TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.1, Jan
uary 20
14
, pp. 286 ~
291
DOI: http://dx.doi.org/10.11591/telkomni
ka.v12i1.3969
286
Re
cei
v
ed
Jun
e
23, 2013; Revi
sed Aug
u
st
17, 2013; Accepted Sept
em
ber 16, 20
13
Sensitivity of Support Vector Machine Classification
to Various Training Features
Nanh
ai Yang
, Shuang Li*, Jing
w
e
n Liu
,
FulingBian
Internatio
na
l Schoo
l of Soft
w
a
re, W
uhan U
n
iversit
y
# 37 Lu
o
y
u R
o
ad, W
uhan, C
h
ina,
43
00
79, F
a
x: +
86-
27-6
8
7
782
21
*corres
pon
di
ng
author, e-mai
l
: lishu
ang
12
9@
gmail.c
o
m
A
b
st
r
a
ct
Re
mote se
nsi
ng i
m
a
ge cl
a
ssificatio
n
is one of
the
most i
m
p
o
rtan
t techniq
ues i
n
imag
e
interpr
e
tatio
n
, w
h
ich
ca
n be u
s
ed
for
env
iro
n
m
e
n
tal
mon
i
tor
i
ng, eval
uati
on and
pre
d
icti
on.
Many
al
gorit
h
m
s
have
be
en
dev
elo
ped
for i
m
a
ge cl
assificati
o
n
in t
he l
i
teratu
re. Supp
ort ve
ctor machi
ne (
SVM) is a ki
nd
o
f
superv
i
sed c
l
a
ssificatio
n
that
has b
e
e
n
w
i
del
y used
r
e
cently
. T
he classific
a
tion acc
u
racy p
r
oduc
ed by
SV
M
m
a
y sh
o
w
va
ria
t
i
o
n
de
pe
nd
i
n
g
o
n
th
e
ch
o
i
ce
o
f
tra
i
n
i
n
g
fea
t
u
r
e
s
. In
th
i
s
p
a
p
e
r
, SVM w
a
s u
s
e
d
fo
r land
cover c
l
assific
a
tion
usi
n
g
Quickbir
d
i
m
ag
es. Spectr
a
l
and
textura
l
f
eatures
w
e
re
extracted
for
th
e
classificati
on
a
nd th
e res
u
lts
w
e
re an
aly
z
e
d
thor
oug
hly.
Results s
how
e
d
that th
e n
u
m
b
e
r of fe
atur
es
empl
oyed i
n
S
V
M w
a
s not the more the b
e
tter. Different
features ar
e su
itabl
e for differ
ent type of la
n
d
cover extracti
o
n
. T
h
is study v
e
rifies th
e effe
ctiveness
and
robustn
ess of
SVM in the cl
a
ssificatio
n
of hi
gh
spatia
l resol
u
ti
on re
mote se
n
s
ing i
m
ages.
Ke
y
w
ords
: Re
mote s
ensi
ng, i
m
a
ge cl
assific
a
tion, su
p
port
vector mac
h
i
n
e
,
feature extraction
Copy
right
©
2014 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
High
spatial
resol
u
tion remote se
nsi
ng image
s
have played
an importa
nt role in
mappin
g
,
u
r
b
an
pl
anni
ng, defen
se and
military,
land
use
an
d
surv
eys, an
d m
a
n
y
other area
s [1
-
3]. As the im
provem
ent of
spatial
re
sol
u
tion, sin
g
le l
and
cover sh
ows a lot of
different
spe
c
tral
value, whi
c
h i
n
creasing the probab
ility of miscl
assification. The simi
lar spect
r
al characteristi
c
s of
different land
covers often lead to co
nfusin
g
in cl
assificatio
n
, su
ch a
s
sh
a
dows an
d water
bodie
s
, me
ad
ows a
nd t
r
ee
s, a
r
e
often
mixed in
spe
c
tral
value. T
hus, it i
s
ha
rd to o
b
tain
hi
gh
cla
ssifi
cation accuracy whe
n
only
the sp
ectral
in
forma
t
ion is u
s
e
d
.
