Int
ern
at
i
onal
Journ
al of Ele
ctrical
an
d
Co
mput
er
En
gin
eeri
ng
(IJ
E
C
E)
Vo
l.
8
, No
.
6
,
Decem
ber
201
8
, p
p.
4584
~
4592
IS
S
N:
20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v
8
i
6
.
pp
4584
-
45
92
4584
Journ
al h
om
e
page
:
http:
//
ia
es
core
.c
om/
journa
ls
/i
ndex.
ph
p/IJECE
MetOp
Satellit
es Data
Processin
g for Ai
r Pollu
tion
Monitori
ng
in Mo
rocco
Moham
ed
Akr
am
Z
ayt
ar,
Chaker El
A
mrani
Facul
t
y
of
Sci
en
ce
and Technolo
g
y
in Ta
ngi
er,
A
bdel
m
al
ek
Essaa
di
Univer
si
t
y
,
Laborat
or
y
of
Infor
m
at
ic
s
S
y
s
te
m
s a
nd
Te
l
ec
om
m
unic
ations (L
IST
)
,
Mo
ro
cc
o
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
J
un
9
, 201
8
Re
vised
Ju
l
1
,
201
8
Accepte
d
J
ul
20
, 2
01
8
Thi
s
pap
er
pre
s
ent
s
a
data
pro
ce
ss
ing
s
y
st
em
base
d
on
an
arc
hi
te
c
tur
e
comprised
of
m
ult
ipl
e
st
ac
ke
d
lay
ers
of
co
m
puta
ti
onal
pro
ce
ss
es
tha
t
tra
nsform
s
Raw
Bina
r
y
Poll
uti
on
Dat
a
co
m
ing
dire
ctl
y
from
Two
EUMETSAT
MetOp
sat
el
l
it
es
t
o
our
serve
rs,
i
nto
re
ad
y
to
interpre
t
and
visual
ise
cont
in
uous
dat
a
stre
a
m
in
nea
r
re
al
t
i
m
e
using
te
chniques
var
y
ing
from
ta
sk
aut
om
at
ion,
data
pre
proc
essing
and
dat
a
ana
l
y
sis
to
m
ac
hine
le
arn
ing
using
fee
dforward
artif
ic
i
al
neur
a
l
netw
orks.
The
proposed
sy
st
em
handl
es
th
e
ac
qu
isit
ion,
clea
n
ing,
proc
essing,
nor
m
al
iz
ing
,
and
pr
edi
c
ti
ng
of
Pollut
ion
D
at
a
i
n
our
a
rea of int
ere
st of
Morocc
o
.
Ke
yw
or
d:
Air p
olluti
on
Data ag
gregati
on
Data analy
sis
Data p
r
ocessin
g
Deep l
ear
ning
Copyright
©
201
8
Instit
ut
e
o
f Ad
vanc
ed
Engi
n
ee
r
ing
and
S
cienc
e
.
Al
l
rights re
serv
ed
.
Corres
pond
in
g
Aut
h
or
:
Moh
am
ed
A
kra
m
Zayt
ar,
Dep
a
rtm
ent o
f Info
rm
at
ic
s,
Lab
or
at
ory
of
I
nfor
m
at
ic
s System
s an
d
Tel
ec
omm
un
ic
at
ion
s (
L
IS
T
),
Abdelm
al
ek
Essaadi
Un
i
ver
sit
y, PO. Bo
x 4
16, T
an
gier, M
orocc
o.
Em
a
il
:
Me
dA
kram
Zaytar@gm
ai
l.co
m
1.
INTROD
U
CTION
Ov
e
r
the
la
st
deca
de,
Air
Po
ll
utio
n
e
nv
i
ronm
ental
thre
at
s
sign
ific
a
nt
ly
increased
[1
]
-
[
4],
a
nd
Cl
i
m
a
te
chan
ge
eff
ect
s
beca
m
e
m
any
and
wide
rangin
g
[
5].
The
re
is
no
do
ub
t
t
hat
excessive
le
vels
of
ai
r
po
ll
utio
n
a
re
c
ausin
g
a
lot
of
dam
age
to
hu
m
an
an
d
anim
al
healt
h
as
we
ll
as
to
the
wi
der
e
nvir
on
m
ent.
