Indonesian J
ournal of Ele
c
trical Engin
eering and
Computer Sci
e
nce
Vol. 2, No. 3,
Jun
e
201
6, pp. 703 ~ 71
1
DOI: 10.115
9
1
/ijeecs.v2.i3.pp70
3-7
1
1
703
Re
cei
v
ed Ma
rch 1
7
, 2016;
Re
vised April
29, 2016; Accepte
d
May 1
4
, 2016
Image F
u
sion in Hyperspectral Image Classification
using Genetic Algorithm
B Saichand
a
n
a*
1
, K Sriniv
as
2
, R Kiran Kum
a
r
3
1
Departme
n
t of CSE, JN
T
U
Kakin
ada
2
Department of CSE, VR Siddhart
ha Engineering Coll
ege, Vijay
a
w
ada
3
Departme
n
t of CS, Krishna U
n
iversit
y
, Mach
ilip
atnam
*Corres
p
o
ndi
n
g
author, em
ail
:
bschand
an
a
@
gmai
l.com
A
b
st
r
a
ct
Hyper sp
ectral
remote sens
or
s collect i
m
a
g
e
data
for a larg
e nu
mb
er of na
rrow
,
adjacent
spectra
l
ban
ds. Every pixel i
n
hypers
pectral i
m
a
ge
involv
es a co
n
t
inuo
us spectr
um that is use
d
to classify th
e
obj
ects w
i
th great detai
l an
d precis
i
on. T
h
is
pap
er pres
ents
hypersp
ectr
al imag
e
classific
a
tion mec
h
a
n
is
m
usin
g gen
etic a
l
gorit
hm w
i
th e
m
p
i
rica
l mod
e
deco
m
pos
iti
on
and i
m
age fusi
on use
d
in pre
p
rocess
ing sta
g
e
.
2-D Empiric
a
l
mo
de d
e
co
mp
ositio
n met
hod
is used to
remove any n
o
isy compo
nents in
each ba
nd of t
h
e
hypers
pectral
data. After f
ilterin
g, imag
e fusio
n
is p
e
rfor
me
d o
n
the h
y
persp
ectral b
ands to s
e
lecti
v
e
l
y
mer
ge th
e
max
i
mu
m p
o
ssib
l
e
features fro
m
t
he so
urce
i
m
a
ges to for
m
a s
i
ngl
e i
m
ag
e. T
h
is fuse
d i
m
ag
e
i
s
classifie
d
usi
n
g gen
etic alg
o
r
i
thm. Differe
nt indic
e
s,
such a
s
K-mea
n
s (KMI), Davies-Bo
u
ldi
n
Index (D
BI),
and Xi
e-Be
ni Index (XBI) are
used as
o
b
j
e
cti
v
e functions. T
h
is metho
d
incr
eases cl
assific
a
tion acc
u
racy
of
hypers
pectral i
m
a
ge.
Ke
y
w
ords
:
Image C
l
assific
a
ti
on, Empiric
a
l
Mode D
e
co
mp
osit
io
n, Geneti
c
Algorith
m
, Hy
persp
ectral Image
Copy
right
©
2016 In
stitu
t
e o
f
Ad
van
ced
En
g
i
n
eerin
g and
Scien
ce. All
rig
h
t
s reser
ve
d
.
1. Introduc
tion
The pro
c
e
ss of acquirin
g
information
about an obj
ect on the earth usi
ng satellites
without ma
kin
g
any physi
ca
l conta
c
t is ca
lled rem
o
te
sensi
ng [1]. The cla
ssifi
catio
n
of object
s
o
n
the earth by usin
g elect
r
o
m
agneti
c
rad
i
ations refl
ect
ed or emitted
by the surface is the mai
n
goal of remo
te sen
s
ing te
chn
o
logy [2]. Ne
w opp
ort
unities to u
s
e rem
o
te se
nsin
g data h
a
ve
arisen, with the increa
se o
f
spatial and
spe
c
tral
re
sol
u
tion of recen
t
ly launche
d satellite
s
. Image
cla
ssifi
cation is
a key step in
rem
o
te
se
nsin
g
ap
p
lica
t
ions [19]. In
remote
se
nsi
ng, se
nsors
are
available that
can gen
erate hy
perspe
c
t
r
al data, invo
lving many narrow ba
nd
s in which e
a
c
h
pixel ha
s
a
continuo
us refl
ectan
c
e
spe
c
trum. Uns
upe
rvised
ima
ge
cla
ssifi
cation
is a
n
imp
o
rta
n
t
resea
r
ch topi
c in hype
rspe
ctral ima
g
ing,
with t
he aim
to develop eff
i
cient alg
o
rith
ms that provide
high cl
assification accu
ra
cy.
