TELKOM
NIKA
, Vol.13, No
.1, March 2
0
1
5
, pp. 146~1
5
4
ISSN: 1693-6
930,
accredited
A
by DIKTI, De
cree No: 58/DIK
T
I/Kep/2013
DOI
:
10.12928/TELKOMNIKA.v13i1.1318
146
Re
cei
v
ed Se
ptem
ber 30, 2014; Revi
se
d De
ce
m
ber
11, 2014; Accepted Janu
ary 9, 2015
An Improved AP-Wishart Classifier for Polarimetric
SAR Images by Incorporating a Textu
r
al Feat
ure
Chen J
u
n
1
*, Du Peijun
2
, Tan Kun
1
1
Jiangsu Ke
y
Lab
orator
y of
Reso
urces an
d
Environm
ent
al
Information En
gin
eeri
ng, Ch
in
a Univ
ersit
y
of
Minin
g
an
d T
e
chno
log
y
,
Xuz
hou C
i
t
y
, Jia
n
g
s
u Provinc
e
, 2211
16, Ch
ina
2
Jiangsu Prov
i
n
cial K
e
y
La
bo
rator
y
of Geogr
aph
ic Informati
on Scie
nce a
n
d
T
e
chnolo
g
y
,
Nanj
in
g
Univers
i
t
y
, N
a
n
jing cit
y
, Jia
ngs
u Provinc
e
, 21
009
3, Chi
n
a
Corresp
on
din
g
author, e-mai
l
: stud
yias@
163
.com
A
b
st
r
a
ct
An i
m
prove
d
cl
assifier
is
pres
ented
by
i
m
p
o
s
ing
a text
ural
feature to
so
lv
e the
pr
obl
e
m
s
of va
gu
e
initia
l cluster
i
n
g
results, low
cl
assificati
on
accuracy a
n
d
unc
han
ge
ab
le class
nu
mb
er in the
itera
t
ive
classifier,
base
d
on
H/Alp
ha
d
e
co
mp
ositio
n a
nd the
co
mp
le
x W
i
shart distri
butio
n for ful
l
y
pol
ari
m
etric SA
R
(Synthetic Ap
erture R
adar)
imag
es. F
i
rst, w
a
velet transformatio
n
is
used to
ext
r
act texture fr
o
m
pol
ari
m
etric S
A
R images. S
e
con
d
, an AP (Affinity Propa
gat
io
n) alg
o
rith
m is ap
pli
ed to create the in
itia
l
clusteri
ng res
u
lt. T
h
is result is then a
p
p
lie
d to t
he iterat
ive class
i
fier b
a
sed
on the c
o
mpl
e
x W
i
sha
r
t
distrib
u
tion
to
obtai
n the
fin
a
l
result. T
w
o P
A
LSAR
(P
hase
d
Array typ
e
L-
ban
d Synth
e
tic
Apertur
e
R
a
d
a
r)
imag
es fro
m
ALOS (Adva
n
c
ed
Lan
d Obs
e
rving S
a
tel
lite)
are
use
d
for
the ex
peri
m
ent
s carrie
d
out
on
exper
imenta
l
p
l
ots in Binh
ai
Prefecture, Ya
nche
ng Cit
y, Jian
gsu Provi
n
ce. T
he results show
that the
improve
d
class
i
fier has so
me
mer
i
ts, includ
in
g clear in
it
ia
l clusterin
g
results
,
flexible class
nu
mb
er and h
i
g
h
classificati
on
a
ccuracy. T
he i
m
pr
ove
d
classi
fier has b
e
tter overa
ll perf
o
rmance th
an the
origi
n
a
l
, and c
a
n
be effectively
a
ppli
ed to the cl
assificati
on of pol
ari
m
etric SA
R imag
es.
Ka
ta
k
unc
i:
POLSAR, H/Alp
ha-w
i
shart clas
sifier, AP
cluster, texture, wa
velet transform
1. Introduc
tion
There are two main types of machin
e learni
ng alg
o
r
ithm in cla
s
sificatio
n
of PolSAR
(Polari
m
etri
c Synthetic
Ap
erture
Rada
r) image
s, nam
ely sup
e
rvis
e
d
and
un
sup
e
rvise
d
lea
r
ni
ng.
The forme
r
requires a
ce
rtain num
ber of trai
ning sample
s; the latter doe
s n
o
t need train
i
ng
sampl
e
s. Cl
a
ssifi
cation a
c
curacy of su
p
e
rvise
d
learni
ng depe
nd
s sub
s
tantially on the quality of
training
sa
mp
les. However,
a PolSAR im
age i
s
differe
nt to an o
p
tical image. It b
e
long
s to a
c
ti
ve
microwave remote sen
s
i
ng and i
s
greatly affected
by the interfere
n
ce of different wave
s. It
therefo
r
e ha
s short
c
omin
g
s
su
ch a
s
se
riou
s
sp
eckle
noise, low
contra
st betwe
en obje
c
ts a
nd
backg
rou
nd,
edge bl
urring
and othe
r un
certai
n fa
cto
r
s. Although
some mea
s
u
r
es
can b
e
taken
to eliminate
such
noi
se, im
age q
uality is signifi
c
antly
degrade
d at the same time
. So, after bei
ng
contraste
d
wi
th the opti
c
al
image, the
PolSAR ima
g
e
is h
a
rder to i
n
terp
ret. The
r
e may b
e
so
me
difficulty in using a su
pervi
sed le
arni
ng
algorith
m
dire
ctly on a PolSAR image.