F
or
these
reas
ons,
caref
ul
sci
entif
ic
researc
h
a
nd
m
on
it
or
in
g
of
ai
r
poll
utants
i
s
a
neces
sit
y
that
m
us
t
be
e
xerci
se
d
with a
great
d
e
al
o
f
att
ention
and preci
sio
n.
Nowa
days,
as
m
uch
as
we
wan
t
t
o
quic
kl
y
evaluate
an
d
co
ncl
ud
e
from
existi
ng
poll
ution
a
nd
cl
i
m
at
e
data,
m
os
t
of
the
prob
le
m
s
we
fac
e
center
ar
ound
prep
arin
g,
cl
eanin
g,
proces
sing,
an
d
tra
nsfo
rm
ing
the
la
rg
e
am
ou
nts
of
ra
w
en
vi
ronm
ental
data
we
receive
fro
m
sat
ellit
es
in
near
real
tim
e.
In
ou
r
case,
th
e
raw
data
ta
kes
m
ulti
ple
pr
im
it
iv
e
form
at
s
su
ch
as
BUFR
(
Bi
nar
y
U
niv
e
r
sal
Fo
rm
fo
r
t
he
Re
prese
ntati
on
of
m
et
eor
ologica
l
data),
GR
IB
2,
HRIT/LR
IT,
HRPT/LR
PT
.
in
this
pa
per,
w
e
are
goin
g
to
pr
ese
nt
a
syst
em
for
proc
essi
ng
BU
FR
base
d
bi
na
ry
file
s
c
om
ing
directl
y
f
ro
m
the
sat
el
li
te
’s
sens
or
s
an
d
tra
ns
f
or
m
it
into
a
data
set
that is rea
dy
f
or
data anal
ysi
s sp
eci
fic ta
sk
s li
ke
s in
fer
e
nce a
nd v
is
ualisa
ti
on
.
The
m
ai
n
so
urce
of
the
dat
a
we
process
is
EUMETS
A
T.
EUME
TS
A
T
is
a
n
inte
rgov
e
r
nm
ental
op
e
rati
onal
sat
el
li
te
agen
cy
w
it
h
a
total
of 3
0
Eu
ropea
n
Me
m
ber
Stat
es.
The
organ
iz
at
io
n’
s
m
issi
on
sta
tem
ent
is
to
gather
acc
ur
at
e
an
d
reli
a
ble
sat
el
li
te
dat
a
on
w
eat
her,
c
lim
a
te
and
the
env
i
ronm
ent
aro
un
d
the
cl
oc
k,
and
to d
el
ive
r
t
hem
to it
s m
e
m
ber
an
d co
operati
ng stat
es, i
nter
na
ti
on
al
partne
r
s,
a
nd to use
rs world
-
wi
de [6]
.
The
data
we
ar
e
m
os
t
interest
ed
in
com
es
di
rectl
y
from
a
t
ype
of
sat
el
li
te
s
nam
ed
Me
to
p.
Me
t
op
is
a
series
of
thre
e
po
la
r
orbiti
ng
m
et
eor
ologica
l
sat
el
li
te
s,
we
cur
ren
tl
y
get
da
ta
from
two
of
them
,
Me
top
-
A
a
nd
Me
top
-
B
,
they
bo
th
a
re
in
a
lowe
r
pola
r
or
bit,
at
an
al
ti
tu
de
of
ap
prox
i
m
at
ely
81
7
kil
om
et
res,
they
pro
vid
e
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N:
20
88
-
8708
Met
Op Satel
li
t
es Data
Pr
oces
sing f
or
Air P
ol
luti
on
M
onit
ori
ng
i
n
M
orocc
o
(
Mo
hame
d A
kram Z
aytar
)
4585
detai
le
d
obser
va
ti
on
s
of
the
gl
ob
al
at
m
os
ph
e
re,
oc
eans
a
nd
con
ti
ne
nts.
T
he
la
st
sat
el
l
it
e,
Me
top
-
C
,
is
pl
ann
e
d
to b
e
lau
nched
in 20
18.
The
syst
em
transfo
rm
s
the
data
from
i
ts
pr
im
itive
BUF
R
fo
rm
at
,
wh
ic
h
is
a
bin
ary
data
form
at
m
ai
ntained
by
the
w
or
l
d
m
et
e
orolo
gical
orga
nizat
ion
,
t
o
c
om
m
a
separ
at
ed
file
s
(CSV).