Comp
ared
wi
th the traditio
nal
cla
ssifi
cation
method
s, Su
pport V
e
cto
r
Machi
ne
(SV
M
) p
o
sse
s
se
s the
me
rits
of learning
with
small
sa
mple
s, hig
h
anti
-
n
o
ise
pe
rform
ance, et
c. M
o
reove
r
, SV
M also h
a
s the a
d
vantag
es
of
high lea
r
nin
g
and promotio
n efficien
cy. Therefor
e, S
V
M classification sh
owed g
ood pe
rform
a
nce
in remote
sen
s
ing ima
ge in
formation extraction [4-6].
In this stu
d
y, SVM wa
s u
s
ed fo
r lan
d
cover
cla
s
sification
of Wu
han di
stri
ct in Chi
n
a
usin
g Qui
c
kbird ima
g
e
s
. Variou
s sp
ectral
and
t
e
xtural featu
r
es
we
re ex
tracted fo
r
SVM
cla
ssifi
cation
pro
c
e
s
s a
nd
cla
ssifi
cation
perfo
rman
ce
s we
re
an
alyzed tho
r
o
ughl
y. It shoul
d b
e
pointed o
u
t that the sel
e
ction of fe
ature
s
ha
s
an effect on
the perfo
rmance of SVM.
Determinatio
n of th
eir
o
p
timum
com
b
ination
s
i
s
rega
rd
ed
as critical
for
the succe
ss of
cla
ssif
i
cat
i
on.
2. Support V
ector M
achi
n
e Algorith
m
Suppo
rt vect
or ma
chi
ne (SVM) is
sup
e
rvis
e
d
he
uri
s
tic al
gorith
m
based on
st
atistical
learni
ng the
o
ry [7]. The aim
of SVM for cl
assificati
o
n
is to determi
ne
a hyper
plan
e
that optimall
y
sep
a
rate
s two cla
s
ses. A
n
optimum h
y
per pla
ne
is determin
ed
usin
g trainin
g
data set
s
an
d is
verified usi
n
g
test data set
s
.
Assu
me dat
a set
11
,
,
...,
,
,
...
,
,
,
1
,
1
ii
N
N
i
xy
x
y
x
y
y
, where
N
is the
numbe
r of sa
mples,
i
x
is the training
sam
p
le,
i
y
is the cla
s
s label of
i
x
. Optimum hyper plane
is used to ma
ximize the m
a
rgin b
e
twe
e
n
cla
s
ses. Th
e hyper pla
n
e
is defined a
s
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 2302-4
046
TELKOM
NIKA
TELKOM
NIKA
Vol. 12, No
. 1, Janua
ry 2014: 286 – 2
9
1
287
0
wx
b
(1)
whe
r
e
x
is a
poi
nt lying on th
e hype
r plan
e
,
w
determin
e
s
the ori
entatio
n of the hyp
e
r
plan
e,
b
is the
bias that indi
cate
s the di
stan
ce
betwe
en hy
per
plane
an
d the o
r
igin.
For the
linea
rly
sep
a
ra
ble ca
se, the hyper
plane i
s
defin
ed as
1
ii
yw
x
b
(2)
The trai
ning
data poi
nts
on the hyp
e
r plane
s a
r
e
parall
e
l to th
e optimum
h
y
per pl
ane.
The
sup
port ve
cto
r
sa
re
define
d
by the fun
c
ti
on
1
i
wx
b
. If a hyper plane
exist
s
and
sati
sfies
Eq. (2), the
cl
asse
s are lin
early
sepa
ra
b
l
e. Therefore,
the margin b
e
twee
n the
s
e
plane
s i
s
eq
ual
to
2/
w
. Thus, the
optimum hyp
e
r pl
a
ne can
be found
by minimizi
ng
2
w
under th
e con
s
traint
Eq. (2).