The pola
r
ime
t
ric features
of a PolSAR
im
age are expre
s
sed by a
cohe
rent scattering
matrix, whi
c
h has
a com
p
lex Wi
sh
art probability density distri
buti
on. Therefore, a
classifier
develop
ed u
nder the
s
e
ci
rcu
m
sta
n
ces
may achiev
e
highe
r
cla
ssi
fication a
c
cu
racy. Le
e [1]-[
4
]
prop
osed
an
un
supe
rvise
d
cl
assifier a
fter co
mbini
n
g H/A/Alpha
decompo
sitio
n
[5]-[7] with
a
compl
e
x Wishart dist
ributi
on, calle
d an
H
/
A
/
Alpha
-
W
i
s
ha
rt
cla
ssif
i
e
r
.
The H/A/Alph
a-Wish
art cl
a
ssifie
r
is wi
d
e
ly
used in remote se
nsi
n
g. Laure
n
t et al. [8
]
utilized it for
cla
ssifi
cation
of a multi-fre
quen
cy
PolS
AR imag
e. Many improve
m
ents h
a
ve
been
made to
the
H/A/Alpha-Wi
sha
r
t cl
assifi
er: Kimura et
al. [9] and
Cao
et al. [1
0] adde
d
spa
n
to
form the
H/Al
pha/span
-Wi
shart
cla
ssifie
r
. Wu
et
al. [1
1] used fo
ur
comp
one
nt d
e
com
p
o
s
ition
to
divide a pixel
into a fou
r
ba
se
scatter-typ
e and
a
com
poun
d scatte
r-type, and th
ey cla
ssifie
d
t
he
image u
s
in
g
the Wi
sha
r
t classifier. Yan
g
et al.
[12] imposed a
n
o
p
timal co
rrela
t
ion co
efficie
n
t
and
spa
n
to
the cla
s
sifier to automati
c
ally extr
act road a
nd
soil.
Zhang
et al.
[13] com
b
in
ed
H/A/Alpha de
comp
ositio
n
and the M
a
rkov Ran
dom
Field for
cla
s
sifying the im
age. Yang
et al.
[14] combi
n
e
d
H
/
Alpha
d
e
com
p
o
s
ition
with a pol
arimetri
c
whi
t
ening filter
to the imag
ery
cla
ssifi
cation.
Yang et al. [15] pro
p
o
s
ed
a weig
hted u
n
su
pervi
sed
Wishart
cla
s
sifier. Zhao et
al.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
An Im
proved
AP-Wi
shart Classifier for P
o
larim
e
tr
ic S
A
R Im
age by Incorpor
ating
.... (Chen Jun)
147
[16] use
d
a li
kelih
ood
-ratio
test to cal
c
u
l
ate
the dist
a
n
ce
between
each pixel a
nd every
cla
s
s
cente
r
; and L
ang et et al. [17] adde
d
H
,
A
and Free
m
an de
comp
osition to the classificatio
n
.
All of these i
m
provem
ents have bee
n b
a
se
d sol
e
ly o
n
pola
r
imetri
c information.
But at
the sa
me tim
e
there a
r
e p
l
enty of
other feature
s
, su
ch a
s
texture
,
in PolSAR i
m
age.
Ho
w to
combi
ne th
ese features wit
h
pola
r
imet
ric informatio
n f
o
r
cla
ssifi
cati
on of PolSA
R
image
is
still
a
hotly debate
d
topic.
Liu et
al. [18] unite
d texture
with
SVM into th
e cl
assification an
d a
c
hiev
ed
high perform
ance; however, te
xture is
still less com
m
only used i
n
improvement of the
H
/
Alpha
-
Wishart
cla
s
sifier. The
r
ef
ore, in this
article
a ne
w un
sup
e
rvi
s
ed
cla
ssifi
cation sche
m
e
is
prop
osed, by imposi
ng texture to
improve the accuracy of the
H
/
Alpha
-Wisha
rt cl
assifier.
2. H/Alpha-Wishart Classifier
Clou
de & Pot
t
ier [6] develo
ped
H
/
Alpha
decompo
sitio
n
to extra
c
t
polarimetric
c
h
arac
ters
from
scatterin
g
information.
They
define
d
Entropy
(
H
) as the
domi
n
ant scatte
ring
me
chani
sm
i
n
each pixel. Its
func
tion is
as follows
[6]:
3
1
3
log
i
i
i
P
P
H
(1)
Her
e
i
P
is co
m
puted a
s
follo
ws:
3
1
j
j
i
i
P
(2)
and defin
e:
3
3
2
2
1
1
P
P
P
(3)
(
Alpha
) he
re
stand
s fo
r th
e
scatteri
ng
an
gle; it exp
r
e
s
se
s the
me
an
scatteri
ng
de
gree
of the object.