T
he
BUFR
form
at
is
a
so
m
ewh
at
co
ntr
ov
e
rsial
a
nd
a
hard
-
to
-
wor
k
-
with
da
ta
for
m
at
becau
se
of
the
dif
ficult
y
of
m
anipu
la
ti
ng
an
d
exp
e
rim
enting wit
h
it
s e
ncode
d values
.
Our
pro
posed
so
luti
on
is
a
s
of
t
war
e
syst
em
co
m
po
sed
of
m
ulti
ple
st
acked
la
ye
rs.
The
first
one
deco
m
pr
esses
and
processes t
he
BUFR b
i
nar
y data, d
eco
de
s it
, s
tructur
es
and
c
om
bin
es it
s d
ecoded m
es
sages
unde
r
the CS
V
(
com
m
a separ
at
ed
value
s)
f
orm
at
, an
d
final
ly
n
or
m
al
iz
es i
t. Bec
ause d
ee
p
le
arn
i
ng m
od
el
s ar
e
us
e
d
in
dif
fer
e
nt
cl
i
m
at
e
rela
te
d
pro
blem
s
[7
]
-
[
9],
we
trai
ned
a
nd
m
easur
e
d
the
pe
rf
orm
ance
of
an
ANN
base
d
arc
hitec
ture
w
he
n
fill
in
g
m
issi
ng
val
ue
points
a
nd
in
te
rpolat
ing
ne
w
on
es
.
T
he
sy
stem
pr
od
uces
a
nea
r
con
ti
nu
ous
data stream
o
n
t
he
2
-
D su
rf
ace
of
our area
of int
erest.
The
s
of
tw
are
so
luti
on
propo
sed
by
this
pa
per
is
a
syst
e
m
that
c
an
be
directl
y
plug
ged
i
nto
the
endp
oin
ts
of
th
e
nea
r
real
ti
m
e
data
st
ream
,
i
t
will
al
low
f
or
fast
e
xperim
e
ntati
on
an
d
vis
ualiz
at
ion
of
al
read
y
processe
d
r
aw
data
points
c
om
ing
directl
y
f
ro
m
the
Me
top
-
X
sat
el
li
te
s
series,
it
will
al
so
res
ult
in
s
pa
ce
an
d
tim
e
red
uction
and
optim
iz
at
i
on
si
nce
it
fo
c
us
es
on
inte
res
t
areas,
we
l
ook
f
orward
f
or
our
s
olu
ti
on
to
further
i
m
pr
ove a
nd a
ccel
erate t
he re
search
pr
ocess done
on to
p o
f t
he
E
UMETC
AS
T
data st
rea
m
p
ipeli
ne.
2.
PRO
CED
U
R
E
2.1.
Data Pr
ocessi
ng
The follo
wing
fig
ur
e
dem
on
st
rates the
proce
dure ta
ken to
pre
-
process
and
norm
al
iz
e the d
at
a
:
Figure
1. Dec
odin
g
B
UF
R
Da
ta
to
Com
m
a separ
at
e
d
m
erg
e
d
m
essages
In
the
first
ste
p,
t
he
syst
em
gets
the
ra
w
ta
r
file
s
t
hro
ugh
the
FTP
prot
oc
ol,
a
fter
e
xtra
ct
ing
t
he
c
om
pr
e
sse
d
file
s
we
get
m
ulti
ple
Bi
nar
y
BUFR
file
s
wh
ic
h
f
ollow
a
stric
t
na
m
ing
c
onve
ntion
in
the
fo
ll
owin
g
form
(INSTR
UME
NT
_ID
-
PR
O
D
UCT_T
Y
PE
-
P
ROCESS
ING
_L
EV
EL
-
S
PA
C
ECR
AF
T
_ID
-
SEN
S
I
NG_ST
ART
-
SEN
S
I
NG_E
N
D
-
P
R
OCES
SING
_MO
DE
-
D
IS
P
OSITI
O
N_M
OD
E
-
PR
OCESSIN
G_
T
IM
E)
that
co
rr
e
sp
on
ds
to
m
ulti
ple
i
m
portant
va
riabl
es
su
c
h
as
t
he
instru
m
ent
ide
ntifie
r,
orbit,
a
nd
ti
m
e
fr
a
m
e,
the
syst
em
filters
th
e
data
do
wn
t
o
get
poll
utio
n
f
il
es
in
the
ti
m
e
the
sat
el
li
t
e
is
scan
ning
the
area
of
in
te
rest
us
in
g
re
gu
la
r
expressi
on
s
on
the
nam
es
of
the
ex
tr
act
ed
fil
es
(un
der
the
poll
ution
c
ode
nam
e
of
”TR
G”
).