Determination of
o
p
timum hype
r plane i
s
e
qui
valent to solv
e optimi
z
atio
n problem
given
by:
2
1
mi
n
2
w
(3)
As no
nline
a
rl
y sep
a
ra
ble
data i
s
the
case
in va
riou
s
cla
ssifi
catio
n
s
of re
mote
sen
s
in
g ima
g
es,
the SVM techniqu
e can b
e
extende
d to allo
w fo
r
n
online
a
r d
e
ci
sion
su
rfaces by introd
uci
n
g
penalty para
m
eter
C
and sl
ack variabl
e
:
2
1
1
mi
n
2
N
i
i
wC
(4)
subj
ect to co
nstrai
nts,
1
1
,
2
,
...,
,
0
ii
i
i
yw
x
b
iN
(5)
whe
r
e pe
nalt
y
paramete
r
C
allows stri
king
a balan
ce be
tween two co
mpeting crite
r
ia of margi
n
maximizatio
n
and
erro
r mi
nimizatio
n
, where
a
s the
sl
ack va
riable
i
indicate the
di
stan
ce
of the
inco
rrectly cl
assified point
s
from
the
o
p
t
imal hype
r pl
ane. T
he l
a
rg
er th
e
C
value, the hi
ghe
r the
penalty asso
ciated to miscl
assified samp
les.
Whe
n
it i
s
no
t possibl
e to
define th
e hy
per pl
an
e
by
linear eq
uatio
ns, the
d
a
ta
may be
map
ped
into a hig
h
e
r
dimen
s
ion
a
l space through
som
e
no
nlin
ear m
appi
ng
function
. The
input poi
nt
x
can be rep
r
e
s
ente
d
by
x
in high-dimen
s
i
onal
spa
c
e.
The time
-con
sumin
g
com
putation of
i
x
x
is
red
u
ce
d
by usi
ng
a
ke
rn
el fun
c
tion. T
hus, th
e
cla
s
sificatio
n
d
e
ci
sion
fun
c
tion
is
defined a
s
:
1
sg
n
N
ii
i
i
f
xy
K
x
x
b
(6)
whe
r
e
sgn
is the sign fu
ncti
on,
K
is the kernel fu
nctio
n
and the m
agnitud
e
of
i
is
determi
ned
by the p
a
ra
meter
C
. The
widely
used
ke
rnel
fun
c
tion in
clud
es linea
r
ke
rn
el,
polynomial
ke
rnel, sigm
oid kernel and
G
aussia
n
radi
a
l
ba
sis
k
e
rn
el
.
B
y
cont
r
a
st
t
e
st
s,
t
h
e
line
a
r
kernel fun
c
tio
n
obtaine
d be
tter results in
our stu
d
y.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 2302-4
046
Sensitivit
y of Suppo
rt Vector Ma
chin
e Classifi
catio
n
to Variou
s Tra
i
ning Featu
r
e
s
(Shu
ang Li
)
288
3. Spectral a
nd Textural Feature Extr
action
s
Many al
gorit
hms have
b
een
develo
p
ed fo
r ima
g
e
cla
s
sificatio
n
u
s
ing
SVM
. Several
factors (e.g. t
r
ainin
g
features,
ke
rn
el function
s, win
d
o
w si
ze
s) ha
ve significant
impact
s
on t
h
e
cla
ssifi
cation
perfo
rman
ce,
which sh
ould
be con
s
id
er
e
d
carefully by the analyst. The sel
e
ctio
n
of
approp
riate t
r
ainin
g
featu
r
es d
epe
nd
s
on the
kn
o
w
l
edge
of land
cove
r types pre
s
e
n
t in the
image by
g
eograph
ers. Thus,
trai
nin
g
features
selectio
n play
an imp
o
rta
n
t role i
n
the
cla
ssif
i
cat
i
on
ac
cur
a
cy
[
8
]
.