Anis
otropy
[19,20] (
A
) is d
e
fined a
s
follows:
3
2
3
2
A
(4)
Anis
otropy
d
enote
s
a
diff
erent
scatteri
ng m
e
chani
sm othe
r th
an
the m
ean
scatterin
g
one. So
Entropy
(
H
),
(Alph
a
) and
Anis
ot
ropy
(
A
) can
be used to re
place scatteri
ng ch
ara
c
ters
of groun
d obj
ects.
On the
ba
si
s of
H
/
Alpha
decompo
sitio
n
, Clo
ude
an
d Pottier [6]
prop
osed
an
H
/
Alpha
cla
ssifi
cation
schem
e. Th
ey defined
a
n
H
/
Alpha
pl
ane to
divid
e
groun
d ob
jects into ei
ght
different types. The plan
e is sh
own in Figure 1:
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No. 1, March 2
015 : 146 – 1
5
4
148
Figure 1.
H
/
Alpha
p
l
an
e
In Figu
re
1,‘Zone
1’
stan
ds fo
r
high
e
n
tropy even scattering,
‘Z
one 2’ stand
s
fo
r
hi
g
h
entropy
multiple
scatte
ring
, ‘Zone
3’ st
and
s for me
dium e
n
tropy
multiple
sca
ttering, ‘Zo
n
e
4’
stand
s fo
r m
edium e
n
tro
p
y
vegetation
scattering,
‘Z
one 5’
stan
d
s
for m
ediu
m
entro
py su
rface
scattering, ‘Z
one 6’ stan
d
s
for low ent
ropy multip
le scattering, ‘Z
one 7’ stan
d
s
for low ent
rop
y
dipole
scattering, and ‘
Z
on
e 8’ sta
n
d
s
f
o
r lo
w e
n
trop
y surfa
c
e
sca
tter [6]. If a prope
r th
re
sho
l
d,
namely
Aniso
t
ropy
(
A
), is chosen, gro
u
n
d
obje
c
ts can
be divided in
to sixteen furt
her types.
With the
cla
ssifi
cation
re
sult of the
H
/
Alpha
initial
cla
ssifie
r
, cl
uster ce
nters can
be
gathered by averagi
ng
th
e
f
eatures wi
thin ea
ch
cla
ss. After
this,
these cl
uste
r cente
r
s
ca
n
be
importe
d to the Wishart iterative classifie
r
to obtain the final c
l
ass
i
fication res
u
lt.
The
H
/
Alpha
-Wishart
u
n
su
pervised cla
s
sifier can
fulf
il
l the cla
ssifi
cation autom
atically. It
is used in so
me are
a
s of
PolSAR imag
e cla
ssifi
catio
n
, but it has the followi
ng d
i
sadva
n
tage
s:
The
cla
ss
nu
mber i
s
u
n
ch
ange
able. If the H/Al
ph
a is use
d
, 8 cl
asse
s can b
e
o
b
tained,
or, if the H/A/Alpha is use
d
, 16 classe
s can be atta
i
ned. Ho
weve
r, in reality, the cla
ss n
u
m
ber
can
cha
nge a
t
any time.
The H/Alpha cla
ssifi
cation is
ba
se
d
o
n
the
H/Alp
ha p
l
ane, but
real
gro
und
obje
c
ts
a
r
e
not st
rictly
re
flected i
n
th
e
H/Al
ph
a pl
a
ne, pa
rticularl
y
whe
n
H i
s
too hig
h
(H
>0.7) to
u
s
e
the
H/Alpha pla
n
e
.
The
cla
s
sification a
c
curacy of the
H/Alpha-Wi
sha
r
t cla
ssifie
r
i
s
d
i
rectly affecte
d
by
the
result of the
H/Alpha i
n
itia
l cla
ssifie
r
,
so
it is
n
e
cessa
r
y to imp
r
ove
the cl
assifica
tion accu
ra
cy
of
the
H
/
Alpha
classifier.
3. Impro
v
ed
AP-Wishar
t
Classifie
r
Con
s
id
erin
g the
H
/
Alpha
-Wishart
cla
s
sifier’
s
di
sad
v
antage
s, an
improve
d
AP-Wi
sha
r
t
cla
ssifie
r
is
prop
osed —
one which combine
s
the
AP algorith
m
with the Wishart itera
t
ive
cla
ssifi
cation
to achi
eve th
e final result. The flo
w
cha
r
t of the AP-Wishart
cla
ssifier is
sh
own
in
Figure 2.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
An Im
proved
AP-Wi
shart Classifier for P
o
larim
e
tr
ic S
A
R Im
age by Incorpor
ating
.... (Chen Jun)
149
Figure 2. Flowchart of imp
r
oved AP-Wishart cla
s
sifier
The process
of the improv
ed cla
s
sificati
on schem
e is as follows:
A wavelet
ex
traction
alg
o
ri
thm is u
s
ed
to
extra
c
t texture
from
the
PolSAR ima
g
e. The
wavelet
as a
kin
d
of i
n
formation extra
c
tion m
e
thod
is
wid
e
ly u
s
ed in
si
gnal
pro
c
e
ssi
ng
a
n
d
image analy
s
is [21],[22]. After Daubechi
es’
wa
velet basi
s
function is
chosen, a py
ramid
algorith
m
fro
m
Mallat i
s
use
d
to extract te
xture
from a
pol
ari
m
etric
sp
an
image. Ve
rtical,
hori
z
ontal a
n
d
diago
nal inf
o
rmatio
n are
extracted
a
s
t
he texture ch
ara
c
ter of the
PolSAR image.
