I
nd
o
ne
s
ia
n J
o
urna
l o
f
E
lect
rica
l En
g
ineering
a
nd
Co
m
pu
t
er
Science
Vo
l.
3
7
,
No
.
3
,
Ma
r
ch
20
2
5
,
p
p
.
1
7
1
2
~
1
725
I
SS
N:
2
502
-
4
7
52
,
DOI
: 1
0
.
1
1
5
9
1
/ijee
cs
.v
3
7
.
i
3
.
pp
1
7
1
2
-
1
7
2
5
1712
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ee
cs
.
ia
esco
r
e.
co
m
Deep learni
ng
t
ec
hniques
for sa
telli
te
ima
g
e clas
sifica
tion
Su
re
s
h K
um
a
r
M
u
s
a
li
1
,
Ra
j
esh
wa
ri
J
a
ntha
k
a
l
1
,
Nuv
v
us
et
t
y
Ra
j
a
s
ek
ha
r
2
,3
1
D
e
p
a
r
t
me
n
t
o
f
I
n
f
o
r
mat
i
o
n
S
c
i
e
n
c
e
a
n
d
E
n
g
i
n
e
e
r
i
n
g
,
D
a
y
a
n
a
n
d
a
S
a
g
a
r
C
o
l
l
e
g
e
o
f
E
n
g
i
n
e
e
r
i
n
g
,
B
a
n
g
a
l
o
r
e
,
V
i
s
v
e
s
v
a
r
a
y
a
T
e
c
h
n
o
l
o
g
i
c
a
l
U
n
i
v
e
r
s
i
t
y
,
B
e
l
a
g
a
v
i
,
I
n
d
i
a
2
D
e
p
a
r
t
me
n
t
o
f
I
n
f
o
r
mat
i
o
n
Te
c
h
n
o
l
o
g
y
,
G
o
k
a
r
a
j
u
R
a
n
g
a
r
a
j
u
I
n
st
i
t
u
t
e
o
f
E
n
g
i
n
e
e
r
i
n
g
a
n
d
Te
c
h
n
o
l
o
g
y
,
JN
TU
,
H
y
d
e
r
a
b
a
d
,
I
n
d
i
a
3
V
i
s
v
e
s
v
a
r
a
y
a
T
e
c
h
n
o
l
o
g
i
c
a
l
U
n
i
v
e
r
s
i
t
y
,
B
e
l
a
g
a
v
i
,
I
n
d
i
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
J
u
n
3
,
2
0
2
4
R
ev
is
ed
Sep
19
,
2
0
2
4
Acc
ep
ted
Oct
7
,
2
0
2
4
Be
c
a
u
se
o
f
it
s
wid
e
ra
n
g
e
o
f
u
se
s
in
c
o
m
p
u
ter
v
isio
n
a
p
p
li
c
a
ti
o
n
s,
in
c
lu
d
in
g
ima
g
e
re
tri
e
v
a
l,
re
m
o
te
se
n
sin
g
,
o
b
jec
t
re
c
o
g
n
it
io
n
,
sc
e
n
e
a
n
a
ly
sis,
a
n
d
su
rv
e
il
lan
c
e
,
ima
g
e
c
las
sifica
ti
o
n
h
a
s
a
tt
ra
c
ted
a
l
o
t
o
f
a
tt
e
n
ti
o
n
.
As
sig
n
in
g
a
p
p
ro
p
riate
c
las
s
lab
e
ls
to
ima
g
e
s
a
c
c
o
rd
in
g
t
o
t
h
e
ir
c
o
n
ten
ts
is
t
h
e
p
rima
ry
o
b
jec
ti
v
e
o
f
ima
g
e
c
las
sifica
ti
o
n
.
In
t
h
e
d
o
m
a
in
o
f
re
m
o
te
se
n
si
n
g
,
ima
g
e
c
las
sifica
ti
o
n
a
n
d
a
n
a
ly
sis
p
lay
c
ru
c
ial
ro
les
in
b
o
th
m
il
it
a
ry
a
n
d
c
iv
i
l
a
p
p
li
c
a
ti
o
n
s.
Co
n
v
e
n
ti
o
n
a
l
m
e
th
o
d
s
fo
r
sc
e
n
e
a
n
a
ly
sis
a
n
d
re
m
o
te
se
n
sin
g
d
e
p
e
n
d
e
d
o
n
lo
w
-
lev
e
l
re
p
re
se
n
ta
ti
o
n
s
o
f
fe
a
tu
re
s,
su
c
h
a
s
t
h
o
se
o
f
c
o
lo
r
a
n
d
tex
tu
re
.
Ho
we
v
e
r,
re
c
e
n
t
a
d
v
a
n
c
e
m
e
n
ts
h
a
v
e
sh
ift
e
d
to
wa
rd
s
t
h
e
u
se
o
f
c
o
n
v
o
lu
ti
o
n
a
l
n
e
u
ra
l
n
e
tw
o
rk
s
(C
NN
s),
wh
ich
h
a
v
e
sh
o
wn
p
r
o
m
isi
n
g
re
su
lt
s
in
re
m
o
te
se
n
sin
g
a
n
d
sc
e
n
e
c
las
sifica
ti
o
n
tas
k
s.
I
n
l
ig
h
t
o
f
e
ffe
c
t
iv
e
n
e
ss
of
d
e
e
p
lea
rn
i
n
g
(DL)
m
o
d
e
ls,
t
h
is
r
e
se
a
rc
h
a
ims
to
d
e
v
e
lo
p
fo
u
r
DL
m
o
d
e
ls b
y
fin
e
t
u
n
i
n
g
a
lrea
d
y
e
x
isti
n
g
D
L
m
o
d
e
ls
-
CNN
s,
re
si
d
u
a
l
n
e
u
ra
l
n
e
two
r
k
(Re
sN
e
t),
v
isu
a
l
g
e
o
m
e
try
g
r
o
u
p
(VG
G
-
1
9
),
n
e
two
rk
m
o
b
i
le
n
e
t
V2
b
a
se
d
m
o
d
e
l
a
n
d
c
las
sifies
sa
telli
te i
m
a
g
e
s o
f
RS
I
-
CB
2
5
6
d
a
ta se
t
i
n
t
o
f
o
u
r
c
las
se
s
n
a
m
e
ly
c
lo
u
d
y
,
d
e
se
rt,
g
re
e
n
_
a
re
a
a
n
d
wa
ter.
F
o
r
t
h
e
RS
I
-
CB
2
5
6
d
a
tas
e
t,
a
p
p
ro
p
riate
n
e
two
r
k
str
u
c
tu
re
s
a
re
e
x
p
lo
re
d
i
n
t
h
is
re
se
a
rc
h
to
g
e
t
g
o
o
d
p
e
rf
o
rm
a
n
c
e
.
T
h
e
CNN
,
Re
sN
e
t
a
n
d
VG
G
-
1
9
b
a
se
m
o
d
e
ls
a
c
h
iev
e
d
a
n
a
c
c
u
ra
c
y
o
f
9
0
.
4
8
,
9
2
.
6
8
a
n
d
9
1
.
1
8
re
sp
e
c
ti
v
e
ly
.
Wh
il
e
t
h
e
m
o
b
i
le
n
e
t
V2
b
a
se
d
m
o
d
e
l
o
u
t
p
e
rfo
rm
e
d
t
h
e
o
t
h
e
r
th
re
e
m
o
d
e
ls wi
th
9
6
.
8
3
%
a
c
c
u
ra
c
y
.
K
ey
w
o
r
d
s
:
Dee
p
lear
n
in
g
I
m
ag
e
class
if
icatio
n
Neu
r
al
n
etwo
r
k
s
R
em
o
te
s
en
s
in
g
Satellite im
ag
es
T
h
is i
s
a
n
o
p
e
n
a
c
c
e
ss
a
rticle
u
n
d
e
r th
e
CC B
Y
-
SA
li
c
e
n
se
.
C
o
r
r
e
s
p
o
nd
ing
A
uth
o
r
:
Su
r
esh
Ku
m
ar
Mu
s
ali
Dep
ar
tm
en
t o
f
I
n
f
o
r
m
atio
n
Scien
ce
an
d
E
n
g
in
ee
r
i
n
g
,
Day
a
n
an
d
a
Sag
ar
C
o
lleg
e
o
f
E
n
g
in
e
er
in
g
,
B
a
n
g
a
l
o
r
e
V
i
s
v
e
s
v
a
r
a
y
a
T
e
c
h
n
o
l
o
g
i
c
a
l
U
n
i
v
e
r
s
i
t
y
B
e
l
a
g
a
v
i
-
5
9
0
0
1
8
,
K
a
r
n
a
t
a
k
a
,
I
n
d
i
a
E
m
ail:
s
u
r
esh
k
u
m
ar
-
is
e@
d
ay
a
n
an
d
asag
ar
.
e
d
u
1.
I
NT
RO
D
UCT
I
O
N
T
h
er
e
h
asn
'
t
b
ee
n
m
u
ch
h
ig
h
-
r
eso
lu
tio
n
s
atellite
p
ictu
r
e
class
if
icatio
n
an
d
an
aly
s
is
d
o
n
e
wit
h
tr
ad
itio
n
al
tech
n
iq
u
es.
T
h
is
is
b
ec
au
s
e
th
e
im
ag
er
y
h
as
s
o
m
an
y
in
t
r
icate
d
etails.
Sp
ec
tr
al
s
ig
n
atu
r
es,
in
tr
icate
tex
tu
r
e
an
d
s
h
ap
e
,
s
p
atial
lin
k
ag
es,
an
d
tem
p
o
r
al
s
h
if
ts
ar
e
s
o
m
e
o
f
th
e
ch
ar
ac
ter
is
tics
th
at
d
ef
in
e
th
ese
im
ag
es.
T
h
is
r
esu
lt
d
em
o
n
s
tr
ates
h
o
w
d
ee
p
lear
n
in
g
alg
o
r
ith
m
s
ca
n
co
m
p
r
eh
en
d
th
e
co
m
p
lex
an
d
n
o
n
-
u
n
if
o
r
m
f
ea
tu
r
es
o
f
h
i
g
h
-
r
eso
l
u
tio
n
r
em
o
te
s
en
s
in
g
im
ag
e
s
.
Den
s
eNe
t1
2
1
an
d
R
esNet1
0
1
,
two
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
s
(
C
NN
)
-
b
ased
ar
ch
itectu
r
es,
o
u
tp
e
r
f
o
r
m
I
n
c
ep
tio
n
V3
,
E
f
f
icien
tNet,
an
d
V
GG1
6
,
s
co
r
in
g
o
v
e
r
9
0
%
in
all
ass
ess
m
en
t
cr
iter
ia
o
n
th
e
E
u
r
o
SAT
d
ataset,
wh
ich
in
clu
d
es
class
if
icatio
n
m
ea
s
u
r
es
in
clu
d
in
g
ac
cu
r
ac
y
,
r
ec
all,
p
r
ec
is
io
n
,
an
d
F1
-
s
co
r
e
[
1]
.