T
h
e
h
y
pe
rs
pe
c
t
r
a
l ima
ges
s
u
ffe
r
fr
om n
o
i
s
e
s d
u
e
to di
sturb
ance of tran
smissio
n
medium
in t
he atmo
sp
he
re o
r
deg
rad
a
tion of
sen
s
ors etc lea
d
i
ng to
affect
the a
c
cura
cy
of
cla
ssifi
cation
algorith
m
s. T
h
is pa
pe
r pre
s
ent
s hy
pe
rspectral imag
e
cla
ssifi
cation
using EM
D a
nd
Image fusio
n
.
2-D Empiri
cal mode de
compo
s
ition
m
e
thod is u
s
e
d
to divide the hyperspe
c
t
r
al
image
belo
n
g
ing to
a
sp
ecific ban
d i
n
to finite
nu
mber of com
pone
nts
calle
d intrin
si
c m
ode
function
s. Th
e last com
p
onent is call
ed a re
sid
u
e
.
The first IMF is filtere
d
usin
g wav
e
lets
shri
nkage
de
noisi
ng m
e
th
od. The
sum
m
ation of filt
ered
IMF a
n
d
re
maini
ng I
M
Fs
plu
s
residue
gives the d
e
-noised ima
g
e
.
The sam
e
p
r
ocedu
re i
s
repeate
d
for a
ll the band
s.
After filtering
the
band
s are fused into
a singl
e image for ap
plic
atio
n ori
ented visu
ali
z
ation, effective
interp
retation,
extra
c
tion of
useful featu
r
es,
and
to p
r
ovide
a b
e
tter d
e
scriptio
n
of the
sce
n
e
usin
g redu
ce
d data
sets.
After fusio
n
, t
he ima
g
e
is
classified
usi
n
g ge
netic alg
o
rithm
with th
ree
different obj
ective functi
ons. Thi
s
m
e
thod in
cre
a
s
e
s
the cla
ssifi
cation a
c
cura
cy both
in
qualitative an
d quantitative
analysi
s
.
This p
ape
r is
st
ru
ct
ure
d
a
s
f
o
llow
s
:
se
ct
i
on 2 p
r
e
s
e
n
ts filtering u
s
ing bi
-dim
e
n
sio
nal
empiri
cal m
o
de de
com
p
o
s
ition, sectio
n
3 pre
s
e
n
ts i
m
age fu
sion
techni
que, se
ction 4 p
r
e
s
e
n
ts
Geneti
c
al
gorithm for i
m
ag
e cl
assificatio
n
, se
ct
ion
5
sho
w
s exp
e
ri
mental
re
sult
s a
nd
se
ction
6
repo
rt co
ncl
u
sion
s.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 25
02-4
752
IJEECS
Vol.
2, No. 3, Jun
e
2016 : 703
– 711
704
2. Empirical
Mode Deco
mposition
Empirical mo
de de
comp
osition [5] is a sign
al pro
c
e
s
sing met
hod
that nonde
structively
fragme
n
ts a
n
y
non-li
nea
r
and n
o
n
-
stati
onary
sig
nal
into oscillato
ry function
s b
y
mean
s of
a
mech
ani
sm
calle
d shiftin
g
pro
c
e
ss.
These osc
ill
atory functio
n
s are call
e
d
Intrinsi
c Mode
Functio
n
s
(IMF), and e
a
ch IMF satisfi
e
s two p
r
o
p
e
r
ties, (a
) the
numbe
r of ze
ro cro
ssi
ng
s and
extrema point
s sho
u
ld be e
qual or differ by
one.
(b
)
S
y
mmetric
env
elope
s (zero mean
)
inte
rp
re
t
by local
maxi
ma and
mini
ma [6]. The
signal after de
comp
ositio
n
usin
g EMD i
s
non
-de
s
tru
c
ti
ve
mean
s that th
e ori
g
inal
sig
nal can b
e
o
b
t
ained
by a
d
d
i
ng the IMF
s
and
re
sidue.
