
n
u
m
b
e
r
o
f
re
m
o
t
e
class
es
to
p
redi
ct
th
e
class
if
icatio
n
o
f
n
o
v
el
sa
m
p
le
p
o
in
t.
KN
N
class
if
icatio
n
dis
tribu
tes d
ata into
a
te
st set an
d
tr
ain
in
g
sets
.
Then
th
e
K
n
e
a
rest
trainin
g
set
o
b
ject
s
are
o
rigin
ated
f
o
r
ev
ery
sin
g
le
row
o
f
th
e
test
set
an
d
t
h
e
p
rocess
o
r
task
o
f
class
if
icatio
n
is
p
erf
o
r
m
ed
b
y
p
redo
m
in
an
ce
v
o
te
with
co
n
n
ectio
n
s
wh
ich
can
b
e
b
rok
en
at
an
y
m
o
m
en
t.
The K
-
n
ea
rest neig
h
b
o
r
alg
o
rith
m
(K
NN)
c
an
be su
m
m
arised p
ri
m
a
ril
y
as
per the f
o
llo
win
g
step
s:
(i)
A po
sitiv
e nu
m
b
er
k is stated
,
with
a
n
ew sa
m
p
le
(ii)
The k
ite
m
s
ar
e sel
ected f
ro
m
the d
at
ab
ase th
at is nex
t t
o
the n
ew sa
m
p
le
(iii)
The u
t
m
o
st
m
u
tu
al
classif
icatio
n
of
s
elected en
tries
is deter
m
in
ed
.
(iv)
Res
u
lted
Class
if
ic
atio
n
is of
f
ered to t
h
e new sa
m
p
le
.
Artif
icial Neu
ral
Netwo
rk (
ANN
)
ANN
g
o
t
th
e
ris
e
o
f
p
o
p
u
larity
a
s
an
alternativ
e
t
o
statis
tical
m
eth
o
d
s
f
o
r
Re
m
o
te
Sen
sin
g
Data
class
if
icatio
n
.
The p
o
p
u
larity
is d
ep
en
d
en
t on
the activ
e natu
re
an
d
f
au
lt toleran
ce
.
The
m
o
d
els
o
f
ANN
are
ap
p
licab
le
t
o
all
ty
p
es
o
f
d
ata,
an
d
n
o
m
ath
e
m
ati
cal
relation
sh
ip
ap
p
lies
to
d
ata
an
d
its classes
.
It
g
en
erall
y
has
thr
ee div
erse c
o
m
p
o
n
en
ts with
I
/P l
ay
e
r,
hid
d
en
-
lay
er
with
th
e O/P
la
y
er
.
ANN
h
as
th
e
o
n
e
I/P
lay
e
r,
o
n
e
O/P
lay
er
h
av
in
g
v
aried
n
eu
ron
s
an
d
t
h
ere
is
a
v
ariation
o
f
h
id
d
en
lay
er
with
the v
ariation
o
f
neu
ron
s.
The n
eu
ron
s o
f
I
/P to h
id
d
en
and
hidden
to O/P no
d
e are lik
e via weigh
t.
Initially,
th
e
w
eig
h
ts
are
arbitra
ry
b
u
t
wh
en
ANN
is
tr
ai
n
ed
th
en
each
o
f
th
e
weig
h
ts
are
acc
u
sto
m
ed
an
d
acco
m
o
d
ate to th
e
d
ata.
Bas
ic ele
m
en
ts o
f
NN
ar
e the p
rocess
in
g
no
d
e corr
esp
o
n
d
in
g
to h
u
m
an
brai
n
neu
ron
s.
Each
p
rocess
in
g
n
o
d
e
ad
m
its
an
d
su
m
s
a
set
o
f
in
p
u
t
v
alu
es,
an
d
trans
f
ers
th
is
v
alu
e
th
roug
h
an
activ
atio
n
f
u
n
ctio
n
w
ith
th
e
o
u
tp
u
t
v
alu
e
o
f
th
e
n
o
d
e
th
at
g
en
erate
s
o
n
e
o
f
t
h
e
in
p
u
ts
to
a
n
o
d
e
th
at
ex
ecut
es
th
e
n
ex
t
lay
er
of
ANNs.
5.
RELATE
D
W
ORK
It
is
a
resea
rc
h
fiel
d
in
w
hich
the
people
are
w
orkin
g
on
va
ried
s
or
ti
ng
m
od
el
s,
featu
r
e
extracti
on
,
dim
ension
reducti
on,
an
d
accuracy.
