Indonesi
an
Journa
l
of El
ect
ri
cal Engineer
ing
an
d
Comp
ut
er
Scie
nce
Vo
l.
23
,
No.
1
,
Ju
ly
20
21
,
pp.
3
87
~
3
9
5
IS
S
N: 25
02
-
4752, DO
I: 10
.11
591/ijeecs
.v
23
.i
1
.
pp
3
87
-
3
9
5
387
Journ
al h
om
e
page
:
http:
//
ij
eecs.i
aesc
or
e.c
om
Semanti
c featur
e e
xtracti
on meth
od for h
yperspe
ctr
al crop
classific
ation
M.
C. Gir
ish
Babu
1
,
P
ad
m
a M.
C
.
2
1
As
sistant
Profess
or
Depa
rtment
of
Com
pute
r
Sci
enc
e
&
Engi
n
ee
r
ing,
PES
Co
ll
eg
e
of Engin
ee
ring
,
Karn
at
ak
a, I
ndi
a
2
Profess
or,
Depa
rtment
of
Com
pute
r
Sc
ie
nc
e & E
ngine
er
ing, PES
Coll
ege of
Enginee
ring
,
Karn
ata
ka,
Ind
ia
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Ja
n 1
9
, 2
021
Re
vised
Jun
7
,
2021
Accepte
d
J
un
1
4
, 202
1
H
y
per
spe
ct
ra
l
i
m
agi
ng
(HS
I)
i
s
compos
ed
of
seve
ral
hundre
d
of
nar
row
bands
(NB)
wi
th
high
spec
tral
cor
relati
on
an
d
is
widely
used
in
cro
p
cl
assifi
ca
t
ion;
th
us,
induc
es
t
ime
and
spac
e
co
m
ple
xity
,
r
esult
i
ng
in
hig
h
computat
ion
al
over
hea
d
and
Hughes
pheno
m
eno
n
in
proc
essing
the
se
images.
Dim
ensiona
l
red
uc
ti
on
t
ec
hniqu
e
such
as
band
sele
ct
ion
and
fea
ture
ext
ra
ct
ion
p
lay
a
n
important
par
t
in
enha
nc
ing
pe
rform
anc
e
of
h
yper
spec
tr
a
l
image
c
la
ss
ifi
c
a
ti
on.
How
ev
er,
exi
sting
m
et
hod
are
no
t
ef
ficien
t
when
pu
t
forth
in
no
is
y
a
nd
m
ixe
d
pixel
envi
ronm
ent
wit
h
d
y
n
amic
i
ll
um
ina
ti
on
an
d
cl
imatic
cond
it
io
n.
Here
th
e
proposed
Sem
at
ic
Fe
at
ure
Repr
ese
nt
a
ti
on
base
d
HS
I
(SF
R
-
HS
I)
cro
p
class
ifi
c
at
i
on
m
et
hod
first
emplo
y
Im
age
Fus
ion
(IF)
m
et
hod
for
findi
ng
m
ea
ningful
fea
tur
es
from
r
aw
HS
I
spec
tra
lly
.
Se
cond,
t
o
ext
ra
ct
inhe
r
ent
feature
s
th
at
k
ee
ps
spat
ia
l
l
y
m
ea
ningful
r
epr
ese
ntation
of
diffe
ren
t
cro
ps
b
y
el
imina
ti
ng
shading
el
ements.
The
n,
th
e
m
ea
ningfu
l
fea
tur
e
se
t
ar
e
used
for
tr
ai
n
in
g
using
Suppor
t
ve
ct
or
m
a
chine
(SV
M).
Expe
riment
out
c
om
e
shows
prop
osed
HS
I
cro
p
cl
assific
a
ti
on
m
odel
ac
h
ie
ves
m
uch
bet
t
er acc
ura
cies a
nd
Kap
pa
co
eff
i
ci
en
t
p
e
rform
anc
e
.
Ke
yw
or
d
s
:
Crop classi
fica
ti
on
Deep l
ear
ning
Dim
ension
r
e
duct
ion
Feat
ur
e
ex
tr
act
ion
Feat
ur
e
selec
ti
on
Hype
rsp
ect
ral im
age
Ma
chine
le
a
rn
i
ng
This
is an
open
acc
ess arti
cl
e
un
der
the
CC
B
Y
-
SA
l
ic
ense
.
Corres
pond
in
g
Aut
h
or
:
M.
C.
Girish B
abu
Assistant
Profe
sso
r
, Depa
rtm
e
nt of C
om
pu
te
r
Science a
nd E
ng
i
neer
i
ng
PES
C
ollege
of E
ng
i
neer
i
ng
Ma
nd
ya
,
K
a
rnat
aka, India
Em
a
il
: girishb
a
buphd@
gm
ail.co
m
1.
INTROD
U
CTION
Hype
rsp
ect
ral
i
m
aging
,
li
ke
oth
e
r
sp
ect
ral
i
m
aging
,
c
ollec
ts
and
pr
oces
ses
inf
or
m
at
ion
acr
os
s
the
el
ect
ro
m
agn
et
ic
sp
ect
r
um
,
usual
ly
in
visible
,
nea
r
in
fr
a
red,
an
d
s
hort
-
wave
inf
rar
e
d
wavel
eng
th
s.
Re
ce
ntly
,
with
t
he
dev
el
op
m
ent
of
hype
rsp
ect
ral
se
nsors
,
it
has
bec
om
e
po
s
sible
to
go
beyo
nd
tra
diti
on
al
R
GB
i
m
ages
and
ca
pture
hundre
ds
of
sp
ect
ral
ba
nds
sa
m
pled
with
na
rrow
wa
velen
gth
inter
vals.T
her
e
fore,
ta
king
adv
a
nt
age
of
c
on
ti
gu
ous
na
rrow
ba
nd
s
,
thes
e
HS
I
sens
ors
enab
le
d
the
a
na
ly
sis
of
the
c
hem
ic
al
pr
ope
r
ti
es
of
scene
m
at
erials
rem
otely
fo
r
the
purpose
of
detect
ion
,
i
den
ti
ficat
io
n
a
nd
c
hem
ic
al
c
om
po
sit
ion
stu
dy
of
obj
ect
s
i
n
a
pa
rtic
ular
a
rea.
Hen
ce
,
HSI
ca
ptured
u
si
ng
ai
rborne
sens
or
and
eart
h
obse
rv
at
io
n
s
at
el
li
t
es
ha
ve
been
inc
reasin
gly
ve
ry
hel
pful
f
or
urba
n
pl
ann
in
g,
en
vir
on
m
ent
m
on
ito
ri
ng,
a
nd
ag
r
ic
ultur
e
e
nvir
onm
ent.
Howe
ver,
this
researc
h
work
pr
e
dom
inantly
fo
c
us
es
on
ag
ricult
ur
e
dom
ai
n
.
This
is
because
by
y
ear
20
50
th
e
popula
ti
on
is
exp
ect
e
d
to
touc
h
9.6
bill
ion
[
1].
Along
with
,
the
world
wide
fo
od
dem
and
is
exp
an
ding,
and
i
n
this way, the a
ccessi
bili
ty
o
f
exact, co
nveni
ent, and
ti
m
e
ly d
at
a ab
ou
t a
gri
cultural on a
ne
ighbor
hood a
nd
als
o
on
w
or
l
dw
i
de
[2
]
,
is
fu
ndam
ental
[3
]
f
or
guara
nteei
ng
th
at
an
incr
easi
ng
po
pu
la
ti
on
c
an
be
ser
ve
d.
So
as
t
o
address
t
he
iss
ues
of
unpredi
ct
abili
ty
of
the
foo
d
m
ark
et
or
m
eet
ing
co
un
t
ry
foo
d
sec
ur
it
y,
Hy
per
s
pe
ct
ral
i
m
aging
tec
hn
i
qu
e
g
i
ve
a
wide
scope
of pr
ospect
s to
m
easur
e these
di
ff
ic
ul
ti
es [
2
]
,
[
4].
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
23
, N
o.
1
,
Ju
ly
20
21
:
3
8
7
-
3
9
5
388
Crop
ide
ntific
at
ion
is
one
of
the
m
ai
n
el
em
ent
of
ag
ric
ultur
e
cr
op
m
on
it
ori
ng
by
ut
il
iz
ing
HSI
ob
ta
ine
d
from
sat
el
li
te
i
m
ager
y.
Cr
op
m
app
in
g
th
rou
gh
HS
I
cl
assifi
cat
ion
ai
d
in
m
a
king
va
rio
us
de
ci
sio
n
m
aking
in
a
gr
i
culture
e
nv
i
ronm
ent
su
ch
as
yi
el
d
fo
recast
ing
,
c
rop
area
assessm
ent,
and
va
rio
us
othe
r
crop
m
anag
em
ent
app
li
cat
io
ns
[
5
]
,
[
6
].
Acc
ur
at
e
crop
m
app
in
g
is
ver
y
im
po
r
ta
nt
an
d
im
pac
t
crop
ide
ntific
at
ion
app
li
cat
io
ns
.
N
on
et
heless,
li
m
it
at
ion
that
exi
st
in
cr
op
i
dent
ific
at
ion
us
i
ng
HSI
m
us
t
be
addresse
d
[
7].
First,
the
possible
hi
gh
dim
ension
siz
e
of
in
pu
t
HS
I
data
is
sti
ll
an
op
e
n
issu
e.
Seco
nd,
H
S
I
data
resem
ble
high
si
m
il
arities
of
te
xtu
res
,
s
hap
e
s
and
sp
ect
ral
sign
at
ur
es
am
on
g
dif
fer
e
nt
cr
op.
Last
ly
,
the
pr
ese
nce
of
no
ise
in
HS
I
sig
nifica
ntly
i
m
pact the accur
aci
es
of e
xi
sti
ng
HS
I
cr
op
classi
ficat
ion
m
od
el
.
The
s
pectral
r
esolutio
ns
have
been
sig
nifi
cantl
y
enh
a
nc
ed
with
gro
wt
h
of
H
SI
m
eth
od
ologies.
analy
zi
ng
HSI
is
diff
ic
ult.
F
ur
t
her
,
co
ns
ide
rin
g
the
behav
iour
of
HS
I
t
he
re
exist
hi
ghe
r
co
rr
el
at
io
n
a
m
on
g
neig
hbori
ng
pix
el
an
d
s
pectra
l
band
set
s
[
8
]
-
[
10]
.