T
h
is
r
esear
ch
p
ap
er
'
s
p
r
im
ar
y
co
n
tr
i
b
u
tio
n
s
ca
n
b
e
s
u
m
m
ed
u
p
as f
o
llo
ws:
−
I
t o
f
f
er
s
a
th
o
r
o
u
g
h
e
x
am
in
ati
o
n
o
f
d
ee
p
lea
r
n
in
g
tech
n
iq
u
es f
o
r
class
if
y
in
g
s
atellite
im
ag
e
s
.
−
I
t su
g
g
ests
o
p
tim
ized
d
ee
p
lea
r
n
in
g
m
o
d
els f
o
r
s
atellite
im
ag
es in
ter
p
r
etatio
n
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2
5
0
2
-
4
7
52
Dee
p
lea
r
n
in
g
tech
n
iq
u
es fo
r
s
a
tellite ima
g
e
cla
s
s
ifica
tio
n
(
S
u
r
esh
K
u
ma
r
Mu
s
a
li
)
1713
−
V
i
s
u
al
i
z
at
i
o
n
s
o
f
tr
ain
in
g
lo
s
s
an
d
v
alid
atio
n
lo
s
s
o
f
d
if
f
er
en
t
tech
n
iq
u
es a
r
e
also
p
r
esen
ted
.
I
n
o
r
d
e
r
to
class
if
y
s
atellite
p
h
o
to
s
in
to
f
o
u
r
d
is
tin
ct
lo
ca
tio
n
s
d
eser
ts
,
clo
u
d
y
ar
ea
s
,
g
r
e
en
s
p
ac
es,
an
d
wate
r
b
o
d
ie
s
th
is
r
esear
ch
p
ap
er
p
r
o
p
o
s
es
an
im
ag
e
in
ter
p
r
etatio
n
tech
n
i
q
u
e.
Af
ter
th
en
,
th
er
e
ar
e
n
u
m
er
o
u
s
p
o
s
s
ib
le
u
s
es
f
o
r
th
is
in
ter
p
r
etatio
n
.
B
etter
ac
c
u
r
ac
y
is
p
r
o
v
id
ed
b
y
th
e
s
u
g
g
ested
m
o
d
els,
an
d
em
p
ir
ical
f
i
n
d
in
g
s
s
u
p
p
o
r
t
t
h
is
.
Fo
r
im
ag
e
class
if
icatio
n
,
th
is
p
ap
er
p
r
esen
ts
th
e
Mo
b
ilet
NetV2
ar
ch
itectu
r
e
,
wh
ich
h
as
b
ee
n
s
h
o
wn
t
o
b
e
m
o
r
e
ef
f
ec
tiv
e
t
h
an
o
th
e
r
wid
ely
u
s
ed
m
o
d
els.
T
h
is
cr
ea
tiv
e
m
eth
o
d
lo
wer
s
th
e
co
m
p
u
tatio
n
al
b
u
r
d
e
n
o
f
tr
a
in
in
g
th
e
m
o
d
el
wh
ile
s
im
u
ltan
eo
u
s
ly
im
p
r
o
v
in
g
class
if
icatio
n
ac
cu
r
ac
y
.
I
d
en
tify
in
g
an
d
ca
teg
o
r
izin
g
th
in
g
s
f
r
o
m
s
atellite
im
ag
es
is
es
s
en
tial
f
o
r
n
u
m
er
o
u
s
u
s
es,
s
u
ch
as
lan
d
p
lan
n
in
g
,
ec
o
lo
g
y
,
m
ilit
ar
y
,
a
n
d
m
ar
in
e
m
o
n
ito
r
in
g
.
R
ich
in
t
em
p
o
r
al
an
d
s
p
atial
in
f
o
r
m
atio
n
,
s
atellite
im
ag
es
ar
e
u
s
ed
in
n
u
m
er
o
u
s
way
s
to
ad
d
r
ess
p
r
ac
t
ical
r
em
o
te
s
en
s
in
g
is
s
u
es.
T
h
er
e
ar
e
n
u
m
e
r
o
u
s
d
if
f
icu
lties
in
class
if
y
in
g
s
atellite
im
ag
es.
T
h
ese
d
if
f
icu
lties
in
clu
d
e
th
e
d
is
tr
ib
u
tio
n
,
q
u
an
tity
,
q
u
ality
,
an
d
av
ailab
ilit
y
o
f
d
ata.
T
h
ese
p
r
o
b
lem
s
m
ak
e
t
h
e
s
tu
d
y
o
f
s
atellite
im
ag
e
s
m
o
r
e
ch
allen
g
in
g
.
Po
s
it
iv
e
an
d
e
n
co
u
r
a
g
in
g
r
esu
lts
ar
e
o
b
tain
ed
o
n
a
c
h
allen
g
in
g
d
ataset
with
s
ig
n
if
ican
t
in
tr
a
-
an
d
in
ter
-
class
h
eter
o
g
en
eity
u
s
in
g
th
e
p
r
o
p
o
s
ed
tech
n
iq
u
e.
T
h
e
R
SI
-
C
B
2
5
6
d
a
taset sh
o
ws th
at
th
e
r
ec
o
m
m
en
d
ed
m
eth
o
d
h
as g
o
o
d
ac
cu
r
ac
y
[
2
]
.
Dee
p
lear
n
in
g
tec
h
n
iq
u
es
ar
e
b
ec
o
m
i
n
g
m
o
r
e
an
d
m
o
r
e
s
u
cc
ess
f
u
l
th
an
class
ic
m
ac
h
in
e
lear
n
in
g
tech
n
iq
u
es.
T
h
ese
tech
n
iq
u
es
ca
n
lear
n
f
r
o
m
tex
t,
im
a
g
e,
v
i
d
eo
,
an
d
au
d
io
d
ata.
T
h
is
is
es
p
ec
ially
tr
u
e
in
lig
h
t
o
f
r
ec
e
n
t
ad
v
a
n
ce
m
en
ts
in
h
ar
d
war
e
p
o
wer
.
Ar
c
h
itectu
r
e
is
p
r
ed
icted
to
h
a
v
e
s
im
ilar
ef
f
ec
ts
g
iv
e
n
th
e
ef
f
ec
tiv
en
ess
an
d
a
d
v
an
tag
es
o
f
d
ee
p
lear
n
in
g
ap
p
r
o
ac
h
es
in
o
th
er
d
is
cip
lin
es
with
g
r
o
wi
n
g
am
o
u
n
ts
o
f
d
ata
.
I
n
s
tead
o
f
u
s
in
g
b
r
o
ad
s
tr
o
k
e
s
to
r
e
p
r
esen
t
tex
t
u
r
es,
we
n
a
r
r
o
wed
d
o
wn
to
p
ar
ticu
la
r
in
th
is
s
tu
d
y
.
I
n
th
i
s
d
ir
ec
tio
n
,
th
e
d
ee
p
c
o
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
m
o
d
el
wa
s
u
s
ed
to
class
if
y
a
to
tal
o
f
4
,
5
0
0
s
atellite
p
h
o
to
s
th
at
in
clu
d
ed
cl
o
u
d
s
,
d
eser
ts
,
g
r
ee
n
s
p
ac
es,
an
d
wate
r
b
o
d
ie
s
[
3
]
.
On
e
im
p
o
r
tan
t
a
p
p
licatio
n
o
f
r
em
o
te
s
en
s
in
g
(
R
S)
d
ata
-
d
r
iv
en
ea
r
th
o
b
s
er
v
atio
n
tech
n
o
lo
g
y
is
th
e
class
if
icatio
n
o
f
R
S
im
ag
es.
T
h
is
tech
n
o
lo
g
y
is
wid
ely
u
s
ed
in
b
o
t
h
th
e
m
ilit
ar
y
a
n
d
t
h
e
civ
ilian
wo
r
ld
s
.
Ho
wev
er
,
b
ec
au
s
e
to
th
e
ch
a
r
ac
ter
is
tics
o
f
R
S
d
ata,
s
u
ch
as
h
ig
h
d
im
en
s
io
n
ality
an
d
a
d
ea
r
th
o
f
r
ele
v
an
t
lab
elled
ex
am
p
les
,
R
S
p
ictu
r
e
class
if
icatio
n
f
ac
es
s
u
b
s
tan
tial
p
r
ac
tical
an
d
tech
n
ical
ch
al
len
g
es.
T
h
e
ad
v
e
n
t
o
f
n
o
v
el
d
ee
p
lear
n
in
g
(
DL
)
a
lg
o
r
ith
m
s
in
r
ec
en
t
y
ea
r
s
h
as
r
esu
lted
in
s
ig
n
if
ican
t
p
r
o
g
r
ess
in
m
eth
o
d
s
f
o
r
R
S
im
ag
e
class
if
icatio
n
u
s
in
g
DL
.
T
h
ese
ad
v
an
ce
m
e
n
ts
o
p
en
u
p
n
ew
d
ir
ec
tio
n
s
f
o
r
th
e
s
tu
d
y
a
n
d
ad
v
a
n
ce
m
en
t
o
f
R
S im
ag
e
class
if
icat
io
n
[
4
]
.
T
h
e
d
is
cip
lin
e
o
f
au
to
m
atic
im
ag
e
ca
teg
o
r
izatio
n
an
d
an
al
y
s
is
is
a
th
r
iv
in
g
an
d
ac
tiv
e
ar
ea
o
f
s
tu
d
y
with
g
r
ea
t
s
ig
n
if
ican
ce
th
r
o
u
g
h
o
u
t
co
m
p
u
ter
v
is
io
n
d
o
m
ai
n
s
.
Ap
p
licatio
n
s
f
o
r
it
ca
n
b
e
f
o
u
n
d
in
a
wid
e
r
an
g
e
o
f
f
ield
s
,
in
clu
d
in
g
v
id
eo
p
r
o
ce
s
s
in
g
,
r
em
o
te
s
en
s
in
g
,
o
b
je
ct
r
ec
o
g
n
itio
n
,
f
ac
e
r
ec
o
g
n
itio
n
,
s
p
atial
m
ap
p
in
g
,
au
to
m
atic
d
i
s
ea
s
e
d
etec
tio
n
,
te
x
tile
im
ag
e
a
n
aly
s
is
,
an
d
p
atte
r
n
r
ec
o
g
n
itio
n
.
I
n
o
r
d
e
r
to
f
ac
i
litate
ac
cu
r
ate
an
d
ef
f
ec
tiv
e
p
ictu
r
e
ca
teg
o
r
izatio
n
in
v
ar
i
o
u
s
d
o
m
ain
s
,
r
esear
c
h
er
s
an
d
p
r
ac
titi
o
n
er
s
ar
e
co
n
s
tan
tly
in
v
esti
g
atin
g
an
d
d
e
v
elo
p
in
g
n
ew
m
eth
o
d
s
.