A
ta
bula
r
com
par
is
on
of
the
pre
vi
ou
sly
carrie
d
researc
hers
ha
s
been
pr
ese
nted
in
t
hi
s secti
on
of th
e p
a
per
as s
h
own
in
Table
5
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
9
, N
o.
3
,
June
2019
:
17
10
-
1731
1726
Table
5
.
A Gla
nce at E
xisti
ng
Tech
niques
Au
th
o
r
&
y
ear
of
stu
d
y
Metho
d
us
ed
in
stu
d
y
Prop
o
sed
work
Co
n
clu
d
in
g
R
em
a
r
k
s
Yo
n
g
g
u
a
n
Xiao
et
al
,
(20
0
4
)
[
1
]
NDVI
(Nor
m
a
lized
Dif
f
erence
Veg
etatio
n
Ind
ex
)
Edg
e thin
n
in
g
an
d
div
isio
n
alg
o
rith
m
PC
-
d
ep
en
d
en
t incre
m
en
tal
sy
ste
m
h
as b
een p
rop
o
sed
f
o
r
b
u
ild
in
g
extraction
as well
as r
o
ad
ext
raction
task
s f
o
r
a
si
m
ilar
pack
ag
e an
d
u
tilizes “
IKONO
S stereo p
air
” f
o
r
g
en
erating
3
-
D cit
y
m
o
d
el
.
The aut
h
o
r
h
as in
v
estig
ated
the
u
sag
e of
so
m
e i
m
a
g
e set in wh
ich
th
e ob
liq
u
e sig
h
t o
f
the site is u
sed
f
o
r
th
e extraction
of
bu
ild
in
g
s.
Tr
ad
itio
n
al
-
N
DVI
has
been
us
ed
f
o
r
lo
catin
g
the v
eg
etat
e
area
s f
o
r
u
tilizin
g
f
o
r
m
ask
in
g
as well
as f
o
r
ap
p
ly
in
g
the
Can
n
y
op
erator
f
o
r
p
an
ch
ro
m
atic i
m
a
g
es f
o
r
th
e pu
rpo
se
o
f
edg
e detectio
n
.
Fo
r
th
e enh
an
ce
m
e
n
t of
the d
etected ed
g
es, edg
e thin
n
in
g
an
d
div
isio
n
alg
o
rith
m
is us
ed
.
Fo
r
h
an
d
lin
g
bu
ild
in
g
o
f
varied si
zes, a
m
u
lti
-
lev
el
ap
p
roach
has
been
us
ed
f
o
r
m
ak
in
g
3
-
D bu
ild
in
g
m
o
d
el
m
o
re
ef
f
ectiv
ely
.
Gio
rgo
s
Mou
n
trakis
et al,
(20
1
1
)
[
2
]
SVM
(Sup
p
o
rt
Vector
m
a
ch
in
e)
I
m
p
le
m
en
tatio
n
of
r
e
m
o
te s
en
sin
g
has
been
r
ev
iewed b
y
SVMs which
is
con
sid
ered as
m
ach
in
e lear
n
in
g
m
e
th
o
d
o
lo
g
y
.
The SV
M
class
if
ie
rs,
self
-
ad
ap
tab
ility
ch
ara
c
terization
,
f
ewe
r
requ
ire
m
en
ts
with
swif
t lear
n
in
g
p
ace
o
n
tr
ain
in
g
size has b
een
p
rov
ed
as the con
sis
ten
t
m
e
th
o
d
o
lo
g
y
f
o
r
in
tellig
en
t
p
rocess
in
g
f
o
r
th
e data taken
f
ro
m
re
m
o
te s
en
sin
g
.
An
e
m
p
iric
al stu
d
y
has
been
pro
v
id
ed
with th
e varied
ap
p
licatio
n
s.
D.
L
U and
Q.
W
E
,
(20
0
7
)
[
3
]
No
n
-
p
ara
m
et
ric
class
if
iers
The p
rob
le
m
s,
cu
rr
en
t practices of
the
i
m
ag
e
class
if
icati
o
n
co
n
cept
hav
e been
exa
m
in
ed
.
It
h
as b
een s
u
g
g
ested
to d
esig
n
an
accu
rate
i
m
ag
e pro
cess
in
g
pro
cess
as a pre
-
requ
isite f
o
r
th
e classif
y
in
g
su
ccessf
u
lly
the re
m
o
t
ely
sen
sed
d
ata in th
e the
m
at
i
c
m
ap
.
The sig
n
if
ican
t issu
es th
at aff
ect the classif
icatio
n
p
erfo
r
m
an
ce a
re
d
i
scu
ss
ed
.
Grae
m
e
G.
