I
A
E
S
I
n
t
e
r
n
at
io
n
al
Jou
r
n
al
of
A
r
t
if
ic
ia
l
I
n
t
e
ll
ig
e
n
c
e
(
I
J
-
AI
)
V
ol
.
10
, N
o.
1
,
M
a
r
c
h
202
1
, pp.
121
~
130
I
S
S
N
:
2252
-
8938
,
D
O
I
:
10.11591/
ij
a
i.
v
10
.i
1
.pp
121
-
130
121
Jou
r
n
al
h
om
e
page
:
ht
tp
:
//
ij
ai
.
ia
e
s
c
or
e
.c
om
S
at
e
l
l
i
t
e
i
m
age
i
n
p
ai
n
t
i
n
g w
i
t
h
d
e
e
p
ge
n
e
r
at
i
ve
ad
ve
r
sar
i
al
n
e
u
r
al
n
e
t
w
or
k
s
M
oh
am
e
d
A
k
r
am
Z
ayt
ar
,
C
h
ak
e
r
E
l
A
m
r
an
i
Department of Compu
ter Engineering, Facul
ty of Sciences and Techno
logies
,
Tangier, Morocco
A
r
t
ic
le
I
n
f
o
A
B
S
T
R
A
C
T
A
r
ti
c
le
h
is
to
r
y
:
R
e
c
e
iv
e
d
A
ug
2
0
, 20
20
R
e
vi
s
e
d
J
a
n 29
, 20
21
A
c
c
e
pt
e
d
F
e
b 11
, 20
2
1
This
work
addresses
the
problem
of
recovering
lost
or
damaged
s
atellite
image
pixels
(gaps)
caused
by
sensor
processing
errors
or
by
natural
phenomena
like
cloud
presence.
Such
errors
decrease
our
ability
to
monitor
regions
of
interest
and
significant
ly
increase
the
average
revisit
time
for
all
satellites.
This
paper
presents
a
novel
neural
system
based
on
con
ditional
deep
generative
adversarial
networks
(cGA
N)
optimized
to
fill
s
atellite
imagery ga
ps using surro
unding
pixel value
s
and static
high
-
resolutio
n visual
priors.
Experimental
results
show
that
the
proposed
system
outpe
rforms
traditional
and
neural
network
baselines.
It
achieve
s
a
normalize
d
least
abs
olute
deviations
error
of
1
=
0
.
33
(
21%
and
60%
decrease
in
error
compared
with
the
two
baselines
)
and
a
mean
squared
error
l
oss
of
ℒ
2
=
0
.
15
(
29%
and
73%
decrease in er
ror)
over the test
set. The model
can
be
deployed
within
a
remote
sensing
data
pipeline
t
o
reconstruct
missing
pixel
measurements
for
near
-
real
-
time
monitoring
and
inference
purposes,
thus
empowering
policymakers
and
users
to
make
environmentally
in
formed
decisions.
K
e
y
w
o
r
d
s
:
A
ir
pol
lu
ti
on
G
e
ne
r
a
ti
ve
a
dve
r
s
a
r
ia
l
ne
ts
I
m
a
ge
i
npa
in
ti
ng
N
e
ur
a
l
n
e
twor
ks
S
a
te
ll
it
e
i
m
a
ge
r
y
This is an
open
acce
ss artic
le unde
r the
CC BY
-
SA
license.
C
or
r
e
s
pon
di
n
g A
u
th
or
:
M
oha
m
e
d A
kr
a
m
Z
a
yt
a
r
D
e
pa
r
tm
e
nt
of
C
om
put
e
r
E
ngi
ne
e
r
in
g
F
a
c
ul
ty
of
S
c
ie
nc
e
s
a
nd
T
e
c
hnol
ogi
e
s
R
out
e
Z
ia
te
n, T
a
ngi
e
r
, M
or
oc
c
o
E
m
a
il
:
m
.z
a
yt
a
r
@
ua
e
.a
c
.m
a
1.
I
N
T
R
O
D
U
C
T
I
O
N
C
li
m
a
te
c
ha
nge
pos
e
s
s
e
r
io
us
c
ha
ll
e
nge
s
th
a
t
th
r
e
a
te
n
hum
a
ni
ty
'
s
lo
ng
-
te
r
m
s
a
f
e
ty
[
1]
.
A
ddr
e
s
s
in
g
th
e
s
e
c
ha
ll
e
nge
s
de
pe
nds
on br
e
a
kt
hr
oughs
i
n
e
nvi
r
onm
e
nt
a
l
p
ol
ic
y a
nd c
li
m
a
te
s
c
ie
nc
e
[
2]
.
C
li
m
a
te
r
e
s
e
a
r
c
h
is
ke
y
to
unde
r
s
ta
ndi
ng
th
e
lo
ng
-
te
r
m
e
f
f
e
c
ts
of
gl
oba
l
w
a
r
m
in
g
on
a
gr
ic
ul
tu
r
e
[
3
-
4]
,
f
ood
s
e
c
ur
it
y
[
5]
,
a
i
r
qua
li
ty
[
6]
,
a
nd
w
e
a
th
e
r
c
ondi
ti
ons
[
7]
.
O
n
th
e
ot
he
r
ha
nd,
o
ne
of
th
e
pr
im
a
r
y
da
ta
s
our
c
e
s
th
a
t
e
m
pow
e
r
e
nvi
r
onm
e
nt
a
l
r
e
s
e
a
r
c
h
i
s
s
a
te
ll
it
e
im
a
ge
r
y
[
8]
.
S
a
te
ll
it
e
pr
ogr
a
m
s
s
uc
h
a
s
l
a
nds
a
t
[
9]
,
s
e
nt
in
e
l,
a
qua
/t
e
r
r
a
,
a
m
ong
ot
he
r
s
,
pr
ovi
de
a
w
e
a
lt
h
of
f
r
e
e
ly
a
va
il
a
bl
e
da
ta
s
e
ts
f
or
th
e
m
a
s
s
e
s
.
T
hi
s
tr
e
m
e
ndous
pr
ogr
e
s
s
ha
s
unl
oc
ke
d
m
a
ny
in
no
va
ti
ons
th
a
t
e
xt
r
a
c
t
va
lu
a
bl
e
in
s
ig
ht
s
[
1
0
-
11]
f
r
o
m
s
a
te
ll
it
e
im
a
ge
r
y
us
in
g
bi
g
da
ta
pi
pe
li
ne
s
a
nd a
dva
n
c
e
d m
a
c
hi
ne
l
e
a
r
ni
ng s
ys
te
m
s
[
12]
.
S
a
te
ll
it
e
s
e
ns
or
s
a
r
e
li
m
it
e
d
by
th
e
ir
S
pa
ti
o
-
te
m
por
a
l
r
e
s
ol
ut
io
n.
A
s
a
t
e
ll
it
e
'
s
te
m
por
a
l
r
e
s
ol
ut
io
n
r
e
pr
e
s
e
nt
s
th
e
d
ur
a
ti
on
of
ge
tt
in
g
in
f
o
r
m
a
ti
on
a
bout
th
e
s
a
m
e
poi
nt
on
e
a
r
th
.
O
n
th
e
ot
he
r
ha
nd,
s
pa
ti
a
l
r
e
s
ol
ut
io
n
s
pe
c
if
ie
s
th
e
s
ur
f
a
c
e
s
iz
e
of
1
pi
x
e
l
of
in
f
or
m
a
ti
on
(
e
x.
S
e
nt
in
e
l
-
2
ha
s
a
n R
G
B
s
pa
ti
a
l
r
e
s
ol
ut
io
n
of
10
×
10
m
pe
r
pi
xe
l)
.
D
ue
to
va
r
io
us
r
e
a
s
ons
,
m
os
t
s
a
te
l
li
te
im
a
ge
r
y
c
ont
a
in
s
"
hol
e
s
"
or
"
ga
ps
"
o
f
m
is
s
in
g
pi
xe
l
va
lu
e
s
.
C
lo
uds
a
r
e
th
e
pr
im
a
r
y
c
ont
r
ib
ut
or
to
s
uc
h
noi
s
e
.
S
a
te
ll
it
e
noi
s
e
is
c
ha
ll
e
ngi
ng
be
c
a
u
s
e
it
w
or
s
e
ns
th
e
s
a
te
ll
it
e
'
s
te
m
por
a
l
r
e
s
ol
ut
io
n
a
nd
in
tr
oduc
e
s
unc
e
r
ta
in
ty
in
to
a
tm
os
phe
r
ic
m
oni
to
r
in
g
pi
pe
li
ne
s
.
M
a
ny
ha
v
e
r
e
s
or
te
d
to
u
s
in
g
I
oT
s
e
n
s
or
s
[
13]
th
a
t
pr
ovi
de
a
hi
gh
e
r
-
qua
li
ty
gr
ound
-
le
ve
l
s
tr
e
a
m
of
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
,
V
ol
.
10
, N
o.
1
,
M
a
r
c
h
20
2
1
:
121
–
130
122
m
e
a
s
ur
e
m
e
nt
s
[
14]
.
H
ow
e
ve
r
,
gr
ound
s
e
ns
or
s
c
a
n
onl
y
gi
ve
i
nf
or
m
a
ti
on
a
bout
a
s
pe
c
if
ic
lo
c
a
ti
on
a
nd,
a
s
a
r
e
s
ul
t,
do not ha
ve
t
he
ge
ogr
a
phi
c
c
ove
r
a
g
e
t
ha
t
s
a
te
ll
it
e
s
ha
ve
(
m
os
t
pol
a
r
s
a
te
ll
it
e
s
c
ove
r
t
he
w
hol
e
e
a
r
th
)
.
R
e
m
ot
e
s
e
ns
or
s
a
r
e
th
e
pr
im
a
r
y
da
ta
s
our
c
e
f
or
la
r
ge
-
s
c
a
le
a
tm
os
phe
r
ic
m
oni
to
r
in
g
a
nd,
m
or
e
s
pe
c
if
ic
a
ll
y,
a
ir
qua
li
ty
m
oni
to
r
in
g.
E
nha
nc
in
g
th
e
S
pa
ti
o
-
te
m
por
a
l
r
e
s
ol
ut
io
n
of
s
a
te
ll
it
e
s
e
n
s
or
s
i
s
of
c
r
it
ic
a
l
im
por
ta
nc
e
s
in
c
e
it
e
na
bl
e
s
gr
e
a
te
r
vi
s
ib
il
it
y
ove
r
th
e
s
ta
te
of
p
la
ne
t
e
a
r
th
.
F
or
th
is
r
e
a
s
on,
th
is
s
tu
dy
f
oc
us
e
s
on
in
pa
in
ti
ng
s
a
te
ll
it
e
N
O
2
im
a
ge
s
.
E
a
c
h
N
O
2
pi
xe
l
m
e
a
s
ur
e
s
th
e
a
t
m
os
phe
r
ic
N
O
2
ve
r
ti
c
a
l
de
ns
it
y
(
in
D
obs
on
uni
ts
)
ove
r
th
e
pi
xe
l.
N
O
2
is
a
tr
a
c
e
g
a
s
th
a
t
ne
ga
ti
ve
ly
a
f
f
e
c
ts
a
ir
qua
li
ty
a
nd
th
e
c
li
m
a
te
.
I
t
is
li
nke
d
to
r
oa
d
tr
a
f
f
ic
a
nd
in
dus
tr
ia
l
a
c
ti
vi
ti
e
s
s
uc
h
a
s
f
os
s
il
f
ue
l
c
o
m
bus
ti
on
[
15]
.
A
hi
gh
N
O
2
c
onc
e
nt
r
a
ti
on
c
a
n
c
a
us
e
nume
r
ous
r
e
s
pi
r
a
to
r
y di
s
e
a
s
e
s
[
16]
.
T
hi
s
pa
pe
r
pr
opos
e
s
a
ge
ne
r
a
ti
ve
a
dve
r
s
a
r
ia
l
s
ys
te
m
us
e
d t
o f
il
l
th
e
m
is
s
in
g ga
ps
i
n i
m
a
ge
s
ba
s
e
d on
th
e
im
a
ge
'
s
c
ont
e
nt
(
a
va
il
a
bl
e
pi
xe
l
va
lu
e
s
)
a
nd
hi
gh
-
r
e
s
ol
ut
io
n
vi
s
ua
l
pr
io
r
s
.
T
he
s
ys
te
m
'
s
nov
e
lt
y
li
e
s
in
it
s
us
e
of
a
di
f
f
e
r
e
nt
da
ta
m
oda
li
ty
(
hi
ghe
r
-
r
e
s
ol
ut
io
n
s
ta
ti
c
R
G
B
im
a
ge
s
)
e
nc
ode
d
by
a
c
ondi
ti
ona
l
la
ye
r
to
pr
ovi
de
a
uxi
li
a
r
y
f
e
a
tu
r
e
s
to
th
e
c
om
pl
e
to
r
ne
twor
k.