T
h
e
s
ig
n
if
ica
n
ce
o
f
im
a
g
e
cla
s
s
if
icatio
n
ad
v
an
ce
s
is
b
r
o
ad
,
allo
win
g
task
s
lik
e
o
b
ject
r
ec
o
g
n
itio
n
in
im
ag
e
s
,
f
ea
tu
r
e
ex
tr
ac
tio
n
,
r
etr
iev
in
g
r
elev
an
t
im
ag
es
f
r
o
m
lar
g
e
d
atasets
,
d
etec
tin
g
d
is
ea
s
es
o
r
an
o
m
alies,
m
ap
p
in
g
g
eo
g
r
a
p
h
ical
r
eg
io
n
s
,
an
d
en
h
an
cin
g
v
id
e
o
an
aly
s
is
.
T
h
e
co
n
tin
u
o
u
s
p
r
o
g
r
ess
in
th
is
f
ield
p
a
v
es
th
e
way
f
o
r
n
u
m
er
o
u
s
in
n
o
v
ativ
e
ap
p
licatio
n
s
an
d
c
o
n
tr
ib
u
tes
to
th
e
d
ev
elo
p
m
en
t
o
f
in
tellig
en
t sy
s
tem
s
th
at
ca
n
u
n
d
er
s
tan
d
an
d
in
ter
p
r
et
v
is
u
a
l in
f
o
r
m
atio
n
with
g
r
ea
ter
p
r
ec
is
io
n
[
5
]
−
[
7
].
Ass
ig
n
in
g
clas
s
lab
els
to
im
ag
es
is
th
e
p
r
im
ar
y
o
b
jectiv
e
o
f
m
o
d
els
b
ased
o
n
im
a
g
e
class
if
icatio
n
.
I
n
o
r
d
er
to
d
o
th
is
,
a
tr
ain
in
g
d
ataset
a
co
llectio
n
o
f
im
a
g
es
u
s
ed
as
tr
ai
n
in
g
s
am
p
les
is
u
s
ed
to
t
r
ain
a
class
if
icatio
n
m
o
d
el.
T
h
e
tr
ai
n
in
g
p
r
o
ce
s
s
in
v
o
lv
es
lear
n
in
g
th
e
d
is
tin
g
u
is
h
in
g
f
ea
tu
r
es
an
d
p
atter
n
s
in
th
e
im
ag
es
th
at
co
r
r
esp
o
n
d
to
d
i
f
f
er
en
t
class
es.
On
ce
t
h
e
m
o
d
el
is
tr
ain
ed
,
a
s
ep
a
r
ate
test
d
ataset
is
u
s
ed
t
o
ev
alu
ate
its
p
er
f
o
r
m
a
n
ce
an
d
p
r
ed
ict
th
e
class
lab
els
o
f
th
e
test
im
ag
es.
T
h
r
o
u
g
h
th
e
ex
am
in
ati
o
n
o
f
th
e
f
o
r
ec
asts
o
n
th
e
test
d
ataset,
th
e
im
ag
e
s
ca
n
b
e
ar
r
an
g
ed
i
n
a
co
h
er
en
t
an
d
s
ig
n
if
ica
n
t
s
eq
u
en
ce
,
en
ab
lin
g
ef
f
icien
t
ca
teg
o
r
izatio
n
an
d
r
et
r
iev
al
o
f
s
im
ilar
im
a
g
es.
I
n
o
r
d
er
to
im
p
r
o
v
e
th
e
p
er
f
o
r
m
a
n
ce
of
class
if
icatio
n
-
b
ased
s
y
s
tem
s
,
it
is
cr
u
cial
to
s
elec
t
d
is
c
r
i
m
i
n
a
t
i
v
e
an
d
u
n
iq
u
e
f
ea
tu
r
es.
T
h
ese
f
ea
tu
r
es
s
h
o
u
ld
ca
p
tu
r
e
th
e
d
is
tin
ctiv
e
ch
ar
ac
ter
is
tics
o
f
ea
ch
class
,
en
ab
lin
g
ac
cu
r
ate
class
if
icatio
n
.
T
h
e
ch
o
ice
o
f
ap
p
r
o
p
r
iat
e
f
ea
tu
r
es
[
8
]
c
o
n
tr
ib
u
tes
s
ig
n
i
f
ican
tly
to
im
p
r
o
v
in
g
th
e
s
y
s
tem
'
s
ca
p
ac
ity
t
o
d
if
f
er
en
tiate
b
etwe
en
class
es
an
d
im
p
r
o
v
e
o
v
er
all
class
if
icatio
n
ac
cu
r
ac
y
.
R
esear
ch
er
s
co
n
tin
u
o
u
s
ly
ex
p
l
o
r
e
an
d
d
e
v
elo
p
n
o
v
el
f
ea
tu
r
e
ex
tr
ac
tio
n
tech
n
iq
u
es
to
ex
tr
a
ct
th
e
m
o
s
t
r
elev
an
t
an
d
in
f
o
r
m
ativ
e
f
ea
tu
r
es
f
r
o
m
im
a
g
es,
th
er
eb
y
en
h
an
cin
g
th
e
p
er
f
o
r
m
an
ce
o
f
im
a
g
e
class
if
icatio
n
m
o
d
els.
As
a
s
u
b
s
et
o
f
m
ac
h
i
n
e
lear
n
in
g
,
d
ee
p
lear
n
i
n
g
[
9
]
r
ef
er
s
t
o
a
class
o
f
m
o
d
els
th
at
ca
n
ef
f
icien
tly
r
ep
r
esen
t
in
p
u
t
at
d
i
f
f
er
en
t
lev
els
o
f
ab
s
tr
ac
tio
n
b
y
u
s
in
g
n
u
m
er
o
u
s
p
r
o
ce
s
s
in
g
lay
e
r
s
.
On
e
o
f
t
h
e
k
e
y
ac
h
iev
em
en
ts
o
f
d
ee
p
lea
r
n
in
g
is
its
r
em
ar
k
ab
le
s
u
cc
ess
in
o
b
ject
d
etec
tio
n
an
d
class
if
icatio
n
task
s
,
p
r
im
a
r
ily
ac
h
iev
ed
th
r
o
u
g
h
th
e
u
s
e
o
f
C
NNs
an
d
th
e
p
ar
allel
co
m
p
u
tin
g
p
o
wer
o
f
g
r
ap
h
ical
p
r
o
ce
s
s
in
g
u
n
its
(
GPUs
)
.
W
e
s
p
ec
if
ically
in
v
esti
g
ate
h
o
w
well
-
s
u
ited
s
ev
en
p
o
p
u
lar
d
ee
p
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
s
—
Den
s
eNe
t1
2
1
,
I
n
ce
p
tio
n
V3
,
VGG1
6
,
VGG1
9
,
Xce
p
tio
n
,
R
esNet5
0
,
an
d
I
n
ce
p
tio
n
R
esNet
V2
ar
e
f
o
r
m
ap
p
i
n
g
wetlan
d
s
in
C
an
ad
a.
T
h
e
to
p
th
r
ee
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
s
ar
e
I
n
ce
p
tio
n
R
esNet
V2
,
R
esNet5
0
,
an
d
Xce
p
tio
n
,
wh
ich
o
f
f
er
s
tate
-
of
-
th
e
-
ar
t
class
if
icatio
n
ac
cu
r
ac
i
es
o
f
9
6
.
1
7
%,
9
4
.
8
1
%,
an
d
9
3
.
5
7
%,
r
esp
ec
tiv
ely
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
52
In
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
3
7
,
No
.
3
,
Ma
r
ch
20
2
5
:
1
712
-
1
7
2
5
1714
Su
p
p
o
r
t
v
ec
to
r
m
ac
h
in
e
(
SVM)
an
d
r
an
d
o
m
f
o
r
est
(
R
F)
y
ield
ed
m
u
ch
lo
wer
class
if
icatio
n
ac
cu
r
ac
ies
th
a
n
C
NNs,
at
7
4
.
8
9
%
an
d
7
6
.
0
8
%
,
r
esp
ec
tiv
ely
.
C
r
u
cially
,
I
n
ce
p
tio
n
R
es
NetV2
co
n
s
is
ten
tly
o
u
tp
er
f
o
r
m
s
all
o
th
er
co
n
v
n
ets,
s
u
g
g
esti
n
g
th
at
th
e
I
n
ce
p
tio
n
an
d
R
esNet
m
o
d
u
les
to
g
eth
er
p
r
o
v
id
e
a
u
s
e
f
u
l
ar
ch
itectu
r
e
f
o
r
class
if
y
in
g
co
m
p
lex
r
em
o
te
s
e
n
s
in
g
s
ce
n
ar
io
s
s
u
ch
as we
tlan
d
s
[
1
0
]
.
On
e
o
f
th
e
p
r
im
ar
y
d
if
f
icu
lti
es
ar
is
es
f
r
o
m
t
h
e
f
ac
t
th
at
o
b
jects
in
r
em
o
te
s
en
s
in
g
im
a
g
es
ca
n
b
e
r
o
tated
with
in
th
e
v
iew,
m
a
k
in
g
th
eir
id
e
n
tific
atio
n
an
d
class
if
icatio
n
m
o
r
e
co
m
p
le
x
.
Ad
d
itio
n
ally
,
t
h
e
b
ac
k
g
r
o
u
n
d
in
r
em
o
te
s
en
s
in
g
im
ag
es
is
o
f
ten
ch
ar
ac
ter
ized
b
y
a
h
ig
h
le
v
el
o
f
co
m
p
lex
ity
,
f
u
r
th
er
a
d
d
in
g
to
th
e
ch
allen
g
e
o
f
ac
cu
r
ate
cl
ass
if
icatio
n
.
T
o
g
ath
er
im
a
g
e
d
atasets
f
o
r
r
em
o
te
s
en
s
in
g
r
esear
ch
,
v
ar
io
u
s
p
latf
o
r
m
s
ar
e
u
tili
ze
d
,
in
clu
d
in
g
ae
r
ial
s
y
s
tem
s
,
u
n
m
an
n
e
d
ae
r
ial
v
eh
icles
(
UAVs)
,
an
d
s
atellite
s
.
T
h
ese
p
latf
o
r
m
s
ca
p
t
u
r
e
im
a
g
es
f
r
o
m
d
if
f
er
e
n
t
p
e
r
s
p
ec
tiv
es
an
d
at
v
ar
y
in
g
s
p
atial
r
eso
lu
tio
n
s
,
p
r
o
v
id
in
g
v
alu
a
b
le
d
ata
f
o
r
ev
al
u
atin
g
an
d
ad
v
a
n
cin
g
th
e
f
ield
o
f
r
em
o
te
s
en
s
in
g
[
11
].