W
ilk
in
so
n
(20
0
4
)
[
4
]
Neu
ral
n
et
wo
rk
No
n
-
n
eu
ral
ap
p
roach
es
The stu
d
y
of
f
if
teen
y
ears
of
pu
b
lish
e
d
peer
-
revie
wed
stu
d
y
on
the class
if
icatio
n
of
satellite i
m
ag
e has
be
en
carr
ied
ou
t.
It
h
as b
een co
n
clu
d
ed
that n
o
d
e
m
o
n
strativ
e
i
m
p
rov
e
m
en
t has
b
een s
een f
ro
m
the
last 15
y
ea
rs.
This
st
u
d
y
has
dea
lt with ass
ess
in
g
the p
rog
ress d
eg
ree
i
n
th
e
m
atic
m
ap
p
in
g
via class
if
icatio
n
algo
rith
m
d
ev
elo
p
m
en
t.
The Kapp
a coef
f
icien
t valu
e of
arou
n
d
0.6
5
5
7
having
a
stan
d
ard
d
ev
iatio
n
of
0.1
9
8
0
.
5
0
0
r
ep
o
rted classi
f
icatio
n
exp
eri
m
en
ts h
av
e been
q
u
an
tit
ativ
ely
anal
y
zed.
The co
m
p
ariso
n
ha
s b
een d
rawn f
o
r
N
N and
no
n
-
n
eu
ral
ap
p
roach
es.
Mano
j
Pan
d
y
a et
al,
(20
1
3
)
[
5
]
K
-
M
EANS
,
ISO
D
ata
Clas
sif
ier,
Mini
m
u
m
Distan
ce
Clas
sif
ier,
Parallelepip
ed
Clas
sif
ier,
Maxi
m
u
m
Likelih
o
o
d
Clas
sif
ier,
Seed
ed
Reg
io
n
Growin
g
Techn
iq
u
e,
Enh
an
ced
Seed
ed
Reg
io
n
Growin
g
The co
m
p
ariso
n
of
dif
f
erent classif
icatio
n
techn
iq
u
es for
satellite d
ata has b
een p
rop
o
sed
.
It
h
as b
een o
b
serv
ed
that th
e
d
if
f
erent techn
iq
u
es o
f
class
if
icatio
n
f
o
r
th
e satellite i
m
ag
e
h
av
e div
erse a
ccur
acy
.
The
h
y
b
rid
m
eth
o
d
s o
f
Mahalan
o
b
is with
the SRG
tech
n
iq
u
es h
av
e r
esu
lted
in
m
o
re
accurac
y
.
The techn
iq
u
es f
o
r
th
e un
su
p
ervis
ed
class
if
icatio
n
like
ISO
D
ATA
(I
te
rati
v
e Self
-
Organ
izin
g
),
K
-
m
e
an
s an
d
f
o
r
su
p
ervis
ed
classif
i
catio
n
like Maxi
m
u
m
and
m
in
i
m
u
m
lik
eli
h
o
o
d
classif
ie
r,
para
llelep
ip
ed
,
e
n
h
an
ced an
d
seed
e
d
regio
n
gro
win
g
techn
iq
u
e
Pu
sh
p
en
d
ra
Sin
g
h
Siso
d
ia
et al,
(20
1
4
)
[
6
]
Mahalan
o
b
is,
MLC
,
Parallelepip
ed
class
if
icatio
n
,
m
in
i
m
u
m
d
istan
ce
Land
sat MSS
,
T
M
an
d
E
TM
+
i
m
ag
es o
f
19
7
2
,
1
9
9
8
and
2
0
1
3
ar
e utilize
d
f
o
r
th
e gen
eration
of
L
an
d
u
se
class
if
icatio
n
with Mahalan
o
b
is, M
L
C (M
ax
i
m
u
m
Likelih
o
o
d
classif
i
catio
n
);
Paral
lelep
i
p
ed
su
p
ervis
ed
class
if
icatio
n
and
m
in
i
m
u
m
dis
tan
ce.
ML
C g
iv
es
m
o
st
a
ccurate
resu
lts
f
o
r
re
m
o
tel
y
sen
se
d
i
m
ag
es
I
m
ag
e
en
h
an
ce
m
e
n
t and
ge
o
m
et
ric
c
o
rr
ectio
n
techn
iq
u
es
h
av
e been
assig
n
ed
f
o
r
en
h
an
ced trainin
g
sig
n
atu
re
for
trainin
g
the class
if
iers.