A
s
a
r
e
s
ul
t,
th
e
ne
ur
a
l
s
y
s
te
m
in
pa
in
t
s
a
ll
f
ut
ur
e
im
a
ge
s
a
nd
pus
he
s
th
e
s
e
n
s
or
'
s
te
m
por
a
l
r
e
s
ol
ut
io
n
to
i
ts
th
e
or
e
ti
c
a
l
li
m
it
(
nul
li
f
yi
ng
th
e
e
f
f
e
c
ts
of
c
lo
uds
or
s
e
ns
or
y
e
r
r
or
s
)
. T
he
pa
pe
r
'
s
m
a
in
c
ont
r
ib
ut
io
ns
a
r
e
out
li
ne
d a
s
:
−
A
c
G
A
N
-
ba
s
e
d ne
ur
a
l
s
ys
te
m
f
or
i
npa
in
ti
ng mul
ti
-
s
pe
c
tr
a
l
s
a
te
l
li
te
i
m
a
ge
r
y.
−
A
f
ul
l
de
s
c
r
ip
ti
on of
t
he
pr
e
-
in
f
e
r
e
nc
e
da
ta
pr
e
pr
oc
e
s
s
in
g pi
pe
li
ne
.
−
C
a
s
e
s
tu
dy:
th
e
m
e
th
od
is
e
va
lu
a
t
e
d
on
2
pol
lu
ti
on
im
a
ge
s
f
or
ne
a
r
-
r
e
a
l
-
ti
m
e
a
ir
qua
li
ty
m
oni
to
r
in
g,
s
how
c
a
s
in
g t
he
pot
e
nt
ia
l
of
f
us
in
g m
ul
ti
-
m
oda
l
s
a
te
ll
it
e
da
ta
us
in
g ne
ur
a
l
a
ppr
oxi
m
a
to
r
s
.
T
he
r
e
s
t
of
th
e
p
a
pe
r
is
s
tr
uc
tu
r
e
d
a
s
;
"
R
e
la
te
d
w
or
ks
"
de
s
c
r
ib
e
s
th
e
m
os
t
not
a
bl
e
r
e
s
e
a
r
c
h
e
f
f
or
ts
th
a
t
ta
c
kl
e
im
a
g
e
in
pa
in
ti
ng.
"
R
e
s
e
a
r
c
h
m
e
th
od"
in
tr
oduc
e
s
th
e
ne
ur
a
l
s
ys
t
e
m
a
r
c
hi
te
c
tu
r
e
,
th
e
tr
a
in
in
g
a
lg
or
it
hm
,
a
nd
th
e
da
ta
s
e
t.
"
R
e
s
ul
ts
a
nd
di
s
c
u
s
s
io
n"
de
s
c
r
ib
e
s
s
ynt
h
e
ti
c
noi
s
e
ge
ne
r
a
ti
on,
in
tr
oduc
e
s
th
e
pe
r
f
or
m
a
nc
e
m
e
tr
ic
s
,
a
nd
pr
e
s
e
nt
s
th
e
f
in
a
l
r
e
s
ul
ts
.
F
in
a
ll
y,
it
pr
ovi
de
s
a
n
in
tu
it
iv
e
unde
r
s
ta
ndi
ng
of
th
e
e
f
f
e
c
ts
of
pr
io
r
s
a
nd t
he
ir
l
im
it
a
ti
ons
. "
C
onc
lu
s
io
n"
s
um
m
a
r
iz
e
s
t
he
pa
pe
r
a
nd de
s
c
r
ib
e
s
f
ut
ur
e
w
or
k.
2.
R
E
L
A
T
E
D
WORKS
T
he
e
xi
s
ti
ng
r
e
s
e
a
r
c
h
li
te
r
a
tu
r
e
on
im
a
ge
in
p
a
in
ti
ng
c
a
n
be
gr
oupe
d
in
to
two
m
a
in
pa
r
ts
.
N
on
-
le
a
r
ni
ng
m
e
th
ods
s
uc
h
a
s
di
f
f
us
io
n/
pa
tc
h
-
ba
s
e
d
a
lg
or
it
hm
s
,
a
nd
th
e
r
e
la
ti
ve
ly
r
e
c
e
nt
w
or
k
th
a
t
a
tt
e
m
pt
s
to
le
a
r
n
in
pa
in
ti
ng
by
tr
a
in
in
g
c
onvolut
io
na
l
ne
ur
a
l
ne
twor
k
-
ba
s
e
d
a
r
c
hi
te
c
tu
r
e
s
(
C
N
N
s
)
.
T
hi
s
s
e
c
ti
on
out
li
ne
s
th
e
m
os
t
not
a
bl
e
e
f
f
or
ts
f
r
om
bot
h s
id
e
s
.
2
.
1
.
D
if
f
u
s
io
n
or
p
at
c
h
-
b
as
e
d
m
e
t
h
od
s
T
he
e
a
r
ly
s
u
c
c
e
s
s
in
im
a
ge
in
pa
in
ti
ng
is
a
tt
r
ib
ut
e
d
to
in
f
or
m
a
ti
on
pr
opa
ga
ti
on
te
c
hni
que
s
th
r
oug
h
pa
tc
h
s
im
il
a
r
it
y
or
va
r
ia
ti
ona
l
m
e
th
ods
.
E
f
r
os
a
nd
L
e
ung
[
17]
pr
opos
e
d
to
m
ode
l
im
a
g
e
te
xt
ur
e
s
a
s
M
a
r
kov
r
a
ndom
f
ie
ld
s
th
e
n
us
e
s
im
il
a
r
it
y
s
e
a
r
c
h
to
f
il
l
th
e
m
is
s
in
g
pi
xe
ls
.
O
th
e
r
e
f
f
or
ts
[
18
-
19
]
w
e
r
e
di
r
e
c
te
d
to
w
a
r
d
in
pa
in
ti
ng
im
a
ge
s
th
r
ough
th
e
ir
te
xt
ur
e
a
nd
s
tr
uc
tu
r
e
u
s
in
g
s
e
a
r
c
h
-
ba
s
e
d
gui
de
d
pr
opa
ga
ti
on
th
a
t
s
ynt
he
s
iz
e
s
pa
tt
e
r
ns
r
e
s
e
m
bl
in
g ot
he
r
i
m
a
ge
r
e
gi
ons
or
ot
he
r
i
m
a
ge
s
w
it
hi
n
a
s
e
a
r
c
ha
bl
e
d
a
ta
ba
s
e
.
V
a
r
ia
ti
ona
l
m
e
th
ods
a
r
e
a
ls
o
pr
e
s
e
nt
in
[
20]
th
a
t
us
e
f
e
a
tu
r
e
e
xt
r
a
c
to
r
s
s
u
c
h
a
s
pa
tc
h
s
ta
ti
s
ti
c
s
,
c
ol
or
s
,
a
nd
gr
a
di
e
nt
s
to
s
ynt
he
s
iz
e
th
e
m
is
s
in
g
im
a
ge
ga
p
s
.
L
a
s
tl
y,
out
-
of
-
s
a
m
pl
e
in
p
a
in
ti
ng
w
a
s
a
c
hi
e
ve
d
by
[
21]
us
in
g
a
n
e
xt
e
ns
iv
e
da
ta
ba
s
e
of
im
a
ge
s
.
I
ts
a
lg
or
it
hm
in
pa
in
ts
m
is
s
in
g
r
e
gi
ons
of
a
n
im
a
ge
by
f
in
di
ng
s
im
il
a
r
im
a
ge
s
th
e
n
di
f
f
us
in
g
e
xt
r
a
c
te
d
lo
w
-
le
ve
l
f
e
a
tu
r
e
s
.
U
nl
i
ke
ot
he
r
s
,
th
is
te
c
hni
que
c
a
n
s
ugge
s
t
m
ul
ti
pl
e
c
om
pl
e
ti
ons
ba
s
e
d on th
e
c
hos
e
n d
a
ta
ba
s
e
i
te
m
.
T
hi
s
c
l
a
s
s
of
m
e
th
ods
w
or
ks
w
e
ll
on
im
a
ge
s
th
a
t
c
ont
a
in
r
e
pe
a
t
e
d
or
s
ta
ti
c
pa
tt
e
r
ns
(
e
x
a
m
pl
e
s
:
s
a
nd,
gr
id
,
pa
pe
r
)
bu
t
f
a
il
s
on
im
a
ge
s
w
it
h
r
ic
h
s
e
m
a
nt
ic
c
ont
e
nt
.
F
u
r
th
e
r
m
or
e
,
a
ut
om
a
ti
c
non
-
le
a
r
ni
ng
a
lg
or
it
h
m
s
c
a
nnot
in
pa
in
t
a
b
s
tr
a
c
ti
ons
th
a
t
m
a
ke
c
om
pl
e
x
im
a
ge
s
c
ohe
s
iv
e
in
th
e
ir
c
ont
e
nt
,
a
nd
th
e
ir
us
e
of
out
-
of
-
s
a
m
pl
e
i
nf
or
m
a
ti
on i
s
l
im
it
e
d due
t
o t
he
ir
l
oc
a
l
de
pe
nde
nc
ie
s
.
2
.
2
.
L
e
ar
n
in
g
-
b
as
e
d
ap
p
r
oac
h
e
s
O
ne
of
th
e
e
a
r
li
e
s
t
e
f
f
or
ts
to
us
e
r
e
pr
e
s
e
nt
a
ti
on
le
a
r
ni
ng
f
or
im
a
ge
in
pa
in
ti
ng
pr
opos
e
d
a
m
ul
ti
-
la
ye
r
pe
r
c
e
pt
r
on
(
M
L
P
)
a
r
c
hi
te
c
tu
r
e
to
f
il
l
m
is
s
in
g
pi
xe
ls
in
gr
a
y
-
s
c
a
le
im
a
ge
s
by
m
in
im
iz
in
g
th
e
r
e
c
ons
tr
uc
ti
on
lo
s
s
[
22]
.
T
he
pa
pe
r
e
s
ta
bl
is
he
d
th
e
im
por
ta
nc
e
of
m
a
s
ki
ng
m
is
s
in
g
pi
xe
ls
a
nd
th
e
pot
e
nt
ia
l
of
ne
ur
a
l
ne
twor
ks
(
N
N
s
)
in
im
a
ge
c
om
pl
e
ti
on.
F
ur
th
e
r
m
or
e
,
X
u
e
t
al
.
[
23]
us
e
d
a
C
N
N
a
r
c
hi
te
c
tu
r
e
to
pr
opos
e
a
ge
ne
r
a
l
m
e
th
od f
or
s
ol
vi
ng t
hr
e
e
t
a
s
ks
:
im
a
ge
i
npa
in
ti
ng, de
noi
s
in
g, a
nd i
m
a
ge
de
gr
a
da
ti
on r
e
c
ove
r
y
.
R
e
c
e
nt
ly
, ne
ur
a
l
ne
twor
ks
t
r
a
in
e
d us
in
g pi
xe
l
-
w
is
e
r
e
c
ons
tr
uc
ti
on e
r
r
or
a
nd a
dve
r
s
a
r
ia
l
lo
s
s
r
e
por
te
d
pr
om
is
in
g
r
e
s
ul
ts
.
T
h
e
w
or
k
of
[
24]
in
tr
oduc
e
d
c
ont
e
xt
e
n
c
ode
r
s
to
f
il
l
la
r
ge
hol
e
s
in
im
a
ge
c
e
nt
e
r
s
.
Y
a
ng
e
t
al
.
[
25]
e
na
bl
e
d
hi
gh
-
r
e
s
ol
ut
io
n
im
a
ge
in
pa
in
ti
ng
b
y
pr
opos
in
g
jo
in
t
c
ont
e
nt
a
nd
te
xt
ur
e
lo
s
s
e
s
.
X
u
e
t
al
.
[
26]
c
om
bi
ne
d
lo
c
a
l
a
nd
gl
oba
l
di
s
c
r
im
in
a
to
r
s
in
to
o
ne
ne
twor
k
a
nd
us
e
d
c
onvolut
io
ns
a
nd
di
la
te
d
c
onvolut
io
ns
t
o i
npa
in
t
im
a
ge
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
Sat
e
ll
it
e
i
m
age
i
npai
nt
in
g w
it
h de
e
p ge
ne
r
at
iv
e
ad
v
e
r
s
ar
ia
l
ne
u
r
al
ne
tw
or
k
s
(
M
ohame
d A
k
r
am
Z
ay
ta
r
)
123
Yu
e
t
al
.
[
27]
im
pr
ove
d
th
e
pr
e
vi
ous
a
r
c
hi
te
c
tu
r
e
by
di
vi
di
n
g
t
he
ge
ne
r
a
ti
on
pr
oc
e
s
s
in
to
two s
ta
ge
s
.