T
he
n,
be
cause
of
hi
gh
dim
ension
al
na
ture
of
hy
per
s
pe
ct
ral
i
m
age
there
is
high
pro
bab
il
it
y
of
i
ncr
ea
se
i
n
s
pace,
ti
m
e
and
com
pu
ta
ti
onal
T
hu
s
,
ov
e
r
head
le
adin
g
t
o
well
-
known
H
ughe
s
phe
nom
eno
n
[11].
Hen
ce
,
reducin
g
dim
ensio
n
of
re
dunda
nt
featu
re
in
H
SI
is
pr
i
m
aril
y
necessa
ry
f
or
processi
ng
hy
per
s
pectral
im
age.
Feat
ur
e
s
el
ect
ion
/
Ba
nd
Sele
ct
ion
[
12
]
-
[
14
]
is
on
e
of
the
eff
ic
ie
nt
dim
ensio
n
re
duct
io
n
te
ch
nique
th
at
rem
ov
e
redunda
nt
featu
re
s.
The
basic
noti
on
of
ba
nd
sel
ect
ion
m
et
ho
dolo
gies
is
to
cho
os
e
m
os
t
eff
ic
ie
nt
ba
nd
s
ubset
that
i
s
com
po
sed
of
inf
or
m
at
ion
within
ori
gin
al
ba
nds
.
On
the
oth
e
r
side,
the
feat
ur
e
extracti
on
m
e
t
hodolo
gies
[15
]
red
uces
dim
ensio
n
in
acco
r
dan
ce
with
co
m
plex
featur
e
t
ran
s
f
orm
ation
.
T
hu
s
,
featur
e
sel
ect
ion
m
et
ho
ds
a
r
e
easi
er
to
unde
rstan
d
an
d
ca
n
be
us
e
d
in
pr
act
ic
al
pur
po
se
.
The
f
eat
ur
e
sel
ect
io
n
m
et
ho
d
can
c
hoos
e
bands
only
within
hy
pe
rsp
ect
ral
i
m
a
ges.
On
the
othe
r
side
,
the
feat
ur
e
e
xtracti
on
m
et
ho
d
can
util
iz
e
th
e
HSI
ba
nd
set
s
f
or
ge
ne
rati
ng
bette
r
disc
rim
inati
ng
feat
ur
e
set
s.
In
La
placi
an
discrim
inant
a
na
ly
sis
of
h
ype
rs
pectral
im
ager
y
[15], p
re
sent
ed
joint featur
e
sel
ect
ion
an
d
f
eat
ure
extracti
on
m
eth
od
ologies
f
or
hype
rsp
ect
ral
i
m
age
cl
assifi
cat
ion
.
Sim
il
a
r
to
[
15]
this
pap
e
r
f
oc
us
e
s
on
dev
el
op
i
ng
t
he
di
m
ension
al
it
y
red
uctio
n
te
chn
i
qu
e
by
ex
tract
ing
sem
antic
al
ly
us
efu
l
f
eat
ur
e
set
fro
m
raw
hype
rsp
ect
ral im
age.
Nu
m
erous
ty
pe
s
of
hype
rsp
e
ct
ral
i
m
age
di
m
ension
al
re
duct
ion
m
et
ho
dolo
gies
that
i
nc
lud
es
both
su
pe
r
vised
a
nd
un
s
uper
vised
te
chn
iq
ue,
has
been
pr
ese
nted
in
recent
tim
e
.
This
w
ork
f
oc
us
es
on
supe
r
vised
m
et
ho
dolo
gies
su
ch
as
li
near
discrim
inant
a
naly
sis
[1
6],
a
nd
al
so
unsupe
rv
ise
d
m
e
tho
ds
su
ch
as
ind
e
pe
nd
e
nt
com
po
ne
nt
an
al
ysi
s
(I
CA)
[
17
]
an
d
pr
inci
pal
com
po
ne
nt
analy
sis
(P
CA)
[18].
From
analy
sis
it
is
seen
that
pr
i
ncipal
c
ompone
nt
a
naly
sis
m
e
tho
dolo
gi
es
extract
m
uch
bette
r
fe
at
ur
e
w
hen
c
om
par
ed
with
ot
he
r
m
et
ho
dolo
gies.
The
PC
A
fe
at
ur
e
e
xtracti
on
t
echn
i
qu
e
can
guara
ntee
that
m
or
e
relevan
t
featur
e
of
H
SI
can
be
k
ept
with
m
inim
al
siz
e
of
use
fu
l
pri
nci
pal
com
po
ne
nts.
On
the
oth
e
r
s
ide
,
inde
pe
nde
nt
com
po
ne
nt
analy
sis
base
d
m
et
ho
dolo
gies
ca
n
guara
ntee
that
the
tra
ns
f
or
m
ed
com
po
ne
nts
are
in
depend
e
nt
as
m
uch
a
s
con
cei
vab
le
.
N
on
et
heless,
IC
A
ba
sed
m
et
hodo
l
og
ie
s
i
nduc
e
hi
gh
c
om
pu
ta
ti
on
al
over
he
ad
due
t
o
it
s
com
plex
com
pu
ta
ti
on
.
Fu
rt
her,
the
I
CA
m
et
ho
dolo
gies
do
es
not
consi
der
s
patia
l
con
te
xt
in
for
m
at
ion
f
or
ext
racti
ng
featur
e
as
t
hey
treat
a
nd
pro
cess
eac
h
pix
e
l
ind
e
pe
nd
e
ntly
.
Ma
jority
of
existi
ng
HSI
dim
ension
reducti
on
t
echn
iq
ue
ca
nn
ot
directl
y
util
iz
e
sp
at
ia
l
inf
orm
ation
of
hy
per
s
pectral
im
ages
a
nd
feat
ure
ext
racti
on
m
et
ho
d
us
es
only
sp
ec
tral
inf
or
m
at
ion
of
e
ver
y
i
ndividu
al
pix
el
wh
ic
h
is
sta
te
d
to
be
e
ff
ic
ie
nt
f
or
e
nha
ncing
the
cl
assifi
cat
ion
a
ccur
aci
es
[
19
]
,
[
20]
.
T
he
ea
rt
h
e
nv
ir
onm
ent
is
com
po
se
d
of
dif
fer
e
nt
ty
pes
of
co
rps
s
uch
as
so
il
,
w
heat,
c
orn
et
c.
hy
pe
rsp
ect
ral
i
m
age
cl
assifi
cat
ion
m
et
ho
d
ai
m
s
at
identify
in
g
these
m
eaningf
ul
represe
ntati
on
of cr
ops
.
T
hus,
it
is n
ecessa
ry
to ex
t
ract m
ea
ningf
ul s
patia
l feat
ur
e
set
s
.
Ther
e
f
or
e,
f
or
extracti
ng
m
eanin
gful
s
patia
l
inf
or
m
at
ion
f
r
om
hyper
s
pect
ral
im
age
that
com
po
sed
of
m
ixed
pi
xel,
it
is
i
m
po
rtant
t
o
captu
re
in
here
nt
pr
op
e
rtie
s
of
physi
cal
char
a
ct
erist
ic
s
of
d
if
fer
e
nt
ob
j
ect
s.
Th
us
by
extracti
ng
sp
at
ia
l
inform
ation
f
ro
m
adj
ac
ent
obj
ect
can
ai
d
in
increasi
ng
the
acc
ur
ac
y
of
dif
fer
e
ntiat
ing
diff
e
re
nt
crops
pr
ese
nt
in
ea
r
th
en
vir
on
m
ent.
The
in
her
e
nt
hype
rsp
ect
ral
i
m
age
are
inh
e
ren
t
feat
ur
e
set
an
d
sh
a
ding
c
om
po
ne
nt.
Ext
racti
ng
these
com
po
ne
nt
play
s
ve
ry
im
po
rtant
r
ole
in
cl
assifi
c
at
ion
ta
s
k
a
nd
at
th
e
sam
e
tim
e
it
i
s
chall
en
ging
to
extract
t
hes
e
com
po
ne
nt
f
ro
m
sing
le
hy
per
s
pectral
im
age.
T
his
is
be
cause
inh
e
ren
t
featu
r
e
set
s
reli
es
on
obj
ect
featu
re
s
of
the
eart
h
a
nd
va
ries
with
resp
ect
t
o
il
lu
m
inati
on
a
nd
c
lim
a
te
conditi
on.
T
hu
s,
f
or
extr
act
ing
m
or
e
m
ea
ningf
ul
s
patia
l
featu
re
of
di
ff
e
ren
t
c
rop,
this
w
ork
pres
ent
an
eff
ic
ie
nt in
her
e
nt f
eat
ur
e e
xtra
ct
ion
m
et
ho
d
nam
el
y
SFR
-
HSI
that keeps
sp
a
ti
al
l
y
m
eaning
f
ul r
ep
rese
ntati
on
of
diff
e
re
nt
cr
op
s
.
Furthe
r,
for
reducin
g
the
hy
per
s
pectral
ba
nd
dim
ension
that
is
eff
ic
ie
nt
agai
ns
t
nois
e
an
d
m
ixed
pi
xel
en
vir
on
m
ent,
thi
s
w
ork
pr
ese
nt
an
dim
ension
reducti
on
te
c
hniq
ue
us
in
g
fusion
te
c
hniq
u
e
[2
1
]
.
Last
ly
,
di
m
ension
al
ly
reduce
d
featu
re
set
s
are
trai
ned
usi
ng
m
ulti
-
cl
as
s
su
pp
or
t
vect
or
m
achine
(SVM).
Using
SV
M
ai
d
in
trai
ning
m
od
el
with
le
s
s
trai
nin
g
sam
ple
and
ca
n
achieve
go
od
preci
sion
with
m
ini
m
al
com
pu
ta
ti
on
c
om
plexity
w
he
n
c
om
par
ed wi
th d
ee
p
le
a
rn
i
ng
base
d
m
et
ho
d [2
2
]
-
[
2
6
]
.
The
m
anu
scri
pt
is
arti
culat
ed
as
desc
rib
ed
bel
ow
.
T
he
propose
d
hy
per
s
pectral
im
age
crops
cl
assifi
cat
ion
m
od
el
is
pr
esented
in
sect
io
n
II
.