Er
d
as I
m
ag
in
e 9.1
So
f
tware
is u
tilize
d
f
o
r
th
e su
p
ervis
ed
class
if
icatio
n
.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
A revie
w of r
e
mo
te
ly
se
ns
e
d satel
li
te
ima
ge
cl
as
sif
ic
ation
(
Sa
k
sh
i
D
hingr
a
)
1727
Table
5
.
A Gla
nce at E
xisti
ng
Tech
niques
(
c
on
ti
nue
)
Au
th
o
r
&
y
ear
of
stu
d
y
Metho
d
us
ed
in
stu
d
y
Prop
o
sed
work
Co
n
clu
d
in
g
R
e
m
a
r
k
s
Nu
rAnis
Mah
m
o
n
et al,
(20
1
4
)
[
7
]
ANN
Maxi
m
u
m
Likelih
o
o
d
(M
L
)
ISODA
TA
This
r
esea
rch h
as
d
ealt with th
e su
rv
ey
of
ANN
capab
ility
f
o
r
class
if
y
in
g
the satellite
i
m
ag
e with
d
if
f
erent ANN
ap
p
roac
h
es.
The accura
cy
ca
m
e ou
t to b
e 89
.3%
A r
ev
iew
of
dif
f
er
en
t techn
iq
u
es f
o
r
th
e satellite i
m
ag
e
class
if
icatio
n
of
th
e satellite
i
m
ag
es h
as b
een co
m
p
ared
with
the ANN
on
the Accurac
y
bas
is.
So
u
m
ad
ip
Gh
o
sh
et al,
(20
1
4
)
[
8
]
ML
P B
PNN;
Su
p
p
o
rt
Vector
Mach
in
e
(SVM
)
K
-
NN
The wo
rk is d
ep
en
d
en
t on
data classif
icatio
n
con
cept
.
ML
P B
PNN
Prov
id
es
m
o
st
acc
u
ra
te
resu
lts with
an ac
c
u
racy
of
89
.3%
,
Initially,
a
m
o
d
el/c
lass
if
ier
h
as b
een d
ep
icted
by
m
eans
o
f
pred
ef
in
e
d
classes
,
it is kn
o
wn
as t
h
e tr
ain
in
g
ph
ase in
wh
ich
the class
if
icatio
n
techn
iq
u
e ha
s d
ev
elo
p
ed
the
class
if
ier
b
y
le
arni
n
g
f
ro
m
the d
atas
e
t of
tr
ain
in
g
and
f
ro
m
its su
b
seq
u
en
t class lab
el.
In th
e su
b
seq
u
en
t
p
h
ase (te
stin
g
ph
ase) has esti
m
ated
accurac
y
f
o
r
class
i
f
icatio
n
f
o
r
th
e classif
icatio
n
m
o
d
el
b
y
u
tilizin
g
the test set f
o
r
th
e tr
ain
in
g
datas
et.
Pre
-
p
rocess
in
g
tec
h
n
iq
u
es su
c
h
as data selectio
n
,
d
ata
cleanin
g
and
data transf
o
r
m
a
tio
n
ar
e
p
rop
o
sed
to th
e
u
n
iq
u
e data bef
o
re
the p
rocess
of
classif
icatio
n
f
o
r
th
e
i
m
p
rov
e
m
en
t of
cl
ass
if
icatio
n
m
eth
o
d
acc
u
rac
y
.
Mahyat
Sh
af
ap
o
u
r
Tehrani et
al,
(20
1
3
)
[
9
]
Su
p
p
o
rt
Vector
Machin
es
,
Decisio
n
Tr
ees
The d
e
m
o
n
stratio
n
of
varied
m
eth
o
d
s with
sp
ectra
l
as
well as sp
atial
attr
i
b
u
tes f
o
r
th
e extraction
of
L
ULC of
th
e inn
er
-
city
ar
ea,
so
il,
wate
r
b
o
d
ies an
d
dif
f
erent pl
an
ts
h
as tak
en
place.
The classi
f
icatio
n
acc
u
rac
y
ca
m
e
ou
t
to
be h
ig
h
er
f
o
r
K
-
NN classif
ier
o
f
90
an
d
91
% f
o
r
2
0
0
3
an
d
20
1
0
co
rr
esp
o
n
d
in
g
ly
.
Kap
p
a coef
f
icien
t is 0.8
7
f
o
r
th
e data.
The SPO
T
5
i
m
ag
es as o
f
20
0
3
to 2
0
1
0
hav
e been
taken
f
o
r
Ma
lay
si
a and
S
elan
g
o
r
with
the cl
ass
if
icatio
n
has
b
een carr
ied
ou
t us
in
g
KNN,
S
VM
and
DT
classif
iers
.