T
he
f
ir
s
t
out
put
s
a
bl
ur
r
y
im
a
ge
opt
im
iz
e
d
w
it
h
s
pa
ti
a
l
di
s
c
ount
e
d
ℒ
1
r
e
c
ons
tr
uc
ti
on
lo
s
s
,
a
nd
th
e
s
e
c
ond
r
e
f
in
e
s
a
nd
out
put
s
th
e
f
in
a
l
im
a
ge
.
T
he
a
ut
hor
s
us
e
d
th
e
ne
t
w
or
k'
s
out
put
a
s
in
put
to
th
e
gl
oba
l
a
nd
lo
c
a
l
di
s
c
r
im
in
a
to
r
s
a
nd
c
ho
s
e
w
a
s
s
e
r
s
te
in
G
A
N
s
(
W
G
A
N
)
to
tr
a
in
th
e
ne
ur
a
l
s
ys
te
m
(
W
G
A
N
s
t
a
bi
li
z
e
s
th
e
ove
r
a
ll
opt
im
iz
a
ti
on
pr
oc
e
s
s
)
.
F
in
a
ll
y,
X
u
e
t
al
.
[
28]
im
pr
ov
e
d
th
e
pr
e
vi
ous
a
r
c
hi
te
c
tu
r
e
by
in
c
or
por
a
ti
ng
c
ont
e
xt
ua
l
a
tt
e
nt
io
n a
nd dil
a
te
d ga
te
d c
onvolut
io
n
s
in
to
bot
h t
he
c
oa
r
s
e
a
nd r
e
f
in
e
m
e
nt
ne
twor
ks
.
A
lt
hough
th
e
m
e
nt
io
ne
d
ne
ur
a
l
s
ys
te
m
s
pr
ovi
de
im
pr
e
s
s
iv
e
in
pa
in
ti
ngs
a
nd
pr
e
di
c
t
hi
gh
-
qua
li
ty
vi
s
ua
l
s
e
m
a
nt
ic
s
,
none
ha
ve
e
xp
e
r
im
e
nt
e
d
w
it
h
pr
io
r
s
or
e
xt
e
nde
d
th
e
ge
ne
r
a
to
r
/d
is
c
r
im
in
a
to
r
w
it
h
a
c
ondi
ti
ona
l
la
ye
r
.
F
ur
th
e
r
m
or
e
,
a
ll
of
th
e
m
e
nt
io
ne
d
m
e
th
od
s
a
s
s
um
e
one
s
our
c
e
da
ta
di
s
tr
ib
ut
io
n
to
be
m
ode
le
d
,
a
s
s
how
n
in
T
a
bl
e
1.
T
hi
s
s
tu
dy
e
s
ta
bl
is
he
s
th
e
im
por
ta
nc
e
of
us
in
g
di
f
f
e
r
e
nt
da
ta
m
oda
li
ti
e
s
a
nd
f
us
in
g
th
e
m
th
r
ough
a
c
ondi
ti
ona
l
la
ye
r
to
s
ol
ve
im
a
ge
in
pa
in
t
in
g
in
ge
ne
r
a
l,
a
nd
a
ir
qua
li
ty
e
s
ti
m
a
ti
on
s
pe
c
if
ic
a
ll
y.
T
a
bl
e
1
.
A
C
om
p
a
r
is
on of
di
f
f
e
r
e
nt
a
lg
or
it
hm
s
t
ha
t
a
im
t
o s
ol
ve
i
m
a
ge
i
npa
in
ti
ng
F
e
a
t
ur
e
s
\
M
e
t
hods
N
on
-
P
a
r
a
m
e
t
e
r
i
c
S
a
m
pl
e
r
[
17]
P
a
t
c
hM
a
t
c
h [
29]
M
a
s
k
-
F
C
I
npa
i
nt
e
r
[
22]
L
oc
a
l
G
l
oba
l
[
26]
O
ur
s
F
r
e
e
-
F
or
m
✓
✓
✓
✓
✓
O
ut
-
of
-
S
a
m
pl
e
✓
✓
✓
✓
S
e
m
a
nt
i
c
s
✓
✓
✓
M
ul
t
i
-
M
oda
l
✓
3.
R
E
S
E
A
R
C
H
M
E
T
H
O
D
T
w
o
C
N
N
-
ba
s
e
d
ne
twor
k
a
r
c
hi
te
c
tu
r
e
s
w
e
r
e
tr
a
in
e
d
w
it
hi
n
a
c
ondi
ti
ona
l
a
dve
r
s
a
r
ia
l
f
r
a
m
e
w
or
k
a
s
s
how
n
in
F
ig
ur
e
1.
T
he
ge
n
e
r
a
to
r
ne
twor
k,
r
e
s
pons
ib
le
f
o
r
f
il
li
ng
th
e
m
is
s
in
g
ga
ps
us
in
g
c
ont
e
xt
u
a
l
in
f
or
m
a
ti
on
a
nd
s
ta
ti
c
pr
io
r
s
,
a
nd
a
n
a
uxi
li
a
r
y
di
s
c
r
im
in
a
to
r
ne
twor
k
tr
a
in
e
d
to
d
is
ti
ngui
s
h
be
twe
e
n
r
e
a
l
a
nd
c
om
pl
e
te
d
pol
lu
ti
on
pa
tc
he
s
.
B
ot
h
ne
twor
ks
a
r
e
c
ondi
ti
one
d
o
ve
r
tr
ue
-
c
ol
or
im
a
ge
r
y
th
a
t
c
o
r
r
e
s
ponds
to
th
e
r
e
gi
on
c
ove
r
in
g
th
e
in
put
pa
tc
h.
T
he
pr
io
r
is
e
nc
ode
d
by
a
c
ondi
ti
ona
l
la
ye
r
(
r
e
duc
e
r
)
.
T
he
in
put
to
th
e
ge
ne
r
a
to
r
c
ons
is
ts
of
a
da
m
a
ge
d
im
a
ge
(
x)
a
nd
it
s
hi
gh
-
r
e
s
ol
ut
io
n
pr
io
r
(
p
)
.
T
he
r
e
duc
e
r
ne
twor
k
c
om
pr
e
s
s
e
s
p
to
th
e
s
a
m
e
s
iz
e
of
x
th
e
n
s
ta
c
k
s
bot
h
f
or
in
pa
in
ti
ng.
T
he
di
s
c
r
im
in
a
to
r
ne
twor
k
ta
ke
s
e
it
he
r
a
he
a
lt
hy
or
a
c
om
pl
e
te
d i
m
a
ge
w
it
h i
ts
e
nc
ode
d
pr
io
r
. T
h
e
di
s
c
r
im
in
a
to
r
j
udge
s
i
f
a
n i
m
a
ge
i
s
r
e
a
l
or
c
om
pl
e
te
d.
F
ig
ur
e
1.
S
ys
te
m
o
ve
r
vi
e
w
3
.1.
C
on
vol
u
t
io
n
al
n
e
u
r
al
n
e
t
w
or
k
s
T
he
r
e
duc
e
r
,
c
om
pl
e
to
r
,
a
nd
di
s
c
r
im
in
a
to
r
ne
twor
ks
a
r
e
b
a
s
e
d
on
c
onvolut
io
na
l
ne
ur
a
l
ne
twor
ks
(
C
N
N
)
.
C
N
N
s
a
r
e
a
s
pe
c
ia
l
ty
pe
of
ne
ur
a
l
ne
twor
k
th
a
t
us
e
s
w
e
ig
ht
s
ha
r
in
g
to
e
xt
r
a
c
t
h
ie
r
a
r
c
hi
c
a
l
vi
s
ua
l
f
e
a
tu
r
e
s
w
it
h
m
in
im
a
l
f
r
e
e
pa
r
a
m
e
te
r
s
a
nd
m
a
xi
m
a
l
lo
c
a
l
c
onne
c
ti
ons
.
K
e
r
ne
l
w
e
ig
ht
s
a
r
e
opt
im
iz
e
d
to
pr
oduc
e
a
c
ti
va
ti
ons
th
a
t
he
lp
in
th
e
f
in
a
l
pr
e
di
c
ti
on
ta
s
k.
C
N
N
s
a
r
e
c
a
pa
bl
e
of
pr
ogr
e
s
s
iv
e
ly
e
xt
r
a
c
ti
ng
hi
ghe
r
-
or
de
r
a
bs
tr
a
c
ti
ons
th
a
t
s
e
r
ve
to
m
in
im
iz
e
a
pr
e
-
de
f
in
e
d
obj
e
c
ti
ve
f
unc
ti
on.
A
s
pe
c
if
ic
a
c
ti
va
ti
on
is
c
a
lc
ul
a
te
d us
in
g (
1)
.
,
=
σ
(
+
∑
∑
,
−
1
=
0
−
1
=
0
+
,
+
)
(
1)
W
it
h
X
r
e
pr
e
s
e
nt
in
g
th
e
in
put
,
K
th
e
ke
r
ne
l
(
m
a
tr
ic
e
s
of
le
a
r
na
bl
e
w
e
ig
ht
s
)
,
s
i
s
th
e
ke
r
ne
l
s
iz
e
,
a
nd
(
m
,
n
)
a
r
e
th
e
in
di
c
e
s
of
th
e
ta
r
ge
t
va
lu
e
Y
m
,
n
in
th
e
a
c
ti
va
ti
on
la
ye
r
.
D
il
a
te
d
c
onvolut
io
na
l
la
ye
r
s
a
r
e
a
l
s
o
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
,
V
ol
.
10
, N
o.
1
,
M
a
r
c
h
20
2
1
:
121
–
130
124
us
e
d t
o s
im
ul
a
te
a
l
a
r
ge
r
r
e
c
e
pt
iv
e
f
ie
ld
w
it
hout
a
ddi
ng mor
e
pa
r
a
m
e
te
r
s
. T
o c
a
lc
u
la
t
e
di
la
te
d a
c
ti
va
ti
ons
, one
pa
r
a
m
e
te
r
i
s
a
dde
d t
o t
he
pr
e
vi
ous
de
f
in
it
io
n:
η
, w
hi
c
h i
s
t
he
di
la
ti
on f
a
c
to
r
a
s
(
2)
.
,
=
σ
(
+
∑
∑
,
−
1
=
0
−
1
=
0
+
η
,
+
η
)
(
2)
3
.
2
.
C
on
d
it
io
n
al
ge
n
e
r
at
iv
e
ad
ve
r
s
ar
ia
l
n
e
t
w
or
k
s
G
e
ne
r
a
ti
ve
a
dve
r
s
a
r
ia
l
ne
twor
ks
(
G
A
N
)
a
r
e
a
c
la
s
s
of
ne
ur
a
l
ne
twor
ks
tr
a
in
e
d
in
a
n
a
dve
r
s
a
r
ia
l
m
a
nne
r
.
A
G
A
N
c
ons
is
ts
of
two
ne
twor
ks
:
a
ge
ne
r
a
ti
ve
ne
tw
or
k
G
(
.
)
th
a
t
le
a
r
ns
th
e
tr
ue
da
ta
di
s
tr
ib
ut
io
n
(
th
e
pr
oc
e
s
s
th
a
t
g
e
ne
r
a
te
d
th
e
tr
a
in
in
g
da
ta
)
a
nd
a
di
s
c
r
im
in
a
ti
ve
ne
twor
k
D
(
.
)
th
a
t
e
s
ti
m
a
te
s
if
a
s
a
m
pl
e
c
a
m
e
f
r
om
t
he
t
r
ue
da
ta
di
s
tr
ib
ut
io
n
or
G
(
.
)
. I
de
a
ll
y, both
G
a
nd
D
a
r
e
t
r
a
in
e
d s
im
ul
ta
ne
ous
ly
,
G
′
s
pa
r
a
m
e
te
r
s
a
r
e
a
dj
us
te
d
to
m
in
im
iz
e
L
o
g
(
1
−
D
(
G
(
z
)
)
)
(
i.
e
.,
to
f
ool
th
e
di
s
c
r
im
in
a
to
r
)
,
a
nd
D
′
s
pa
r
a
m
e
te
r
s
a
r
e
tu
ne
d
to
m
a
xi
m
iz
e
L
o
g
(
D
(
x
)
)
(
i.
e
.,
to
de
te
c
t
f
a
ke
ge
ne
r
a
te
d
in
put
s
)
.
D
a
nd
G
pl
a
y
th
e
f
ol
lo
w
in
g
two
-
pl
a
ye
r
m
in
im
a
x
ga
m
e
w
it
h va
lu
e
f
unc
ti
on
V
(
G
,
D
)
a
s
(
3)
.
m
in
m
a
x
(
,
)
=
∼
(
)
[
(
(
)
)
]
+
∼
(
)
[
(
1
−
(
(
)
)
)
]
(
3)
I
n
th
e
c
ont
e
xt
of
th
is
s
tu
dy,
vi
s
ua
l
pr
io
r
s
a
r
e
of
hi
ghe
r
r
e
s
ol
ut
io
n
th
a
n
pol
lu
ti
on
pa
tc
he
s
.