The texture i
n
formation above
is
used to fulfill the AP initial cluster. The AP algorithm
was proposed by F
r
ey
and Dueck
in 2007 [23],[24]. It is one of t
he best
cl
ust
e
ring al
gorithms
at
pre
s
ent. Mo
st of the curren
t cluste
ring al
gorithm
s a
r
e
based on initi
a
l clu
s
ter
cen
t
ers, which are
comm
only
rel
a
ted to
cl
assi
fication
accu
racy. Howe
ver, the AP al
gorithm is differe
nt. It con
s
id
ers
each pie
c
e o
f
data as a
p
o
tential clu
s
t
e
r cente
r
, wh
ich i
s
call
ed t
he ‘exempl
a
r’
. Con
s
equ
ent
ly,
the re
sult of
the AP alg
o
rithm i
s
not
affected
by initial clu
s
te
r centers. T
he AP algo
ri
thm
cal
c
ulate
s
si
milarity bet
ween
different
data
point
s
by a
defined
dista
n
ce m
e
asu
r
e,
and
then
c
h
an
g
e
s
da
ta va
lu
es
in
to
N
*N (
N
i
s
th
e
size
of the d
a
ta)
simila
ri
ty matrix S. By iterative up
dat
ing
of,
Responsi
b
ility
(
R
) and
Availability
(
A
), an exact exempla
r
can
be found an
d con
s
ide
r
e
d
as
the final clu
s
ter ce
nter. Th
e pro
c
e
ss of t
he AP algorit
hm is a
s
follows:
Algorithm init
ializatio
n: Ca
lculatin
g the sim
ilarity matrix using a
Euclidea
n distan
ce
measure,
S
(
i
,
j
), 0<
i
<
N
,
0<
j
<
N
, the formula is
as
follows
:
22
2
2
11
(
)
()
(
)
()
i
j
i2
j2
i3
j3
i
n
j
n
d
=
T
T
TT
T
T
TT
(5)
T
i
stand
s fo
r
‘texture of
sa
mple
i
;
T
j
sta
nds for ‘texture
of sampl
e
j
,
n
r
e
p
r
es
en
ts
the
d
i
me
ns
io
ns
of te
xtu
r
e
.
Iterative upda
te of
R
and
A
,
R
(
i
,
k
)
sho
w
s that data p
o
int
k
is
suita
b
le for the ca
ndidate
exempla
r
of point
i
—it re
flects the a
ccumulated
eviden
ce ab
out
how
well-suit
ed point
k
is
to
serve a
s
the
exempla
r
for point
i
[23].The formula i
s
:
(,
)
(
,
)
m
a
x
(
,
)
(,
)
R
i
k
Si
k
A
i
j
Si
j
(6)
subj
ect to
j
{1,
2
,……,
N
, but
j
≠
k
}.
A
(
i
,
k
)
stand
s
for the deg
re
e that point
i
will sel
e
ct point
k
as it
s ca
ndidate exe
m
plar, an
d
reflect
s
the
a
c
cumulate
d e
v
idence on
h
o
w a
p
p
r
op
riat
e it wo
uld b
e
for poi
nt
i
to c
h
oo
se
po
in
t
k
as its exempl
ar
A
[23],[24]
.
The formula i
s
:
(,
)
m
i
n
0
,
(
,
)
m
a
x
(
0
,
(
,
)
)
j
Ai
k
R
k
k
R
j
k
(7)
PolSAR image
Texture extra
c
tion u
s
ing
wavelet algorit
hm
AP initial clustering
Wishart iterative classification
Cla
ssif
i
cat
i
on
result
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No. 1, March 2
015 : 146 – 1
5
4
150
subj
ect to
j
{1,
2,
……
N
, but
j
≠
i
and
j
≠
k
}.
Maximize
R
(
i
,
k
)+
A
(
i
,
k
), if
k
=
i
th
e
n
co
ns
id
er
i
a
s
a
n
exempla
r
, el
se
co
nsi
der
k
as
the
exempla
r
of point
i
. In this
way the clu
s
t
e
r ce
nters ca
n be obtain
e
d
.
Usi
ng the
clu
s
ter
ce
nters
shown ab
ove
and Eu
clide
a
n
dista
n
ce, e
v
ery data p
o
i
n
t can
be
clu
s
tere
d to its own ce
nters.
Above is the
prin
ciple
an
d
pro
c
e
s
s of th
e AP cl
uste
ri
ng al
gorithm.
The AP al
gori
t
hm ca
n
be u
s
ed
on t
he texture fe
ature extract
ed from th
e
wavelet extra
c
tion al
gorith
m
to find ea
ch
c
l
us
te
r
c
e
n
t
er.