T
he
sect
io
n
III
disc
us
ses
about
the
res
ul
t
at
ta
ined
by
propose
d
SFR
-
H
SI
cro
p
cl
a
ssific
at
ion
m
od
el
ov
er
va
rio
us
existi
ng
HS
I
c
r
op
cl
ass
ific
at
ion
m
od
el
.
In
la
st
sect
ion
t
he
sign
ific
a
nce
of
pro
posed
SFR
-
HSI
crop
cl
as
sific
at
ion
m
odel
is
co
nclu
de
d
with
f
ut
ur
e
directi
on
of
re
search
work.
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Sema
ntic f
eatu
re ext
ra
ct
io
n m
et
hod
fo
r
hyp
e
rsp
ect
r
al crop
cl
as
sif
ic
ation
(
M. C. Giri
sh B
abu
)
389
2.
SEMANTI
C FE
ATU
RE
R
EPRESE
NT
A
TION
FO
R
H
YPER
SPECT
RA
L
C
RO
P
CLASSIFI
C
A
TION
This
work
pre
sent
an
ef
fici
ent
cr
op
cl
assif
ic
at
ion
m
et
ho
d
na
m
ely
SFR
-
HSI
us
in
g
hype
rsp
ect
ral
i
m
age.
The
b
l
oc
k diag
ram
o
f pro
po
se
d
c
r
op
cl
assifi
cat
ion
t
echn
i
qu
e
is s
how
n
i
n
Fi
gure
1.
T
he p
rop
os
e
d
S
FR
-
HS
I
hyperspec
tral
cr
op
cl
ass
ific
at
ion
is
de
scribe
d
i
n
Algorithm
1
.
The
first
phase
of
hy
per
s
pectral
cr
op
cl
assifi
cat
ion
i
s
the
featu
re
re
du
ct
io
n.
F
or
id
entify
ing
us
e
f
ul
featur
e
set
s
f
ro
m
raw
hype
r
sp
ect
ral
im
age
,
this
work
use
s
dim
ensio
nalit
y red
uction t
ech
nique
for
e
nh
a
nci
ng classi
ficat
io
n
acc
ur
acy
.
Figure
1.
Bl
oc
k diag
ram
o
f
pro
posed
SF
R
-
HS
I
cr
op classi
ficat
ion
tec
hn
i
qu
e
Let
us
c
onside
r
the
re
are
bands
in
a
ra
w
hy
per
s
pectral
im
age,
t
his
w
ork
reduces
t
he
ba
nd
s
to
bands
.
Th
e
ba
nd
is
segm
ente
d
int
o
su
b
-
gr
oup
set
s.
T
hen,
m
ean
of
t
hese
su
b
-
group
set
s
is
com
pu
te
d.
F
or
ob
ta
ini
ng
dim
e
ns
io
nalit
y
redu
ced
hype
rs
pectral
im
age
featu
re
m
os
t
of
the
existi
ng
m
et
ho
d
us
e
d
PC
A.
Using
PCA
can
guar
antee
of
pr
e
se
rv
i
ng
hype
rspect
ral
i
m
age
featur
es
in
a
s
m
al
le
r
qu
antit
y
of
us
e
fu
l
pri
ncipal
com
po
ne
nts.
H
ow
e
ve
r,
us
in
g
PCA
ca
nnot
as
su
re
that
the
s
pe
ct
rall
y
us
ef
ul
featur
e
is
retai
ned.
F
o
r
a
d
d
r
e
s
s
i
n
g
,
t
h
e
d
i
m
e
n
s
i
o
n
a
l
i
t
y
r
e
d
u
c
e
d
h
y
p
e
r
s
p
e
c
t
r
a
l
i
m
ag
e
f
e
a
t
u
r
e
i
s
o
b
t
a
i
n
e
d
b
y
u
s
i
n
g
f
u
s
i
o
n
t
e
c
h
n
i
q
u
e
[
2
1
]
m
e
t
h
o
d
a
s
(1)
=
∑
=
(
−
1
)
+
1
,
=
⌊
⌋
,
(1)
wh
e
re
an
d
de
picts
the
band
ind
ic
es,
de
picts
the
ℎ
dim
ension
of
act
ual
sp
e
ct
ral
band
(
),
dep
ic
t
s
the
ℎ
dim
ension
of
the
reduce
d
,
dep
ic
ts
each
sub
-
gro
up
band
siz
e,
⌊
∙
⌋
rounds
∙
to
the
cl
os
est
val
ue
towa
rd
ne
gativ
e
infi
nity
,
an
d
de
picts
the
redunda
nt
in
form
at
ion
a
nd
nois
y
pix
el
of
eac
h
ba
nd
set
s.
T
he
(1)
is use
d
in
eli
m
i
nating re
dunda
nt in
form
at
ion
an
d n
oisy
pix
el
s f
r
om
each
s
ubgr
oup.
Algorithm
1.
Pr
op
os
e
d
SFR
-
HSI
Hy
per
s
pe
ct
ral
crop
cl
assifi
c
at
ion
us
i
ng
dim
ension
r
edu
ct
io
n
a
nd
m
achine
le
arn
in
g
te
c
hn
i
qu
e
.
Inp
ut. Hyper
spe
ct
ral i
m
age col
le
ct
ed
from
satel
li
te
∈
∗
∗
with
∗
pi
xels
by
bands.
Ou
t
pu
t.
A cl
as
sifie
d ou
tc
om
e
(
i.e.
, labell
ed
Hype
rsp
ect
ral im
age
).
Step
1.
Start.
Step
2.
Dim
ension
re
du
ct
io
n o
f
ba
nd
s
to
usi
ng E
q.
(1).
Step
3.
For
∈
{
1
,
}
do
Step
4.
By
it
er
at
ing
E
q. (
5), t
he
is dec
om
posed
i
nto
an
d
.
Step
5.
En
d for
.
Step
6.
Vector
i
se
to
=
{
1
,
2
,
…
,
}
∈
∗
wh
e
re
=
∗
f
or
ass
ur
i
ng
eac
h
pix
el
can
be
represe
nted
as a
−
dim
ension
al
d
at
a poi
nt.
Step
7.
For
obta
ining l
abels
∈
s
upport
vecto
r m
achine lear
nin
g al
gorithm
is u
ti
li
zed.
Step
8.
Rec
ons
truct
to
be
a
hy
per
s
pectral i
m
age
∈
∗
.
Step
9.
Sto
p
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
23
, N
o.
1
,
Ju
ly
20
21
:
3
8
7
-
3
9
5
390
Fu
rt
her,
it
is
i
m
po
rtant
to
e
xt
ract
inhe
re
nt
pro
per
ti
es
(i.e.
,
t
he
m
ini
m
u
m
i
nfor
m
at
ion
re
quire
d
to
ho
l
d
the
releva
nt
in
form
ation
in
a
ny
giv
e
n
im
ag
e
with
m
ixed
pix
el
is
firm
ly
known
as
the
i
nh
e
re
nt
pro
pert
ie
s)
of
hype
rsp
ect
ral
im
age.
The
in
her
e
nt
featu
re
set
s
rely
on
obj
ect
(i.e.
,
surf
ace
)
featu
res
of
the
earth
.
These
featur
e
s
are
i
nhere
nt
with
re
sp
ect
to
il
lum
i
nation
an
d
cl
i
m
at
e
con
diti
on.
T
his
work
pr
ese
nts
a
n
ef
fici
ent
m
et
ho
d
f
or
e
xt
racti
ng
in
he
rent
featur
e
set
s
and
el
im
inate
t
he
sh
a
ding
co
m
po
nen
t
in
hy
per
s
pectral
im
age
to
ob
ta
in
sem
antic
al
ly
(i.e.,
m
e
anin
gful)
s
patia
l
inf
orm
ation
.
Let
∈
∗
de
pict
th
e
inte
ns
it
y
pa
r
a
m
et
er
of
a
hype
rsp
ect
ral
i
m
age,
∈
∗
de
picts
it
s
in
her
e
nt
featur
e
set
el
e
m
ent
an
d
∈
∗
de
picts
it
s
in
herent
featur
e
set
s
s
had
i
ng
el
em
e
nt.
T
he
hy
perspectra
l
im
age
f
or
a
pixe
l
is
re
pr
ese
nt
ed
as
a
pi
xel
-
wis
e
m
ul
ti
plica
ti
ve
of in
her
e
nt f
eat
ur
e
set el
em
ent
s u
si
ng
(
2)
=
,
(2)
wh
e
re
de
picts
the p
i
xel in
dic
es.
Fr
om
(2
),
it
is
seen
that
a
hy
per
s
pectral
im
age
is
com
po
sed
two
in
her
e
nt
featur
e
set
ele
m
ents
,
bu
t
sti
ll
,
it
is
diff
i
cult
to
s
olv
e
t
he
iss
ue
with
t
wo
un
known
pa
ram
et
ers
(i.e.,
an
d
)
with
one
know
n
value
(i.e.,
)
.
I
n
ge
ne
ral,
the
ref
le
ct
ance
is
c
on
sta
nt
withi
n
each
reg
i
on
an
d
al
ong
e
dges
has
a
sh
a
rp
cha
nge
.
Fu
rt
her,
it
is
seen
as
cha
nges
in
intensit
y
par
am
et
er
wi
ll
rese
m
ble
a
ref
le
ct
ance
c
ha
ng
e
a
nd
pix
e
l
with
identic
al
intens
it
y param
et
er w
il
l resem
ble iden
ti
cal
r
e
flect
ance
ou
tc
om
es
. T
hu
s
,
is o
btained
usi
ng
(3)
=
∑
,
∈
(
)
(3)
wh
e
re
and
dep
ic
ts
pix
el
i
nde
xe
s,
dep
ic
t
s
a
featu
re
set
s
com
po
ne
nt
of
a
ff
i
nity
m
atr
ix
(
AM)
.
T
he
m
at
rix
de
picti
ng
a
pair
wise
sim
il
arit
ie
s
(PW
S
)
am
ong
an
d
,
an
d
(
)
de
picts
co
rr
e
spo
nd
i
ng
(i.e
.,
neig
hbour
)
pixe
l
.