D
ow
ns
a
m
pl
in
g
vi
s
ua
l
im
a
ge
r
y
to
f
it
pol
lu
ti
on
pa
tc
he
s
w
il
l
r
e
s
ul
t
in
lo
s
in
g
m
uc
h
of
it
s
e
nc
ode
d
in
f
or
m
a
ti
on.
A
ddi
ti
ona
ll
y,
f
r
om
th
e
pe
r
s
pe
c
ti
ve
of
a
va
ni
ll
a
G
A
N
(
i.
e
.,
a
n
unc
ondi
ti
one
d
G
A
N
)
,
th
e
r
e
is
no
c
ont
r
ol
ove
r
da
ta
m
ode
s
dur
in
g t
he
i
npa
in
ti
ng p
r
oc
e
s
s
. I
nput
ti
ng poll
ut
io
n
i
m
a
ge
s
w
it
hout
pi
xe
l
-
le
ve
l
m
e
ta
-
da
ta
w
il
l
r
e
s
ul
t
in
a
m
ode
l
th
a
t
m
im
ic
s
a
ge
ne
r
a
l
-
pur
pos
e
in
te
r
pol
a
to
r
.
H
ow
e
v
e
r
,
by
c
ondi
ti
oni
ng
th
e
out
put
ove
r
it
s
r
e
gi
on'
s
vi
s
ua
l
im
a
ge
r
y,
th
e
m
ode
l
c
a
n
pr
oduc
e
a
c
c
ur
a
te
in
p
a
in
ti
ngs
by
f
in
di
ng
c
or
r
e
la
ti
ons
be
twe
e
n
pr
io
r
s
a
nd
pol
lu
ti
on pa
tc
he
s
.
A
s
a
r
e
s
ul
t,
th
e
ge
n
e
r
a
ti
ve
a
dve
r
s
a
r
ia
l
ne
twor
k
i
s
e
xt
e
nde
d
w
it
h
a
c
ondi
ti
ona
l
la
ye
r
to
e
nc
ode
th
e
s
ta
ti
c
pr
io
r
s
.
T
he
ge
n
e
r
a
to
r
a
nd
di
s
c
r
im
in
a
to
r
a
r
e
pr
ovi
de
d
w
it
h
hi
gh
-
r
e
s
ol
ut
io
n
e
nc
ode
d
im
a
g
e
r
y
(
pr
io
r
s
:
p
)
.
T
he
obj
e
c
ti
ve
f
unc
ti
on of
t
he
t
w
o
-
pl
a
ye
r
m
in
im
a
x ga
m
e
i
s
upda
te
d a
s
(
4)
.
m
in
m
a
x
(
,
)
=
∼
(
)
[
(
(
|
)
)
]
+
∼
(
)
[
(
1
−
(
(
|
)
|
)
)
]
(
4)
I
n
a
c
ondi
ti
ona
l
ge
ne
r
a
ti
ve
a
dve
r
s
a
r
ia
l
s
e
tu
p,
th
e
s
a
m
e
c
ondi
ti
on
is
pr
ovi
de
d
to
bot
h
th
e
g
e
ne
r
a
to
r
a
nd
di
s
c
r
im
in
a
to
r
ne
twor
ks
.
P
r
io
r
s
a
r
e
pur
pos
e
ly
us
e
d
a
s
c
ondi
ti
ons
to
he
lp
th
e
ge
ne
r
a
to
r
e
nha
nc
e
it
s
c
om
pl
e
ti
ons
.
F
or
e
xa
m
pl
e
,
one
c
a
n
im
a
gi
ne
how
us
e
f
ul
ve
hi
c
l
e
tr
a
f
f
ic
de
ns
it
y
im
a
ge
s
w
oul
d
be
f
or
a
m
ode
l
th
a
t
pr
e
di
c
ts
ne
a
r
-
r
e
a
l
-
ti
m
e
N
O
2
c
onc
e
nt
r
a
ti
ons
.
I
n
th
is
c
a
s
e
,
R
G
B
im
a
ge
s
pr
ovi
de
lo
w
-
le
ve
l
in
f
or
m
a
ti
on
a
bout
ur
ba
n
a
nd
gr
e
e
nne
s
s
de
ns
it
ie
s
.
T
hi
s
s
tu
dy
a
r
gue
s
th
a
t
a
vi
s
ua
l
pr
io
r
c
oul
d
be
us
e
f
ul
to
th
e
ta
s
k
o
f
pr
e
di
c
ti
ng
N
O
2
de
ns
it
ie
s
ove
r
l
a
r
ge
r
e
gi
ons
of
i
nt
e
r
e
s
t.
3
.
3
.
C
om
p
le
t
io
n
n
e
t
w
or
k
T
he
c
om
pl
e
ti
on
ne
twor
k
ta
ke
s
lo
w
-
r
e
s
ol
ut
io
n
N
O
2
im
a
ge
s
th
a
t
c
ont
a
in
th
e
ga
ps
to
be
s
f
il
le
d,
a
nd
a
m
a
s
k
c
ha
nne
l
th
a
t
in
di
c
a
te
s
w
hi
c
h
pi
xe
ls
a
r
e
m
is
s
in
g.
E
a
c
h
da
m
a
ge
d
pa
tc
h
ha
s
it
s
c
or
r
e
s
ponding
hi
gh
-
r
e
s
ol
ut
io
n
R
G
B
im
a
ge
th
a
t
c
ove
r
s
th
e
s
a
m
e
r
e
gi
on
a
nd
pr
ovi
de
s
ga
p
-
f
r
e
e
vi
s
ua
l
in
f
or
m
a
ti
on.
T
w
o
ne
twor
ks
w
e
r
e
tr
a
in
e
d.
T
he
r
e
duc
e
r
a
c
ts
a
s
a
dow
n
-
s
a
m
pl
e
r
th
a
t
in
te
ll
ig
e
nt
ly
r
e
s
iz
e
s
th
e
hi
gh
-
r
e
s
ol
ut
io
n
R
G
B
im
a
ge
(
th
e
pr
io
r
)
to
th
e
s
a
m
e
s
iz
e
a
s
th
e
da
m
a
ge
d
im
a
ge
.
T
a
bl
e
2
pr
e
s
e
nt
s
it
s
la
ye
r
s
in
s
uc
c
e
s
s
iv
e
or
de
r
.
T
he
c
om
pl
e
to
r
ne
twor
k
is
a
f
u
ll
y
c
onvolut
io
na
l
ne
tw
or
k
(
F
C
N
)
th
a
t
a
c
ts
a
s
th
e
m
a
in
in
pa
in
te
r
.
I
t
is
opt
im
iz
e
d
t
o
f
il
l
th
e
m
is
s
in
g
ga
ps
in
th
e
in
put
im
a
ge
.
T
a
bl
e
3
s
pe
c
if
ie
s
it
s
la
ye
r
s
.
T
he
a
c
ti
va
ti
ons
f
or
bot
h
ne
twor
ks
w
e
r
e
pa
s
s
e
d t
hr
ough ba
tc
h nor
m
a
li
z
a
ti
on
a
nd
R
e
L
U
a
f
te
r
e
a
c
h l
a
y
e
r
.
T
a
bl
e
2
.
R
e
duc
e
r
ne
twor
k l
a
ye
r
s
Nº
T
ype
K
e
r
ne
l
S
t
r
i
de
O
ut
put
1
C
onv.
5
1
16
2
C
onv.
3
2
32
3
C
onv.
3
1
64
4
C
onv.
3
2
32
5
C
onv.
3
1
1
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
Sat
e
ll
it
e
i
m
age
i
npai
nt
in
g w
it
h de
e
p ge
ne
r
at
iv
e
ad
v
e
r
s
ar
ia
l
ne
u
r
al
ne
tw
or
k
s
(
M
ohame
d A
k
r
am
Z
ay
ta
r
)
125
T
a
bl
e
3
.
C
om
pl
e
to
r
ne
twor
k l
a
ye
r
s
Nº
T
ype
K
e
r
ne
l
S
t
r
i
de
D
i
l
a
t
i
on
O
ut
put
1
C
onv.
5
1
1
16
2
C
onv.
3
2
1
32
3
C
onv.
3
1
1
64
4
C
onv.
3
2
1
64
5
D
i
c
onv.
3
1
2
64
6
D
i
c
onv.
3
1
3
64
7
C
onv.
3
1
1
64
8
4
2
1
32
9
C
onv.
3
1
1
32
10
4
2
1
16
11
C
onv.
3
1
1
8
12
C
onv.
3
1
1
1
T
he
c
om
pl
e
to
r
ne
twor
k
is
u
s
e
d
to
in
pa
in
t
th
e
m
i
s
s
in
g
r
e
gi
ons
in
th
e
in
put
.
O
n
th
e
ot
he
r
ha
nd,
th
e
r
e
duc
e
r
ne
twor
k
r
e
s
iz
e
s
th
e
pr
io
r
to
th
e
s
a
m
e
s
iz
e
a
s
th
e
i
nput
.
W
it
hout
it
,
th
e
m
ode
l
w
oul
d
le
a
r
n
th
e
unc
ondi
ti
one
d
pol
lu
ti
on
im
a
ge
di
s
tr
ib
ut
io
n,
w
hi
c
h
is
not
opt
i
m
a
l
f
or
lo
c
a
ti
on
-
va
r
ia
nt
pa
tt
e
r
ns
.
U
r
ba
n,
la
nd,
a
nd
ot
he
r
vi
s
ua
l
f
e
a
tu
r
e
s
s
e
r
ve
a
s
s
tr
ong
pr
io
r
s
f
or
th
e
c
om
pl
e
t
or
to
ge
ne
r
a
te
a
c
c
ur
a
t
e
pa
tc
he
s
. T
he
c
om
pl
e
to
r
ne
twor
k w
a
s
t
r
a
in
e
d us
in
g m
e
a
n
s
qua
r
e
d e
r
r
or
l
os
s
(
M
S
E
)
a
ve
r
a
ge
d ove
r
t
he
m
a
s
ke
d (
ga
p)
pi
xe
ls
.
3
.
4
.
D
is
c
r
im
in
at
or
n
e
t
w
or
k
T
he
di
s
c
r
im
in
a
to
r
ne
twor
k
is
tr
a
in
e
d
to
de
te
c
t
c
om
pl
e
te
d
N
O
2
im
a
ge
s
.
A
R
e
s
N
e
t
-
18
[
30]
a
r
c
hi
te
c
tu
r
e
s
how
n
in
F
ig
ur
e
1
i
s
us
e
d
to
e
xt
r
a
c
t
th
e
f
e
a
tu
r
e
ve
c
to
r
w
hi
c
h
is
m
a
ppe
d
to
th
e
pr
oba
bi
li
ty
of
th
e
in
put
be
in
g
c
om
pl
e
te
d
(
f
a
ke
)
or
r
e
a
l.
R
e
duc
e
r
ne
twor
k
w
e
ig
ht
s
a
r
e
f
r
oz
e
n
w
hi
le
opt
im
iz
in
g
th
e
di
s
c
r
im
in
a
to
r
s
in
c
e
t
he
r
e
duc
e
r
i
s
opt
im
iz
e
d f
or
e
f
f
ic
ie
nt
i
npa
in
ti
ng, not d
is
c
r
im
in
a
ti
on. T
he
pr
im
a
r
y r
ol
e
o
f
R
G
B
i
m
a
ge
s
i
n
th
e
c
ont
e
xt
of
th
e
di
s
c
r
im
in
a
to
r
is
to
pr
ovi
de
a
us
e
f
ul
pr
io
r
th
a
t
is
in
de
pe
nde
nt
of
w
he
th
e
r
th
e
in
put
im
a
ge
is
r
e
a
l
or
not
. H
e
nc
e
, t
he
di
s
c
r
im
in
a
to
r
i
s
opt
im
iz
e
d t
o e
s
ti
m
a
te
P
(
in
p
ut
=
c
o
m
p
l
e
t
e
d
|
p
)
.
3
.
5
.
T
r
ai
n
in
g
T
he
c
om
pl
e
to
r
n
e
twor
k
is
d
e
not
e
d:
C
(
x
,
p
)
.
x
r
e
pr
e
s
e
nt
s
a
ba
tc
h
of
N
O
2
im
a
ge
s
w
it
h
m
a
s
k
s
M
c
om
pr
is
e
d
of
0
s
a
nd
1
s
,
w
it
h
1
s
r
e
pr
e
s
e
nt
in
g
th
e
pi
xe
ls
th
a
t
a
r
e
m
i
s
s
in
g
in
x
.
p
a
r
e
th
e
p
r
io
r
s
f
o
r
e
a
c
h
im
a
g
e
in
x
.