T
h
e
s
tu
d
y
'
s
f
o
llo
win
g
s
ec
tio
n
s
a
r
e
ar
r
an
g
e
d
as
f
o
llo
ws:
A
th
o
r
o
u
g
h
r
e
v
iew
o
f
th
e
lite
r
atu
r
e
is
p
r
o
v
id
ed
in
Sectio
n
2
a
n
d
a
d
is
cu
s
s
io
n
o
f
p
er
tin
en
t
s
tu
d
ies
co
n
ce
r
n
in
g
th
e
class
if
icatio
n
o
f
r
em
o
te
s
en
s
in
g
im
a
g
es.
I
n
Sectio
n
3
,
th
e
au
th
o
r
s
d
is
cu
s
s
th
e
d
ata
s
et
u
s
ed
in
t
h
is
r
esear
ch
,
h
ig
h
lig
h
tin
g
its
s
ig
n
if
ican
ce
in
e
v
alu
atin
g
th
e
p
r
o
p
o
s
ed
m
eth
o
d
s
.
Sectio
n
4
p
r
esen
ts
a
d
etailed
d
escr
ip
tio
n
o
f
th
e
m
o
d
els
em
p
lo
y
ed
in
th
e
s
tu
d
y
f
o
r
r
em
o
te
s
en
s
in
g
im
ag
e
class
if
icatio
n
.
Sectio
n
5
p
r
esen
ts
th
e
r
esu
lts
o
f
th
e
ex
p
er
im
en
ts
co
n
d
u
cted
,
al
o
n
g
with
a
th
o
r
o
u
g
h
an
al
y
s
is
,
d
is
cu
s
s
io
n
,
an
d
co
m
p
ar
is
o
n
s
o
f
th
e
o
b
tain
ed
ex
p
er
im
en
tal
v
alu
es.
T
h
e
p
ap
er
is
f
in
ally
co
n
clu
d
e
d
in
Sectio
n
6
,
wh
ich
s
u
m
m
ar
izes
th
e
m
ain
c
o
n
clu
s
io
n
s
an
d
c
o
n
tr
ib
u
tio
n
s
o
f
t
h
e
s
u
g
g
ested
s
tu
d
y
i
n
th
e
f
ield
o
f
r
em
o
te
s
en
s
in
g
im
ag
e
class
if
icatio
n
.
C
h
en
g
et
a
l.
[
1
2
]
d
is
cu
s
s
th
e
in
ter
s
ec
tio
n
o
f
r
em
o
te
s
en
s
in
g
im
ag
e
s
ce
n
e
class
if
icatio
n
an
d
d
ee
p
lear
n
in
g
,
h
i
g
h
lig
h
tin
g
th
e
ch
a
llen
g
es,
m
eth
o
d
s
,
b
en
c
h
m
ar
k
s
,
an
d
o
p
p
o
r
tu
n
ities
in
th
is
f
iel
d
.
T
h
ey
p
r
o
v
id
e
a
co
m
p
r
eh
e
n
s
iv
e
o
v
e
r
v
iew
o
f
th
e
s
tate
-
of
-
th
e
-
ar
t
tec
h
n
iq
u
e
s
an
d
d
atasets
u
s
ed
in
r
em
o
te
s
en
s
in
g
im
ag
e
class
if
icatio
n
.
T
h
e
p
ap
er
co
n
c
lu
d
es
b
y
e
m
p
h
asizin
g
t
h
e
p
o
ten
tial
o
f
d
ee
p
lear
n
i
n
g
a
p
p
r
o
ac
h
es
an
d
th
e
n
ee
d
f
o
r
f
u
r
th
er
r
esear
ch
an
d
co
lla
b
o
r
atio
n
t
o
ad
d
r
ess
th
e
r
em
ai
n
in
g
ch
allen
g
es
an
d
u
n
lo
ck
n
ew
o
p
p
o
r
tu
n
iti
es
in
r
em
o
te
s
en
s
in
g
im
ag
e
an
aly
s
is
.
By
em
p
lo
y
in
g
an
u
n
s
u
p
e
r
v
is
ed
lear
n
in
g
alg
o
r
ith
m
ca
lled
"d
ee
p
b
elief
n
ets,"
t
h
e
au
t
h
o
r
s
o
f
[
1
3
]
d
e
m
o
n
s
tr
ate
th
at
n
e
u
r
al
n
etwo
r
k
s
with
m
an
y
lay
er
s
o
f
h
id
d
e
n
u
n
its
ca
n
lear
n
t
o
ef
f
icien
tly
r
ep
r
esen
t
th
e
s
tatis
tical
s
tr
u
ctu
r
e
of
h
ig
h
-
d
im
en
s
io
n
al
d
ata.
A
g
r
o
u
p
of
r
e
s
e
a
r
c
h
e
r
s
[
1
4
]
p
er
f
o
r
m
e
d
a
s
tate
-
of
-
th
e
-
a
r
t
r
e
v
iew
o
n
D
ee
p
lear
n
i
n
g
a
p
p
r
o
ac
h
es
a
p
p
li
ed
to
r
em
o
te
s
en
s
in
g
d
atasets
.
Data
s
ets
lik
e
th
e
UC
Me
r
ce
d
L
an
d
Use
f
r
o
m
th
e
U.
S.
Geo
lo
g
ical
Su
r
v
e
y
wer
e
u
s
ed
b
y
r
esear
ch
er
s
in
th
e
p
a
s
t.
Dee
p
lear
n
in
g
is
u
s
ed
in
r
em
o
te
s
en
s
in
g
en
v
ir
o
n
m
en
ts
,
an
d
th
e
m
eth
o
d
o
lo
g
ical
ap
p
r
o
ac
h
es
o
f
d
ee
p
lear
n
in
g
h
av
e
g
e
n
er
ated
a
lo
t
o
f
in
ter
est
in
th
e
d
ev
elo
p
m
en
t
o
f
r
em
o
te
s
en
s
in
g
d
o
m
ain
s
.
D
L
p
o
s
s
es
s
es
th
e
ca
p
ac
ity
to
ef
f
ec
ti
v
ely
co
m
p
ete
an
d
o
v
er
co
m
e
th
e
d
iv
e
r
s
e
o
b
s
tacle
s
en
co
u
n
ter
e
d
in
t
h
e
f
ield
o
f
r
em
o
te
s
en
s
in
g
.
T
h
e
p
o
ten
tial
o
f
d
ee
p
lear
n
in
g
in
a
r
e
m
o
te
s
en
s
in
g
co
n
tex
t
is
ex
am
i
n
ed
in
t
h
is
s
p
ec
ial
is
s
u
e,
alo
n
g
with
th
e
m
o
s
t r
ec
en
t a
d
v
an
ce
m
en
ts
an
d
im
p
r
o
v
em
e
n
ts
[
1
5
]
.
T
h
e
au
th
o
r
s
in
[
1
6
]
co
n
d
u
cte
d
a
m
eta
-
a
n
aly
s
is
an
d
r
ev
iew
ed
o
v
er
2
0
0
p
u
b
licatio
n
s
in
t
h
e
f
ield
of
DL
f
o
r
r
em
o
te
s
en
s
in
g
.
Usi
n
g
a
lar
g
e
d
ataset
o
f
1
.
2
m
illi
o
n
h
ig
h
-
r
eso
lu
tio
n
im
a
g
es
f
r
o
m
th
e
I
m
ag
eNe
t
L
SVR
C
-
2
0
1
0
c
o
n
test
,
th
e
r
esear
ch
er
s
o
f
[
1
7
]
tr
ai
n
ed
a
d
ee
p
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
with
th
e
g
o
al
o
f
class
if
y
in
g
th
ese
im
ag
es
in
to
1
0
0
0
d
is
tin
ct
class
es
with
t
o
p
-
1
a
n
d
t
o
p
-
5
er
r
o
r
r
ates
o
f
3
7
.
5
%
a
n
d
1
7
.
0
%,
r
esp
ec
tiv
ely
.
A
g
r
o
u
p
o
f
r
esea
r
ch
er
s
u
s
ed
th
e
I
n
ce
p
tio
n
a
r
ch
itectu
r
e
an
d
ac
h
iev
e
d
e
x
ce
llen
t
p
er
f
o
r
m
an
ce
at
a
co
m
p
ar
ativ
ely
s
m
all
co
m
p
u
tat
io
n
al
co
s
t
[
1
8
]
.
W
ith
C
N
Ns,
d
at
ab
ases
o
f
R
S
im
ag
e
s
m
ay
b
e
r
ec
o
g
n
ized
a
n
d
ca
teg
o
r
ize
d
.
T
h
e
co
n
v
en
ti
o
n
al
C
NN
m
eth
o
d
g
e
n
er
ates
co
ar
s
e
m
ap
s
f
o
r
class
if
y
in
g
im
ag
e
s
o
f
b
r
o
ad
ar
ea
s
.
T
h
e
C
NN
m
eth
o
d
b
ased
o
n
o
b
jects
ca
n
b
e
em
p
lo
y
ed
t
o
tack
le
th
is
p
r
o
b
lem
[
1
9
]
.
I
n
th
is
s
tu
d
y
,
we
i
n
v
esti
g
ate
th
e
r
elatio
n
s
h
ip
b
etwe
en
a
co
n
v
o
l
u
tio
n
al
n
etwo
r
k
'
s
ac
cu
r
ac
y
an
d
d
e
p
th
wh
en
ap
p
lied
to
lar
g
e
-
s
ca
le
im
ag
e
r
ec
o
g
n
itio
n
.
Ou
r
m
ain
co
n
tr
ib
u
tio
n
is
a
th
o
r
o
u
g
h
in
v
esti
g
atio
n
o
f
n
etwo
r
k
s
with
v
ar
y
in
g
d
e
p
th
u
s
in
g
an
ar
ch
itec
tu
r
e
with
r
elativ
ely
tin
y
(
3
×
3
)
c
o
n
v
o
lu
tio
n
f
i
lter
s
.
W
e
s
h
o
w
th
at
a
s
ig
n
if
ican
t
im
p
r
o
v
em
en
t
o
v
e
r
t
h
e
s
tate
-
of
-
th
e
-
ar
t
co
n
f
ig
u
r
atio
n
s
m
ay
b
e
ac
h
iev
e
d
b
y
in
cr
ea
s
in
g
th
e
d
ep
th
t
o
1
6
–
1
9
wei
g
h
t la
y
er
s
[
2
0
]
.
2.
M
E
T
H
O
D
2
.
1
.
Da
t
a
des
cr
iptio
n
T
h
e
R
S
I
-
C
B
2
5
6
s
a
t
e
l
l
i
t
e
i
m
a
g
e
c
l
a
s
s
i
f
i
c
a
t
i
o
n
d
a
t
a
s
e
t
i
s
t
h
e
d
a
t
a
s
e
t
t
h
a
t
w
a
s
c
h
o
s
e
n
f
o
r
t
h
i
s
r
e
s
e
a
r
c
h
[
2
1
]
.