The
nei
ghbori
ng
pi
xel
are
ge
ner
al
ly
a
G
aussian
wi
ndow
(
)
w
hich
ca
n
be
ob
ta
in
e
d
us
in
g
(4)
=
exp
(
−
‖
−
‖
2
2
2
2
)
(4)
and
t
he
siz
e
of
it
is
est
ablished
us
in
g
.
Fur
ther,
def
i
ning
aff
init
y
grap
h
(AG)
play
very
essenti
al
part
in
sem
antic
a
ll
y
e
xtracti
ng
in
he
r
ent
cha
racteri
s
ti
cs.
Using
(
2)
an
d
(4),
we
c
an
est
a
b
li
sh
th
e
li
near
pro
pert
ie
s
of
these to
(
5)
{
=
∑
,
∈
(
)
̃
=
1
,
(5)
wh
e
re
̃
=
1
.
Af
te
r
ob
ta
ini
ng
t
he
approxim
at
ed
ou
tc
om
e
of
and
,
we
ca
n
est
ablish
eve
r
y
pix
el
ref
le
ct
ance
val
ue
(i.e.,
physi
cal
pro
per
ti
es)
of
e
ver
y
obj
e
ct
,
w
her
e
s
hadi
ng
pro
pe
rtie
s
is
not
relat
ed
wit
h
resp
ect
to sem
antic
featu
re
se
ts
prop
e
rtie
s.
Using
in
her
e
nt
featur
e sets
w
e
can
prese
rv
e the
intrinsic prop
e
rtie
s
of
m
ixed
pi
xel an
d
el
im
inate
t
he
use
le
ss sp
at
ia
l i
nf
or
m
at
ion pr
ese
nted
i
n
s
had
i
ng
elem
ent. Thus,
we wil
l get a
sp
at
ia
ll
y
m
eani
ngf
ul info
rm
ati
on of eac
h o
bj
e
ct
s/
cro
p
.
Last
ly
,
fo
r
cl
a
ssific
at
ion
of
op
ti
m
iz
ed
hyper
sp
ect
ral
featur
e
set
s
this
work
c
onside
r
ed
pix
el
wise
SV
M
cl
assifi
cat
ion
m
et
ho
d.
Suppor
t
Vecto
r
Ma
chine
,
al
so
know
n
as
S
VM,
is
popula
r
for
s
ol
ving
pro
blem
s
about
cl
assifi
cat
ion
,
detect
io
n
an
d
regressi
on.
In
s
upport
vecto
r
m
achin
e,
the
m
od
el
con
st
ru
ct
s
a
hy
perplane
or set o
f
hype
r
planes
in
highe
r dim
ension
s
pa
ce. F
ro
m
, [
2
7
]
the str
uctur
e
of SV
M
is obta
it
ned
as
(
6)
a
rgm
a
x
,
,
1
2
‖
‖
2
+
∑
=
1
.
(
(
)
+
)
≥
1
−
;
≥
0
;
=
1
,
2
,
…
,
(6)
wh
e
re
(
)
is
tra
nsfo
rm
ed
feat
ure
s
pace
a
nd
th
e
va
riable
is
the
re
gu
la
rizat
ion
par
am
et
er
to
c
ontr
ol
t
he
trade
off
betwe
en
m
arg
in
an
d
the
t
olera
nce
of
m
isc
la
ssifi
cat
ion
.
Usi
ng
SV
M
m
od
el
[
2
7
]
,
can
br
in
g
good
trade
off
bet
we
en
m
arg
in
a
nd
the
tolera
nce
of
m
isc
la
ssificati
on
.
In
a
ddit
ion
,
i
n
the
dat
aset
is
non
-
li
ne
arly
separ
a
ble,
a
pp
l
yi
ng
ke
rn
el
tri
ck
can
tra
ns
f
orm
the
m
to
hig
he
r
li
near
dim
ension,
su
c
h
as
Gau
ssia
n
Kernel.
Since
the
SVM
pr
oble
m
is
a
convex
opti
m
iz
at
ion
pr
obl
e
m
,
the
m
od
el
can
al
ways
ob
ta
in
a
global
op
ti
m
u
m
from
the
m
od
el
.
W
it
h
the
op
t
i
m
al
decisi
on
bounda
ry,
the
m
od
e
can
us
e
it
to
cl
assify
hyper
s
pectral
da
ta
set
into
dif
fer
e
nt
la
bels.
The
pro
po
s
ed
hy
per
s
pe
ct
ral
i
m
age
c
la
ssific
at
ion
m
od
el
at
ta
in
s
be
tt
er
accuracy
wh
e
n
com
par
ed wit
h exist
in
g hyper
sp
ect
ral im
age classi
ficat
ion
m
od
el
w
hich
is
experim
ental
l
y show
n belo
w
.
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Sema
ntic f
eatu
re ext
ra
ct
io
n m
et
hod
fo
r
hyp
e
rsp
ect
r
al crop
cl
as
sif
ic
ation
(
M. C. Giri
sh B
abu
)
391
3.
RESU
LT
S
A
ND
DI
SCUS
S
ION
S
This
sect
ion
di
scusses
ab
out
the
resu
lt
achieve
d
us
in
g
pro
po
se
d
im
pr
ov
e
d
dim
ension
reducti
on
base
d
hype
rs
pe
ct
ral
crop
cl
assifi
cat
ion
m
eth
od
a
doptin
g
m
achine
le
arni
ng
te
ch
nique.
Fu
rt
her,
the
resu
lt
ob
ta
ine
d
is
c
om
par
ed
with
e
xisti
ng
hy
per
s
pectral
c
rop
cl
assifi
cat
ion
a
doptin
g
m
achine
an
d
dee
p
le
arn
i
ng
te
chn
iq
ues
.
F
or
ca
rr
yi
ng
out
exp
e
rim
ent
the
data
is
co
ll
ect
ed
us
in
g
A
VIR
IS
se
nsor
[
28
]
.
This
w
ork
s
el
ect
ed
the
I
nd
ia
n
Pi
nes
dataset
for
eval
uating
c
rop
cl
assifi
cat
ion
m
od
el
be
cause
it
co
ntains
of
tw
o
-
t
hird
of
agr
ic
ultur
e
sce
ne
an
d
rest
is
com
po
sed
of
f
or
est
a
nd
ot
her
veg
et
at
io
n
c
rops.
T
he
dataset
is
com
po
sed
of
16
cl
asses wh
ic
h
is descri
bed
in Ta
ble 1
[
29
]
. T
he
I
nd
ia
n
Pine
s d
at
aset
that com
po
sed of
22
0
sp
ect
ral re
fle
ct
ance
bands
with
wa
velen
gth
ra
nge
of
0
.
4
–
2
.
5
*
10
−
6
m
et
ers
w
her
e
twe
nty
w
at
er
a
bs
orpti
on
hy
per
s
pectral
ba
nds
rangin
g
f
r
om
10
4
–
108,
150
–
163,
an
d
220
a
r
e
rem
ov
ed
a
nd
us
e
d
f
or
e
xper
i
m
ent
analy
sis
si
m
il
ar
to
[2
1
]
,
and
[
29
]
.
T
he
I
nd
ia
n
pi
ne
im
ag
e
captu
red
usi
ng
AVIRI
S
s
ens
or
a
nd
it
s
gro
und
tr
uth
value
is
s
ho
wn
i
n
Fig
ure
2
(a
),
and
Fi
g
ure
2
(
b),
r
especti
vel
y.
The
dataset
is
com
po
se
d
of
10
249
sam
ples.
Ex
per
im
ent
ar
e
cond
ucted
to
evaluate
the
pe
rfor
m
ance
of
pr
op
os
e
d
hype
rsp
ect
ral
crop
cl
assifi
cat
ion
m
od
el
ov
er
va
rio
us
existi
ng
hy
per
s
pectral
cr
op
cl
assifi
cat
ion
rangin
g
from
featur
e
extracti
on
m
et
ho
d,
feat
ure
sel
ect
ion
,
m
a
chin
e
le
arn
in
g
an
d
de
ep
le
arn
i
ng
m
et
hod
.
T
he
pe
r
form
ance
is
evaluated
in
te
rm
s
of
ov
e
rall
accuracy
(
OA),
a
ver
a
ge
accuracy
(AA
),
a
nd
Ka
pp
a
coeffic
ie
nt
(
K)
.
The
over
al
l
accuracies
perform
ance
dep
ic
ts
the
correct
ly
cl
assifi
ed
featur
e
ove
r
total
te
st
featu
re.
Av
e
ra
ge
acc
uraci
es
re
pr
ese
nt
s,
the
ave
r
ag
e
of
eac
h
i
ndividu
a
l
cl
asses.
Kapp
a
coef
fici
ent
involves
both
com
m
issi
on
and
om
issi
on
loses
and
pro
vid
es
a
n
ef
fici
ent
inf
or
m
at
ion
of
the
cl
assifi
cat
ion
m
od
el
rob
us
tness
perf
orm
ances.
For
al
l
the
per
f
or
m
ance
m
et
ric,
hig
her
the
value de
picts s
up
e
rio
r
cl
ass
ifi
cat
ion
ou
tc
om
e
s.
Table
1.
T
otal
sam
ples o
f
eac
h
cl
asses
of
Indian
Pines
HSI
data
Nu
m
b
e
r
Clas
ses
Total Sa
m
p
l
es
1)
Alf
alf
a
46
2)
Co
rn n
o
till
1428
3)
Co
rn
m
in
till
830
4)
Co
rn
237
5)
Grass p
astu
re
483
6)
Grass tr
ees
730
7)
Grass p
astu
re
m
o
v
ed
28
8)
Hay
win
d
rowed
478
9)
Oats
20
10)
So
y
b
ean n
o
till
972
11)
So
y
b
ean
m
in
till
2455
12)
So
y
b
ean clean
593
13)
wh
eat
205
14)
wo
o
d
s
1265
15)
Bu
ild
in
g
s Grass Trees
Dr
iv
es
386
16)
Sto
n
e Steel
Tow
er
s
93
Figure
2. I
nd
ia
n
Pi
nes data
se
t
;
(a)
Indian
Pi
nes
HIS,
(b) Gr
ound tr
ut
h
in
f
or
m
at
ion
of Ind
ia
n pine
HS
I
and
(c)
C
orres
pond
ing
la
bel in
for
m
at
ion
of
diff
e
ren
t c
rop o
f
the In
dian Pi
ne
s H
S
I
3.1.