T
he
y
c
ons
is
t
of
m
a
ny
R
G
B
im
a
ge
s
f
or
th
e
c
or
r
e
s
ponding
R
O
I
s
.
S
im
il
a
r
ly
,
D
(
x
̅
,
p
)
de
s
ig
na
te
s
th
e
di
s
c
r
im
in
a
to
r
ne
twor
k,
x
̅
r
e
pr
e
s
e
nt
s
th
e
pol
lu
ti
on
im
a
ge
s
(
r
e
a
l
or
c
om
pl
e
te
d)
,
a
nd
p
th
e
e
nc
ode
d
p
r
io
r
s
ov
e
r
x
̅
'
s
r
e
gi
ons
of
i
nt
e
r
e
s
ts
(
R
O
I
s
)
.
M
e
a
n
s
qua
r
e
d
e
r
r
or
lo
s
s
(
ℒ
2
)
is
a
n
in
pa
in
ti
ng
lo
s
s
c
hoi
c
e
th
a
t
r
e
s
ul
ts
in
bl
ur
r
e
d
e
s
ti
m
a
ti
ons
ove
r
th
e
ga
ps
. I
t
a
ve
r
a
ge
s
t
he
s
qua
r
e
d di
f
f
e
r
e
nc
e
s
b
e
twe
e
n ga
p pi
x
e
l
pr
e
di
c
ti
ons
a
nd t
a
r
ge
ts
a
s
(
5)
.
ℒ
2
(
,
̂
)
=
|
|
⊙
(
(
̅
,
)
−
)
)
|
|
2
(
5)
O
n t
he
ot
he
r
ha
nd, a
dve
r
s
a
r
ia
l
lo
s
s
c
a
n be
f
or
m
ul
a
te
d a
s
(
6)
.
m
in
m
a
x
[
(
(
|
)
)
+
(
1
−
(
(
̅
|
)
|
)
)
]
(
6)
C
is
th
e
c
om
pl
e
to
r
ne
twor
k,
D
is
th
e
di
s
c
r
im
in
a
to
r
,
x
is
th
e
in
put
,
x
̅
is
th
e
da
m
a
ge
d/
he
a
lt
hy
in
put
,
a
nd
p
th
e
pr
io
r
.
ℒ
2
a
nd a
d
ve
r
s
a
r
ia
l
lo
s
s
e
s
a
r
e
c
om
bi
ne
d t
o f
or
m
a
li
z
e
t
he
ge
n
e
r
a
l
opt
im
iz
a
ti
on pr
obl
e
m
a
s
(
7)
.
m
in
m
a
x
[
ℒ
2
(
,
,
)
+
(
(
|
)
)
+
(
1
−
(
(
|
)
|
)
)
]
(
7)
G
A
N
s
a
r
e
c
ha
ll
e
ngi
ng
to
tr
a
in
due
to
th
e
in
s
ta
bi
li
ty
be
twe
e
n
th
e
ge
ne
r
a
to
r
a
nd
di
s
c
r
im
in
a
to
r
ne
twor
ks
in
th
e
e
a
r
ly
tr
a
in
in
g
pha
s
e
.
F
or
th
is
r
e
a
s
on,
th
e
tr
a
in
in
g
lo
op
is
ba
la
nc
e
d
a
s
de
s
c
r
ib
e
d
in
A
lg
or
it
hm
1.
T
he
m
e
th
od
pr
opos
e
d
in
[
26]
is
c
ho
s
e
n
a
s
th
e
ne
ur
a
l
b
a
s
e
li
n
e
.
I
ts
m
ode
l
w
a
s
tr
a
in
e
d
to
pr
oduc
e
vi
s
ua
ll
y
a
ppe
a
li
ng
c
om
pl
e
ti
ons
f
or
a
va
r
ie
ty
of
na
tu
r
a
l
s
c
e
n
e
i
m
a
ge
s
.
I
t
s
e
r
ve
s
a
s
a
good
be
nc
hm
a
r
k
b
e
c
a
us
e
th
e
pr
opos
e
d a
r
c
hi
te
c
tu
r
e
i
s
a
n
e
xt
e
ns
io
n of
t
he
ba
s
e
li
ne
'
s
m
odu
la
r
de
s
ig
n.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
,
V
ol
.
10
, N
o.
1
,
M
a
r
c
h
20
2
1
:
121
–
130
126
A
lg
or
it
h
m
1
:
T
C
&
T
D
s
e
r
ve
t
o pr
e
-
tr
a
in
e
a
c
h ne
twor
k s
e
pa
r
a
te
ly
b
e
f
or
e
c
onduc
ti
ng a
dve
r
s
a
r
ia
l
tr
a
in
in
g
Data
:
,
,
,
H
y
p
e
r
p
a
r
a
m
e
t
e
r
s
:
=
10
,
0
0
0
,
=
100
,
=
100
,
η
=
0
.
0
0
1
,
η
′
=
0
.
01
Begin
F
o
r
∈
[
0
,
]
do
S
a
m
p
l
e
m
i
n
i
b
a
t
c
h
⊂
G
e
n
e
r
a
t
e
c
o
m
p
l
e
t
o
r
m
a
s
k
s
for
∀
∈
G
e
n
e
r
a
t
e
d
i
s
c
r
i
m
i
n
a
t
o
r
m
a
s
k
s
for
∀
∈
If
<
−
then
⟵
−
η
∇
2
(
,
,
)
E
l
s
e
i
f
<
−
then
⟵
−
η
′
∇
(
̅
,
,
)
E
l
s
e
U
p
d
a
t
e
w
i
t
h
j
o
i
n
t
l
o
s
s
U
p
d
a
t
e
w
i
t
h
b
i
n
a
r
y
c
r
o
s
s
e
n
t
r
o
p
y
l
o
s
s
3
.
6
.
D
at
a
T
he
e
ur
ope
a
n
or
ga
ni
z
a
ti
on
of
th
e
e
xpl
oi
ta
ti
on
of
m
e
te
or
ol
o
gi
c
a
l
s
a
te
ll
it
e
s
(
E
U
M
E
T
S
A
T
)
is
a
n
in
te
r
na
ti
ona
l
s
a
te
ll
it
e
a
g
e
nc
y
r
e
s
pon
s
ib
le
f
or
a
c
qui
r
in
g,
pr
e
p
r
oc
e
s
s
in
g,
a
nd
di
s
tr
ib
ut
in
g
r
e
li
a
bl
e
w
e
a
th
e
r
,
c
li
m
a
te
,
a
nd
e
nvi
r
onm
e
nt
a
l
d
a
ta
.
I
ts
lo
w
-
or
bi
ti
ng
s
a
te
ll
it
e
,
M
e
t
O
p,
c
ont
in
uous
ly
de
li
ve
r
s
c
r
it
ic
a
l
c
li
m
a
t
e
da
ta
.
E
U
M
E
T
S
A
T
a
ls
o
di
s
tr
ib
ut
e
s
da
ta
f
r
om
ot
he
r
pa
r
tn
e
r
s
s
uc
h
a
s
th
e
na
ti
ona
l
oc
e
a
nogr
a
phi
c
a
nd
a
tm
os
phe
r
ic
a
dm
in
is
tr
a
ti
on (
N
O
A
A
)
.
O
f
f
li
ne
E
U
M
E
T
S
A
T
'
s
da
ta
pr
oduc
ts
a
r
e
f
r
e
e
a
nd
a
va
il
a
bl
e
f
or
r
e
s
e
a
r
c
h pur
pos
e
s
.
M
e
tOp
is
a
s
e
r
ie
s
of
3
pol
a
r
-
or
bi
ti
ng
m
e
te
or
ol
ogi
c
a
l
s
a
te
ll
it
e
s
de
ve
lo
pe
d
by
th
e
e
ur
ope
a
n
s
pa
c
e
a
ge
nc
y
(
E
S
A
)
a
nd
ope
r
a
te
d
by
E
U
M
E
T
S
A
T
.
M
e
tOp
ta
ke
s
90
m
in
ut
e
s
to
or
bi
t
th
e
e
a
r
th
,
to
ta
li
ng
14
ti
m
e
s
a
da
y.
H
a
vi
ng
th
r
e
e
s
a
t
e
ll
it
e
s
e
nha
n
c
e
s
th
e
t
e
m
por
a
l
r
e
s
ol
ut
io
n
of
M
e
tOp
a
s
a
da
ta
pr
ovi
de
r
.
M
e
tOp
c
a
r
r
i
e
s
a
pa
yl
oa
d
of
11
s
c
ie
nt
if
ic
in
s
tr
um
e
nt
s
.
A
f
te
r
tr
a
ns
f
e
r
r
in
g
th
e
da
t
a
,
it
ge
ts
pr
e
pr
oc
e
s
s
e
d
in
to
m
ul
ti
pl
e
le
ve
ls
a
nd
f
e
d
in
to
num
e
r
ic
a
l
s
im
ul
a
to
r
s
f
or
w
e
a
th
e
r
f
or
e
c
a
s
ti
ng
a
nd
e
nv
ir
onm
e
nt
a
l
m
oni
to
r
in
g.
M
a
ny
of
it
s
a
va
il
a
bl
e
da
ta
pr
oduc
ts
pr
ovi
de
ve
r
ti
c
a
l
d
e
ns
it
y m
e
a
s
ur
e
m
e
nt
s
.
G
a
s
tr
a
c
e
s
w
e
r
e
a
c
qui
r
e
d
f
r
om
th
e
gl
oba
l
o
z
one
m
oni
to
r
in
g
e
x
pe
r
im
e
nt
-
2
(
G
O
M
E
-
2)
in
s
tr
um
e
nt
,
a
s
c
a
nni
ng
s
p
e
c
tr
om
e
te
r
th
a
t
pr
ovi
de
s
gl
oba
l
m
oni
to
r
in
g
c
ove
r
a
ge
.
T
he
n
e
a
r
r
e
a
l
-
ti
m
e
to
ta
l
c
ol
um
n
(
N
T
O
)
pr
oduc
t
pr
ovi
de
s
c
onc
e
nt
r
a
ti
on
m
e
a
s
ur
e
m
e
nt
s
f
or
f
our
ty
pe
s
o
f
a
tm
os
phe
r
ic
tr
a
c
e
ga
s
e
s
:
O
3
,
N
O
2
,
S
O
2
,
a
nd
H
C
H
O
.
T
he
pr
oduc
t
is
ope
r
a
ti
ona
l
s
in
c
e
01/
12/
2007,
ha
s
a
s
pe
c
t
r
a
l
r
e
s
ol
ut
io
n
of
0
.
26
−
0
.
51
nm
,
a
nd
pr
ovi
de
s
gl
oba
l
ge
ogr
a
phi
c
c
ove
r
a
ge
w
it
h a
s
pa
ti
a
l
r
e
s
ol
ut
io
n o
f
40
×
4
0
km
(
M
e
tOp
-
A
)
.
O
n
th
e
o
th
e
r
ha
nd,
M
e
te
os
a
t
is
a
s
e
r
ie
s
of
ge
os
ta
ti
ona
r
y
m
e
te
or
ol
ogi
c
a
l
s
a
te
ll
it
e
s
ope
r
a
te
d
by
E
U
M
E
T
S
A
T
.
M
e
t
e
os
a
t
s
e
c
ond
ge
n
e
r
a
ti
on
(
M
S
G
)
pr
ovi
de
s
im
a
ge
s
of
th
e
f
ul
l
e
a
r
th
di
s
c
a
nd
da
ta
f
or
w
e
a
th
e
r
f
or
e
c
a
s
ts
.
I
t
ha
s
a
te
m
por
a
l
r
e
s
ol
ut
io
n
of
15
m
in
ut
e
s
.
T
he
s
pi
n
ni
n
g
e
nha
nc
e
d
vi
s
ib
le
a
nd
in
f
r
a
r
e
d
im
a
ge
r
(
S
E
V
I
R
I
)
in
s
tr
um
e
nt
c
a
pt
ur
e
s
th
e
tr
ue
-
c
ol
or
im
a
ge
s
in
a
s
pa
ti
a
l
r
e
s
ol
ut
io
n
of
1
×
1
km
.
T
he
pr
io
r
'
s
(
p
)
im
a
ge
r
y i
s
c
ol
le
c
te
d a
nd pr
e
pr
oc
e
s
s
e
d f
r
om
M
S
G
'
s
S
E
V
I
R
I
i
ns
t
r
um
e
nt
.
M
S
G
c
ove
r
s
th
e
R
O
I
of
M
or
oc
c
o.