The first IMF
is
a hig
h
frequ
e
n
cy
com
pone
nt and
the
su
bse
que
nt IM
Fs
co
ntain f
r
om n
e
xt high
freq
uen
cy to
the
low freq
uen
cy compon
ent
s. The shiftin
g
pro
c
e
ss [7]
[12] used to obtain IMFs on a 2-D
sig
nal
(imag
e
) is
su
mmari
zed a
s
follow
s
:
a)
Let I(x,y) be
a Remote Se
nsin
g Imag
e
use
d
for EM
D d
e
compo
s
i
t
ion. Find
all l
o
cal
maxima
and lo
cal mini
ma points in I
(
x,y).
b)
Upp
e
r envel
o
pe Up
(x,y) is cr
e
a
ted by interpol
ating the maxima p
o
ints an
d lower envelo
p
e
Lw(x,y)
i
s
created by
int
e
rpol
ating mi
nima
p
o
ints.
This interpol
ation i
s
ca
rri
ed o
u
t u
s
in
g
cubi
c spline i
n
terpol
ation
method.
c)
Comp
ute the mean of lowe
r and up
pe
r
e
n
velope
s de
n
o
ted by Mean
(x,y).
((
,
)
(
,
)
)
(,
)
2
Up
x
y
L
w
x
y
Mean
x
y
(1)
d)
This me
an si
gnal is
subtra
cted from the
input sig
nal.
(,
)
(
,
)
(,
)
Sub
x
y
I
x
y
M
e
an
x
y
(2)
e)
If Sub(x,y) satisfies the IM
F prop
ertie
s
, then an IMF is obtain
ed .
(,
)
(
,
)
i
I
M
F
xy
S
u
b
xy
(3)
f)
Subtra
ct the extracted IMF
from the i
npu
t signal. No
w
the value of I(x,y) is
(,
)
(
,
)
(,
)
i
I
xy
I
x
y
I
M
F
xy
(4)
Rep
eat the a
bove step
s (b
) to (f
) for the
gene
ration of
next IMFs.
g)
This p
r
o
c
e
s
s is rep
eated
until I(x,y) does
n
o
t hav
e maxima or minima poin
t
s to create
envelop
es.
Origin
al Imag
e can b
e
re
co
nstru
c
ted by i
n
verse EMD
given by
1
(,
)
(
,
)
(,
)
n
i
i
I
xy
I
M
F
x
y
r
e
s
xy
(5)
Image De
noi
sing u
s
in
g EMD:
The me
cha
n
i
s
m of de-noi
sing usi
ng EM
D is summa
ri
zed a
s
follo
ws
a)
Apply 2-D E
M
D for
ea
ch
band i
n
the h
y
per
spe
c
tral
image to
obt
ain IMFi (i
=1,
2, …k). The
k
th
IMF is call
ed re
sidu
e.
b)
The first intri
n
si
c mode f
unctio
n
(IMF
1)
co
ntain
s
high freq
uen
cy comp
one
nts and it is
suitabl
e for denoi
sing. T
h
is IMF1 is
denoi
se
d wit
h
wavelet sh
rinkage d
eno
ising meth
od
pre
s
ente
d
in [18]. This de
-n
oise
d IMF1 is represented
with DNIMF1.
c)
The ne
w ban
d is re
con
s
tru
c
ted by the summa
tion of
FIMF and re
maining IMF
s
given by
2
1
k
i
i
R
ID
N
I
M
F
I
M
F
(6)
Whe
r
e RI i
s
the re
con
s
tructed band.
Evaluation Warning : The document was created with Spire.PDF for Python.
IJEECS
ISSN:
2502-4
752
Im
age Fusion
in Hyp
e
rspe
ctral Im
age Classificatio
n
u
s
ing G
eneti
c
Algorithm
(B Saicha
nda
na
)
705
3. Image Fusion Techniq
u
e
The hyp
e
rsp
e
ctral
data
prese
n
t ab
unda
nt multid
ime
n
s
ion
a
l info
rm
ation that
co
n
t
ains fa
r
more im
age
band
s than
those that
can b
e
displ
a
yed on the
standa
rd tri
s
timulu
s di
sp
lay.
Therefore,
a
n
efficie
n
t a
nd a
p
p
r
op
ria
t
e mea
n
s of
visuali
z
atio
n
of the
hype
rsp
e
ct
ral
dat
a is
need
ed [8].
Let u
s
con
s
id
er I
1
, I
2
, …. I
k
be
a
set of
hyperspe
c
tral
ban
ds,
co
ntaining
K con
s
e
c
utive
band
s. We
want to fuse these
ban
ds to ge
nerate a high
contrast
re
sultant image
for
visuali
z
ation.