Fe
ature
ext
r
act
i
on
The
Fig
ur
e
2
(
a)
shows
the
a
ct
ual
sat
el
lite
i
m
age
captur
e
d
of
Indian
pin
e
data.
Figure
2
(b
)
s
hows
the
ref
e
re
nce
im
age
i.e.,
the
gro
und
tr
uth
va
lue
of
I
nd
ia
n
pin
e
im
age.
From
Figu
re
2
(c
)
it
can
be
see
n
the
re
are
diff
e
re
nt
ki
nd
of
c
r
op
w
hi
ch
is
re
pr
ese
nt
ed
by
diff
e
re
nt
col
or
.
First,
both
act
ual
a
nd
gro
und
tr
ut
h
i
m
age
value
are
ob
ta
i
ned
(i
.e.,
loa
de
d).
Total
there
are
16
cl
asses.
Then,
us
in
g
gro
und
tr
uth
val
ue
inf
or
m
at
ion
of
a
par
ti
cula
r
cr
op
the
featu
re
se
t
are
extracte
d
in
rand
om
m
a
nn
e
r
f
r
om
resp
ect
ive
sou
rce
i
m
age
(
Fig
ur
e
2
(a
)
)
wh
ic
h
a
re
la
te
r
us
ed
f
or
trai
ning
pur
po
se
.
In
sam
e
way
the
featu
re
set
are
e
xtracted
f
or
al
l
16
cl
ass
es
in
rand
om
m
ann
e
r
by
va
ryi
ng
spa
ce
and
range
par
am
et
ers.
Th
en,
re
st
of
im
a
ge
will
be
use
d
for
te
sti
ng
pur
po
s
e.
Pr
io
r
to
pe
r
for
m
i
ng
featu
re
e
xtracti
on,
fi
rst
the
ba
nd
siz
e
of
hype
rs
pectral
data
is
reduce
d
us
i
ng
Im
age
fu
si
on
te
chn
iq
ue
[21]
by
el
i
m
inati
n
g
re
dundant
f
eat
ur
es
us
in
g
neig
hbori
ng
pi
xel
correla
ti
on.
The
n,
f
or
fe
at
ure
extracti
on
co
rresp
onding
cr
op
(i.e
.,
obj
ect
)
inh
e
ren
t
fe
a
tu
res
are
e
xtract
ed
usi
ng
abs
or
ption
a
nd
re
fle
ct
ance
pro
per
ti
es
by
va
ryi
ng
sp
ace
a
nd
ra
ng
e
pa
ra
m
et
er.
That
is
each
obj
ect
with
dif
fe
ren
t
pixe
l
intensit
y
will
have
diff
e
re
nt
ref
le
ct
ance
wa
vele
ng
t
h
(t
his
is
because
so
m
e
obj
ect
li
ke
water
will
ab
so
r
b
s
om
e
po
r
ti
on
of
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
23
, N
o.
1
,
Ju
ly
20
21
:
3
8
7
-
3
9
5
392
wav
el
e
ng
t
h
a
nd
s
om
e
obj
ect
does
n’
t
a
bsor
b).
F
ro
m
Fig
ure
2
it
ca
n
be
seen
the
ed
ge
s
of
eac
h
c
r
op
is
com
po
sed
of
di
ff
ere
nt
sh
a
ding
el
e
m
ent
(n
oi
se)
an
d
are
sep
arated
by
blac
k
colo
r.
T
hese
sh
a
ding
com
ponen
ts
are
extracte
d
usi
ng
inte
ns
it
y
par
am
et
er
(i.e
.,
ref
le
ct
ance
pro
per
ti
es
of
ob
j
ect
).
I
n
ou
r
w
ork
sh
a
ding
el
e
m
ent
will
hav
e
m
uch
le
sser
sea
rc
h
s
pace
val
ue
wh
e
n
c
om
par
ed
with
i
ntens
it
y
value.
The
n,
the
se
feat
ures
ar
e
com
par
ed wit
h o
bj
ect
i
nhere
nt
p
r
operti
es, i
f i
t
var
ie
s,
t
hey a
re eli
m
inate
d
or else
feat
ure a
re r
et
ai
ne
d.
3.2.
Clas
si
ficat
i
on
A
T
he,
featu
re s
et
extracte
d
of
al
l
16
cl
asses/
crops
are
trai
ne
d
us
in
g
S
VM
cl
assifi
cat
ion
m
et
ho
d.
F
or
evaluati
ng
pe
r
form
ance
this
w
ork
c
onsid
ered
10
-
f
old
cro
ss
validat
ion.
T
he
cl
ass
ific
at
ion
ou
tc
om
e
is
evaluate
d
in
te
rm
s o
f ov
e
rall
accur
acy
,
K
a
pa
Coeffic
ie
nt,
a
nd av
e
ra
ge
acc
ur
acy
.
3.
3
.
E
xp
eri
m
ent
p
ar
ame
ter
o
p
timi
z
at
ion
fo
r
h
yp
erspec
tra
l cr
op cl
as
s
ific
at
io
n
The
factors
s
uc
h
as
il
lum
inatio
n,
sh
a
ding,
a
nd
noise
,
t
he
s
pectral
feat
ur
e
s
of
di
ff
e
ren
t
l
and
co
vers
in
natu
r
al
sce
nes
are
us
ually
s
ubj
ect
to
so
m
e
de
gr
ee
of
disto
r
ti
on
.
I
n
gen
e
ra
l,
the
ref
le
ct
an
ce
is
co
ns
ta
nt
within
each
re
gi
on
a
nd
al
on
g
e
dg
es
has
a
s
ha
rp
ch
ang
e
.
F
ur
t
her,
it
is
seen
as
c
hanges
i
n
inte
ns
it
y
par
am
et
e
r
will
resem
ble
a
ref
l
ect
ance
cha
nge
an
d
pi
xel
wi
th
identic
al
inten
sit
y
par
am
et
er
will
resem
ble
i
den
ti
cal
re
flect
anc
e
ou
tc
om
es.
Thus,
us
i
ng
a
ff
i
nity
pairw
ise
sim
il
arit
y
m
at
rix
a
m
on
g
neig
hb
orh
ood
pi
xels
we
can
sem
ant
ic
al
ly
extract
i
nh
e
rent
char
act
e
risti
cs.
T
he
n
we
ca
n
est
ablish
eve
r
y
pix
el
ref
le
ct
a
nce
val
ue
(i.e
.,
physi
cal
pro
pe
rtie
s)
of
e
ver
y
obj
ect
,
w
her
e
s
ha
ding
pr
op
e
rtie
s
is
no
t rela
te
d
wit
h
res
pect
to
s
e
m
antic
featur
e sets
pro
per
ti
es.
Using
inh
e
ren
t
featu
r
e
set
s
we
ca
n
pr
ese
r
ve
the
i
nt
rinsic
pro
pe
rtie
s
of
m
ixed
pix
el
an
d
el
im
inate
the
us
el
ess
sp
at
ia
l
inf
or
m
at
ion
presented
i
n
s
ha
ding
el
em
ent.
Th
us
,
we
will
get
a
s
patia
ll
y
m
eaning
f
ul
inf
or
m
at
ion
of
each
obj
ect
s/
cr
op
. An
im
po
rtant
thi
ng
to
be
note
d i
s
the p
ai
r
wise si
m
il
arity
is
co
m
pu
te
d
us
i
ng
t
wo
dista
nce
m
et
ric
s
su
c
h
as
ra
ng
e
distance
am
ong
intensit
y
an
d
intensit
y
an
d
ot
her
m
et
ric
is
sp
ace
distan
ce
a
m
on
g
a
pi
xel
and
it
s
a
djace
nt
pix
el
.
T
hus,
we
hav
e
var
i
ed
t
he
pa
ram
e
te
r
a
nd
e
xp
e
ri
m
ent
is
c
ondu
ct
ed
t
o
e
valuat
e
it
s
influ
e
nce
on
c
la
ssific
at
ion
ac
cur
acy
.
First,
we
ha
ve
e
valu
at
ed
the
in
flue
nce
of
s
pace
par
am
et
er
by
fixi
ng
range
pa
ram
eter
.
I
n
si
m
il
ar
m
ann
er
the
in
f
luence
of
ra
nge
par
am
et
er
is
analy
zed
by
ke
epin
g
sp
ace
pa
ram
et
er
fixe
d.
F
ur
t
her,
it
is
no
ti
ced
ke
epin
g
hi
gh
e
r
va
lue
of
ra
nge
pa
ram
et
er
aff
ect
the
cl
assi
ficat
ion
acc
uracy
.
T
his
is
because
us
in
g
i
m
age
fu
si
on
base
d
rec
ursiv
e
filt
er
with
hi
gh
e
r
range
val
e
will
ne
glect
us
ef
ul
e
dg
e
fe
at
ur
es.
Th
us
,
s
om
e
cl
a
sses
are
not
id
entifi
ed
pro
perl
y.
On
sim
il
ar
t
erm
s
keep
in
g
bo
t
h
s
pace
a
nd
range
t
o
sm
al
l
valu
e
will
no
t
yi
e
ld
good
outc
om
e
s.
This
is
beca
us
e
durin
g
feat
ur
e
e
xtracti
on
proce
dure
ver
y
a
sm
a
ll
er
nu
m
ber
of
local
sp
at
ia
l
inform
ation
w
ou
ld
ha
ve
bee
n
c
on
si
der
e
d.
Fin
al
ly
,
fr
om
e
m
p
iric
al
analy
sis
i
n
w
ork
the
s
pa
ce
and
range
par
am
et
er
are
set
to
20
0 an
d 0.1
resp
ec
ti
vely
.
3.
4
.
Co
m
par
at
i
ve an
aly
sis
of SF
R
-
HSI o
ver s
patial
-
s
p
ectral
feature
-
ba
sed
and
ma
chine l
eanin
g
-
ba
sed
HSI
cr
op cl
assi
ficat
i
on
m
odel
This
sect
io
n
pr
esents
perf
or
m
ance
e
valuati
on
acc
ordin
g
to
pa
ram
et
er
defi
ned
in
[
30
]
,
[
28
]
.
I
n
[30
]
pr
ese
nted
a
m
ulti
scal
e
j
oin
t
colla
borati
ve
r
epr
ese
ntati
on
with
local
ly
adap
ti
ve
dicti
onary
(MLJCR
C).