T
h
e
5
,
3
6
1
im
a
g
e
s
in
th
is
co
lle
ctio
n
ar
e
s
p
lit
u
p
in
to
4
ty
p
es
an
d
ar
e
a
co
m
b
i
n
atio
n
o
f
Go
o
g
le
Ma
p
s
n
ap
s
h
o
ts
an
d
s
en
s
o
r
im
a
g
es.
E
v
er
y
im
a
g
e
h
as
a
s
ize
o
f
2
5
6
×
2
5
6
.
C
lo
u
d
y
(
1
5
0
0
)
,
d
eser
t
(
1
1
3
1
)
,
g
r
e
en
ar
ea
(
1
5
0
0
)
,
a
n
d
wate
r
(
1
5
0
0
)
ar
e
th
e
class
if
icatio
n
s
ass
ig
n
ed
to
th
e
im
ag
e
s
.
Sam
p
le
im
ag
e
s
f
r
o
m
th
e
co
lle
ctio
n
ar
e
s
h
o
wn
in
Fig
u
r
e
1
.
Als
o
s
am
p
le
s
atellite
im
ag
es f
r
o
m
d
atas
et
ar
e
s
h
o
w
n
in
Fig
u
r
e
2.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2
5
0
2
-
4
7
52
Dee
p
lea
r
n
in
g
tech
n
iq
u
es fo
r
s
a
tellite ima
g
e
cla
s
s
ifica
tio
n
(
S
u
r
esh
K
u
ma
r
Mu
s
a
li
)
1715
Fig
u
r
e
1
.
R
SI
-
C
B
2
5
6
d
ata
s
e
t
s
am
p
le
im
ag
es
Fig
u
r
e
2.
R
SI
-
C
B
2
5
6
d
a
t
as
e
t
-
s
am
p
les
of
s
atellite
im
ag
es
[
2
]
2
.
2
.
Da
t
a
a
u
g
me
n
t
a
t
i
o
n
As
th
e
d
ataset
co
n
tain
s
o
n
ly
5
,
3
6
1
im
a
g
es
it
is
n
o
t
s
u
itab
l
e
to
wo
r
k
with
DL
m
o
d
els.
Hen
ce
th
is
wo
r
k
in
c
r
ea
s
ed
th
e
n
u
m
b
er
o
f
im
ag
es
in
th
e
d
ataset
u
s
in
g
a
u
g
m
en
tatio
n
tech
n
iq
u
es.
An
8
0
:2
0
r
atio
was
u
s
ed
to
s
p
lit
th
e
d
ata
s
et
in
to
tr
ain
i
n
g
a
n
d
v
alid
atio
n
s
ets.
Data
a
u
g
m
en
tatio
n
p
r
o
v
id
es
a
wid
e
r
v
ar
iety
o
f
tr
ain
i
n
g
d
ata
to
m
ac
h
in
e
lear
n
i
n
g
m
o
d
els,
wh
ich
en
h
an
ce
s
th
e
m
o
d
e
l's
p
er
f
o
r
m
an
ce
.
T
h
e
in
p
u
t
im
ag
e
I
is
m
ap
p
ed
t
o
d
im
en
s
io
n
s
o
f
Mx
Nx
C
,
s
u
ch
th
at
M
is
n
u
m
b
er
o
f
r
o
ws,
N
is
n
u
m
b
er
o
f
co
lu
m
n
s
an
d
C
is
n
u
m
b
er
o
f
ch
an
n
els.
R
escale
s
an
in
p
u
t in
[
0
,
2
5
5
]
r
a
n
g
e
to
b
e
in
[
0
,
1
]
r
an
g
e
.
R
S(I
(
i,j)
)
=I
(
i,j)
/m
ax
I
(
s
,
t)
wh
e
r
e
i r
an
g
es f
r
o
m
1
to
M,
j
r
an
g
es f
r
o
m
1
t
o
N,
C
b
elo
n
g
s
to
R
,
G,
B
.
R
o
tatio
n
r
an
g
e
R
(
I
(
i,j)
)
=I
(
i*
co
s
–
j*
s
in
,
i*
s
in
+
j*
co
s
)
Sh
ea
r
r
an
g
e
Sh
u
(
I
(
i,j)
)
=
I
(
i+u
*
j,j)
Z
o
o
m
r
a
n
g
e
Z
v
(
I
(
i,j)
)
=I
(
i*
v
,
j*
v
)
Ho
r
izo
n
tal
f
lip
HFw(
I
(
i,j)
)
=I
(
M
-
i
-
1
,
j)
Ver
tical
f
lip
VFx
(
I
(
i,j
)
)
=I
(
i,N
-
j
-
1)
2
.
3
.
P
r
o
po
s
ed
m
et
ho
d
T
h
e
o
b
jectiv
e
o
f
th
is
r
esear
ch
is
to
ca
teg
o
r
ize
s
atellite
im
ag
es.
Usi
n
g
th
e
s
atell
ite
im
ag
e
as
in
p
u
t,
th
e
s
u
g
g
ested
m
eth
o
d
o
lo
g
y
d
iv
i
d
es
th
e
to
p
o
g
r
ap
h
y
in
to
f
o
u
r
class
if
icatio
n
s
.
T
h
e
in
p
u
t
im
ag
e
s
ar
e
f
ir
s
t
s
u
b
jecte
d
to
p
re
-
p
r
o
ce
s
s
in
g
,
an
d
th
en
t
h
ey
ar
e
class
if
ied
.
All
ex
p
er
im
en
ts
wer
e
co
n
d
u
cte
d
o
n
h
o
s
te
d
en
v
ir
o
n
m
en
t
u
s
in
g
g
r
ap
h
ical
p
r
o
ce
s
s
in
g
u
n
its
as
well
as
ce
n
tr
al
p
r
o
ce
s
s
in
g
u
n
its
.
GPU
P1
0
0
was u
s
ed
.
C
UD
A
s
o
f
twar
e
was u
s
ed
to
ac
ce
s
s
th
e
GPU’
s
.
Pack
ag
es
lik
e
Nu
m
Py
[
2
2
]
,
Pan
d
as
[
2
3
]
,
Ma
tp
lo
tlib
[
2
4
]
,
T
en
s
o
r
Fl
o
w
[
2
5
]
,
Ker
as
[2
6
]
an
d
Sk
lear
n
[
2
7
]
wer
e
u
s
ed
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
52
In
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
3
7
,
No
.
3
,
Ma
r
ch
20
2
5
:
1
712
-
1
7
2
5
1716
2
.
3
.
1
.
P
ro
po
s
ed
CNN
ba
s
ed
m
o
d
e
l
T
h
e
p
r
o
p
o
s
ed
s
eq
u
en
tial
C
NN
b
ased
m
o
d
el
h
as
1
4
lay
er
s
th
at
in
clu
d
e
th
r
ee
2
D
co
n
v
o
lu
tio
n
lay
er
s
,
th
r
ee
m
ax
p
o
o
lin
g
2
D
lay
e
r
s
,
f
o
u
r
d
en
s
e
lay
er
s
an
d
th
r
ee
d
r
o
p
o
u
t
lay
er
s
.
Fig
u
r
e
3
s
h
o
ws
C
NN
ar
ch
itectu
r
e.
2
D
co
n
v
o
lu
tio
n
lay
er
is
u
s
ed
t
o
ex
tr
ac
t f
ea
tu
r
es
f
r
o
m
im
a
g
es.
M
ax
p
o
o
lin
g
2
D
la
y
er
is
u
s
e
d
to
r
ed
u
ce
th
e
s
ize
o
f
th
e
f
ea
tu
r
e
m
ap
s
wh
ile
p
r
eser
v
in
g
t
h
e
m
o
s
t
im
p
o
r
tan
t
f
ea
t
u
r
es.
I
t d
o
wn
s
ca
les
in
p
u
t b
y
u
s
in
g
p
o
o
ls
ize
(
2
,
2
)
.
T
h
e
in
p
u
t
d
ata
is
class
if
ied
o
r
p
r
e
d
icted
u
s
in
g
d
en
s
e
l
ay
er
s
.
I
t
em
p
lo
y
s
1
2
8
,
5
1
2
,
2
5
6
,
a
n
d
4
u
n
its
co
n
s
ec
u
tiv
ely
.
I
n
o
r
d
er
to
a
v
o
id
o
v
er
f
itti
n
g
,
th
e
d
r
o
p
o
u
t
lay
er
s
ar
e
em
p
lo
y
e
d
last
.
T
h
e
r
at
e
is
s
et
to
0
.
5
.
T
h
e
m
o
d
el
e
m
p
lo
y
ed
So
f
tMa
x
an
d
r
ec
tifie
d
lin
ea
r
u
n
it
(
R
eL
U)
as
ac
tiv
atio
n
f
u
n
ctio
n
s
to
g
eth
er
with
th
e
Ad
am
o
p
tim
izer
.
T
wen
ty
ep
o
c
h
s
in
a
ll we
r
e
ch
o
s
en
f
o
r
th
e
m
o
d
el'
s
tr
ain
in
g
.
Fig
u
r
e
3
.
Pro
p
o
s
ed
C
NN
ar
ch
i
tectu
r
e
T
ab
le
1
d
escr
ib
es
th
e
c
o
m
p
le
te
s
u
m
m
ar
y
o
f
C
NN
ar
ch
itect
u
r
e.
B
lo
ck
r
ef
er
s
to
a
s
et
o
f
o
p
er
atio
n
s
ap
p
lied
s
eq
u
en
tially
in
a
p
ar
ticu
lar
lay
er
o
r
g
r
o
u
p
o
f
lay
er
s
in
th
e
C
NN.
Op
er
ato
r
s
(
F)
ar
e
th
e
ty
p
es
o
f
o
p
er
atio
n
s
p
er
f
o
r
m
ed
in
th
e
b
lo
ck
.
B
lu
e
co
l
o
u
r
e
d
lay
er
r
ep
r
esen
ts
co
n
v
o
lu
ti
o
n
al
lay
e
r
s
.
R
ed
co
lo
u
r
e
d
lay
er
r
ep
r
esen
ts
m
ax
p
o
o
lin
g
lay
er
s
.
Gr
ee
n
c
o
lo
u
r
e
d
lay
e
r
r
e
p
r
esen
ts
d
en
s
e
lay
er
s
o
r
f
u
ll
y
co
n
n
ec
ted
lay
er
s
.
R
eso
lu
tio
n
s
(
H
×
W)
ar
e
th
e
d
i
m
en
s
io
n
s
o
f
t
h
e
f
ea
tu
r
e
m
ap
s
af
ter
th
e
o
p
er
atio
n
s
in
th
e
b
l
o
ck
.
T
h
ey
d
escr
ib
e
th
e
h
eig
h
t
(
H)
an
d
wid
th
(
W
)
o
f
th
e
f
ea
tu
r
e
m
ap
s
at
th
at
s
tag
e
in
th
e
n
etwo
r
k
.
C
h
an
n
el
s
(
C
)
in
d
icate
s
th
e
n
u
m
b
er
o
f
f
ea
tu
r
e
m
a
p
s
(
o
r
c
h
an
n
els)
p
r
o
d
u
ce
d
b
y
th
e
c
o
n
v
o
lu
tio
n
al
lay
er
.