T
he
ti
le
s
w
e
r
e
c
l
ip
pe
d
us
i
ng
th
e
r
e
gi
on
of
in
te
r
e
s
t
a
nd
m
e
r
ge
d
th
r
ough pixe
l
-
a
ve
r
a
gi
ng t
o s
to
r
e
a
s
in
gl
e
(
m
os
a
ic
)
hi
gh
-
qua
li
ty
i
m
a
ge
ove
r
t
he
R
O
I
.
I
n
th
is
s
tu
dy,
M
or
oc
c
o
w
a
s
c
hos
e
n
a
s
a
r
e
gi
on
of
in
te
r
e
s
t.
A
ll
im
a
ge
s
w
e
r
e
f
il
te
r
e
d
to
be
in
th
e
bounding
box
[
(
−
5
.
39
,
35
.
54
)
,
(
−
5
.
3
4
,
35
.
54
)
,
(
−
5
.
34
,
35
.
59
)
,
(
−
5
.
39
,
35
.
59
)
,
(
−
5
.
39
,
35
.
54
)
]
in
(
la
ti
tu
de
,
lo
ngi
tu
de
)
c
oor
di
na
te
s
.
A
ll
S
E
V
I
R
I
ti
le
s
w
e
r
e
ta
ke
n
f
r
om
06/
2017
to
01/
2018.
F
or
pol
lu
ti
on
im
a
ge
s
,
T
il
e
s
w
e
r
e
a
c
qui
r
e
d
a
nd
f
il
te
r
e
d
f
or
th
e
s
a
m
e
R
O
I
th
a
t
r
a
nge
f
r
om
03/
2018
to
09/
2018.
T
he
pr
io
r
s
w
e
r
e
s
a
m
pl
e
d
f
r
om
a
pr
e
vi
ous
ti
m
e
f
r
a
m
e
be
c
a
us
e
th
e
y
w
il
l
be
us
e
d
to
in
pa
in
t
f
ut
ur
e
pa
tc
he
s
.
S
ynt
he
ti
c
m
a
s
k
ge
ne
r
a
ti
on
is
e
xpl
a
in
e
d i
n t
he
"
R
E
S
U
L
T
S
A
N
D
D
I
S
C
U
S
S
I
O
N
"
s
e
c
ti
on.
T
is
de
not
e
d
a
s
th
e
s
e
t
of
a
c
qui
r
e
d
ti
le
s
f
or
th
e
r
e
gi
on
of
in
te
r
e
s
t.
f
or
e
a
c
h
t
i
∈
T
,
i
t
is
pr
oc
e
s
s
e
d
i
s
by
as
:
−
t
i
is
pr
oj
e
c
te
d i
nt
o a
pr
e
-
de
f
in
e
d s
ta
ti
c
s
pa
ti
a
l
gr
id
t
o nor
m
a
li
z
e
p
ix
e
l
pos
it
io
ns
.
−
is
s
pl
it
i
nt
o
64
×
64
pa
tc
he
s
, de
not
e
d
.
−
F
or
e
a
c
h i
m
a
ge
x
ij
∈
x
i
, i
ts
256
×
256
R
G
B
pr
io
r
(
)
i
s
c
ol
le
c
te
d. I
t
c
ove
r
s
t
he
s
a
m
e
r
e
gi
on a
s
.
−
{
,
,
}
a
r
e
s
pl
it
in
to
two
s
e
ts
.
T
he
f
ir
s
t
c
ont
a
in
s
he
a
lt
hy
pol
lu
ti
o
n
pa
tc
he
s
(
a
ll
pi
xe
l
va
lu
e
s
a
r
e
a
va
il
a
bl
e
)
. T
he
s
e
c
ond s
e
t
hol
ds
t
he
da
m
a
g
e
d pa
tc
he
s
.
−
P
ix
e
l
va
lu
e
s
w
e
r
e
nor
m
a
li
z
e
d a
nd
s
ta
nda
r
di
z
e
d
be
tw
e
e
n
[
0
,
1
]
f
or
bot
h i
nput
pa
tc
he
s
a
nd pr
io
r
s
.
T
he
f
in
a
l
da
ta
s
e
t
w
a
s
c
li
ppe
d
a
t
1
M
N
O
2
pa
tc
he
s
.
90%
of
th
e
da
ta
s
e
t
w
a
s
us
e
d
to
tr
a
in
th
e
m
ode
l
a
nd
th
e
r
e
m
a
in
in
g
100
,
000
im
a
ge
s
f
or
te
s
ti
ng.
A
ti
m
e
-
s
e
r
ie
s
tr
a
in
/t
e
s
t
s
pl
it
w
a
s
c
onduc
te
d
w
he
r
e
a
ll
of
th
e
10%
te
s
t
s
a
m
pl
e
s
c
a
m
e
f
r
om
a
f
ut
ur
e
t
im
e
-
in
te
r
va
l.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
Sat
e
ll
it
e
i
m
age
i
npai
nt
in
g w
it
h de
e
p ge
ne
r
at
iv
e
ad
v
e
r
s
ar
ia
l
ne
u
r
al
ne
tw
or
k
s
(
M
ohame
d A
k
r
am
Z
ay
ta
r
)
127
4.
R
E
S
U
L
T
S
A
N
D
D
I
S
C
U
S
S
I
O
N
4
.
1
.
N
oi
s
e
ge
n
e
r
at
io
n
A
s
oppos
e
d
to
th
e
ta
s
k
of
na
tu
r
a
l
im
a
ge
in
pa
in
ti
ng,
w
he
r
e
no
t
m
uc
h
c
ons
id
e
r
a
ti
on
is
gi
ve
n
to
th
e
s
ha
pe
of
th
e
ga
p
m
a
s
k,
th
e
ge
om
e
tr
y
of
m
is
s
in
g
r
e
gi
ons
in
s
a
te
ll
it
e
im
a
ge
r
y
f
ol
lo
w
s
s
tr
ong
pa
tt
e
r
ns
(
c
lo
uds
,
hol
e
s
,
a
nd
li
ne
s
)
.
T
h
e
m
a
s
k
g
e
ne
r
a
to
r
s
houl
d
r
e
pr
oduc
e
th
e
s
e
p
a
tt
e
r
ns
in
tr
a
in
in
g
ti
m
e
.
A
s
a
r
e
s
ul
t,
a
s
e
pa
r
a
te
G
A
N
G
n
w
a
s
tr
a
in
e
d
us
in
g
th
e
is
ol
a
te
d
noi
s
e
d
pa
tc
he
s
to
le
a
r
n
t
he
na
tu
r
a
l
noi
s
e
di
s
tr
ib
ut
io
n.
G
n
s
e
r
ve
s
a
s
a
noi
s
e
m
a
s
k
ge
ne
r
a
to
r
th
a
t
is
us
e
d
to
s
a
m
pl
e
a
r
ti
f
ic
ia
l
ga
ps
dur
in
g
tr
a
in
in
g.
F
or
e
ve
r
y
he
a
lt
hy
pa
tc
h
x
i
∈
X
,
a
64
×
64
ga
p
m
a
s
k
is
s
a
m
pl
e
d
f
r
om
G
n
,
a
c
opy
of
x
i
is
pur
pos
e
ly
da
m
a
g
e
d
us
in
g
th
e
m
a
s
k.
T
he
m
a
s
k
is
s
ta
c
ke
d
on
to
p
of
th
e
d
a
m
a
ge
d
im
a
g
e
;
th
e
f
in
a
l
2
-
c
h
a
nne
l
im
a
ge
r
e
pr
e
s
e
nt
s
th
e
in
put
.
T
he
or
ig
in
a
l
im
a
ge
x
i
be
c
om
e
s
t
he
t
a
r
ge
t
u
s
e
d i
n l
os
s
c
a
lc
ul
a
ti
on.
4
.
2
.
R
e
s
u
lt
s
T
he
pr
opos
e
d
m
ode
l
w
a
s
be
n
c
hm
a
r
ke
d
a
ga
i
ns
t
two
a
lg
or
it
hm
s
,
P
a
tc
hM
a
tc
h
[
29]
,
w
hi
c
h
r
e
pr
e
s
e
nt
s
th
e
c
la
s
s
ic
a
l
s
ta
te
-
of
-
th
e
-
a
r
t
m
e
th
od,
a
nd
th
e
L
oc
a
lG
lo
ba
l
[
26]
ne
ur
a
l
in
pa
in
te
r
.
F
or
tr
a
in
in
g,
in
put
im
a
ge
s
of
s
iz
e
64
×
64
pi
xe
ls
a
nd pr
io
r
i
m
a
ge
s
of
s
iz
e
256
×
256
pi
xe
ls
w
e
r
e
us
e
d a
s
i
nput
t
o t
he
m
ode
l.
T
he
t
r
a
in
in
g
a
nd t
e
s
ti
ng da
ta
s
e
ts
w
e
r
e
e
xt
r
a
c
te
d us
in
g a
t
im
e
-
s
e
r
ie
s
s
pl
it
.
A
s
oppos
e
d
to
na
tu
r
a
l
s
c
e
ne
c
om
pl
e
ti
on,
w
he
r
e
M
A
E
a
nd
M
S
E
a
r
e
not
c
ons
id
e
r
e
d
good
m
e
tr
ic
s
be
c
a
us
e
m
a
ny
c
om
pl
e
ti
ons
c
a
n
be
c
onc
e
pt
ua
ll
y
pos
s
ib
le
,
in
th
e
c
a
s
e
of
pol
lu
ti
on
im
a
ge
s
,
s
e
ns
or
m
e
a
s
ur
e
m
e
nt
s
a
r
e
uni
que
ta
r
ge
ts
.
H
e
nc
e
,
th
e
m
ode
l
w
a
s
e
va
lu
a
te
d
us
in
g
two
r
e
gr
e
s
s
io
n
m
e
tr
ic
s
;
le
a
s
t
a
bs
ol
ut
e
de
vi
a
ti
ons
l
o
s
s
a
s
(
8)
.
ℒ
1
=
1
∑
|
⊙
(
(
̅
,
)
−
)
|
=
1
(
8)
ℒ
1
a
ve
r
a
ge
s
th
e
a
bs
ol
ut
e
di
f
f
e
r
e
nc
e
s
be
twe
e
n
th
e
ta
r
ge
t
a
nd
pr
e
di
c
te
d
ga
p
pi
xe
ls
.
M
e
a
n
s
qua
r
e
d
e
r
r
or
(
ℒ
2
)
is
a
ls
o
r
e
por
te
d.
B
ot
h
m
e
tr
ic
s
m
e
a
s
ur
e
e
r
r
or
a
s
th
e
di
s
ta
nc
e
be
tw
e
e
n
th
e
m
ode
l'
s
pr
e
di
c
ti
ons
a
nd
th
e
gr
ound
-
tr
ut
h
nor
m
a
li
z
e
d
ta
r
ge
ts
.
T
he
m
ode
l
a
c
hi
e
ve
s
a
le
a
s
t
a
b
s
ol
ut
e
de
vi
a
ti
ons
e
r
r
or
of
ℒ
1
=
0
.
33
(
21%
a
nd
60%
de
c
r
e
a
s
e
in
e
r
r
or
w
it
h
r
e
s
pe
c
t
to
th
e
two
ba
s
e
li
ne
s
)
a
nd
a
n
M
S
E
of
ℒ
2
=
0
.
15
(
29%
a
nd
73%
de
c
r
e
a
s
e
in
e
r
r
or
,
r
e
s
pe
c
ti
ve
ly
)
ove
r
th
e
te
s
t
s
e
t
.
T
he
s
e
r
e
s
ul
t
s
r
e
pr
e
s
e
nt
a
s
ig
ni
f
ic
a
nt
pe
r
f
or
m
a
nc
e
in
c
r
e
a
s
e
ov
e
r
th
e
P
a
tc
hM
a
tc
h
a
nd
L
oc
a
lG
lo
ba
l
in
pa
in
te
r
s
.
W
he
n
c
onve
r
te
d
to
D
ob
s
on
U
ni
ts
(
D
U
)
,
w
e
ge
t
ℒ
1
=
0
.
56
DU
a
nd
ℒ
2
=
0
.
41
DU
.
1
c
or
r
e
s
ponds
to
a
c
ol
um
n
de
ns
it
y
of
2
.
8
×
10
16
−
2
,
th
e
r
a
w
m
e
a
s
ur
e
m
e
nt
s
r
a
nge
be
twe
e
n
0
a
nd 2.5 DU
.
4
.
3
.
D
is
c
u
s
s
io
n
T
he
m
a
in
c
ont
r
ib
ut
or
to
th
e
in
c
r
e
a
s
e
in
pe
r
f
or
m
a
nc
e
i
s
th
e
c
on
di
ti
ona
l
la
ye
r
.
A
n
a
bl
a
ti
on
s
tu
dy
w
a
s
c
onduc
te
d
by
r
e
m
ovi
ng
th
e
c
ondi
ti
ona
l
la
ye
r
a
nd
tr
a
in
in
g
th
e
G
A
N
w
it
hout
vi
s
ua
l
pr
io
r
s
.