Less accura
cy
of
DT
classif
ier
has
6
8
.06
% acc
u
racy
with
the k
ap
p
a
co
ef
f
icien
t o
f
0.6
f
o
r
2
0
1
0
.
SVM
c
lass
if
icatio
n
acc
u
rac
y
r
an
g
es
f
ro
m
DT
as wel
l as
ob
ject
-
o
riented
class
if
ier
8
0
.06
% h
av
in
g
Kapp
a
co
ef
f
icien
t f
o
r
0
.75
in 2
0
0
3
with th
e
7
8
.15
% has
kap
p
a
coef
f
icien
t f
o
r
0
.72
in 2
0
1
0
.
Su
n
ith
a
Ab
b
u
ru an
d
Su
resh
Bab
u
Go
lla (
2
0
1
5
)
[
1
0
]
Clas
sif
icatio
n
M
eth
o
d
f
o
r
satellite
i
m
ag
e
class
if
icatio
n
A su
m
m
a
r
y
of
auto
m
ated
satellite
i
m
ag
e classif
icatio
n
m
e
th
o
d
s h
as b
een p
rov
id
ed
with a com
p
ariso
n
of
d
if
f
erent r
ev
iews.
The re
v
iew
assis
ts
th
e r
esear
ch
ers
with
the selectio
n
o
f
su
itab
le satellite
i
m
ag
e classif
icatio
n
techn
iq
u
e or
m
e
th
o
d
s o
n
the b
asis
of
r
eq
u
ire
m
en
ts
.
An
ev
alu
atio
n
of
p
erfo
r
m
an
ce
with
the v
aried datas
ets
h
as b
een p
rov
id
ed
with
the d
if
f
erent datas
ets.
Ro
y
a Abed
i
et al,
(20
1
5
)
[
1
1
]
k
-
Near
est
Neig
h
b
o
u
r
(k
-
NN)
The stu
d
y
has
inv
estig
ated
the p
o
ss
ib
i
lity
of
KNN
tech
n
iq
u
e f
o
r
estim
a
tin
g
m
a
jo
r
f
o
rest attr
ib
u
tes
lik
e
h
eig
h
t,
d
en
sity,
ba
sal are
a
,
tr
ees DB
H and
f
o
rest cov
er
ty
p
e on
the b
asis
of
the ex
istin
g
f
ield
o
f
f
o
rest inv
en
to
ry
f
o
r
h
av
in
g
a
m
o
re
accurate
class
if
icatio
n
.
KNN prov
id
es acc
u
rate
class
if
icatio
n
an
d
f
easib
le the
m
a
tic
m
ap
s
Bas
h
ir
Ah
m
ed
et al,
(20
1
5
)
[
1
2
]
Back
Prop
ag
atio
n
Neu
ral
Netwo
rk
(BPNN)
NN has
been
utiliz
ed
f
o
r
so
lv
in
g
m
o
r
e variet
y
of
task
s f
o
r
so
lv
in
g
ord
i
n
ary
r
u
le
-
b
ased
pro
g
ra
m
m
in
g
with
sp
eech
recog
n
itio
n
and
co
m
p
u
te
r
v
isio
n
.
BPNN is self
-
ad
ap
tiv
e dyna
m
ic
syste
m
an
d
bes
t f
o
r
th
e classif
icatio
n
satellite i
m
ag
es th
a
n
oth
er
tech
n
iq
u
es
The p
rocess
in
g
no
d
e has
r
eceived
an
d
calculates an
inp
u
t
v
alu
e set with tran
sf
ers the su
m
via a
ctiv
atio
n
of
the
f
u
n
ctio
n
f
o
r
g
iv
in
g
the n
o
d
e ou
tp
u
t va
lu
e that d
ev
elo
p
s
o
n
e of
the in
p
u
ts to
wards p
rocess
in
g
no
d
e in
th
e
su
b
seq
u
e
n
t lay
er
.
An
an
d
Up
ad
h
y
ay
et
al,(
2
0
1
5
)[
1
3
]
Feed
-
Fo
rwar
d
Neu
ral
Netwo
rk
(FFNN
)
In class
if
icatio
n
,
L
ISS
-
I
II
satellite
i
m
ag
e,
th
e varied
b
an
d
s are
g
ath
ered and
RGB h
as b
ee
n
gen
erate
d
.
The classi
f
icatio
n
acc
u
rac
y
of
ANN
d
ecre
ases
with th
e
in
crea
se in
the
n
u
m
b
er
of
neu
ron
s
.