The
prim
ary
aim of
imag
e fusi
on i
s
to
sel
e
ctively
merg
e the
m
a
ximum p
o
ssible
feature
s
from
the
so
urce
i
m
age
s to
form a
si
n
g
le i
m
age.
Hype
rspe
ctral
ima
g
e
ba
nd
s
are
the
result of
sam
p
ling
a conti
nuou
s
sp
ectrum at n
a
rro
w
wavelen
g
t
h
interval
s where
the no
minal
band
width of
a single b
a
n
d
is 10 nm [
9
]. The spe
c
t
r
al re
sp
on
se
of the sce
ne
varies g
r
a
d
u
a
lly
over the sp
ectrum, and thu
s
, the su
ccessive ban
ds
in
the hyperspe
c
tral ima
ge h
a
ve a signifi
cant
correl
ation.
Therefore, fo
r an effici
en
t fusion,
we
sho
u
ld be
able to extract the
spe
c
ific
informatio
n containe
d in a particula
r ban
d [13].
The p
r
ocess
of combi
n
ing
a few hu
ndre
d
band
s into
a singl
e imag
e may involve readi
ng
all the band
s in the input hyperspe
c
tral i
m
age at
on
ce
, computing t
he weig
hts, a
nd gen
eratin
g
a
resultant fuse
d image a
s
the linea
r co
mbination
of
all the input band
s. This
one time re
a
d
ing
and combini
n
g all the imag
e band
s have
the following
sho
r
tco
m
ing
s
[8].
1.
Fra
c
tional
Weights: Thi
s
result
s in assi
gning ve
ry small fractio
n
a
l
weight
s to the location
s in
each of the image ba
nd
s that might lea
d
to
washing
out some of t
he mino
r deta
ils.
2.
Memory
re
qu
ireme
n
ts: A
s
the
size of
hypersp
e
c
tral
data i
s
ve
ry
larg
e, it req
u
ire
s
h
uge
amount of me
mory to read
the data and
merg
e.
In ord
e
r to o
v
erco
me th
ese issu
es,
a h
i
era
r
chical m
e
thod fo
r ima
ge fu
sion i
s
use
d
fo
r
hyperspe
c
tral
data. Instea
d of
fusing t
he entire set
of bands
at a single tim
e
, this meth
od
cre
a
tes pa
rtitions of the
data into
su
bset
s
of
hyp
e
rspe
ctral
ba
nds.
For a
given ima
g
e
of
dimen
s
ion
s
(X × Y × K),
contai
ning K band
s, form
P subsets of dimen
s
ion
s
(X × Y × M) from
contig
uou
s b
and
s of the data,
where P is given by P =
K
M
.
The first sta
ge imag
e fusion
schem
e
may be em
ployed to ea
ch of the
s
e
sub
s
et
s
indep
ende
ntly to obtain
P
fused
imag
es. The
s
e P im
age
s form th
e ba
se
imag
es
(o
r in
puts) for
the second
-stage fu
sio
n
. T
h
is
procedu
re is repe
at
ed
in
a hi
era
r
ch
ical
mann
er to ge
ne
rate th
e
final result of fusion in a few sta
g
e
s
. The sa
m
e
pro
c
e
ss
contin
u
e
s till the sin
g
le fuse
d image
whi
c
h re
presents the com
b
ining of
the
entire data i
s
gene
rated. Fi
gure 2
sho
w
s the sch
emati
c
of
the hie
r
a
r
chical sch
e
me
of fusio
n
. The
fuse
d ima
ge
F at ea
ch
sta
ge
can
be
re
pre
s
ente
d
a
s
a
linear
com
b
in
ation of input image
s I
k
, k = 1 to M as sh
own b
e
lo
w:
1
1
(,
)
(
,
)
(,
)
(,
)
1
,
(
,
)
M
kk
k
M
i
k
Fx
y
w
x
y
I
x
y
and
wx
y
x
y
(7)
whe
r
e w
k
(x, y) is the weig
ht for the pixel at location
(x, y) in the k
-
th obser
vation, F(
x,y)
is
th
e
fused ima
ge
and
the su
m of all weig
hts at any sp
atia
l locatio
n
equal
s unity, i.e., normalized
weig
hts.
4. Genetic
Algorithm
Geneti
c
Algo
rithms [1
4] bel
ong to
the
cla
s
s of evol
utio
nary al
go
rith
ms th
at are b
a
se
d o
n
prin
ciple
s
of natural
sele
ct
ion and ge
ne
tics. It
is a search techni
que used in
comp
uting true
solutio
n
s to
optimizatio
n
probl
em
s tha
t
is drive
n
b
y
natural
evolution p
r
o
c
e
ss.