T
he
r
e
s
ul
ti
ng
ℒ
1
a
nd
ℒ
2
s
c
or
e
s
w
e
r
e
s
li
ght
ly
w
or
s
e
t
ha
n
L
oc
a
lG
lo
ba
l
[
26]
, l
a
s
t
c
ol
um
n of
T
a
bl
e
4
.
T
a
bl
e
4
.
B
e
nc
hm
a
r
ks
of
m
e
a
n
ℒ
1
a
nd me
a
n s
qua
r
e
d l
o
s
s
on t
he
t
e
s
t
s
e
t.
M
e
t
r
i
c
s
\
M
e
t
hod
s
P
a
t
c
hM
a
t
c
h [
29]
L
oc
a
l
G
l
oba
l
[
26]
O
ur
s
O
ur
s
(
no pr
i
or
s
)
ℒ
1
0.83
0.42
0.33
0.44
ℒ
2
0.55
0.27
0.15
0.28
F
ig
ur
e
2
s
how
c
a
s
e
s
th
e
e
f
f
e
c
t
s
th
a
t
vi
s
ua
l
pr
io
r
s
c
a
n
ha
ve
on
pol
lu
ti
on
pr
e
di
c
ti
ons
.
T
he
m
ode
l
pr
e
di
c
ts
hi
ghe
r
N
O
2
c
onc
e
nt
r
a
ti
ons
in
th
e
c
it
y
of
C
a
s
a
bl
a
nc
a
w
it
hout
pr
oc
e
s
s
in
g
hi
gh
-
de
ns
it
y
ne
ig
hbor
in
g
pi
xe
ls
,
to
p
r
ow
in
F
ig
ur
e
2.
I
t
pr
e
di
c
te
d
hi
gh
N
O
2
c
onc
e
nt
r
a
ti
ons
by
r
e
ly
in
g
s
ol
e
ly
on
th
e
R
G
B
pr
io
r
.
S
uc
h
pa
tt
e
r
ns
a
r
e
not
ic
e
a
bl
e
in
ot
he
r
ur
ba
n
a
nd
in
dus
tr
ia
l
c
it
ie
s
in
M
or
oc
c
o.
H
o
w
e
ve
r
,
a
t
th
e
bot
to
m
r
ow
,
th
e
m
ode
l
f
a
il
e
d
to
pr
e
di
c
t
a
hi
gh
pol
lu
ti
on
c
onc
e
nt
r
a
ti
on
ove
r
th
e
r
e
gi
on
of
T
a
our
ir
t.
T
h
a
t
c
oul
d
h
a
ve
b
e
e
n
th
e
r
e
s
ul
t
of
th
e
te
m
po
r
a
l
na
tu
r
e
o
f
in
dus
tr
ia
l
a
c
ti
vi
ti
e
s
a
nd
th
e
no
is
e
th
a
t
is
in
he
r
e
nt
to
s
e
ns
or
y
m
e
a
s
ur
e
m
e
nt
s
.
T
he
a
ve
r
a
ge
in
c
r
e
a
s
e
in
pe
r
f
or
m
a
nc
e
in
di
c
a
te
s
th
a
t
th
e
m
ode
l
le
a
r
ne
d
to
a
s
s
oc
ia
te
c
e
r
ta
in
vi
s
ua
l
f
e
a
tu
r
e
s
to
hi
gh/
lo
w
pol
lu
ti
on de
ns
it
ie
s
. T
hi
s
s
tu
dy s
how
c
a
s
e
s
t
he
be
ne
f
it
s
of
m
ul
ti
-
m
oda
l
le
a
r
ni
ng i
n c
om
put
e
r
vi
s
io
n.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
,
V
ol
.
10
, N
o.
1
,
M
a
r
c
h
20
2
1
:
121
–
130
128
F
ig
ur
e
2. S
a
m
pl
e
:
nor
m
a
li
z
e
d pr
e
di
c
ti
ons
ove
r
nor
th
e
r
n M
or
oc
c
o. F
r
om
l
e
f
t
to
r
ig
ht
:
pr
io
r
, i
nput
im
a
ge
,
ge
ne
r
a
te
d i
m
a
ge
, a
nd gr
ound
-
tr
ut
h i
m
a
ge
(
hi
ghe
r
va
lu
e
s
r
e
pr
e
s
e
nt
gr
e
a
te
r
c
onc
e
nt
r
a
ti
ons
)
D
e
s
pi
te
th
e
pr
io
r
not
pr
ovi
di
ng
popula
ti
on,
ti
m
e
-
de
pe
nde
nt
i
ndus
tr
y
a
c
ti
vi
ty
e
s
ti
m
a
te
s
,
or
tr
a
f
f
ic
in
f
or
m
a
ti
on,
it
in
c
r
e
a
s
e
d
th
e
ove
r
a
ll
pe
r
f
or
m
a
nc
e
by
s
im
pl
y
p
r
ovi
di
ng
hi
gh
-
qua
li
ty
vi
s
ua
l
in
f
or
m
a
ti
on.
T
he
r
e
duc
e
r
a
ls
o
pl
a
ye
d
a
c
r
it
ic
a
l
r
ol
e
in
tr
a
ns
f
or
m
in
g
th
e
pr
io
r
s
in
a
w
a
y
th
a
t
w
a
s
us
e
f
ul
to
in
pa
in
t
th
e
m
is
s
in
g
r
e
gi
ons
.
T
he
pr
opos
e
d
m
ode
l
c
a
n
a
ls
o
be
us
e
d
a
s
a
da
ta
e
n
a
h
nc
e
r
f
or
dow
ns
tr
e
a
m
ta
s
ks
by
r
e
m
vi
ng
c
lo
u
d
e
f
f
e
c
ts
.
D
ow
ns
tr
e
a
m
a
ppl
ic
a
ti
ons
in
s
c
lu
de
obj
e
c
t
de
te
c
ti
o
n
[
31]
,
l
a
ndc
ove
r
ge
ne
r
a
ti
on
[
32]
,
la
ndus
e
c
la
s
s
if
ic
a
ti
on [
33]
, c
ha
nge
de
te
c
ti
on, c
r
op moni
to
r
in
g, f
ie
ld
a
nd ur
ba
n m
a
ppi
ng a
m
ong othe
r
s
.
5.
C
O
N
C
L
U
S
I
O
N
T
hi
s
pa
pe
r
pr
opos
e
d
a
ne
ur
a
l
im
a
ge
in
pa
in
te
r
c
a
pa
bl
e
o
f
f
il
li
ng
m
e
a
s
ur
e
m
e
nt
ga
ps
us
in
g
ne
ig
hbor
in
g
pi
xe
l
va
lu
e
s
a
nd
s
ta
ti
c
pr
io
r
s
.
I
t
de
s
c
r
ib
e
d
th
e
pr
e
pr
oc
e
s
s
in
g
d
a
ta
pi
pe
li
ne
,
th
e
C
N
N
-
ba
s
e
d
s
ub
-
ne
twor
ks
,
a
nd
th
e
tr
a
in
in
g
pr
oc
e
s
s
.
T
h
e
ne
ur
a
l
s
y
s
te
m
w
a
s
s
u
c
c
e
s
s
f
ul
ly
tr
a
in
e
d
to
f
il
l
pol
lu
ti
on
ga
ps
in
th
e
r
e
gi
on
of
M
or
oc
c
o.
I
t
out
pe
r
f
or
m
e
d
two
pr
io
r
-
le
s
s
ba
s
e
li
ne
s
a
nd
s
how
e
d
th
e
pot
e
nt
ia
l
of
da
ta
f
u
s
io
n
f
or
s
a
te
ll
it
e
im
a
ge
in
pa
in
ti
ng.
T
he
de
s
c
r
ib
e
d
ne
ur
a
l
s
ys
te
m
c
a
n
be
de
pl
oye
d
w
it
hi
n
a
r
e
m
ot
e
s
e
ns
in
g
pi
pe
li
ne
to
f
il
l
in
c
om
in
g
s
a
te
ll
it
e
pa
tc
he
s
,
r
e
s
ul
ti
ng
in
gr
e
a
te
r
ne
a
r
-
r
e
a
l
-
ti
m
e
vi
s
ib
il
it
y
ove
r
w
e
a
th
e
r
,
a
tm
os
phe
r
ic
,
a
nd
c
li
m
a
te
c
ondi
ti
ons
.
T
he
s
ys
te
m
in
te
ll
ig
e
nt
ly
nul
li
f
ie
s
th
e
e
f
f
e
c
ts
of
s
e
ns
or
pe
r
tu
r
ba
ti
ons
a
nd
c
lo
ud
e
f
f
e
c
ts
.
H
ow
e
ve
r
,
one
li
m
it
a
ti
on
of
th
e
s
ys
te
m
is
th
a
t
te
m
por
a
l
r
e
s
ol
ut
io
n
e
nha
nc
e
m
e
nt
is
not
e
nough
to
of
f
e
r
a
c
om
pe
ti
ti
ve
a
lt
e
r
na
ti
ve
to
I
oT
-
ba
s
e
d
d
e
vi
c
e
s
f
or
s
m
a
ll
-
s
c
a
le
m
oni
to
r
in
g,
th
e
li
m
it
e
d
s
pa
ti
a
l
r
e
s
ol
ut
io
n
of
s
a
te
ll
it
e
s
e
ns
or
s
r
e
m
a
in
s
th
e
m
os
t
im
por
ta
nt
ope
n
c
ha
ll
e
nge
in
r
e
m
ot
e
s
e
n
s
in
g.
H
ow
e
v
e
r
,
th
e
m
ode
l
c
a
n
b
e
m
odi
f
ie
d
to
ta
c
k
le
s
upe
r
-
r
e
s
ol
ut
io
n
th
r
ough
pr
io
r
le
a
r
ni
ng.
S
uc
h
s
ys
te
m
c
a
n
e
nh
a
nc
e
th
e
s
a
t
e
ll
it
e
'
s
s
pa
ti
a
l
r
e
s
ol
ut
io
n by us
in
g l
ow
-
r
e
s
ol
ut
io
n s
our
c
e
i
m
a
ge
r
y a
nd high
-
r
e
s
ol
ut
io
n s
ta
ti
c
pr
io
r
s
.
R
E
F
E
R
E
N
C
E
S
[1]
C.
D.
Thomas,
A.
Cameron,
R.
E.
Green,
M.
Bakkenes,
L.
J.
Beaum
ont
,
Y.
C.
Collingham,
B.
F.
Erasmus,
M.
F.
De
Siqueira,
A.
Grainger,
L.
Hannahet
al.,
"
Extinction
risk
from
cli
mate
change,"
Nature
,
vol.
427,
no.
6970,
pp.
145
–
148, 2004. https://doi.org/10.1038/nature02121.
[2]
E.
Shove,
"
Beyond
the
abc:
climate
change
policy
and
theories
of
s
ocial
change,"
Environment
and
Planning
A,
vol. 42, no. 6, pp. 1273
–
1285, 2010. https://doi.org/10.1068/a42282.
[3]
M.
Dimyati,
K.
Kustiyo,
and
R.
D.
Dimyati,
"
Paddy
field
classific
ation
with
modis
-
terra
multi
-
temporal
image
transforma
tion
using
phenologica
l
approac
h
in
java
island,"
Intern
ational
Journal
of
Electrical
and
Compute
r
Engineering (IJECE)
, vol. 9, no. 2, pp. 1346
-
1358, 2019. DOI: http://doi.org/10.11591/ijece.v9i2.pp1346
-
1358.
[4]
S.
B
.
Jadhav,
"
Convolutional
neural
networks
for
leaf
image
-
based
pl
ant
disease
classification,"
IAES
International
Journal
of
Artifi
cial
Intelli
gence
(IJAI)
,
vol.
8,
no.
4,
p.
328,
2019.
DOI:
http://doi.org/10.11591/ijai.v8.i4.pp328
-
341.
[5]
O. Olaniyi, E. Dani
ya
, J. Kolo, J.
Bala, and A. Olanrewaju, "
A comput
er vision
-
based weed control system for low
-
land
rice
precision
farming,"
Internati
onal
Journal
of
Advances
in
Ap
plied
Sciences
(IJAAS)
,
vol.
9,
no.
1,
pp.
51
–
61, 2020. DOI: http://doi.org/10.11591/ijaas.v9.i
1.pp51
-
61.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
Sat
e
ll
it
e
i
m
age
i
npai
nt
in
g w
it
h de
e
p ge
ne
r
at
iv
e
ad
v
e
r
s
ar
ia
l
ne
u
r
al
ne
tw
or
k
s
(
M
ohame
d A
k
r
am
Z
ay
ta
r
)
129
[6]
S.