For
trai
ning
m
od
el
from
each
cl
ass,
10
%
of
the
sam
ples
are
sel
ect
ed
in
ra
nd
om
m
ann
er
an
d
rem
ai
nin
g
sa
m
ples
are
us
e
d
for
te
sti
ng
pur
po
s
e.
The
cl
assif
ic
at
ion
accura
cy
at
tained
by
p
roposed
hy
per
s
pectral
im
age
cl
assifi
cat
ion
pe
rfor
m
ance
over
existi
ng
,
MLJC
RC
cl
assifi
cat
ion
m
od
el
[30
]
is
gr
ap
hic
al
ly
sh
ow
n
in
Figure
3.
It
ca
n
be
se
en
t
hat
S
VM
m
od
el
[30
]
at
ta
in
a
n
OA,
A
A,
a
nd
Ka
pp
a
coeffic
ie
nt
perform
ance
of
82.
79%,
77.83%
,
an
d
80.
33%,
r
e
sp
ect
ively
.
LJCR
C
m
od
el
[
30
]
a
n
OA,
A
A,
a
nd
Kappa
c
oeffic
ie
nt
perf
or
m
ance
of
95.18%
,
91.
22%,
an
d
94.
5%,
resp
ect
ively
.
MLJC
RC
m
od
el
[
30
]
at
ta
in
an
O
A,
A
A,
a
nd
Ka
pp
a
c
oeff
ic
ie
nt
perform
ance
of
96.8
%
,
91.
63%,
an
d
96.
34
%,
r
especti
vel
y.
The
propos
ed
m
od
el
at
ta
ins
a
n
O
A,
A
A,
a
nd
Kappa
c
oe
ff
ic
i
ent
pe
rfo
rm
ance
of
97.98%
,
97.7%,
a
nd
97.69%,
res
pecti
vely
.
F
ro
m
overall
res
ult
at
ta
ined
it
can
be
see
n
th
e
propose
d
SFR
-
HSI
m
od
el
at
ta
in
m
uch
bette
r
cl
assifi
cat
ion
pe
rfor
m
ance
wh
e
n
c
om
par
ed
wi
t
h
oth
e
r
sta
te
-
of
-
a
rt sp
at
ia
l
a
nd
m
achine lear
nin
g ba
sed
classi
ficat
ion
al
gorithm
.
3.
5
.
Co
m
par
at
i
ve an
aly
sis
ov
er
deep
lear
ning base
d hy
perspectr
al cr
op
cl
as
sific
at
i
on
m
od
el
This
sect
io
n
pr
ese
nts
perform
ance
evalua
ti
on
acc
ordin
g
to
pa
ram
et
er
def
i
ned
in
[
24
]
.
In
[
24
]
pr
ese
nted
a
S
horten
Sp
at
ia
l
-
sp
ect
ral
RN
N
wit
h
Pa
rall
el
-
GRU
(S
t
-
SS
-
pGR
U1)
f
or
hype
rs
pectral
crop
cl
assifi
cat
ion
.
Fo
r
trai
ning
m
od
el
17
65
sam
ples
are
sel
ect
ed
in
ra
ndom
m
ann
er
a
nd
re
m
ai
nin
g
sam
pl
es
are
us
e
d
f
or
te
sti
ng
purpose
.
T
he
cal
ci
ficat
ion
a
ccur
acy
at
ta
ine
d
by
pro
posed
hype
rsp
ect
ral
im
age
cl
assifi
cat
ion
perform
ance
ov
er
e
xisti
ng,
MLJC
RC
cl
as
sific
at
ion
m
od
el
[2
4]
is
grap
hical
ly
sh
own
in
Fig
ur
e
4.
It
can
be
seen
t
hat
St
-
SS
-
GR
U
m
od
el
[
24
]
at
ta
in
an
O
A
perf
or
m
a
nce
of
87.
16%
.
St
-
SS
-
pGRU
m
od
el
[
24
]
a
n
O
A
perform
ance
of
9
0.3
5%
.
GR
U
m
od
el
[2
4]
at
ta
in
an
OA
of
77.
01%.
T
he
pro
posed
m
od
el
attai
ns
an
O
A
perform
ance
of
98.47%.
F
rom
ov
erall
res
ul
t
at
ta
ined
it
c
an
be
see
n
the
pro
posed
m
od
el
at
ta
in
m
uch
bette
r
cl
assifi
cat
ion
pe
rfor
m
ance wh
en
c
om
par
ed w
it
h
oth
e
r dee
p
l
earn
i
ng classi
fi
cat
ion
al
gorith
m
.
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Sema
ntic f
eatu
re ext
ra
ct
io
n m
et
hod
fo
r
hyp
e
rsp
ect
r
al crop
cl
as
sif
ic
ation
(
M. C. Giri
sh B
abu
)
393
Figure
3. The
c
la
ssific
at
ion
outc
om
e att
a
ined by
pro
po
se
d SFR
-
HS
I
classi
ficat
ion m
od
el
over
existi
ng
HS
I
classi
ficat
ion m
et
ho
d
Figure
4. The
c
la
ssific
at
ion
outc
om
e att
a
ined by
pro
po
se
d SFR
-
HS
I
classi
ficat
ion m
od
el
over
existi
ng d
e
e
p
l
earn
i
ng HSI cl
assifi
cat
ion
m
et
hod
4.
CONCL
US
I
O
N
In
t
his
w
ork,
we
ha
ve
c
onduced
dee
p
r
oo
te
d
analy
sis
of
var
i
ou
s
hype
r
sp
ect
ral
cr
op
c
la
ssific
at
ion
m
od
el
.
Fr
om
extensi
ve
analy
sis,
it
is
no
te
d
that
crop
ide
ntific
at
ion
a
nd
cl
assifi
cat
ion
us
i
ng
sat
el
li
te
i
mages
is
chall
eng
i
ng.
V
ario
us
m
et
ho
dolo
gies
f
or
fea
ture
e
xtracti
on
,
featu
re
sel
ect
ion
,
an
d
cl
assi
ficat
ion
m
od
el
us
in
g
m
achine
le
ar
nin
g
an
d
dee
p
le
arn
i
ng
m
od
el
by
ex
plo
it
in
g
s
patia
l
and
s
pec
tral
inf
or
m
at
ion
has
been
pre
sented
for
crop
cl
assif
ic
at
ion
.
Each
m
od
el
has
it
s
o
wn
a
dv
a
ntage
s
and
disa
dv
a
nt
age.
E
xisti
ng
dim
ension
al
reducti
on
te
chn
iq
ue
featur
es
ca
nnot
as
su
re
t
hat
the
s
pectrall
y
us
ef
ul
featur
es
a
re
r
et
ai
ned
.
F
or
im
pr
ov
in
g
cl
ass
ific
at
ion
accuracy
c
onsi
der
i
ng
noisy
a
nd
m
ixed
pi
xe
l
env
i
ronm
ent,
existi
ng
m
et
ho
d
ad
opte
d
de
ep
le
ar
ning
a
ppr
oac
h
rather t
ha
n
m
a
chine
le
a
rn
i
ng
m
od
el
. Thus,
t
hese m
od
el
in
duces
high c
ompu
ta
ti
on
ov
e
r
he
ad
a
nd r
e
qu
i
r
es large
nu
m
ber
of
la
be
le
d
trai
ni
ng
s
a
m
ple.
The
refor
e
it
is
im
po
rtant
to
br
i
ng
good
tra
deoff
a
m
on
g
cl
assifi
cat
ion
accurac
y
an
d
com
pu
ta
ti
on
al
com
plexity
.
Fo
r
a
ddressi
ng
these
issues
,
this
work
presente
d
an
ef
fici
ent
dim
ension
al
re
du
ct
io
n
te
ch
ni
qu
e
t
hat
retai
n
m
eaningfu
l
s
patia
l
inform
ation
.
Alon
g
wit
h,
ca
n
ad
dress
no
isy
and
m
ixed
pixe
l
prob
le
m
affe
ct
ing
cr
op
cl
assifi
cat
ion
by
le
arn
in
g
(i.e.,
captu
rin
g)
i
nhere
nt
pr
op
e
rtie
s
of
ph
ysi
cal
cha
ra
ct
erist
ic
s
of
dif
fer
e
nt
obj
ect
s
unde
r
dif
fer
e
nt
il
lu
m
inati
on
and
cl
im
atic
con
diti
on.
T
he
pr
opos
e
d
hype
rsp
ect
ral
c
rop
cl
assi
ficat
ion
m
od
el
is
e
f
fici
ent
as
it
re
quires
ve
ry
le
ss
trai
ning
data
w
hen
c
om
par
ed
with
deep
le
a
rn
i
ng
base
d
cr
op
cl
a
ssific
at
ion
m
eth
od.
E
xperim
e
nt
are
co
nduct
ed
on
sta
nda
rd
dataset
.
The
e
xisti
ng
MLJC
RC
ob
ta
in
an
O
A,
A
A
,
an
d
Kappa
c
oeffici
ent
pe
rfor
m
ance
of
96.8
%
,
91.
63%,
and
96.
34%
w
hich
i
s
m
uch
higher
t
han
othe
r
e
xis
ti
ng
cr
op
cl
ass
ific
at
ion
m
od
e
l.
The
pro
po
se
d
m
od
el
at
ta
in
m
uch
s
up
e
rio
r
O
A,
AA, and K
ap
pa
co
eff
ic
ie
nt perf
or
m
ance o
f 97.9
8%
, 9
7.7
%, an
d 97.69%
w
hich bett
er th
an
MLJC
RC
. Fur
t
her,
exp
e
rim
ent
are
co
nducted
f
or
c
om
par
in
g
pro
po
se
d
m
odel
existi
ng
de
ep
le
ar
ning
ba
sed
m
o
dels.
T
he
S
S
-
pG
R
U
m
od
el
a
tt
ai
n
an
OA
pe
rfor
m
ance
of
90.
35
w
hich
is
m
uch
higher
t
ha
n
oth
er
existi
ng
cr
op
cl
assifi
cat
ion
m
od
el
.
The
pr
opos
e
d
m
od
el
at
ta
in
m
uch
superi
or
O
A
perf
or
m
ance
of
98
.47%
w
hich
be
tt
er
than
SS
-
pGRU.