3.1. Spectral
Featur
e Extr
action
The wi
dely u
s
ed
sp
ectral
feature i
s
me
an val
ue a
n
d
the metri
c
d
e
rived from spectral
value, i.e. Normali
z
e
d
Dif
f
eren
ce Ve
g
e
tati
on Index
(NDVI), Ratio Index (RI), Soil Adjust
ed
Vegetation Index (SAVI), Normali
z
ed
Diff
erence
Water Index (NDWI
)
. The m
e
tric
is described i
n
Table 1.
Table 1. The
spe
c
tral featu
r
es u
s
e
d
in the study
Metric Equation
Description
NDVI
Re
Re
NI
R
d
NI
R
d
B
and
B
a
nd
B
and
B
a
nd
It is used to extra
c
t vegetation, i.e. grassland.
RI
Re
d
N
IR
B
and
B
and
It is used to extra
c
t high density
v
egetation, i.e. tre
e
s.
SAVI
Re
Re
1
NI
R
d
NI
R
d
B
and
B
and
L
B
and
B
and
L
It is used to extra
c
t soil w
i
th low
v
egetation cover.
NDWI
G
r
een
N
I
R
G
r
een
N
I
R
B
a
n
d
B
and
B
a
n
d
B
and
It is used to extra
c
t w
a
te
r from
land covers.
3.2. Textur
al Features Ex
traction
The G
r
ay L
e
ver
Co-occu
r
rence Matrix
(GLCM
)
i
s
p
r
o
posed by
Ha
ralick in
197
0s, whi
c
h
is a
n
imp
o
rta
n
t tech
niqu
e
to analy
z
e im
age textu
r
e.
The
GL
CM i
s
ba
sed
on
th
e second
ord
e
r
combi
nation of
probability density
functi
on, by calcul
ating the
co
rrelation between two
point
s in
the e
s
timated
image
s [9,10
]. The texture
features
are
derived
from
GLCM,
i.e. Angula
r
Se
co
n
d
Moment (AS
M
), Contrast,
En
tropy and
Correlation.
Let
,
Gi
j
be the element in GL
CM and
the size of the matrix be
*
kk
, the metri
c
are
descri
bed in
Table 2.
Table 2. The
textural features u
s
ed in th
e study
Metric Equation
Description
ASM
2
11
,
kk
ij
Gi
j
It denotes the im
age gra
y
uniform
ity
and te
xture
coarseness.
Contrast
1
2
||
0
,
k
ij
n
i
nG
i
j
It reflects the text
ure clarit
y
.
Entrop
y
11
,l
g
(
,
)
kk
ij
Gi
j
G
i
j
It measures the
amount of inform
ation contained in
the image.
Correl
a
ti
on
11
11
11
2
2
11
2
2
11
*,
,,
,
,
,
kk
ij
ij
ij
kk
k
k
ij
ij
i
j
kk
ii
ij
kk
jj
ij
ij
G
i
j
U
U
SS
Ui
G
i
j
U
j
G
i
j
SG
i
j
i
U
SG
i
j
j
U
It describes the periodicity
of t
e
xtu
r
e element in a
certain positional relationship.