FFNN and
back
p
r
o
p
ag
atio
n
o
f
er
ror
h
av
e been
des
ig
n
e
d
with
ANN
too
lb
o
x
.
The re
su
lt has
bee
n
calculated
acc
o
r
d
in
g
to th
e kap
p
a
co
ef
f
icien
t and
the
acc
u
rac
y
.
Nik
ita
Ag
g
arwa
l
et al,
(20
1
6
)
[
1
4
]
Pix
el
-
Bas
ed
Clas
sif
icatio
n
(
PBC
)
an
d
Ob
ject
-
Bas
ed
Clas
sif
icatio
n
(
OBC
)
The stu
d
y
has
f
o
cu
ss
ed
on
classif
icatio
n
app
roach
ed
on
th
e bas
is o
f
ob
ject
-
o
riented
and
pix
el orien
ted
.
OBC
pro
v
id
es
m
o
r
e Opti
m
al
and
ef
f
icien
t Class
if
ica
tio
n
than
PBC.
Ob
ject
-
b
ased
has
s
p
ectra
l,
sp
atial
and
te
m
p
o
ral
p
rop
ertie
s th
at can be
u
tilized
f
o
r
h
ig
h
r
eso
lu
tio
n
r
e
m
o
tel
y
sen
sed
data f
o
r
o
v
erco
m
in
g
the p
ix
el
-
b
ased
classif
icatio
n
li
m
itatio
n
s o
n
the
b
asis
of
the sp
ectra
l data v
alu
e.
It
h
as b
een co
n
clu
d
ed
that
class
if
icatio
n
of
o
b
ject
-
b
ased
tech
n
iq
u
e is
ex
act as c
o
m
p
ared to
p
ix
el
-
b
ased
classif
icatio
n
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
9
, N
o.
3
,
June
2019
:
17
10
-
1731
1728
Table
5
.
A Gla
nce at E
xisti
ng
Tech
niques
(
c
on
ti
nue
)
Au
th
o
r
&
y
ear
of
stu
d
y
Metho
d
us
ed
in
stu
d
y
Prop
o
sed
work
Co
n
clu
d
in
g
R
e
m
a
r
k
s
Say
ali
M
.
Jo
g
et al,
(20
1
6
)
[
1
5
]
FCM
clu
sterin
g
with
bo
th
m
u
lti
-
class
SVM
an
d
KNN
Ad
ap
tiv
e his
to
g
ram
eq
u
alizatio
n
has
been
utilized
f
o
r
th
e enh
an
ce
m
en
t o
f
the i
m
ag
e con
tras
t.
K
-
m
e
an
s with
the
Fu
zzy
C
-
m
e
an
s cl
u
sterin
g
techn
iq
u
e
s
b
ein
g
utilized
f
o
r
o
b
tain
in
g
the seg
m
e
n
ted
i
m
ag
es.
The p
rocess
of
cla
ss
if
icatio
n
has
bee
n
car
ried o
u
t
by
KNN and
SV
M.
FCM with
m
u
lti
-
cl
ass
SVM
sh
o
ws
b
etter
resu
lts th
an
KNN
Kad
irSaban
ci
Et
al,
(20
1
6
)
[
1
6
]
Multilaye
r
Perceptron
(M
L
P)
k
-
NN
J4
8
,
Naiv
e Bay
es
Bayes
,
Net
KStar
ASTE
R
in a f
o
rest
lan
d
of
al
m
o
st 1
3
X1
2
k
m
of
Jap
an
h
as b
een u
sed
.
ML
P
an
d
KNN
has
m
u
ch
better
accurac
y
than
oth
ers and
acc
u
rac
y
rate
is 9
0
.4%
and
8
9
.10
%,
resp
ectiv
ely
Data prod
u
ced f
o
rm
UC
I
f
o
rest t
y
p
e
m
a
p
p
in
g
data set
h
as two
ph
ases
,
lik
e tr
ain
in
g
as well
a
s testin
g
.
3
8
% of
the d
ata is f
o
r
trainin
g
and
62
% o
f
data is f
o
r
testin
g.
W
ith
the t
rainin
g
data, su
g
i and
hin
o
k
i average sp
ectr
al
v
alu
es h
av
e been
ob
tain
ed
W
E
KA so
f
tware
has
been
us
ed
f
o
r
th
e classif
icatio
n
of
th
ree
v
aried f
o
rest
ty
p
es, na
m
el
y
,
H
in
o
k
i,
Su
g
i and
m
ix
ed
bro
ad
leaf
.