GA pe
rfo
r
ms
parall
e
l sea
r
ch of the so
lution spa
c
e
rather t
han
point by point search. Geneti
c
Algo
rithm
con
s
i
s
ts of three op
erators namely,
Selection, Crosso
ver and Mutat
i
on.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 25
02-4
752
IJEECS
Vol.
2, No. 3, Jun
e
2016 : 703
– 711
706
The Ge
netic
Algorithm me
cha
n
ism
can
be ab
stra
cted
as follows [1
5].
1)
The initial po
pulation of so
lutions i
s
ran
domly gene
ra
ted across th
e sea
r
ch sp
a
c
e.
2)
Usi
ng a
n
obj
ective fun
c
tio
n
, the fitness of
ea
ch i
n
dividual
solut
i
on in the
p
opulatio
n is
evaluated.
3)
Usi
ng this fitn
ess value
s
, the solutio
n
s in
the popul
atio
n are sele
cte
d
.
4)
Ne
w p
opulati
on i
s
cre
a
te
d from
sele
cted solution
s usi
ng th
e
crossove
r
and
mutation
operators.
5)
The ne
w pop
ulation is repl
ac
e
d
inste
ad
of old popul
ation.
6)
Rep
eat iteratively from (2)
to
(5)
until a stop criterion
is satisf
ie
d.
Each ite
r
atio
n of this GA
pro
c
e
ss i
s
cal
l
ed gen
eratio
n.
GA is a
met
hod of p
a
rall
el se
arch
of the sol
u
tion
spa
c
e
ba
sed
on two
a
s
sumption
s
inspi
r
ed by e
v
olutionary bi
ology. 1) Th
e
measure
of probl
em solving ability by an individu
al in
the popul
atio
n is dete
r
min
ed by its fitness val
ue. 2
)
Ne
w individ
uals
whi
c
h a
r
e obtai
ned
by
combi
n
ing dif
f
erent individ
uals in the p
o
pulation h
a
ve
more p
r
oble
m
solving abil
i
ty.
Image Classi
fication u
s
in
g GA
The Ge
netic
Algorithm is a
pplied a
s
follo
ws.
a)
Assu
me P chrom
o
some
s in the po
pul
ation w
here
P is the
size
of the pop
ul
ation. Each
chromo
som
e
is en
cod
ed wi
th K cluster
centers
that are rand
omly selecte
d
from the image.
b)
Usi
ng
an
obj
ective fun
c
tio
n
, the fitne
ss value
of e
a
c
h
ch
rom
o
so
me i
s
eval
ua
ted. Th
ree
Different in
di
ce
s, su
ch a
s
K-mean
s in
d
e
x (K
MI) [16], Jm mea
s
u
r
e
and Xie-Be
ni
Index (XBI)
[17] are
used
as o
b
je
ctive function
s in
di
vidua
lly. For
comp
uting th
e mea
s
u
r
e
s
, the ce
nters
z
1
, z
2
, ….,
z
k
en
cod
ed i
n
a chromo
so
me a
r
e fi
rst
extracted.
Th
e mem
b
e
r
ship value
s
u
ik
,
i=1,2,….K an
d k=1,2,….n are
comp
uted
[10] as follows:
(8)
Whe
r
e D
(
z
i
,x
k
) is the Eucli
d
ean di
stan
ce
betwe
en two
points x
k
and clu
s
ter cente
r
z
i
.
The ce
nters e
n
co
ded in a
chrom
o
some a
r
e upd
ated u
s
ing the followi
ng equ
ation:
(9)
The XB index is defined as the function of ration of
the total variation to the minimum sep
a
ration
of the cluste
rs, whi
c
h is gi
ven by the followin
g
equati
o
n
(10
)
The J
m
mea
s
ure [18] whi
c
h is to be min
i
mized i
s
defi
ned a
s
(11
)
Evaluation Warning : The document was created with Spire.PDF for Python.
IJEECS
ISSN:
2502-4
752
Im
age Fusion
in Hyp
e
rspe
ctral Im
age Classificatio
n
u
s
ing G
eneti
c
Algorithm
(B Saicha
nda
na
)
707
Whe
r
e m
is t
he fuzzy exp
onent, D de
n
o
tes the
Eucli
dean
dista
n
ce between
two point
s x
j
an
d z
k
and u
kj
de
not
es the mem
b
ership value
s
.