T
h
e
n
u
m
b
er
o
f
ch
an
n
els
o
f
te
n
in
cr
ea
s
es
as
th
e
n
etwo
r
k
g
o
es
d
ee
p
er
,
allo
w
in
g
th
e
n
etwo
r
k
to
lear
n
m
o
r
e
co
m
p
lex
f
ea
t
u
r
e
s
.
Fo
r
in
s
tan
ce
,
th
e
f
ir
s
t b
lo
ck
o
u
tp
u
ts
3
2
ch
an
n
els,
th
e
s
ec
o
n
d
b
lo
ck
o
u
tp
u
ts
6
4
ch
an
n
els,
etc
.
T
ab
le
1
.
C
NN
ar
ch
itectu
r
e
d
etails
B
l
o
c
k
O
p
e
r
a
t
i
o
n
s
(
F
)
I
mag
e
r
e
s
o
l
u
t
i
o
n
(H
×
W)
F
e
a
t
u
r
e
s (
C
)
1
C
o
n
v
(
1
l
a
y
e
r
)
,
R
e
LU
,
M
a
x
p
o
o
l
(
1
l
a
y
e
r
)
2
5
4
×
2
5
4
32
2
C
o
n
v
(
1
)
,
R
e
LU
,
M
a
x
p
o
o
l
(
1
)
1
2
5
×
1
2
5
64
3
C
o
n
v
(
1
)
,
R
e
LU
,
M
a
x
p
o
o
l
(
1
)
60
×
60
1
2
8
Pro
p
o
s
ed
C
NN
a
lg
o
r
ith
m
:
−
Step
1
:
p
r
o
p
o
s
ed
s
eq
u
e
n
tial
C
NN
b
ased
m
o
d
el
h
as 1
4
lay
er
s
.
−
Step
2
:
C
NN
u
s
es 3
co
n
v
o
lu
ti
o
n
al
lay
er
s
to
e
x
tr
ac
t f
ea
tu
r
es
f
r
o
m
im
a
g
es.
−
Step
3
:
u
s
es
3
m
ax
p
o
o
lin
g
la
y
er
s
,
wh
ich
r
ed
u
ce
th
eir
s
ize
an
d
d
im
en
s
io
n
ality
.
−
Step
4
:
f
o
u
r
d
en
s
e
lay
e
r
s
u
s
ed
to
class
if
y
o
r
p
r
e
d
ict
in
p
u
t
d
ata.
−
Step
5
:
th
r
ee
d
r
o
p
o
u
t
lay
er
s
ar
e
u
s
ed
to
p
r
ev
e
n
t
o
v
er
f
itti
n
g
.
−
Step
6
:
f
in
ally
,
th
e
o
u
tp
u
t
o
f
t
h
e
p
o
o
lin
g
lay
er
s
is
f
ed
in
to
a
f
u
lly
c
o
n
n
e
cted
lay
e
r
,
wh
ic
h
p
r
o
d
u
ce
s
th
e
f
in
al
o
u
tp
u
t o
f
t
h
e
n
etwo
r
k
.
−
Step
7
:
R
eL
U
an
d
So
f
tMa
x
ar
e
u
s
ed
as
ac
ti
v
atio
n
f
u
n
cti
o
n
s
an
d
Ad
am
o
p
tim
izer
wer
e
u
s
ed
in
th
e
m
o
d
el.
2
.
3
.
2
.
P
ro
po
s
ed
ResNet
9
ba
s
ed
m
o
d
e
l
T
h
e
p
r
o
p
o
s
ed
R
esNet9
b
ased
m
o
d
el
h
as
9
lay
er
s
th
at
in
clu
d
e
co
n
v
o
lu
tio
n
,
r
esid
u
al
b
l
o
ck
,
m
ax
p
o
o
lin
g
la
y
er
s
an
d
class
if
ier
.
Fig
u
r
e
4
s
h
o
ws
R
esNet9
ar
ch
itectu
r
e.
T
h
e
class
if
ier
co
n
s
is
ts
o
f
an
ad
a
p
tiv
e
m
a
x
p
o
o
lin
g
lay
er
,
a
f
latten
lay
er
,
a
d
r
o
p
o
u
t
la
y
er
,
an
d
a
lin
ea
r
lay
er
.
T
h
e
m
o
d
el
u
s
es
R
eL
U
ac
tiv
atio
n
f
u
n
ctio
n
an
d
Ad
am
o
p
tim
izer
.
T
a
b
le
2
g
iv
es th
e
co
m
p
lete
d
etails o
f
R
esn
et9
ar
ch
itectu
r
e.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2
5
0
2
-
4
7
52
Dee
p
lea
r
n
in
g
tech
n
iq
u
es fo
r
s
a
tellite ima
g
e
cla
s
s
ifica
tio
n
(
S
u
r
esh
K
u
ma
r
Mu
s
a
li
)
1717
Fig
u
r
e
4
.
Pro
p
o
s
ed
R
esNet9
ar
ch
itectu
r
e
T
ab
le
2
.
R
es
N
et9
ar
ch
itectu
r
e
d
etails
B
l
o
c
k
O
p
e
r
a
t
i
o
n
s
(
F
)
I
mag
e
r
e
s
o
l
u
t
i
o
n
(H
×
W)
F
e
a
t
u
r
e
s (
C
)
1
C
o
n
v
(
2
)
,
R
e
LU
,
M
a
x
p
o
o
l
(
1
)
1
1
2
×
1
1
2
64
2
C
o
n
v
(
3
)
,
R
e
LU
,
M
a
x
p
o
o
l
(
1
)
56
×
56
1
2
8
3
C
o
n
v
(
1
)
,
R
e
LU
,
M
a
x
p
o
o
l
(
1
)
28
×
28
2
5
6
4
C
o
n
v
(
2
)
,
R
e
LU
,
M
a
x
p
o
o
l
(
1
)
14
×
14
5
1
2
Pro
p
o
s
ed
R
esNet9
alg
o
r
ith
m
:
−
Step
1
:
p
r
o
p
o
s
ed
R
esNet9
b
ased
m
o
d
el
h
as 9
lay
er
s
.
−
Step
2
: Res
Net9
u
s
es 8
co
n
v
o
l
u
tio
n
al
lay
er
s
to
e
x
tr
ac
t f
ea
tu
r
es f
r
o
m
im
ag
es.
−
Step
3
:
u
s
es
4
m
ax
p
o
o
lin
g
la
y
er
s
,
wh
ich
r
ed
u
ce
th
eir
s
ize
an
d
d
im
en
s
io
n
ality
−
Step
4
:
o
n
e
class
if
ier
lay
er
u
s
e
d
to
class
if
y
o
r
p
r
ed
ict
in
p
u
t d
ata.
−
Step
6
:
f
in
ally
,
th
e
o
u
tp
u
t
o
f
t
h
e
p
o
o
lin
g
lay
er
s
is
f
ed
in
to
a
f
u
lly
c
o
n
n
e
cted
lay
e
r
,
wh
ic
h
p
r
o
d
u
ce
s
th
e
f
in
al
o
u
tp
u
t o
f
t
h
e
n
etwo
r
k
.
−
Step
7
:
R
eL
U
is
u
s
ed
as a
ctiv
atio
n
f
u
n
ctio
n
s
an
d
Ad
am
o
p
tim
izer
wer
e
u
s
ed
in
t
h
e
m
o
d
el.
2
.
3
.
3
.
P
ro
po
s
ed
VG
G
_
1
9
ba
s
ed
mo
d
e
l
T
h
e
p
r
o
p
o
s
ed
s
eq
u
en
tial
VG
G_
1
9
b
ased
m
o
d
el
h
as
a
to
tal
o
f
2
6
lay
er
s
s
u
ch
th
at
1
6
ar
e
2D
co
n
v
o
l
u
tio
n
lay
er
s
,
f
iv
e
m
ax
pool
lay
er
s
,
o
n
e
f
latten
lay
er
,
one
in
p
u
t
lay
er
an
d
th
r
ee
d
e
n
s
e
lay
er
s
.
Fig
u
r
e
5
s
h
o
ws
VGG_
1
9
a
r
ch
itectu
r
e.
So
f
tMa
x
was
u
s
ed
as
th
e
ac
ti
v
atio
n
f
u
n
ctio
n
,
A
d
am
o
p
tim
i
ze
r
an
d
2
0
ep
o
c
h
s
.
T
h
e
m
o
d
el
u
s
es
th
e
p
r
e
-
tr
ain
e
d
weig
h
ts
f
r
o
m
th
e
I
m
ag
eNe
t
d
ataset.
T
ab
le
3
d
esc
r
ib
es
th
e
co
m
p
lete
s
u
m
m
a
r
y
o
f
VGG_
1
9
a
r
ch
itectu
r
e.
Fig
u
r
e
5
.
Pro
p
o
s
ed
VGG_
1
9
a
r
ch
itectu
r
e
T
ab
le
3
.
VGG
_
1
9
ar
ch
itectu
r
e
d
etails
B
l
o
c
k
O
p
e
r
a
t
i
o
n
s
(
F
)
I
mag
e
r
e
s
o
l
u
t
i
o
n
(H
×
W)
F
e
a
t
u
r
e
s (
C
)
1
C
o
n
v
(
2
)
,
R
e
LU
,
M
a
x
p
o
o
l
(
1
)
2
2
4
×
2
2
4
64
2
C
o
n
v
(
2
)
,
R
e
LU
,
M
a
x
p
o
o
l
(
1
)
1
1
2
×
1
1
2
1
2
8
3
C
o
n
v
(
4
)
,
R
e
LU
,
M
a
x
p
o
o
l
(
1
)
56
×
56
2
5
6
4
C
o
n
v
(
4
)
,
R
e
LU
,
M
a
x
p
o
o
l
(
1
)
28
×
28
5
1
2
5
C
o
n
v
(
4
)
,
R
e
LU
,
M
a
x
p
o
o
l
(
1
)
14
×
14
5
1
2
Pro
p
o
s
ed
VGG_
1
9
a
lg
o
r
ith
m
:
−
Step
1
:
p
r
o
p
o
s
ed
VGG
-
1
9
b
ased
m
o
d
el
h
as 2
6
lay
e
r
s
.
−
Step
2
: V
GG
-
19
u
s
es 1
6
c
o
n
v
o
lu
tio
n
al
lay
er
s
to
e
x
tr
ac
t f
ea
t
u
r
es f
r
o
m
i
m
ag
es.
−
Step
3
:
u
s
es
5
m
ax
p
o
o
lin
g
la
y
er
s
,
wh
ich
r
ed
u
ce
th
eir
s
ize
an
d
d
im
en
s
io
n
ality
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
52
In
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
3
7
,
No
.