Go
n
g
Ch
en
g
et al,
(20
1
7
)
[
1
7
]
Han
d
cr
af
ted
-
Featu
re
-
Bas
ed
Metho
d
s
,
Un
su
p
ervis
ed
Featu
re
Lear
n
in
g
-
Bas
ed
Metho
d
s
,
Deep
-
Featu
re
-
Lear
n
in
g
-
Bas
ed
Metho
d
s
Initially,
an in
clu
siv
e evalu
atio
n
of
cu
rr
en
t work
of
re
m
o
te s
en
sin
g
app
roach
es h
as b
een g
iv
en
.
The re
sea
rch can h
elp
the resea
rchers
to
ado
p
t
m
o
r
e inf
lu
en
tial and
m
u
lti
-
v
iew f
eatu
re
rep
res
en
tatio
n
.
Dataset “
NW
PU
-
R
ESI
SC4
5
” has
bee
n
pro
p
o
sed
that
co
n
sis
ted
of
3
1
5
0
0
i
m
ag
es
with
f
o
u
rty
f
iv
e scene
class
es an
d
sev
en
hu
n
d
red i
m
ag
es in
ev
ery
class
.
Dif
f
erent r
ep
resent
ativ
e techn
iq
u
es h
av
e been
exec
u
ted
with
the d
ataset.
Harish
Ku
n
d
ra
an
d
Harsh
Sad
awar
ti
,
(20
1
5
)
[
1
8
]
Cu
ck
o
o
Search
a
nd
Particle
Swar
m
Op
ti
m
izatio
n
CS an
d
PSO
h
as b
een p
resented
f
o
r
r
e
m
o
te sen
sin
g
f
o
r
th
e better e
f
f
ectiv
en
ess
and
op
ti
m
izat
io
n
valu
e.
Better perf
o
r
m
an
c
e of
the p
rop
o
sed
w
o
rk f
o
r
su
b
stitu
ti
n
g
the search
m
e
th
o
d
of
CS f
o
r
d
isco
v
ering
bes
t
h
o
st n
est h
av
in
g
gr
eatest PSO
alg
o
rith
m
po
sitio
n
.
The classi
f
icatio
n
of
an i
m
ag
e
in v
ari
ed
ter
rain f
eatu
res
is th
e
m
ain
ai
m
of
th
e r
esear
ch
.
The wo
rks
hav
e eff
ectiv
ely
extract
ed
th
e
terr
ain
f
eatu
res
f
ro
m
the p
rov
id
ed
d
ataset and
also
su
stain
m
o
re
class
if
icatio
n
acc
u
racy level.
The classi
f
icatio
n
ef
f
icien
cy
has
been
co
m
p
a
red with
th
e pro
p
o
sed
algo
r
ith
m
and
with th
e s
war
m
intellig
en
ce
an
d
so
f
t co
m
p
u
tin
g
techn
iq
u
es.
6.
E
X
ISTI
NG W
ORK
A
N
ALY
SIS
This
sect
io
n
de
fines
t
he
wor
k
of
Ha
rish
K
undra
a
nd
Ha
r
sh
Sa
da
war
ti
[
17
]
e
xp
la
i
ning
the
sat
el
li
te
i
m
age
cl
assifi
cat
ion
us
in
g
natu
re
-
in
sp
ire
d
cl
assifi
cat
ion
al
gorithm
,
nam
el
y,
PSO
and
Cuc
koo
Searc
h.
The
aut
hors
ha
ve
pr
e
sente
d
the
hybri
dizat
io
n
of
op
ti
m
iz
ati
on
al
gorit
hm
s
su
c
h
as
PSO
a
nd
Cuc
koo
Se
arch
to
reduce
t
he
cl
a
ssific
at
ion
e
rro
r
rate
an
d
t
o
i
m
pr
ov
e
t
he
ac
cur
acy
base
d
on
ob
j
ect
ive
f
un
ct
i
on.
T
he
F
igure
3
disp
la
yi
ng
is
a
sat
el
li
te
i
m
a
ge
of
Alwa
r
r
egio
n
in
I
nd
ia
.
This
re
gion
is
sel
ect
ed
be
cause
it
c
ar
ries
so
m
e
su
pe
rio
r
la
nd c
ov
e
r feat
ures li
ke Ur
ba
n,
Ba
rren, Wat
er
, Veg
et
at
ion
, a
nd Ro
cky areas
.
Figure
3
.