Th
us
,
pro
po
se
d
m
od
el
is
m
uch
e
ff
ic
ie
nt
wh
e
n
c
om
par
ed
with
e
xisti
ng
hy
per
s
pectra
l
i
m
age
cl
assifi
cat
ion.
Fu
tu
re
w
ork
would
c
on
si
de
r
inducin
g
art
ific
ia
l
no
ise
i
nto
hy
per
s
pect
ral
i
m
age
and
see
ho
w
the
m
od
el
perform
.
Fu
rth
er,
re
fine
t
he
dim
ension
al
re
duct
ion
m
et
hod
for
obta
inin
g
m
or
e
eff
ic
ie
nt
feat
ur
e
e
xtracti
on
a
nd
al
so
dev
el
op a
n
im
pr
oved
classi
ficat
ion
al
gorithm
that le
arn
s m
ixed pixel
m
or
e eff
ic
ie
ntly
.
REFERE
NCE
S
[1]
P
.
Gerl
and
et
al
.
,
“
W
orld
popula
ti
on
stabi
l
izati
on
unli
kely
th
is
centur
y
,”
Sc
ie
n
ce
,
v
ol.
346,
no.
62
06,
pp.
234
-
237
,
2014,
doi
:
10
.
11
26/sci
en
ce
.
12
57
469
.
[2]
A
.
K.
W
hit
cra
ft
,
I
.
B
.
-
R
eshe
f
a
nd
C
.
O.
Jus
ti
ce
,
“
A
fra
m
ework
for
def
ini
ng
sp
at
i
al
l
y
exp
licit
e
art
h
observa
ti
on
req
uire
m
ent
s
for
a
globa
l
agr
i
cu
lt
ura
l
m
onit
or
in
g
ini
ti
a
ti
v
e
(GEOG
LAM)
,”
Rem
ote
Sens
,
vol.
7,
pp.
1461
-
1481
,
2015
,
doi
:
10.3390/rs
7020
1461
.
[3]
J
.
Fuhrer
and
P
.
J
.
Gregor
y
,
“
Cli
m
at
e
Chang
e
Im
pac
t
and
Adapta
ti
on
in
Agri
c
ult
ura
l
S
y
s
te
m
s:
Soil
Ec
os
y
ste
m
Mana
gement in S
ustai
nabl
e
Agri
cul
tur
e
,
”
W
al
l
in
gford,
CT
,
US
A
:
CABI
,
201
4
.
[4]
A
.
Maha
l
anobi
s
,
B
.
V
.
V
.
Kum
ar
and
S
.
R
.
Sim
s
,
“
Di
stanc
e
-
cl
assifi
er
cor
relati
on
f
il
t
ers
for
m
ult
ic
l
ass
ta
rg
e
t
rec
ogni
ti
on
,”
Ap
pli
ed
Optic
s
,
vol
.
35
,
no
.
17
,
pp
.
3127
-
3133,
201
3
,
doi
:
10
.
1364/
AO
.
35.
003127
.
70
75
80
85
90
95
1
0
0
SVM
LJCR
C
M
LJCR
C
SFR-
HSI
Clas
sif
caitio
n
Per
f
o
rm
an
ce
(%)
Clas
sif
icatio
n
m
o
d
el
Hy
pers
pect
ral
crop cl
ass
i
fi
cat
i
on
perf
orm
ance
OA
(
%)
AA
(
%)
Kap
p
a (
%)
0
20
40
60
80
100
120
Clas
sific
at
ion
perfo
ramnce
Clas
sific
at
ion
mod
el
H
y
perspetral
crop
clas
sifica
tion
o
v
eral
l
a
ccuracy
perf
o
ra
m
a
nce
GRU
[24
]
St
-SS
-GR
U
[24
]
SS-pGR
U [24
]
SFR-HS
I
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
23
, N
o.
1
,
Ju
ly
20
21
:
3
8
7
-
3
9
5
394
[5]
F.
Low,
U.
Miche
l,
S.
Dec
h
and
C.
Conrad,
“
Impa
ct
of
feature
select
ion
on
the
a
cc
ura
c
y
and
spat
ia
l
unc
ert
a
inty
o
f
per
-
field
cro
p
c
l
assific
a
ti
on
usin
g
support
ve
ct
or
m
ac
hine
s,
”
IS
P
RS
J
.
Phot
ogra
mm
et
ry
Re
mot
e
Sens.
,
vo
l.
85,
p
p.
102
-
119,
2013
,
doi:
1
0
.
1016/j
.
is
prsjprs.2013.
08.
007
.
[6]
S.
Murm
u
and
S
.
Biswas,
“
Applic
ation
of
fuz
z
y
l
ogic
and
neur
a
l
net
work
in
cro
p
cl
assifi
ca
t
ion:
A
rev
ie
w,
”
Aquat
i
c
Proce
dia
,
vo
l. 4, pp. 1203
-
1210,
2015
,
doi
:
10
.
10
16/j
.
aqpr
o.
2015
.
02.
153
.
[7]
K.
Ta
tsum
i,
Y
.
Yam
ashiki
,
M.
A.
C.
Torr
es
and
C.
L
.
R.
Taipe
,
“
Crop
cl
assificat
ion
of
upla
nd
fields
using
Rando
m
fore
st
of
t
ime
-
serie
s
La
ndsat
7
ET
M+
d
ata,
”
Comput.
Elec
tron.
Agri
cul
ture
,
vol.
115
,
pp.
17
1
-
179,
2015
,
do
i:
10.
1016/j.c
om
pa
g.
2015.
05
.
001
.
[8]
M.
G
ong,
M.
Zha
ng
and
Y.
Yuan,
“
Uns
uper
vised
Band
Sele
c
ti
on
Based
on
Evol
uti
on
ar
y
Multi
ob
jecti
v
e
Optimiza
ti
o
n
for
H
y
p
erspe
ct
r
al
I
m
age
s,
”
in
IE
EE
Tr
ansacti
ons
on
Geosci
en
ce
and
Re
mot
e
S
ensing
,
vol
.
54
,
no
.
1
,
pp.
544
-
557
,
Jan
.
2016
,
doi
:
10
.
1
109/T
GRS
.
2015.
2461653
.
[9]
W
.
Sun
and
Q.
Du,
“
Graph
-
Reg
ula
ri
ze
d
Fast
an
d
Robust
Principal
Com
ponent
Anal
y
sis
for
H
y
per
spec
tr
al
B
an
d
Sele
c
ti
on,
”
in
IE
EE
Tr
ansacti
ons
on
Geosci
ence
and
Re
mot
e
Sen
sing
,
vol
.
56
,
no
.
6,
pp.
3185
-
31
95,
June
2018
,
doi:
10
.
1109/T
G
RS
.
2018.
279444
3
.
[10]
H.
Zhai,
H.
Zhang,
L
.
Zh
ang
an
d
P.
Li,
“
La
pl
aci
an
-
Regularized
Low
-
Rank
Subs
pac
e
C
luste
r
ing
for
H
y
per
sp
ec
tr
a
l
Im
age
Band
Sele
ction,
”
in
IE
EE
Tr
an
sacti
ons
on
Geosci
enc
e
an
d
Re
mote
Sensin
g
,
vol.
57,
no.
3
,
pp.
1723
-
1740,
Marc
h
2019
,
doi
:
10.
1
109
/T
GRS
.
2018.
2868796
.
[11]
F.
Melga
ni
and
L.
Bruz
zone,
“
Cla
ss
ifi
ca
t
ion
of
h
y
p
erspe
ct
r
a
l
remote
sensin
g
images
with
support
vec
tor
m
ac
hine
s,
”
in
IE
EE
Tr
ansacti
ons
on
Geosci
enc
e
and
Re
mot
e
Sen
sing
,
vol
.
42,
no
.
8,
pp.
1778
-
17
90,
Aug.
2004
,
doi:
10
.
1109/T
G
RS
.
2004
.
831865
.
[12]
M.
Zha
ng,
M.
Gong
and
Y.
Chan,
“
Hy
p
erspe
c
tr
al
band
select
io
n
base
d
on
m
ult
i
-
objecti
v
e
optim
iz
at
ion
with
hi
gh
informati
on
an
d
low
red
undancy
,
”
Applie
d
Soft
Computing
,
vol
.
7
0,
pp.
604
-
6
21,
2018
,
do
i:
10.
1016/j.a
soc
.
2
018.
06.
009
.
[13]
P.
Hu,
X.
Li
u,
Y.
Ca
i
and
Z
.
C
ai
,
“
Band
Sel
ec
t
ion
of
H
y
per
spe
ct
ra
l
Im
age
s
Us
ing
Multi
objecti
v
e
Optimiza
t
ion
-
Based
Sparse
Se
lf
-
Repre
sen
ta
t
io
n,
”
in
IE
EE
Geo
scie
nc
e
and
R
e
mote
Sensing
Le
tt
ers
,
vol
.
16,
no
.
3,
pp
.
452
-
456
,
Marc
h
2019
,
doi
:
10.
1109
/L
GRS
.
2018.
2872540
.
[14]
Q.
W
ang,
F.
Zha
ng
and
X.
Li,
“
Optimal
Cluste
ring
Fram
ework
for
Hy
p
erspe
ct
ra
l
Band
Selecti
on
,
”
in
IEEE
Tr
ansacti
ons
on
Geosci
en
ce
and
Re
mot
e
S
ensing
,
vo
l.
5
6,
no
.
10
,
pp
.
5910
-
5922
,
Oct.
2018
,
do
i
:
10.
1109/T
GRS
.
2018.
2828161
.
[15]
X.
Jiang,
X.
Song,
Y.
Zh
ang,
J.
Jiang,
J
.
Gao,
an
d
Z.
C
ai
,
“
L
apl
a
ci
an
r
egulari
z
ed
spati
al
-
awa
r
e
co
ll
abor
at
iv
e
gra
p
h
for
discri
m
ina
n
t ana
l
y
sis of
h
y
pe
rspec
tral
imag
er
y
,
”
Re
mot
e
S
ensi
ng
,
vol
.
11
,
no
.
1
,
p
p
.
29
,
2019
.
[16]
W
.
Li
ao
,
A.
Pi
zur
ica,
P.
Sch
eu
nder
s,
W
.
Phili
ps
and
Y.