4. Experimental Re
sults
and An
aly
z
e
The test ima
ge is from Q
u
ickbird sen
s
or, wi
th the
spectral ba
nd
rang
es f
r
om
450nm to
900
nm. Th
e
image
si
ze
i
s
4
00*4
00,
which
covers
water (W),
grasslan
d (GL
)
, bare la
nd
(B
L),
blue ro
of (BlR), brown
ro
of (BrR
), ce
ment su
rface
(CS) an
d trees
(T). The
original ima
ge is
s
h
ow
n
in
F
i
gu
r
e
1
.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 2302-4
046
TELKOM
NIKA
TELKOM
NIKA
Vol. 12, No
. 1, Janua
ry 2014: 286 – 2
9
1
289
Figure 1. The
test image wi
th the size of
400*4
0
0
Five types of features are sele
cted for t
he SVM classificatio
n
. The cla
s
sifiers a
r
e
SVM_1 (fo
u
r
feature
s
: Me
an valu
e fo
r t
he fou
r
ban
d
s
), SVM_
2
(six feature
s
: f
our mea
n
val
ue,
NDVI and SA
VI), SVM_3 (eight features: four
mean value, NDVI,
RI, SAVI and NDWI), SVM_4
(twelve featu
r
es: fou
r
me
an value, m
ean val
ue a
nd stan
da
rd
deviation of ASM, Contra
st,
Entropy, Correlation), and SVM_5
(sixt
een features:
four
mean
value, NDVI, RI, SAVI and
NDWI, mea
n
value an
d sta
ndard deviati
on of ASM,
Contra
st, Entro
p
y, Correlatio
n). The
num
b
e
r
of training dat
a is 60 (bl
o
ck) * 7 (cl
a
sse
s
). The cla
ssifie
d
image
s are sho
w
n in Fig
u
re 2.
(a) SVM_1
(b) SVM_2
(c
) SVM_3
(d) SVM_4
(e) SVM_5
Figure 2. The
classified im
age
s of the study area
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 2302-4
046
Sensitivit
y of Suppo
rt Vector Ma
chin
e Classifi
catio
n
to Variou
s Tra
i
ning Featu
r
e
s
(Shu
ang Li
)
290
From Fi
gure
2, it can b
e
se
en that
t
he SVM classifiers pe
rform well on
image
cla
ssifi
cation.
The water,
blue ro
of and bro
w
n ro
of are dete
c
te
d accurately. As it is hard to
disting
u
ish b
e
twee
n
wate
r and
sha
dow, we
cl
assi
fy them
a
s
a cl
ass.
To
com
pare
the
results
from different
SVMs, we
calcul
ate the o
v
erall a
c
cura
cy (OA
)
an
d
Kappa
coefficient by co
nfu
s
ion
matrix, which is sh
own in Table 3 to Tabl
e 7. T
he test data set
s
we
re formed u
s
in
g rand
om pixel
sele
ction
st
ra
tegy with
pro
porti
on
al n
u
m
ber of
sam
p
les for ea
ch
cla
s
s. A tota
l num
ber of
350
pixels are sel
e
cted fo
r testi
ng cla
s
sified i
m
age
s.
Table 3. Co
nfusio
n matrix for SVM_1
SVM_1
W
G
L
BL BlR
BrR
CS
T
W 33
2
0
0
1
5
0
GL
3
27
0
1
1
0
1
BL 0
0
43
0
6
3
1
BlR
0
0
3
12 0
1 0
BrR
0
0
0
0 13
1 0
CS 9
0
6
2
2
161
0
T 0
0
0
0
1
0
12
OA
86%
Kappa
0.8
Table 4. Co
nfusio
n matrix for SVM_2
SVM_2
W G
L
BL
BlR
BrR
CS
T
W 33
2
0
0
1
5
0
GL
3
27
0
1
1
0
1
BL 0
0
42
0
7
3
1
BlR
0
0
3
12 0
1 0
BrR
0
0
0 0
13 1
0
CS 9
0
6
2
2
161
0
T 0
0
0
0
1
0
12
OA
86%
Kappa
0.8
Table 5. Co
nfusio
n matrix for SVM_3
SVM_3
W
G
L
BL BlR
BrR
CS
T
W 34
2
0