Wishart ite
r
at
ive cla
ssifi
cat
i
on: Using
cl
us
ter cente
r
s created
by t
he AP alg
o
rit
h
m, a
Wishart ite
r
at
ive cla
ssifi
cat
i
on ba
sed
on
the Wi
sha
r
t distan
ce
can
be u
s
ed to
obtain the fin
a
l
c
l
as
s
i
fic
a
tion
result. The formula is
as
follows
:
1
d
i
s
t
(
[
]
|
[]
)
l
n
|
[]
|
T
r
(
[]
[
]
)
ii
i
TT
(8)
]
[
i
stand
s for the
average p
o
la
rimetri
c
co
he
rency matrix o
f
the cluster
center
i
,
]
[
T
stand
s for the
polarim
etric
coh
e
re
ncy m
a
trix of an unkno
wn
sampl
e
.
Acco
rdi
ng to the maximum
-
likeliho
od crit
erion, when t
he followi
ng condition i
s
met:
])
[
|
]
([
])
[
|
]
([
j
i
T
dist
T
dist
(9)
The un
kno
w
n
sample i
s
consi
dered to belon
g to
clu
s
ter cente
r
j. An iterative
pro
c
e
s
s
can b
e
used i
n
this way unt
il all unkn
o
wn
sample
s hav
e their clu
s
te
r centers.
4. Experiment and an
aly
s
is
4.1. Rese
arc
h
area and d
a
ta
Two sce
n
e
s
of the PALSAR image from ALOS were u
s
ed a
s
data sou
r
ce
for this
research. The image time
was
Ap
ril 9, 2009.
Res
e
arch areas
ar
e loc
a
ted on the Binhai
wetland
in Yanc
heng
c
i
ty. These ar
e two
blocks of squar
e
area
w
i
th the
s
i
z
e
of 800*800 pixels
.
Ther
e
are five type
s of land
cove
r in ea
ch of th
ese
re
sea
r
ch
area
s: ‘field’,
‘building’, ‘
r
o
ad’, ‘wate
r
’ a
nd
‘sua
eda’, which are disord
erly and un
systematically
located o
n
the re
sea
r
ch
area
s. They
are
thus suitable
for com
p
a
r
ing
the diversity
betwe
en the
origin
al
H
/
Alpha
-Wisha
rt and the imp
r
ov
ed
AP-Wi
sha
r
t algorithm. Ima
ges of sp
an and pauli de
co
mpo
s
ition (p
olarim
etric
ca
libration
) with
a
false color
co
mposite a
r
e
shown in Figu
re 3.
(a)
(b
)
(c
)
(d)
Figure 3. Pauli decom
po
sition and
spa
n
image of two
experim
ental
plots
((a
) and
(b) a
r
e thePauli d
e
com
p
o
s
ition
and sp
an im
age of Re
se
a
r
ch A
r
ea 1;(c) and (d
) are t
he
Pauli decomp
o
sition a
nd span imag
e of Re
sea
r
ch Are
a
2)
4.2. Textur
e extr
action
w
i
th
w
a
v
e
let d
ecomposi
t
ion
T11,T22,T
33,
which are t
he main dia
gonal
el
eme
n
ts of the pol
arim
etric coheren
cy
matrix are
shown in
Figu
re 4:
The
sp
an ima
ge
i
s
extracted
fro
m
T11,T2
2,T
33, an
d wavelet
decompo
sitio
n
is co
ndu
cte
d
on this ima
ge.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
An Im
proved
AP-Wi
shart Classifier for P
o
larim
e
tr
ic S
A
R Im
age by Incorpor
ating
.... (Chen Jun)
151
(a)
(
b
)
(c
)
(
d
)
(e)
(f
)
Figure 4. Main diago
nal el
ements of pol
arimetri
c coh
e
ren
c
e m
a
trix of two experi
m
ental plots
((a
)-(c) are T
11, T22, T33
im
age
s on ex
perim
ental on
e; (d)-(f
) a
r
e
T11, T22, T3
3 image
s on
experim
ental two)
Span image i
s
extracte
d from T11, T22,
T33, and wa
velet decom
p
o
sition i
s
con
ducte
d
with it.
4.3. The AP-Wishar
t
clas
sification
AP cluste
rin
g
wa
s p
e
rfo
r
med
usi
ng
texture extra
c
ted by the
wavelet ex
traction
algorith
m
, the cla
s
s n
u
m
ber
wa
s
set
manually. A
fter the AP
cl
usteri
ng,
clu
s
ter centers
were
acq
u
ire
d
. Ne
xt, a refined
Lee [25] filter wa
s
cond
uc
t
ed to elimi
nat
e sp
eckle
noi
se
with a
win
d
o
w
size of 7.Followin
g
this, Wisha
r
t iterative clu
s
te
rin
g
wa
s implem
e
n
ted to obtai
n the final re
sult.
For a
comp
arison of differe
nt algorithm
s,
SVM
was u
s
ed on texture
and pola
r
ime
t
ric informatio
n
to fulfill the supervi
sed
cla
ssifi
cation. F
o
r conv
eni
en
ce, when eva
l
uating cl
assif
i
cation a
c
cu
racy,
simila
r cla
s
se
s of un
supe
rv
ised
cla
ssifi
cation we
re
m
e
rge
d
and th
e followin
g
cl
assificatio
n
re
sult
image
s we
re
obtaine
d:
4.4. Classific
a
tion res
u
lt and analy
s
is
From Fi
gu
re
5 it can
be
observed th
a
t
the
H
/
Alpha
-Wish
a
rt
ca
nnot di
stingui
sh ‘field’
from ‘wate
r
’ in (b) an
d (c), and this rai
s
e
s
impo
rtan
t question
s a
bout the miscla
ssifi
cation
of
‘field’ into ‘water’. By contras
t, the AP-Wishart is a
b
l
e to avoid this situatio
n an
d perfo
rm bet
ter
on the
cla
s
sification
of ‘fiel
d
’ and
‘water’
.