Pi,
“
Sem
isupervi
se
d
Loc
a
l
Discrimin
ant
Anal
y
sis
for
Feat
ure
Ext
r
ac
t
i
on
in
H
y
p
erspe
c
tra
l
Im
ag
es,
”
in
IEE
E
Tr
ansactions
on
Geosci
e
nce
and
Re
mot
e
Sensing
,
vo
l.
5
1,
no.
1
,
pp
.
184
-
1
98,
Jan
.
2013
,
d
oi:
10
.
1109/T
G
RS
.
2012.
220010
6
.
[17]
A.
Vill
a
,
J.
A.
Bene
dikt
ss
on,
J.
Chanussot
and
C.
Jutte
n
,
“
H
y
p
erspe
ct
r
al
Im
age
Cla
ss
ifi
c
at
ion
w
it
h
Inde
pend
e
nt
Com
ponent
Discriminant
Anal
y
s
i
s,
”
in
IE
EE
Tr
ansacti
ons
on
Geosci
ence
and
Re
mote
Sensing
,
vol
.
49,
no.
12,
pp.
4865
-
4876,
De
c. 2011, doi
:
10.
11
09/T
GRS
.
2011.
2153861
.
[18]
S.
Prasad
and
L
.
M.
Bruc
e,
“
Li
m
it
ations
of
Prin
cipal
Com
ponent
s
Anal
y
s
is
for
H
yper
spec
tr
al
Ta
rg
et
R
ec
ogni
ti
on,
”
in
IE
EE
Geos
c
ie
nc
e
and
Rem
ote
Sensing
Lett
ers
,
vol.
5,
no
.
4
,
pp.
625
-
629,
Oct.
2008
,
d
oi:
10.
1109/L
GRS
.
2008.
2001282
.
[19]
L.
W
ang,
J.
Zhang,
P.
Li
u
,
K.
-
K
.
R.
Choo
and
F
.
Huang,
“
Spectral
–
spat
ia
l
m
ulti
-
fea
tur
e
-
base
d
d
ee
p
l
ea
rning
for
h
y
per
sp
ec
tr
al
r
emote
sensing
image
cl
assifi
ca
t
ion,
”
So
ft
Computing
,
vol
.
21,
pp.
213
-
221,
2016
,
d
oi:
10.
1007/s00500
-
016
-
2246
-
3
.
[20]
J.
Jiang,
C.
Chen,
Y.
Yu,
X.
J
ia
ng
and
J.
Ma,
“
Spati
al
-
Aw
are
Coll
abor
ative
Repre
sent
at
ion
f
or
Hy
p
erspe
c
tral
Remote
Sensing
Im
age
Cla
ss
ific
at
ion
,
”
in
IE
EE
Geosci
ence
and
Re
mote
S
ensing
Lett
ers
,
vol
.
14,
no.
3,
pp
.
404
-
408,
Mar
ch
201
7,
doi
:
10
.
1109/
LGRS.2016.
264
5708
.
[21]
X.
Kang,
S.
Li
and
J.
A.
Ben
edi
ktsson,
“
Feature
Ext
ra
ct
ion
of
Hy
p
erspe
c
tral
Im
age
s
w
it
h
I
m
age
Fus
ion
an
d
Rec
ursive
Fi
lt
er
i
ng,
”
in
IE
EE
Tr
ansacti
ons
on
Geosci
en
ce
and
R
emote
Sensing
,
v
ol.
52,
no.
6
,
pp.
3742
-
3752,
June
2014,
doi
:
10
.
11
09/T
GRS
.
2013.
2275613
.
[22]
A.
Santa
ra
et
al
.
,
“
BA
SS
Net:
Band
-
Adapti
ve
Spe
ct
ra
l
-
Spatial
Fea
ture
Learni
ng
Neura
l
Network
f
or
Hy
p
erspe
c
tral
Im
age
Cla
ss
ifi
cation,
”
in
IE
EE
Tr
ansacti
ons
on
Geosci
ence
and
Re
mote
Sensin
g
,
vol.
55,
no.
9
,
pp.
5293
-
5301,
Sept.
2017
,
do
i:
10.
1109/T
GRS
.
2017.
2705073
.
[23]
X.
Cao,
F.
Zhou
,
L.
Xu,
D.
Meng
,
Z.
Xu
and
J.
Pa
isley
,
"H
y
per
spe
ct
ra
l
Im
age
Cla
s
sific
a
ti
on
W
it
h
Markov
Random
Fiel
ds
and
a
Co
nvolut
ional
Neu
ral
Ne
twork,
"
in
IEE
E
Tr
ansacti
ons
on
Image
Pro
ce
ss
ing
,
vo
l.
27
,
no.
5
,
pp.
2354
-
2367,
Ma
y
2018
,
doi
:
10
.
1109/T
I
P.2018.
2799324
.
[24]
H.
Luo,
“
Shorten
Spat
ia
l
-
sp
ec
tr
al
RNN
with
Para
l
le
l
-
G
RU
for
H
y
pe
rspec
tral
Im
ag
e
Cla
ss
ifica
t
io
n,
”
arXiv
:1810.
1256
3v1
,
2018
.
[25]
Y.
Cai,
X.
Li
u
and
Z
.
Ca
i,
"BS
-
Nets:
An
End
-
to
-
End
Fram
ework
for
Band
Se
l
ec
t
ion
of
H
y
per
spec
tral
Im
age,"
in
IE
EE
Tr
ansa
ct
ions
on
Geosci
en
ce
and
Rem
ote
Sensing
,
vol.
58,
no.
3,
p
p.
1969
-
1984,
Marc
h
2020,
doi:
10.
1109/T
GRS
.
2019.
2951433
.
[26]
Y.
Liang,
X
.
Zh
ao,
A.
J.
X.
Guo
and
F.
Zhu,
"H
y
per
sp
ec
tr
al
Im
a
ge
Cla
ss
ifica
ti
on
W
it
h
Dee
p
M
etric
L
ea
rn
ing
an
d
Condit
ional
Ran
dom
Fiel
d,
"
in
I
EE
E
Geosci
enc
e
and
Re
mote
Sensing
Lett
ers
,
v
ol.
17,
no.
6,
pp.
1042
-
1046,
Jun
e
2020,
doi
:
10
.
11
09/L
GRS
.
2019.
2939356
.
[27]
F.
Li,
K
.
Ergun
,
O.
Kili
n
c
and
Y
.
Qiao
,
“
Vege
t
at
i
on
Cla
ss
ifica
t
ion
Us
ing
H
y
p
erspe
ct
ra
l
Im
age
s
,
”
in
ECE
228
and
SIO209 Mac
hine l
earning
for ph
y
sical
app
li
ca
ti
on
s
,
2018.
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Sema
ntic f
eatu
re ext
ra
ct
io
n m
et
hod
fo
r
hyp
e
rsp
ect
r
al crop
cl
as
sif
ic
ation
(
M. C. Giri
sh B
abu
)
395
[28]
L.
W
ang,
Y.
Fe
ng,
Y.
Gao,
Z
.
W
ang
and
M.
He,
“
Com
pre
ss
ed
Sensing
Rec
onst
ruc
ti
on
of
H
y
pe
rspec
tral
Im
age
s
Based
on
Spectral
Unm
ixi
ng,
”
in
IE
EE
Journal
of
Sel
ecte
d
Topics
in
Appl
i
ed
Earth
Obs
erv
ati
ons
and
Re
mote
Sensing
,
vo
l. 11, no. 4, pp. 1266
-
1284,
April
2
01
8,
doi
:
10
.
1109/J
STARS
.
2017.
2787483
.
[29]
M.
C.
G
.
Babu
and
M.
C.
Padma
,
“
A
Eff
ic
ie
n
t
Soluti
on
for
Clas
sific
at
ion
of
Cr
ops
using
Hy
per
Spect
ral
Sat
el
l
ite
Im
age
s,”
Int
ernati
onal
Journal
of
Innov
ative
Te
chnol
ogy
and
E
xpl
oring
Engi
n
e
ering
(
IJI
TEE)
,
vol.
9
,
no
.
2
,
pp
.
5204
-
5211,
201
9
,
doi
:
10
.
35940
/i
jitee.B7311.
12
9219
.
[30]
J.
Yang
and
J.
Qian,
“
H
y
p
erspe
ct
ra
l
Im
age
Cl
assific
a
ti
on
vi
a
Multi
sca
le
Join
t
C
oll
abor
at
iv
e
Rep
rese
ntation
W
ith
Loc
a
lly
Adapti
v
e
Dict
ion
ar
y
,
”
in
IEEE
Geosci
e
nce
and
Re
mote
Sensing
Lett
ers
,
vol.
15,
no.
1,
pp.
112
-
116,
Jan.
2018,
doi
:
10
.
11
09/L
GRS
.
2017.
2776113
.
BIOGR
AP
HI
ES OF
A
UTH
ORS
Girish
Babu
M
C
recei
ved
th
e
B.
E
Degre
e
in
Com
pute
r
Scie
n
ce
&
Eng
ineeri
n
g
in
2004
,
M
.
Tech
in
2009
fro
m
Visvesvara
y
a
Te
chno
logi
c
al
U
nive
rsit
y
.
He
is
r
ese
arc
h
schol
ar
i
n
the
field
of
Im
age
proc
essing
at Visvesvaraya
T
ec
hnolog
ical
Univer
sit
y
Ind
ia.
Dr.
Padm
a
M
C
rec
e
ive
d
th
e
B.
E
Degre
e
in
Co
m
pute
r
Scie
nc
e
&
Engi
ne
eri
ng
i
n
1990,
Ms
c
(
Tech
)
from
Uni
ver
sit
y
of
M
y
sor
e,
Ind
ia
in
2004
and
th
e
Phd
Deg
ree
in
Im
age
Proce
ss
ing
from
Visvesvara
y
a
Te
chno
logi
c
al
Univer
sit
y
in
2010.
She
h
a
s
publi
shed
m
ore
th
an
45
int
ern
at
ion
al
/n
ational
Journals.
Her
m
ai
n
rese
ar
ch
are
a
inc
lud
es
Natur
al
L
anguage
Proce
ss
in
g,
Data
m
ini
ng
,
Im
age
pro
ce
ss
ing a
nd
Pattern
r
ec
og
nit
ion.
Evaluation Warning : The document was created with Spire.PDF for Python.