0
1
4
0
GL
3
27
0
1
1
0
1
BL 0
0
48
0
4
0
1
BlR
0
0
3
12 0
1 0
BrR
0
0
1
0 12
1 0
CS 9
0
0
2
2
167
0
T 0
0
0
0
1
0
12
OA
89%
Kappa
0.84
Table 6. Co
nfusio
n matrix for SVM_4
SVM_4
W G
L
BL
BlR
BrR
CS
T
W 34
3
0
0
0
4
0
GL
3
27
0
1
1
0
1
BL 0
0
45
0
7
0
1
BlR
0
0
3
12 0
1 0
BrR
0
0
1 0
12 1
0
CS 9
0
3
2
2
164
0
T 0
0
0
0
1
0
12
OA
87%
Kappa
0.82
Table 7. Co
nfusio
n matrix for SVM_5
SVM_5
W
G
L
BL
BlR BrR
CS
T
W 34
3
0
0
0
4
0
GL
3
27
0
1
1
0
1
BL 0
0
49
0
3
0
1
BlR 0
0
3
12
0
1
0
BrR
0
0
0 0
13 1
0
CS 9
0
0
2
2
167
0
T 0
0
0
0
1
0
12
OA
90%
Kappa
0.85
Figure 3. The
accura
cy tre
nd for re
sult
s image
s
To comp
are
the overall a
c
cura
cy a
nd
kap
pa
coefficients fo
r e
a
ch cl
assified i
m
age
s
more o
b
viou
sly, the trend is given in Fig
u
re 3. Fr
o
m
Figure 3, it can
be see
n
that the accuracy
of
SVM_1 and SVM_2 is the same. Ho
we
ver, the numb
e
r
of feature
s
used in SVM_2 is larg
er th
an
SVM_1. It indic
a
tes
that
the NDVI and
SAVI in SVM
_2 do not help to increas
e
the c
l
ass
i
fication
accuracy. Th
e accu
ra
cy from SVM_3 i
s
larg
er th
an S
V
M_2. It indicates that th
e
RI and
NDWI in
SVM_3 incre
a
se the
cla
s
si
fication a
c
curacy. The a
ccura
cy for SVM_4 de
crea
ses, which m
e
ans
the texture f
eature
s
are
not as goo
d
as the
sp
e
c
tral index, i.e. RI, NDWI.
The a
c
curacy for
SVM_5 i
s
the
high
est,
whi
c
h i
s
1% hi
g
her than
t
he
result of SV
M_3. It me
an
s that
the tex
t
ure
adde
d in the cla
ssifie
r
s d
o
not signifi
cant
ly
improve the cla
ssifi
catio
n
accuracy.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 2302-4
046
TELKOM
NIKA
TELKOM
NIKA
Vol. 12, No
. 1, Janua
ry 2014: 286 – 2
9
1
291
5. Conclusio
n
Cla
ssifi
cation
of remote
sen
s
in
g im
age
s is a
n
importa
nt appli
c
ation f
o
r ima
ge
interp
retation.
Suppo
rt vect
or ma
chi
ne (SVM) hav
e b
een recently use
d
for ma
n
y
classificatio
n
probl
em
s. Althoug
h it is re
ported
that S
V
M pro
d
u
c
e
more
accu
rat
e
cla
s
sificatio
n
re
sult
s tha
n
the
conve
n
tional
method
s, the sele
ct
ion of optimum trai
ning features
is one of the most import
ant
issue
s
that
affect thei
r p
e
rform
a
n
c
e.
In this
stu
d
y, five types
of features a
r
e
used i
n
t
h
e
cla
ssifie
r
s. F
r
om the exp
e
r
iment
s, sev
e
ral im
po
rtan
t con
c
lu
sion
s can
be
dra
w
n. Firstly, the
numbe
r
of fe
ature
s
i
s
not
the mo
re th
e
better for th
e cl
assificatio
n
a
c
curacy, i
.
e. SVM_3 a
n
d
SVM_5. Secondly, the RI and NDWI feature
s
pe
rfor
m better than
the texture feature
s
, incl
u
d
ing
ASM, Entropy, Contras
t
and Correlation. This
co
ncl
u
sio
n
s ma
de here a
r
e ba
sed on the limited
tests. More comprehensiv
e tests
will be conducted in the future.