From Fi
gu
re
5 (f) and
(g
)
raise furth
e
r i
s
sue
s
ab
out t
he
H
/
Alpha
-Wi
sh
art cl
assifier,
su
ch a
s
th
e fact that it
can
extract le
ss ‘
r
oad’
an
d ‘bui
lding’ tha
n
th
e
AP-Wi
sha
r
t.
After a field study on these
two experim
ental
plots, some pra
c
tical
measu
r
ed d
a
ta were
acq
u
ire
d
a
n
d
co
nsi
dered
as
sam
p
le
d
a
ta, to pe
rform an
accu
ra
cy a
s
sessme
nt. The
re
sul
t
is
sho
w
n in Ta
b
l
e 1.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No. 1, March 2
015 : 146 – 1
5
4
152
(a)
(
b
)
(c)
(d)
(e
)
(f
)
(
g
)
(
h)
Figure 5. Co
mpari
s
o
n
of
H
/
Alpha
c
l
assifier,
H
/
Alpha
-Wishart
classifier, SVM classifie
d
and
improve
d
AP-Wishart
classifier re
sult in two expe
rime
ntal plots
(a, b, c, d are respe
c
tively result of
H
/
Al
pha
,
H
/
Alpha
-Wishart, SVM, AP-Wis
h
art c
l
as
s
i
fier in
experim
ental
plot one; e, f,
g, h are corre
s
po
ndin
g
re
sult in experim
ental plot two
)
Table 1. Co
m
pari
s
ion
re
sul
t
s from differe
nt classifier o
n
two experi
m
ental plots
Classifier
Experimental plo
t
s
Total accurac
y
(
%
)
Kappa coefficient
H/Alpha-Wishart
one 73.73
0.52
tw
o
79.87
0.68
AP-Wishart
one 85.95
0.77
tw
o
90.38
0.85
SVM
one 86.56
0.78
tw
o
88.53
0.82
Comp
ared with the
H
/
Alpha
-Wi
s
ha
rt cla
ssifie
r
, the AP-Wi
sha
r
t classifier
can be
con
s
id
ere
d
to
greatly en
ha
nce
cla
s
sifica
tion ac
cu
ra
cy from Ta
ble1.
Its accu
ra
cy even exceed
s
that of the popular
sup
e
rvi
s
ed SVM cla
ssifie
r
.
The AP-Wi
sha
r
t is an un
supe
rvi
s
ed
cla
ssifi
ca
tion
scheme; it is
more
conve
n
i
ent and effici
ent than the SVM.
5. Conclusio
n
In this pap
e
r
a
brand
n
e
w AP
-Wi
sh
art
cla
ssifi
ca
tion sch
e
me
wa
s propo
sed. By
synthe
sizi
ng
polarim
etri
c and textua
l informat
ion
,
which
wa
s extra
c
ted
by the wa
velet
decompo
sitio
n
algorith
m
, a high pe
rformance AP
cl
usteri
ng alg
o
r
ithm wa
s co
mbined
with
the
cla
ssi
c Wi
sh
art iterative clu
s
terin
g
al
gorithm
to form a po
we
rful unsupe
rvised AP-Wishart
clu
s
terin
g
al
gorithm. Exp
e
rime
nts
were co
ndu
cted
on two
re
search a
r
e
a
s and the
re
sults
demon
strated
that the AP-Wi
sha
r
t achieved
a hi
gher
cla
ssifi
cation a
c
curacy and Ka
ppa
coeffici
ent than the
H
/
Alpha
-Wisha
rt and the supe
rvised SVM. T
herefore, it
can b
e
con
c
l
uded
that:
Texture extracted by a
wavelet de
comp
osit
io
n
algorith
m
ca
n assist p
o
l
a
rimetri
c
informatio
n to improve cl
assificatio
n
perf
o
rma
n
ce;
An AP algorithm on texture information
can obt
ain b
e
tter initial cl
uster
re
sult than the
H/Alpha cl
assification sch
e
m
e.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
An Im
proved
AP-Wi
shart Classifier for P
o
larim
e
tr
ic S
A
R Im
age by Incorpor
ating
.... (Chen Jun)
153
The AP-Wi
sh
art cl
assification sch
e
me i
s
more
suitabl
e for u
n
supe
rvised
cla
ssifi
cation of
PolSAR imag
e than the
H
/
Alpha
-W
is
hart.
Ackn
o
w
l
e
dg
ements
This
re
sea
r
ch
is
sup
porte
d
by the
Natio
nal Natural S
c
ien
c
e
Fou
n
d
a
tion of
Chin
a und
er
Grant
No.41
1713
23, a
p
r
oje
c
t fun
ded
by the
Prio
rity Acad
emi
c
Prog
ram
Develo
pment
of
Jian
gsu
Hig
her E
d
u
c
atio
n Institution
s
und
er
Gran
t No. SZBF
2011
-6
-B35,
and
the
China
Geolo
g
ical Survey Proje
c
t unde
r Grant No. 121
201
1
1202
29.