3
,
Ma
r
ch
20
2
5
:
1
712
-
1
7
2
5
1718
−
Step
4
:
th
r
ee
d
e
n
s
e
lay
er
s
u
s
ed
to
class
if
y
o
r
p
r
ed
ict
in
p
u
t d
ata.
−
Step
5
:
f
in
ally
,
t
h
e
o
u
t
p
u
t
o
f
th
e
p
o
o
lin
g
lay
er
s
is
f
e
d
in
to
a
f
u
lly
c
o
n
n
ec
te
d
lay
er
,
wh
i
ch
p
r
o
d
u
ce
s
th
e
f
in
al
o
u
tp
u
t o
f
t
h
e
n
etwo
r
k
.
−
Step
6
:
So
f
tMa
x
was u
s
ed
as a
ctiv
atio
n
f
u
n
ctio
n
s
an
d
A
d
am
o
p
tim
izer
wer
e
u
s
ed
i
n
th
e
m
o
d
el.
2
.
3
.
4
.
P
ro
po
s
ed
M
o
bil
eNe
t
V
2
ba
s
ed
m
o
d
e
l
T
h
e
p
r
o
p
o
s
ed
Mo
b
ileNetV
2
b
ased
m
o
d
el
is
a
s
eq
u
en
tial
m
o
d
el
th
at
co
n
tain
s
a
to
tal
o
f
f
iv
e
lay
er
s
.
I
t
co
n
tain
s
two
d
en
s
e
lay
e
r
s
,
o
n
e
g
lo
b
al
av
er
a
g
e
p
o
o
lin
g
lay
er
,
o
n
e
in
p
u
t
a
n
d
o
n
e
f
u
n
ctio
n
al
lay
er
.
Fig
u
r
e
6
s
h
o
ws
Mo
b
ileNetV2
ar
ch
itectu
r
e.
W
e
ad
d
th
e
Mo
b
ileNetV2
b
ase
m
o
d
el
with
o
u
t
th
e
to
p
lay
er
.
R
e
L
u
an
d
So
f
tMa
x
ac
tiv
atio
n
f
u
n
ctio
n
s
ar
e
u
s
ed
.
Ad
a
m
o
p
tim
izer
is
u
s
ed
f
o
r
t
r
ain
in
g
th
e
m
o
d
el.
T
a
b
le
4
d
escr
ib
es
th
e
co
m
p
lete
s
u
m
m
ar
y
o
f
M
o
b
ile
NetV2
ar
ch
itectu
r
e.
Fi
g
u
r
e
6
.
Pro
p
o
s
ed
Mo
b
ileNetV2
ar
ch
itectu
r
e
T
ab
le
4
.
Mo
b
ilen
etV2
ar
ch
itec
tu
r
e
d
etails
B
l
o
c
k
O
p
e
r
a
t
i
o
n
s
(
F
)
I
mag
e
r
e
s
o
l
u
t
i
o
n
(H
×
W)
F
e
a
t
u
r
e
s (
C
)
1
R
e
LU
,
g
l
o
b
a
l
A
v
g
p
o
o
l
(
1
)
N
o
n
e
,
n
o
n
e
1
2
8
0
Pro
p
o
s
ed
Mo
b
ileNetV2
a
lg
o
r
i
th
m
:
−
Step
1
:
p
r
o
p
o
s
ed
Mo
b
ileNet
V2
b
ased
m
o
d
el
h
as 5
la
y
er
s
.
−
Step
2
:
Mo
b
ileNet
V2
u
s
es 1
g
lo
b
al
av
er
ag
e
p
o
o
lin
g
lay
e
r
.
−
Step
3
:
two
d
en
s
e
lay
er
s
u
s
ed
to
class
if
y
o
r
p
r
e
d
ict
in
p
u
t
d
ata.
−
Step
4
:
R
eL
U
an
d
So
f
tMa
x
ar
e
u
s
ed
as
ac
ti
v
atio
n
f
u
n
cti
o
n
s
an
d
Ad
am
o
p
tim
izer
wer
e
u
s
ed
in
th
e
m
o
d
el.
2
.
3
.
5
.
Acc
ura
cy
T
h
e
f
o
llo
win
g
f
o
u
r
ev
al
u
atio
n
s
ar
e
u
s
ed
f
o
r
class
if
icatio
n
est
im
ates [
1
7
]
:
−
I
.
T
r
u
e
p
o
s
itiv
e
(
T
P):
th
ey
ar
e
p
o
s
itiv
e
class
lab
els th
at
h
av
e
b
ee
n
co
r
r
ec
tly
p
r
e
d
icted
b
y
th
e
m
o
d
el.
−
II.
T
r
u
e
n
eg
ativ
e
(
T
N)
:
th
ey
ar
e
n
eg
ativ
e
class
lab
els th
at
h
a
v
e
b
ee
n
c
o
r
r
ec
tly
p
r
ed
icted
b
y
th
e
m
o
d
el.
−
III.
Fals
e
p
o
s
itiv
e
(
FP
)
:
th
ey
a
r
e
p
o
s
itiv
e
class
lab
els th
at
h
av
e
b
ee
n
i
n
co
r
r
ec
tly
p
r
ed
icte
d
b
y
th
e
m
o
d
el.
−
I
V.
Fals
e
n
eg
ativ
e
(
FN)
:
th
ey
ar
e
n
eg
ativ
e
class
lab
els th
at
h
av
e
b
ee
n
i
n
co
r
r
ec
tly
p
r
ed
icte
d
b
y
th
e
m
o
d
el.
A
cc
u
r
ac
y
is
ex
p
r
ess
ed
as th
e
r
atio
o
f
co
r
r
ec
t p
r
e
d
ictio
n
s
in
t
h
e
m
o
d
el
t
o
all
p
r
ed
ictio
n
s
:
A
c
c
ura
c
y
=
+
+
+
+
3.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
3
.
1
.
CNN
ba
s
ed
m
o
del
T
ab
le
5
s
h
o
ws
th
e
tr
ain
in
g
lo
s
s
,
v
alid
atio
n
lo
s
s
an
d
v
alid
atio
n
ac
cu
r
ac
y
o
f
th
e
m
o
d
el
d
esig
n
ed
u
s
in
g
C
NN.
T
h
er
e
wer
e
2
0
ep
o
ch
s
in
t
o
tal.
T
h
e
m
o
d
el'
s
h
ig
h
est
ac
cu
r
ac
y
was
9
0
.
4
8
%.
As
th
e
n
u
m
b
er
o
f
ep
o
ch
s
in
cr
ea
s
ed
f
r
o
m
1
to
2
0
,
th
e
m
o
d
el'
s
lo
s
s
d
ec
r
ea
s
ed
.
Fig
u
r
e
7
s
h
o
ws
th
e
tr
ain
in
g
lo
s
s
an
d
v
alid
atio
n
lo
s
s
as
a
f
u
n
ctio
n
o
f
th
e
n
u
m
b
er
o
f
ep
o
ch
s
.
I
t
s
h
o
wed
co
n
s
is
ten
t
lo
s
s
o
v
e
r
th
e
c
o
u
r
s
e
o
f
th
e
tr
ain
i
n
g
.
Fig
u
r
e
8
s
h
o
w
s
th
e
ac
cu
r
ac
y
a
n
d
e
p
o
ch
v
is
u
aliza
tio
n
o
f
th
e
m
o
d
el.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2
5
0
2
-
4
7
52
Dee
p
lea
r
n
in
g
tech
n
iq
u
es fo
r
s
a
tellite ima
g
e
cla
s
s
ifica
tio
n
(
S
u
r
esh
K
u
ma
r
Mu
s
a
li
)
1719
T
ab
le
5
.
R
esu
lts
of
th
e
C
NN
b
ased
m
o
d
e
l
Ep
o
c
h
.
no
.
Tr
a
i
n
l
o
s
s
V
a
l
i
d
a
t
i
o
n
l
o
s
s
A
c
c
u
r
a
c
y
1
8
1
.
5
2
5
4
.
1
2
7
5
.
8
4
2
5
5
.
0
3
3
6
.
0
2
8
6
.
0
7
3
4
1
.
9
9
6
0
.
0
7
8
4
.
3
0
4
4
1
.
0
4
3
2
.
7
4
8
8
.
3
6
5
4
2
.
7
4
3
4
.
8
3
8
5
.
7
1
6
3
8
.
7
1
3
4
.
5
5
8
3
.
7
7
7
3
9
.
0
7
3
1
.
7
2
9
0
.
1
2
8
3
3
.
4
5
3
7
.
4
3
8
2
.
5
4
9
3
4
.
7
9
3
0
.
0
1
9
0
.
4
8
10
3
3
.
4
9
2
5
.
6
2
8
9
.
2
4
11
3
1
.
0
6
2
6
.
9
6
9
0
.
8
3
12
3
0
.
7
1
3
0
.
1
8
8
7
.
8
3
13
3
2
.
0
7
2
7
.
7
7
8
8
.
8
9
14
3
3
.
1
1
4
2
.
4
5
8
6
.
2
4
15
3
2
.
6
4
2
4
.
4
0
9
0
.
6
5
16
2
8
.
4
6
2
8
.
9
2
8
9
.
4
2
17
2
9
.
8
1
2
1
.
9
9
9
0
.
8
3
18
2
7
.
2
1
2
5
.
5
9
8
9
.
7
7
19
3
4
.
3
2
2
9
.
8
4
9
0
.
6
5
20
3
2
.
0
3
2
8
.
3
5
9
0
.
4
8
Fig
u
r
e
7.
Vis
u
aliza
tio
n
of
t
r
ain
in
g
lo
s
s
an
d
v
alid
atio
n
lo
s
s
vs
n
o.
o
f
e
p
o
c
h
s
Fig
u
r
e
8.
Vis
u
aliza
tio
n
of
ac
c
u
r
ac
y
a
n
d
ep
o
c
h
s
of
C
NN
b
ased
m
o
d
e
l
3
.
2
.
ResNet
9
ba
s
ed
mo
del
T
ab
le
6
s
h
o
ws
th
e
tr
ain
i
n
g
lo
s
s
,
v
alid
atio
n
lo
s
s
an
d
v
alid
atio
n
ac
cu
r
ac
y
o
f
t
h
e
m
o
d
el
d
esig
n
ed
u
s
in
g
R
esNet9
.
T
h
e
n
u
m
b
er
o
f
ep
o
c
h
s
was
s
et
to
2
0
.
T
h
e
m
o
d
el'
s
h
ig
h
est
ac
cu
r
ac
y
was
9
2
.
6
8
%
.
As
th
e
n
u
m
b
er
o
f
ep
o
ch
s
in
c
r
ea
s
ed
f
r
o
m
1
t
o
2
0
,
th
e
m
o
d
el'
s
lo
s
s
d
ec
r
ea
s
ed
.
Fig
u
r
e
9
s
h
o
ws
th
e
r
elatio
n
s
h
ip
b
etwe
en
tr
ain
in
g
lo
s
s
an
d
v
alid
atio
n
lo
s
s
an
d
ep
o
ch
co
u
n
t g
r
ap
h
ically
.