The k-me
an
s index [17] which i
s
used
as the obj
ecti
ve function in
this GA pro
c
ess is defin
e
d
as
follows
:
2
11
1
KN
ik
ki
KM
I
x
z
(12
)
Whe
r
e K num
ber of clu
s
te
rs and
z
k
is th
e clu
s
ter
cent
ers
c)
The sele
ctio
n of ch
rom
o
some
s i
s
do
ne ba
se
d on
the fitness
value u
s
ing
roulette wheel
techni
que.
d)
By applying
cro
s
sove
r an
d mutation
o
perato
r
s with
rate
0.8 an
d
0.07, a
ne
w
popul
ation i
s
prod
uced fro
m
the pare
n
ts. This ne
w po
pulation repla
c
e
s
the old p
opulatio
n.
e)
Maximum nu
mber of iterati
ons i
s
used a
s
stop
ping
cri
t
eria.
After the
exe
c
ution
sto
p
s,
the
highe
st
fit
ness value
ch
rom
o
som
e
is
sele
cted
and
the
values in thi
s
chromo
som
e
rep
r
e
s
ent the
solution to th
e cla
ssifi
catio
n
of image.
5. Experimental Re
sults
The
pro
p
o
s
e
d
meth
odol
og
y is te
sted
on
Pavia
Unive
r
sity and
Indi
a
n
pin
e
s hype
rspe
ctral
image d
a
ta
set. The Pavia
Unive
r
sity d
a
ta set
co
ntains 1
03
spe
c
t
r
al b
and
s a
n
d image
in e
a
ch
band
co
nsi
s
t
s
of 6
10*
340
pixels. T
he I
ndian Pi
ne
s
data
set cont
ains 200
spe
c
tral
ban
ds a
n
d
image in e
a
ch band
co
nsi
s
ts of 14
5*1
45 pixels. Th
e data sets
are
colle
cted
from [11] that
con
s
i
s
t of nine cla
s
ses in
Pavia university dat
a set and sixteen
cl
asse
s in Indi
an Pines
dat
a set
with the geo
metric resolut
i
on is 1.3 met
e
rs. Th
e
expe
rimental resul
t
is cond
ucte
d on ban
d 10
0 in
each of th
e
d
a
ta set an
d i
s
sh
own in
fig
u
re
1.
T
he
sa
me p
r
o
c
ed
ure is exe
c
uted
for
all b
and
s in
the data set i
n
-o
rde
r
to en
han
ce the im
age an
d final
l
y
increa
se th
e cla
ssifi
catio
n
accu
ra
cy. The
qualitative an
alysis of the prop
osed me
thod on Pa
via University and Indi
an Pin
e
s hype
rspe
ctral
data set is shown in Figu
re 1. Tabl
e 1
spe
c
if
ie
s the
quantitative index value
s
of the pro
p
o
s
ed
method comp
ared
with the grou
nd truth i
n
formatio
n available in [11]
.
Table 1. Cla
s
sificatio
n
accura
cy of pr
op
ose
d
method
on two differe
nt data sets
Pavia University
Without EMD
With EMD
Objective Function
CA%
Kappa Coe
fficient CA%
Kappa
Coefficient
KMI 74.51
0.682
79.3
0.742
Jm 84.21
0.829
89.4
0.823
XB
88.52
0.852
94.32
0.904
Indian Pines
Without EMD
With EMD
Objective Function
CA%
Kappa Coe
fficient CA%
Kappa
Coefficient
KMI 69.41
0.682
74.3
0.713
Jm 81.62
0.799
88.4
0.813
XB 84.62
0.842
93.9
0.894
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 25
02-4
752
IJEECS
Vol.
2, No. 3, Jun
e
2016 : 703
– 711
708
Pavia University image ba
nd
100
IMF1 IMF2
IMF3
Image Band
after de-noi
si
ng
usin
g IMF an
d Mean filter
Fuse
d Image
Cla
ssif
i
e
d
U
s
i
ng GA
(XB index)
Figure 1. Hyp
e
rspe
ctral Im
age cl
assifica
tion (
co
nt
.
)
Evaluation Warning : The document was created with Spire.PDF for Python.
IJEECS
ISSN:
2502-4
752
Im
age Fusion
in Hyp
e
rspe
ctral Im
age Classificatio
n
u
s
ing G
eneti
c
Algorithm
(B Saicha
nda
na
)
709
Indian pin
e
s
data set
-
ima
ge
band 1
0
0
IMF1 IMF2
IMF3
Image Band
after de-noi
si
ng
usin
g IMF an
d Mean filter
Fuse
d Image
Cla
ssif
i
e
d
U
s
i
ng GA
(XB index)
Figure 1. Hyp
e
rspe
ctral Im
age cl
assifica
tion
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 25
02-4
752
IJEECS
Vol.