Referen
ces
[1]
Lee JS, Grun
es MR, K
w
ok
R. Classificat
i
on of
Multi-l
o
ok Polar
i
metri
c
SAR Imager
y Bas
ed on
Compl
e
x W
i
sh
art Distributi
on.
Internation
a
l J
ourn
a
l of Re
mote Sensi
ng.
1
994; 1(1
5
): 229
9-23
11.
[2]
Lee JS, Grun
e
s
MR, Ains
w
o
r
t
h T
L
. Unsupe
rvised C
l
assifi
cation Us
in
g Polarim
e
tric De
compos
iti
o
n
and th
e C
o
mpl
e
x W
i
s
hart Cl
a
ssifier.
IEEE Transactions
on
Geosci
ence
and Remote Sensing
. 19
99
;
37(5): 22
49-
22
58.
[3]
Lee JS, Grun
es MR, Pottier E, et al. Unsuper
v
i
sed T
e
rrain Cl
assificati
on Preservi
ng
Polarimetri
c
Scattering Ch
aracteristics.
I
EEE Transactions on Geosc
i
enc
e and Rem
o
te Sens
ing
. 200
4; 42(
4):
722-
731.
[4]
Pottier E, Lee JS.
Appl
ic
ation
of the
H/A/alph
a P
o
lari
metric De
compos
ition T
heor
e
m
for
Unsu
pervis
ed
Classific
a
tio
n
of Fully Polar
i
metric
SAR D
a
ta Based
on
the W
i
shart Distributi
o
n
.
Procee
din
g
of Committee o
n
Earth Observin
g
Satell
ites SAR W
o
rkshop. T
oul
ouse. 1
999.
[5]
Clou
de S
R
, P
o
ttier E. A Re
vie
w
of T
a
rget Decom
positi
o
n T
heorems i
n
Rad
a
r Pol
a
ri
metr
y
.
IEEE
T
r
ansactio
n
s o
n
Geoscie
n
ce
and R
e
mote S
ensi
n
g
. 19
96; 34(2): 49
8-5
1
8
.
[6]
Cloude SR, Pottier E.
An E
n
trop
y B
a
sed
Cl
a
ssificatio
n
Sch
e
me for
La
nd
Appl
icatio
ns
of Pol
a
rimetr
i
c
SAR.
IEEE Tra
n
sactio
ns on G
eosci
ence
and
Re
mote Se
nsi
n
g
. 199
7; 35(1)
: 68-78.
[7]
Pottier
E.
T
he H/A/
α
Polari
metric Dec
o
mp
o
s
ition A
ppr
oac
h Ap
pli
ed to
P
o
lSAR
Data Pr
ocessi
ng
. In
Procee
din
g
s PIERS W
o
rksho
p
Advanc
es in
Rad
a
r Method
s. Baveno. 19
9
8
; (7):120-1
22.
[8]
Laur
ent F
F
,
Pottier E, L
ee J
S
. Unsu
pervis
ed C
l
ass
i
ficati
on
of Multifre
q
uenc
y
an
d F
u
ll
y P
o
l
a
rimetri
c
SAR Images
Based o
n
the
H/A/Alpha
–W
is
hart Class
ifier.
IEEE Transactions on Geoscienc
e and
Rem
o
te Sensing
. 200
1; 39(1
1
): 2332-
23
42.
[9]
Kimura
K,
Ya
maguchi Y,
Yamada H.
PISAR Imag
e
Analysis
Usi
ng
Polarimetric Scattering
Param
e
ters
and Total Power
. 2003 Inter
n
a
t
iona
l Geosci
e
n
ce An
d Rem
o
te Sens
ing S
y
mp
osi
u
m.
T
oulouse. 20
0
3
; 1: 425-4
27.
[10]
Cao
F
,
Ho
ng
W
,
W
u
Y.
An
Improv
ed
Cl
ou
de-Pottier
D
e
c
o
mpos
ition
Usi
ng
H SPA
N a
nd
Co
mp
l
e
x
W
i
shart Cl
assif
i
er for Po
la
rimetric SAR Classification
. Ra
d
a
r, 200
6. CIE '
06. Intern
ation
a
l C
onfere
n
c
e
on T
opic(s): Aerosp
ace, C
o
mmunic
a
tio
n
, Net
w
o
r
ki
ng
& Broadcasti
n
g
, Engin
eer
ed
Materials,
Dielectrics & Plasmas, Fields,
Waves
& Electro- magn
etics. 200
6: 1-4.
[11]
W
u
Z
C
, Ouya
ng QD, Sun
XA, et al. Improved Ite
rative
W
i
shart Class
ificatio
n Alg
o
rith
m Based o
n
Polarim
e
tric Scattering C
har
acteristics Pre
s
ervatio
n
.
Science of Surv
eyin
g and Ma
ppi
ng
. 20
11
;
11(3
6
):16
1-16
3.
[12]
Yang
J, Shi
L, Li P
X.
W
i
shart-H/Alp
ha
Classific
a
tio
n
Based
on
SP
AN an
d B
e
st Polar
i
metric
Coh
e
renc
y.