I
t sh
o
w
ed
co
n
s
is
ten
t lo
s
s
o
v
er
th
e
co
u
r
s
e
o
f
th
e
tr
ain
in
g
.
Fig
u
r
e
1
0
s
h
o
ws th
e
a
cc
u
r
ac
y
an
d
ep
o
ch
s
o
f
th
e
R
esNet9
-
b
ased
m
o
d
el
v
is
u
alize
d
.
Fig
u
r
e
9.
Vis
u
aliza
tio
n
of
tr
ai
n
in
g
lo
s
s
an
d
v
alid
atio
n
lo
s
s
vs
no.
of
e
p
o
c
h
s
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
52
In
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
3
7
,
No
.
3
,
Ma
r
ch
20
2
5
:
1
712
-
1
7
2
5
1720
T
ab
le
6
.
R
esu
lts
of
th
e
R
esNet
9
b
ased
m
o
d
e
l
Ep
o
c
h
n
o
Tr
a
i
n
l
o
s
s
V
a
l
i
d
a
t
i
o
n
l
o
s
s
A
c
c
u
r
a
c
y
1
1
3
5
.
0
6
1
3
1
.
3
0
7
6
.
1
9
2
1
2
7
.
5
7
1
2
4
.
3
5
7
9
.
2
8
3
1
2
0
.
7
5
1
1
7
.
9
7
8
0
.
3
4
4
1
1
4
.
3
9
1
1
2
.
0
1
8
1
.
6
6
5
1
0
8
.
4
7
1
0
6
.
4
5
8
3
.
4
2
6
1
0
2
.
9
4
1
0
1
.
2
7
8
4
.
6
6
7
9
7
.
7
7
9
6
.
4
2
8
6
.
0
7
8
9
2
.
9
5
9
1
.
9
1
8
6
.
7
7
9
8
8
.
4
5
8
7
.
6
9
8
7
.
2
1
10
8
4
.
2
5
8
3
.
7
6
8
8
.
1
8
11
8
0
.
3
3
8
0
.
0
9
8
8
.
7
1
12
7
6
.
6
6
7
6
.
6
5
8
9
.
1
5
13
7
3
.
2
5
7
3
.
4
6
8
9
.
7
7
14
7
0
.
0
6
7
0
.
4
7
9
0
.
7
4
15
6
7
.
0
8
6
7
.
6
7
9
0
.
9
2
16
6
4
.
2
9
6
5
.
0
5
9
1
.
1
8
17
6
1
.
6
7
6
2
.
6
1
9
1
.
4
5
18
5
9
.
2
3
6
0
.
3
0
9
2
.
0
6
19
5
6
.
9
3
5
8
.
1
5
9
2
.
5
0
20
5
4
.
7
7
5
6
.
1
2
9
2
.
6
8
Fig
u
r
e
10.
Vis
u
aliza
tio
n
of
ac
c
u
r
ac
y
a
n
d
ep
o
ch
s
o
f
R
esNet9
b
ased
m
o
d
el
3
.
3
.
VG
G
–
1
9
ba
s
ed
m
o
del
T
ab
le
7
s
h
o
ws
th
e
tr
ain
i
n
g
lo
s
s
,
v
alid
atio
n
lo
s
s
an
d
v
alid
atio
n
ac
cu
r
ac
y
o
f
t
h
e
m
o
d
el
d
esig
n
ed
u
s
in
g
VGG
-
19.
T
h
e
n
u
m
b
e
r
o
f
e
p
o
ch
s
was
s
et
to
2
0
.
T
h
e
m
o
d
el
ac
h
iev
ed
a
m
ax
im
u
m
ac
cu
r
a
cy
o
f
9
1
.
1
8
%.
T
h
e
m
o
d
el
lo
s
s
d
ec
r
ea
s
ed
as
th
e
n
u
m
b
er
o
f
ep
o
c
h
s
in
cr
ea
s
ed
f
r
o
m
1
to
2
0
.
Fig
u
r
e
1
1
,
g
r
a
p
h
ically
d
ep
icts
o
f
tr
ain
in
g
lo
s
s
an
d
v
alid
atio
n
l
o
s
s
VS
No
.
o
f
ep
o
c
h
s
.
I
t
s
h
o
wed
co
n
s
is
ten
t
lo
s
s
o
v
er
th
e
co
u
r
s
e
o
f
tr
ain
in
g
.
v
is
u
aliza
tio
n
o
f
ac
cu
r
ac
y
a
n
d
ep
o
ch
s
o
f
m
o
d
el
ca
n
b
e
s
ee
n
i
n
F
ig
u
r
e
1
2
.
Fig
u
r
e
1
1
.
Vis
u
aliza
tio
n
of
tr
ai
n
in
g
lo
s
s
an
d
v
alid
atio
n
lo
s
s
Vs
n
o.
of
e
p
o
c
h
s
Fig
u
r
e
1
2
.
Vis
u
aliza
tio
n
of
ac
cu
r
ac
y
a
n
d
ep
o
ch
s
of
VGG
b
ased
m
o
d
el
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2
5
0
2
-
4
7
52
Dee
p
lea
r
n
in
g
tech
n
iq
u
es fo
r
s
a
tellite ima
g
e
cla
s
s
ifica
tio
n
(
S
u
r
esh
K
u
ma
r
Mu
s
a
li
)
1721
T
ab
le
7
.
R
esu
lts
o
f
th
e
VGG
m
o
d
el
Ep
o
c
h
no
Tr
a
i
n
l
o
s
s
V
a
l
i
d
a
t
i
o
n
l
o
s
s
A
c
c
u
r
a
c
y
1
1
2
8
.
5
3
1
1
1
.
7
5
4
3
.
2
1
2
9
9
.
3
0
8
2
.
5
3
8
1
.
1
3
3
8
0
.
0
4
6
8
.
7
2
8
5
.
1
9
4
6
7
.
3
0
5
8
.
8
7
8
5
.
5
4
5
5
9
.
8
9
5
3
.
7
2
9
2
.
2
4
6
5
3
.
6
4
5
1
.
5
4
8
9
.
7
7
7
5
0
.
3
5
6
7
.
1
3
6
3
.
4
9
8
4
6
.
1
9
4
1
.
5
6
9
3
.
3
0
9
4
4
.
1
5
4
8
.
2
3
8
5
.
3
6
10
4
1
.
9
4
4
2
.
3
5
8
8
.
8
9
11
3
9
.
1
1
4
5
.
3
5
8
6
.
2
4
12
3
5
.
8
7
3
7
.
9
4
8
9
.
7
7
13
3
5
.
2
5
4
0
.
1
0
9
0
.
8
3
14
3
5
.
3
2
4
0
.
2
5
8
5
.
0
1
15
3
2
.
4
8
3
0
.
1
7
9
4
.
0
0
16
3
1
.
3
5
3
2
.
2
3
9
0
.
6
5
17
2
9
.
8
6
3
1
.
3
8
9
2
.
4
2
18
2
9
.
9
1
4
8
.
1
1
8
0
.
2
5
19
3
0
.
0
2
3
3
.
4
7
9
0
.
6
5
20
2
7
.
8
8
3
0
.
1
0
9
1
.
1
8
3
.
4
.
M
o
bil
eNe
t
V2
ba
s
ed
m
o
del
T
ab
le
8
s
h
o
ws
th
e
tr
ai
n
in
g
lo
s
s
,
v
alid
atio
n
lo
s
s
an
d
v
alid
atio
n
ac
cu
r
ac
y
o
f
th
e
m
o
d
el
d
esig
n
ed
u
s
in
g
Mo
b
ileNetV2
.
T
h
e
n
u
m
b
er
o
f
ep
o
ch
s
was
s
et
to
2
0
.
T
h
e
m
o
d
el
ac
h
iev
ed
a
m
a
x
im
u
m
ac
cu
r
ac
y
of
9
6
.
8
3
%.
T
h
e
m
o
d
e
l
lo
s
s
d
ec
r
ea
s
ed
as
t
h
e
n
u
m
b
er
of
ep
o
c
h
s
in
cr
ea
s
ed
f
r
o
m
1
to
20.
Fig
u
r
e
1
3
,
g
r
ap
h
ically
d
ep
icts
of
tr
ain
in
g
l
o
s
s
a
n
d
v
alid
atio
n
l
o
s
s
VS
n
o
.
o
f
ep
o
ch
s
.
O
v
er
th
e
co
u
r
s
e
o
f
th
e
tr
ain
in
g
,
th
e
r
e
is
n
o
co
n
s
is
ten
t
lo
s
s
.
Vis
u
aliza
tio
n
o
f
ac
cu
r
ac
y
an
d
ep
o
ch
s
o
f
M
o
b
ileNetV2
b
ased
m
o
d
el
ca
n
b
e
s
ee
n
in
Fig
u
r
e
1
4
.
T
ab
le
8
.
R
esu
lts
o
f
th
e
Mo
b
ileNet
V2
b
ased
m
o
d
el
E
p
o
c
h
n
o
Tr
a
i
n
l
o
s
s
V
a
l
i
d
a
t
i
o
n
l
o
s
s
A
c
c
u
r
a
c
y
1
2
6
.
1
4
1
1
.
5
2
9
7
.
0
0
2
9
.
2
5
6
.
4
9
9
8
.
2
4
3
7
.
0
2
9
.
7
4
9
6
.
4
7
4
5
.
9
2
9
.
9
8
9
6
.
6
5
5
5
.
6
6
6
.
4
4
9
8
.
2
4
6
6
.
2
9
1
6
.
6
6
9
4
.
1
8
7
5
.
3
6
1
3
.
8
2
9
5
.
7
7
8
5
.
7
6
3
.
5
9
9
8
.
7
7
9
4
.
5
7
6
.
4
6
9
7
.
7
1
10
3
.
5
8
6
.
6
1
9
7
.
7
1
11
3
.
2
1
9
.
8
7
9
6
.
8
3
12
4
.
5
7
6
.
4
5
9
7
.
5
3
13
4
.
3
2
4
.
2
7
9
8
.
4
1
14
4
.
6
7
6
.
8
8
9
7
.
3
5
15
3
.
2
9
1
1
.
0
4
9
6
.
3
0
16
2
.
8
0
3
.
9
6
9
8
.
0
6
17
3
.
3
0
1
3
.
2
5
9
5
.
2
4
18
3
.
7
1
1
.
8
6
9
9
.
1
2
19
3
.
4
5
2
.
4
8
9
9
.
1
2
20
2
.
3
2
1
3
.
0
2
9
6
.
8
3
Fig
u
r
e
1
3
.
Vis
u
aliza
tio
n
of
tr
ai
n
in
g
lo
s
s
an
d
v
alid
atio
n
lo
s
s
VS
No
.
of
e
p
o
c
h
s
Evaluation Warning : The document was created with Spire.PDF for Python.