2, No. 3, Jun
e
2016 : 703
– 711
710
Figure 2. Hierarchical Sch
e
m
e of Image Fusio
n
7. Conclusio
n
s
In this p
ape
r, hyp
e
rspe
ctral i
m
age
cla
ssifi
cation ba
se
d
on EMD an
d
Image fu
sio
n
is
pre
s
ente
d
. EMD is u
s
ed in
the prep
ro
ce
ssi
ng sta
ge
for rem
o
val of noise in hyp
e
rspe
ctral b
a
nds.
After noise removal, the i
m
age
ban
ds
are fu
se
d int
o
a
single
im
age fo
r visu
a
lization
pu
rpo
s
e.
This fused i
m
age is
cla
s
sified usi
ng
Geneti
c
algo
rithm with three different o
b
jective funct
i
ons.
The experi
m
ental re
sults show that XB index as
objective function in G
A
classifies
the
hyperspe
c
tral
image m
o
re
efficiently. The EMD f
ilte
r
ing me
cha
n
ism increa
se
s
the accu
ra
cy of
cla
ssifi
cation algorith
m
.
Referen
ces
[1]
Yi Ma, Jie Z
h
a
ng an
d Yufe
ng
Gao. “
High R
e
soluti
on R
e
mot
e
Sensi
ng I
m
a
ge Cl
assificati
o
n
of Coasta
l
Z
one an
d
its Auto
matic Rea
l
i
z
a
t
io
n
”. Proc
e
edi
ngs
of IEE
E
Inter
nati
o
n
a
l
Co
nferenc
e
o
n
C
o
mpute
r
Scienc
e an
d Soft
w
a
r
e
Eng
i
n
e
e
rin
g
. 200
8: 82
7-82
9.
[2]
Xi
aofa
ng Liu e
t.al.
“
R
e
mo
te
Se
n
s
in
g
Im
ag
e C
l
a
ssi
fi
ca
ti
on
U
s
i
n
g Fu
n
c
ti
on We
i
g
h
t
ed
FCM C
l
u
ste
r
in
g
Algorit
h
m
”. Proceed
ings of IE
EE Internatio
n
a
l Geosci
ence
s
and Rem
o
te Sensi
ng S
y
mp
osium. 20
07
:
201
0-20
13.
[3]
Nor As
hid
i
M
a
t Isa, Sam
y
A.Salam
ah,
Umi K
a
lthum
Nga
h
. “Ad
apti
v
e F
u
zz
y M
o
ving
K-mea
n
s
Clusteri
ng Al
g
o
rithm for Image Segm
entati
on”.
IEEE Tra
n
sactions on
Cons
um
er Electronics
. 201
2;
55(4): 21
45-
21
53.
[4]
Siti Naraini Sulaiman
, Nor Ashidi Mat Isa. “Deno
isin
g bas
ed
Cluterin
g
Alg
o
r
ithms for Seg
m
entatio
n of
Lo
w
lev
e
l
of Sa
lt an
d Pe
pp
er
Noise
C
o
rrupte
d
Imag
es”.
IEEE Transactions
on Co
nsum
er
Electronics
.
201
0; 56(4).
[5]
JC Nu
nes
et al. “Image
an
al
ysis
b
y
bi
di
mensi
ona
l em
pirica
l mo
de
d
e
comp
ositio
n”.
Im
ag
e
Vi
de
o
Co
mp
uting.
2
0
03; 21(1
2
): 101
9–1
02
6.
[6]
Zeiler A, et.al. “
Emp
i
rica
l M
o
d
e
D
e
co
mp
ositi
on-
an
Introd
uc
tion
”. in Pr
oceedings of
IEEE I
n
ternational
Joint Co
nfere
n
c
e on Ne
ural N
e
t
w
o
r
ks. 201
0: 1-8.
[7]
Alp Erturk, et.al. “H
ypers
pec
tral Image C
l
a
ssi
ficatio
n
Usi
ng Emp
i
rica
l Mode D
e
com
p
ositio
n
w
i
t
h
Spectral
Gradi
ent e
n
h
ancem
ent”.
IEEE Transactions on
GEo
scienc
e and Rem
o
teS
e
nsing
. 20
13
;
51(5): 27
87-
27
90.