Ackn
o
w
l
e
dg
ement
This
re
sea
r
ch wa
s
su
pp
orted
by a
grant f
r
om
973 p
o
rj
ect
in China
(Grant
#
2012
CB71
99
01)
Referen
ces
[1]
F
ood
y GM, Mathur A. A rel
a
tive eva
l
u
a
tio
n
of
multiclas
s
image cl
ass
i
ficatio
n
b
y
s
u
pport vector
machi
nes.
IEEE Transactions
on Geoscie
n
c
e
and R
e
mote
Sensi
n
g
. 20
04;
42: 133
5-13
43
.
[2]
Han F
,
Li H, W
en C, Z
hao
W
.
A Ne
w
Incr
ementa
l
Sup
p
o
r
t Vector Machine Al
gorit
hm.
TEL
K
OMNIKA
Indon
esi
an Jou
r
nal of Electric
al Eng
i
ne
eri
n
g
.
2012; 1
0
: 117
1-11
78.
[3]
Yu Y, Z
hou L. Acoustic Emission Si
gn
a
l
Classific
a
tio
n
base
d
on
Supp
ort Vector Machi
n
e
.
T
E
LKOMNIKA Indon
esi
an Jou
r
nal of Electric
al Eng
i
ne
eri
n
g
.
2012; 1
0
: 102
7-10
32.
[4]
Shao Y, Lu
nett
a
RS. Compar
i
s
on of sup
port vect
or machi
n
e
,
neural
net
w
o
r
k
, and CART
algorithms fo
r
the lan
d
-cov
er classificati
on u
s
ing lim
ited trai
nin
g
data p
o
int
s
.
ISPRS Journal of Phot
ogr
ammetry a
n
d
Rem
o
te Sensing
. 201
2; 70: 7
8
-87.
[5]
Otukei J, Bl
asc
h
ke T
.
Lan
d co
ver ch
ang
e
as
sessment
usin
g d
e
cisi
on tre
e
s
, supp
ort vect
or mac
h
in
es
and ma
ximum
likel
ih
ood cl
as
sificatio
n
alg
o
ri
thms.
Internationa
l Jour
nal of
Appli
ed Earth
Observati
o
n
and Geo
i
nfor
mation
. 20
10; 12
: S27-S31.
[6]
Mountrak
i
s G, Im J, Ogole C. Supp
ort vect
or machi
nes i
n
re
mote sensi
ng:
A revie
w
.
ISPRS Journal of
Photogr
a
m
met
r
y and Re
mote
Sensin
g
. 201
1
;
66: 247-2
59.
[7]
Vapn
ik VN. A
n
overvi
e
w
of statistical
le
arni
n
g
the
o
r
y
.
IEEE
T
r
ansactio
n
s
on N
eur
al N
e
t
w
orks
. 1999;
10: 988-
99
9.
[8]
Kavzog
l
u T
,
C
o
lkes
en I. A
k
e
rne
l
functi
ons
an
al
ysis
for
supp
ort vector
mach
ines
for
la
nd c
o
ve
r
classification.
I
n
ternati
o
n
a
l Jo
urna
l of Ap
pli
e
d Ea
rth Obs
e
rvation
an
d Ge
oinfor
mation
.
2
009; 1
1
: 35
2-
359.
[9]
Haral
i
ck RM,
Shanm
ug
am
K, Dinstei
n
IH. T
e
xtural
features for
imag
e cl
ass
i
ficatio
n
.
IEEE
T
r
ansactio
n
s o
n
Systems, Ma
n and Cy
ber
ne
tics
. 1973; 61
0
-
621.
[10]
Dell'Acq
ua F, Gamba P.
T
e
xture-
bas
ed c
haracter
i
zatio
n
of
urban env
ironme
n
ts
on satellit
e
SA
R
imag
es.
IEEE
Transactio
n
s o
n
Geoscie
n
ce
and R
e
mote S
ensi
n
g
. 20
03; 41: 153-
15
9.
Evaluation Warning : The document was created with Spire.PDF for Python.