Com
par
is
on of t
he Ori
gin
al
im
age (
A
)
a
nd classi
f
ie
d
im
age
(B)
a
nd
U
sin
g sat
el
li
te
i
m
age classi
ficat
ion
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
A revie
w of r
e
mo
te
ly
se
ns
e
d satel
li
te
ima
ge
cl
as
sif
ic
ation
(
Sa
k
sh
i
D
hingr
a
)
1729
The
process
of
sat
el
li
te
im
age
cl
assifi
ca
ti
on
fall
s
under
t
wo
proc
ess
nam
el
y
tr
ai
nin
g
a
nd
cl
assifi
cat
ion
. T
he
trai
ning
process is
fur
t
her cat
egorized
in pre
processi
ng
and opti
m
iz
at
i
on of t
he data
set
.
Algori
th
m
1.
Data Prepr
oc
essing
(Raw
D
ata)
1.
Fo
r
eac
h d in t
he data
ro
w.Da
ta
set
(Raw_Dat
a)
2.
D1
=
Datase
t(
d).all
colum
ns
()
;
//
Taking
al
l
c
olu
m
n
data
as
input
ra
w
data
,
now
it
has
t
o
be
decide
d
t
hat
wh
ic
h
3.
Fo
r
eac
h
el
em
e
nt in D
1
data ha
s to be
us
e
d
a
nd which
not
4.
Popu
la
ti
on_Si
ze=D1
.elem
ent co
unt
5.
Apply o
ptim
al
so
luti
on(
D
1)
6.
Fit
_V
al
ue=
[];
7.
Count=
1;
8.
Fo
r
eac
h D1 e
lem
ent
9.
CurrentEl
em
ent=D1
;
10.
Thr
e
shold=l
am
da;
11.
L=A
pp
ly
Fit
nes
(Thres
ho
l
d,
C
urren
tEl
em
ent);
12.
If
L==1
13.
Fit
_V
al
ue
(
D1.
Value,c
ount
)=
Current;
14.
Count=
Co
un
t+
1;
15.
En
d
16.
En
d Alg
or
it
hm
Algorithm
1
descr
ibes
the
pr
ocessin
g
of
ra
w
data
in
wh
i
ch
it
util
iz
es
a
fitness
functi
on
to
dr
op
or
sel
ect
data
for
processi
ng.
T
he
sel
ect
ion
of
the
th
res
ho
l
d
is
done
on
t
he
basis
of
the
data
pas
sed
t
o
the
op
ti
m
iz
ation
f
un
ct
io
n.
I
f
t
he
data
sat
isfie
s
the
fitness
functi
on
the
n
it
is
sel
ect
ed
el
se
it
is
dr
oppe
d.
The
sel
ect
io
n
of
t
he
fitness
f
un
ct
io
n
is
de
pe
nd
e
nt
on
data
set
util
iz
ed.
T
he
trai
ning
m
e
chan
ism
inclu
des
the
util
iz
at
ion
of
tr
ai
nin
g
functi
on
wh
ic
h
ta
kes
th
e
op
ti
m
iz
ed
set
as
inp
ut
an
d
t
hen
trai
ns
the d
at
a
accor
ding to
the
trai
ning str
uctu
re.
T
he st
ru
ct
ure o
f
the
traini
ng alg
ori
thm
is as foll
o
w
s.
Figure
4
.
Trai
ni
ng
c
onfi
gurati
on for sat
el
li
te
i
m
age
The
tr
ai
ne
d
a
rch
it
ect
ure
set
is
the
n
se
nt
f
or
t
he
cl
ass
ific
at
ion
.
The
cl
assifi
cat
ion
proces
s
is
com
plete
ly
based
on
the
gro
und
tr
uth
val
ue
s.
The
res
ult
will
be
the
m
a
tc
he
d
with
the
gr
ou
nd
tr
ut
h
values
.
The
cl
assifi
cat
i
on alg
or
it
hm
is as foll
ows.
Algori
th
m
2.
Clas
sific
at
i
on
St
ruc
tu
re
(O
p
timi
z
ed St
ruct
ure, Gr
ound
Tr
ut
h
V
alues)
1.
Fo
r
j
=1:
Op
ti
m
i
zedStr
uctu
re
2.
Cl
assifi
ed_
St
r
uctu
re=Cl
assif
y ( Opti
m
iz
ed
Stru
ct
ur
e,
Gr
ound
_
Tr
uth
_
Va
lue )
;
3.
Df
=
Cl
assifi
ed
_S
t
r
uctu
re
–
G
rou
nd_T
ru
t
h_Value;
4.
En
d
5.
Fo
r
eac
h df
i
n Diff val
ues
//
T
akin
g
the
d
i
ff
e
ren
ce
of the
cla
ssifie
d
st
ru
ct
ure
Evaluation Warning : The document was created with Spire.PDF for Python.