[8]
Ketan Kot
w
a
l
, et al. “Visualizati
on of H
y
p
e
rsp
e
ctral Images Usi
n
g
Bilatera
l F
ilterin
g”.
IE
EE
transactio
n
s of Geoscie
n
ce a
n
d
remote Sens
i
n
g
. 201
0; 48(5)
: 2308-2
3
1
9
.
[9]
B Guo, S Gun
n
, B Dam
per
and J
Ne
lso
n
. “
Hypers
pectra
l
i
m
a
ge fus
i
o
n
usin
g sp
ectral
ly w
e
ight
e
d
kernels
”. in Proc. 8th Int. Conf.
Inf.
Fusion,
Ph
ilad
e
l
phi
a, PA. 200
5; 1: 402–
4
08.
[10]
J Harikira
n, et al. Multiple
F
eature F
u
zz
y
C-me
ans
clusteri
ng al
g
o
rithm for segmentati
on o
f
microarray
im
age”.
Internato
n
a
l jo
urna
l of Electrical a
nd Co
mp
uter Eng
i
n
e
e
rin
g
. 201
5; 5: 104
5-10
53.
[11]
http://
w
w
w
.
e
h
u
.
eus/cc
w
i
n
tco/i
nde
x.p
h
p
?
title=
H
y
pers
pectral
_
R
emote
_
Se
nsi
ng_Sc
e
n
e
s.
Evaluation Warning : The document was created with Spire.PDF for Python.
IJEECS
ISSN:
2502-4
752
Im
age Fusion
in Hyp
e
rspe
ctral Im
age Classificatio
n
u
s
ing G
eneti
c
Algorithm
(B Saicha
nda
na
)
711
[12]
B Saic
ha
nda
n
a
et
al. “C
lus
t
ering
Alg
o
rith
m Comb
in
ed
w
i
t
h
Em
piric
a
l
Mod
e
D
e
co
mpositi
on f
o
r
Classification of Remo
te Sens
ing Imag
e”.
IJSER.
2014; 5(
9): 686-6
95.
[13]
Subh
asis Ch
au
dhur
i, Ketan K
o
t
w
a
l
. “H
yp
ers
pectral Imag
e F
u
sion”. Spri
ng
er book.
[14]
J Holl
an
d. “Ad
aptatio
n in
Na
t
u
ral a
nd Artific
i
al S
y
stems”.
Ann Ar
b
o
r, MI: Univ. of Mich
i
gan Pr
ess.
197
5.
[15]
A Erturk et al. “Hyper Sp
ec
tral Image Cla
ssi
ficatio
n
Usi
ng Empiric
a
l
Mode Dec
o
mp
ositio
n
w
i
t
h
Spectral Grad
i
ent enh
anc
em
ent”.
IEEE Tra
n
sactions on
Geosci
ence and Remote Sensing
. 51(5):
278
7-27
98.
[16]
B Sang
hamitr
a
et al. “Multi
o
b
j
ective Ge
netic
Clusteri
ng for
Pixel
Class
ific
ation
in R
e
mot
e
Sens
in
g
Imager
y”.
IEEE Transactions
on Geosci
enc
e
and Re
mote S
ensi
n
g
. 20
07; 45(5): 15
06-
15
12.
[17]
Yeh
F
en, Yang et
al.
“Genetic Algorit
hms for Multispectr
al Image
C
l
assificati
on”
.
http://
w
w
w
.
is
pr
s.org/proce
edi
ngs/
XXXVI/par
t3/singl
ep
apers
/
P_16.pdf.
[18]
Anutam an
d R
a
jni. “Perfo
rma
nce Ana
l
ysis o
f
Image Deno
i
s
ing
w
i
th W
a
v
e
let T
h
reshold
i
ng Metho
d
s
for differe
nt L
e
v
els
of Dec
o
m
positi
on”.
T
h
e Internatio
na
l
J
ourn
a
l of
Mu
lti
m
e
d
i
a
a
n
d
its
app
licati
ons
.
201
4; 6(3): 35-
46.
[19]
B Saic
ha
nda
n
et al. “
Clust
eri
ng A
l
g
o
rithm
Combi
n
e
d
w
i
t
h
Hil
l
Clim
bin
g
f
o
r Cl
assific
a
tio
n
of
Rem
o
t
e
Sensi
ng Imag
e
”
.
Internation
a
l
Journ
a
l of Elec
tr
ical an
d Co
mputer Eng
i
n
eeri
n
g
. 201
4; 4(6): 923-
930.
Evaluation Warning : The document was created with Spire.PDF for Python.