M.
Bernard,
J.
M.
Samet,
A.
Grambsch,
K.
L.
Ebi,
and
I.
Romieu,
"
The
potential
impacts
of
climate
variability
and
change
on
air
pollution
-
related
he
alth
effects
in
the
united
states,"
Environmental
health
perspectives
,
vol.
109,
no. Suppl 2,
pp. 199
–
209, 2001.
https://doi.org/10.2307/3435010.
[7]
W.
Suparta
and
W.
S.
Putro,
"
Comparison
of
tropical
thunderstorm
e
stimation
between
multiple
linear
regression,
Dvorak,
and
anfis,"
Bulletin
of
Electrical
Engineering
and
Informati
cs
(BEEI)
,
vol.
6,
no.
2
,
pp.
149
–
158,
2017.
DOI: 10.11591/eei.v
6i2.648.
[8]
E. C. Barrett and L. F. Curti
s, “Introduction to environmental remote sensin
g,”
Psychology Press
, 1999.
[9]
B.
C.
Naik
and
B.
Anuradha,
"
Extraction
of
water
-
body
area
from
hig
h
-
resolutio
n
Landsat
imagery,"
Inter
nationa
l
Journal
of
Electrica
l
and
Comput
er
Engineeri
ng
(IJECE)
,
vol
.
8,
no.
6,
pp.
4111
-
4119,
2018.
DOI
:
http://doi.org/10.11591/ijece.v8i6.pp4111
-
4119.
[10]
S.
Dhingra
and
D.
Kumar,
"
A
review
of
remotely
sensed
satellite
i
mage
classification,"
Internati
on
al
Journal
of
Electrical
and
Computer
Engineering
(IJECE)
,
vol.
9,
no.
3,
pp.
1720
-
1731,
2019.
DOI
:
http://doi.org/10.11591/ijece.v9i3.pp1720
-
1731.
[11]
M.
Dimyati,
A.
Fauzy,
and
A.
S.
Putra,
"
Remote
sensing
techn
ology
for
disaster
mitigation
and
regiona
l
infra
structure
planning
in
an
urban
area:
a
review,"
Telkom
nika
Tel
ecomm
unicatio
n
Computin
g
Electr
onics
and
Control
, vol. 17, no. 2, pp. 601
–
608, 2019. DOI: http://dx.doi.org/10.12928/telkomnika.v17i2.12242.
[12]
B.
-
R.
B.
Semlali,
C.
El
Amrani,
and
S.
Denys,
"
Develo
pment
of
a
jav
a
-
based
application
for
environmental
remote
sensing
data
processing,"
Internati
onal
Journal
of
Electrical
and
Computer
Engineering
(IJECE)
,
vol.
9,
no.
3,
pp.
1978
-
1986, 2019. DOI: http://doi.org/10.11591/ijece.v9i3.pp1978
-
1986.
[13]
I.
M.
Shirbh
ate
and
S.
S.
Barve,
"
Solar
panel
monitoring
and
energy
prediction
for
smart
solar
system,"
Internati
onal
Journal
of
Advances
in
Applied
Sciences
(IJAAS)
,
vol.
8,
no.
2,
pp.
136
-
142,
2019.
DOI
:
http://doi.org/10.11591/ijaas.v8.i2.pp136
-
142.
[14]
H.
F.
Hawari,
A
.
A.
Zainal,
and
M.
R.
Ahmad,
"
Development
of
real
-
time
internet
of
things
(IoT)
based
air
quality
monitoring syste
m,"
Indonesian
Journal
of Elect
rical
Engineeri
ng a
nd Compu
ter
Science
(IJEECS)
, vol.
13, no
.
3, pp. 1039
–
1047, 2019. DOI: http://doi.org
/10.11591/ijee
cs.v13.i3.pp1
039
-
1047.
[15]
J. J. Quackenbo
ss, J.
D. Spengler,
M. S.
Kanarek, R.
Letz, and
C. P. D
uffy, "
Personal exposure
to nitrogen
dioxide
:
relationshi
p to ind
oor/outdo
or air quali
ty and acti
vity patt
erns,"
Environmental science & technology
,
vol. 20, no. 8,
pp. 775
–
783, 1986. https://doi.org/10.1021/es00150a003.
[16]
A.
Chauhan,
M.
Krishna,
A.
Frew,
and
S.
Holgate,
"
Exposure
to
ni
trogen
dioxide
(no2)
and
respiratory
disease
risk."
Reviews on environmental health
, vol. 13, no. 1
-
2, pp. 73
–
90, 1998.
[17]
A.
A.
Efros
and
T.
K.
Leung,
"
Texture
synthesis
by
non
-
parametric
s
ampling"
in
Proceedings
of
the
seventh
IEEE
international
conferenc
e
on
computer
vision
,
IEEE,
vo
l.
2.
1999,
pp.
1033
–
1038.
DOI
:
10.1109/ICCV.1999.790383.
[18]
A.
A.
Efros
and
W.
T.
Freeman,
"
Image
quilting
for
texture
synthe
sis
and
transfer,"
in
Proceedings
of
the
28th
annual
conference
on
Computer
graphics
and
interactive
techniques
.
ACM,
pp.
341
–
346,
2001.
https://doi.org/10.1145/383259.383296.
[19]
A.
Criminisi,
P.
Perez,
and
K.
Toyama,
"
Region
filling
and
object
re
moval
by
exemplar
-
based
image
inpainting,"
IEEE Trans
actions o
n image pro
cessing
, vol. 13, no. 9, pp. 1200
–
1212, 2004. DOI: 10.1109/TIP.2004.833105.
[20]
C.
Ballester,
M.
Bertalmio,
V.
Caselles
,
G.
Sapiro,
and
J.
Verdera,
"
Filling
-
in
by
joint
interpolation
of
vector
fields
and
gray
levels,"
IEEE
transactions
on
image
processing
,
vol.
10,
no.
8,
pp.
1200
–
1211,
2001.
DOI:
10.1109/83.935036.
[21]
J.
Hays
and
A.
A.
Efros,
"
Scene
completion
using
millions
of
ph
otographs,"
ACM
Transactions
on
Graphics
(TOG), vol. 2
6, no. 3,
p. 4, 2007
. https:
//doi.org
/10.1145/
1276377.12
76382.
[22]
R.
Kohler,
C.
Schuler,
B.
Scholkopf,
and
S.
Harmeling,
"
Mask
-
speci
fic
inpainting
with
deep
neural
networks"
in
German Conference on P
attern Recognition
. Springer, 2014, pp. 523
–
534.
[23]
L.
Xu,
J.
S.
Ren,
C.
Liu,
and
J.
Jia,
"
Deep
convolutional
neural
netw
ork
for
image
deconvolution,"
in
Advances
in
neural information processing systems
, 2014, pp. 1790
–
1798.
[24]
D.
Pathak,
P.
Krahenbuhl,
J.
Dona
hue,
T.
Darrell,
and
A.
A.
Efr
os,
"
Context
encoders:
Feature
learning
by
inpainting,"
in
Proceedings
of
the
IEEE
conference
on
computer
visi
on
and
pattern
recognition
,
2016,
pp.
2536
–
2544.
[25]
C.
Yang,
X.
Lu,
Z.
Lin,
E.
Shechtman,
O.
Wang,
and
H.
Li,
"
High
-
r
es
olution
image
inpainting
using
multi
-
scale
neural
patch
synthesis,"
in
Proceedings
of
the
IEEE
Conference
on
Computer
Vision
and
Pattern
Recognition
,
2017, pp. 6721
–
6729.
[26]
S.
Iizuka,
E.
Simo
-
Serra,
and
H.
Ishikawa,
"
Globally
and
locally
cons
istent
image
com
pletion,"
ACM
Transactions
on Graphics
(ToG)
, vol. 36, no. 4, p. 107, 2017. https://doi.org/10.1145/3072959.3073659.
[27]
J. Yu, Z. Lin, J. Yang
, X. Shen, X. Lu, a
nd T. S. Huang, "
Generative
i
mage inpainting with co
ntextual attention,"
in
Proceedings
of
the
IEE
E
Conference
on
Computer
Vision
and
Patte
rn
Recognition
,
2018,
pp.
5505
–
5514.
DOI
:
10.1109/CVPR.2018.00577.
[28]
Yu, J., Lin, Z.
,
Yang, J.,
Shen, X.,
Lu, X.,
& Huang, T.
S., "
Free
-
form
image inpainting with
gated convolution,"
i
n
Proceedings of the IEEE Interna
tional Conferenc
e on Compute
r Vision
, 2019, pp. 4471
–
4480.
[29]
C.
Barnes,
E.
Shechtman,
D.
B.
Goldman,
and
A.
Finkelstein,
"
The
generalized
patchmatch
correspondence
algorit
hm,"
in
European
Conference
on
Computer
Vision
.
Springer,
2010,
pp.
29
–
43.
DOI:
https:/
/doi.org/10.100
7/978
-
3
-
642
-
15558
-
1_3.
[30]
He, K., Zhang, X., R
en, S., and
Sun, J., “Deep resid
ual
learning for im
age recognition,” In
Proceedings of the IEE
E
Conferenc
e on Comp
uter Vision
and Patte
rn Reco
gnition
, 2016, pp. 770
-
778. DOI: 10.1109/CVPR.2016.90
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
,
V
ol
.
10
, N
o.
1
,
M
a
r
c
h
20
2
1
:
121
–
130
130
[31]
J.
Sherme
yer
and
A.
Van
Etten,
"
The
Effe
cts
of
Super
-
Resolution
o
n
Object
Detect
ion
Perfor
mance
in
Satellite
Imagery,"
2019
IEEE/CVF
Conference
on
Computer
Vision
and
Pattern
Recognition
Workshops
(CVPRW)
,
pp.
1432
-
1441, 2019. DOI: 10.1109/CVPRW.2019.00184.
[32]
Jayanetti,
J.
M.,
Meedeniya,
D.
A.,
Dilini,
M.
D.
N.,
Wickramapa
la,
M.
H.,
&
Madushanka,
J.
H.,
“Enhanced
land
cover
and
land
use
information
generation
from
satellite
imagery
and
foursquare
data,”
In
Proceedings
of
the
6th
Internati
onal
Conference
on
Soft
ware
and
Computer
Applications
,
pp.
149
-
153,
2017.
https://doi.org/10.1145/3056662.3056681.
[33]
Meedeniya,
D.
A.,
Jayanetti,
J.
M.,
Dilini,
M.
D.
N.,
Wickramapa
la
,
M.
H.,
and
Madushanka,
J.
H.,
“Land‐Use
Classific
ation
with
Integra
ted
Data,”
Machine
Vision
Ins
pectio
n
Systems:
Image
Processing,
Concepts,
Methodologi
es and Appl
ications
, 2020, ch. 1, pp. 1
-
36. https://doi.org/10.1002/9781119682042.ch1.
B
I
O
G
R
A
P
H
I
E
S
O
F
A
U
T
H
O
R
S
Mohamed
Akram
Zaytar
is
a
Ph.D.
candidate
at
the
faculty
of
sciences
and
technologie
s
Tangier,
working
at
the
intersection
of
representation
learning
and
environmental
science.
He
obtained
his
Master'
s
Degree
in
Information
Systems
and
Networking
from
the
Faculty
of
Scienc
es
and
Techn
ologies
(Tang
ier)
in
2017.
He
is
intere
st
ed
in
machine
learning,
representation
learning,
and
climate
change.
He
has
research
experie
nce
in
weather
forecasting,
air
quality
monitoring,
precision
agriculture,
and
satellite
imagery
a
pplications.
His
work
has
been featured on interna
tional venues such
as ICLR
2020, Deep
Learn
ing Indab
a 2018,
and EGU
2018.
Professor Chaker El Amrani
is a Doctor
in Mathema
tical Modellin
g and Nume
rical Simulatio
n
from
the
University
of
Liege,
Belgium
(2001).
He
lectures
in
distribu
ted
systems
and
promotes
HPC
educati
on
at
the
University
of
Abdelmaalek
Essaadi.
His
resear
ch
interests
include
Cloud
Computing, Big
Data M
ining, and
Environme
ntal Scie
nce. D
r. El Amr
ani has
served
as an
active
volunteer
in
IEEE
Morocco.
He
is
curr
ently
Vice
-
Chair
of
IEEE
Communication
and
C
omputer
Societie
s
Morocc
o
Chapte
r,
and
advisor
of
th
e
IEEE
Computer
Socie
ty
Student
Branc
h
Chapte
r
at
Abdelmalek
Essaadi
University.
He
is
the
NATO
Partner
Countr
y
Director
of
the
real
-
time
remote sensi
ng init
iative for earl
y warning and
mitigat
ion of di
sa
sters & epidemics in Mor
occo.
Evaluation Warning : The document was created with Spire.PDF for Python.