Ge
omatics an
d Informatio
n
Scien
c
e of W
uhan U
n
iversity
. 20
12;
1(37): 22-2
5
.
[13]
Z
hang B, Ya
n
g
R, Xi
e
X, et al. Polar
i
metric
SAR Image C
l
assificati
on B
a
sed o
n
Pol
a
ri
metric T
a
rget
Decom
positi
o
n
and W
M
RF
.
Geomatics an
d Information
Scienc
e of W
uhan U
n
ivers
i
ty
. 2011; 3(3
6
):
297-
300.
[14]
Yang J, La
n
g
F
K
, Li DR
. A Polarimet
r
ic SAR Image Cl
assificati
on Util
izin
g
Clou
de-P
o
ttier
Decom
positi
o
n
and Po
larim
e
tric W
h
itenin
g
F
ilte.
Geoma
t
ics and Infor
m
ati
on Sci
enc
e of W
uhan
Univers
i
ty
. 201
1; 1(36): 10
4-1
07.
[15]
Yang
L, Li
u W
,
W
ang Z
G. W
e
ighte
d
-bas
ed
Unsu
per
vis
ed
W
i
shart Class
ificatio
n of F
u
ll
y Polarim
e
tri
c
SAR Image.
Jo
urna
l of Electro
n
ics an
d Informati
on T
e
ch
no
logy
. 20
08; 12(
30): 282
7-2
830
.
[16]
Zhao L
L
, YANG J, Li PX
. Statistical C
l
assif
i
catio
n
of Wea
k
Backscatteri
ng Scatterers
of PolSAR
Image.
Journ
a
l
of Remote Se
nsin
g
. 201
3; 17(2): 307-
31
8.
[17]
Lan
g F
K
, Yang J, Z
hao LL. Rese
arch o
n
Classi
fic
a
tio
n
of SAR Image
Based o
n
F
r
eeman Scatt
e
r
Entrop
y an
d A
n
isotro
p
y
.
Acta
Geodaetic
a et Cartogr
aph
ica
Sinic
a
.
201
2; 41(4): 556-
56
2.
[18]
Liu M, Z
h
a
ng
H, W
ang C.
A
pplyi
ng t
he L
o
g
-Cu
m
u
l
a
n
ts o
f
T
e
xture Para
meter t
o
F
u
lly
Polari
metri
c
SAR Class
ific
ation
Usin
g
Supp
ort Vector Machi
nes
Classifi
er
. 2011 IEEE CIE
International
Confer
ence
on
Radar. 20
11;
11:72
8-7
31.
[19]
Pottier
E.
U
n
superv
i
sed
Cl
as
sificatio
n
Sch
e
m
e
an
d T
o
po-
grap
hy D
e
rivat
i
on
of Po
lSAR
Data B
a
se
d
on the H/Al
p
ha/A Pol
a
ri
me
tric Deco
mp
os
ition T
h
e
o
re
m
.
Proceed
in
gs
of the 4th Internati
o
n
a
l
W
o
rkshop o
n
Rad
a
r Polar
i
m
e
tr
y
.
19
98; 7: 5
35-5
48.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No. 1, March 2
015 : 146 – 1
5
4
154
[20]
Pottier E, Lee
JS.
Appl
icatio
n
of the
H/A/
α
Polari
metric D
e
co
mp
ositio
n
T
heore
m
for U
n
sup
e
rvise
d
Classific
a
tio
n
of Fully Po
lari
metric SA
R D
a
ta Base
d o
n
the Wishart
Distributi
o
n
. In
Procee
din
g
Committee o
n
Earth observ
i
n
g
Satell
it
es SAR
w
o
rksh
op. T
oul
ouse. 1
999.
[21]
Cho
y
SK, T
ong CS. Statistic
a
l W
a
v
e
let S
u
bba
nd
Ch
aract
e
rizati
on B
a
se
d o
n
Gen
e
ra
li
zed Gamm
a
Densit
y a
nd Its Applic
atio
n i
n
T
e
xture Retr
ieval.
IEEE Transactions. on Im
age Processing
. 20
10
;
19(2): 28
1-2
8
9
.
[22]
Z
hou SR, Yin
JP. Lbp T
e
xtur
e F
eat
ure Bas
ed o
n
Ha
ar Ch
aracteristics.
J
ourn
a
l of Softw
are
. 20
13; 8:
190
9-19
26.
[23]
Fre
y
BJ, Du
ec
k D. Clusteri
n
g
b
y
P
a
ssing M
e
ssages Bet
w
een Dat
a
Points.
Science.
2
0
0
7
; 315(
581
4):
972-
976.
[24]
Xi
ao Y, Y
u
J. Semi-Su
pervi
sed C
l
uste
ri
ng
Based
on Af
finit
y
Prop
ag
ation A
l
gor
ithm.
Jo
u
r
na
l
of
Software
. 2008
; 19(11): 28
03-
281
3.
[25]
Lee JS, Grun
es MR, Gran
di GD. Pol
a
ri
metric
SAR S
peckl
e F
ilteri
n
g an
d Its implicati
on for
Classification.
IEEE Transactions on Geosci
ence and Remote Sensing
. 1
999; 37(
5): 236
3-23
73.
Evaluation Warning : The document was created with Spire.PDF for Python.