Wh
at
we
fin
al
ly
get are
m
ulti
pl
e BUFR
poll
ution
file
s corre
s
pondin
g
to
the
area
of
i
nterest
that are
r
ea
dy
to b
e
d
ec
oded
.
In
the
seco
nd
ste
p,
t
he
syst
e
m
us
es
a
thir
d
pa
rty
softwa
r
e
so
l
ution
na
m
ed
BUFR
E
xt
ract
[
10
]
t
o
decode
t
he
B
UF
R
file
s
i
nto
bul
ks
of
e
xport
ed
m
essag
es,
each
m
ess
age
c
on
ta
ini
ng
a
desc
riptio
n
of
it
s
colum
ns
an
d
t
he
val
ues
in
each
on
e
i
n
a
te
xt
file
for
m
at
.
In
the
t
hi
rd
ste
p,
the
s
yst
e
m
per
form
s
fast
m
erg
e/
sel
ect
ion
te
ch
niques
t
o
com
bin
e
al
l
of
t
he
m
essages
into
tw
o
c
omm
a
separ
at
ed
f
il
es
cor
res
pond
ing
to
the
scan
ning
ti
m
efr
am
e,
on
e
for
the
Me
t
op
-
A
sat
el
li
te
and
the
seco
nd
f
or
Me
top
-
B
.
Bot
h
CS
V
file
s
c
onta
in
the foll
owin
g
c
olu
m
ns
of
inte
r
est
:
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
201
8
:
4584
-
4592
4586
Table
1.
E
xtrac
te
d
Feat
ures
No
.
Featu
re
Un
it
1
Year
Integ
er
2
Mon
th
Integ
er
3
4
5
6
7
8
9
10
11
Day
Ho
u
r
Minu
te
Seco
n
d
latitu
d
e
Lon
g
itu
d
e
CH4
Dens
ity
CO2
Dens
ity
N2
O Dens
ity
Integ
er
Integ
er
Integ
er
Integ
er
DEGRE
E
DEGRE
E
k
m
.
m
-
2
k
m
.
m
-
2
k
m
.
m
-
2
Af
te
r
ex
portin
g
the
neces
sar
y
values
i
nto
m
ul
ti
ple
structur
e
d
CS
V
file
s,
the
syst
em
groups
r
ow
s
by
locat
ion
points
an
d
the
e
xact d
at
e
(
Year
-
Mo
nth
-
Day
-
H
our
-
Mi
nu
te
-
Sec
ond)
a
nd
a
pp
li
es the
m
ean
functi
on
on
the
poll
utant
va
lues
to
ta
ke
the
ave
rag
e
of
po
s
sible
re
dundant
m
easur
e
m
ents.
In
the
f
ourth
ste
p,
the
sy
stem
deals
with
cl
ea
ning
data
point
s
that
are
substanti
vely
unreas
on
a
ble
us
i
ng
l
og
ic
al
co
ndit
ion
s
on
data
po
i
nts
of
CH
4
,
CO
2
a
nd
N
2
O
us
i
ng
Z
-
s
cor
es
.
Last
ly
,
t
he
syst
em
no
r
m
al
iz
es
al
l
po
ll
ution
points
in
to
va
lues
i
n
[−
1,
1]
to
acce
le
rate co
nver
ge
nce in t
he
t
rainin
g p
hase
s,
usi
ng the
fol
lowing
f
or
m
ula fo
r
al
l t
hree
num
erical
v
ariables
:
←
−
(
)
(
)
−
(
)
∀
∈
{
1
,
2
,
3
}
,
∈
{
1
,
.
.
.
,
}
As
a
ge
ne
ral
de
scriptio
n
of
t
he
pro
cess,
ea
ch
hal
f
an
hour,
the
syst
em
receives
one
c
om
pr
essed
ta
r
file
throu
gh
th
e
serv
e
rs’
e
nd
po
i
nts,
the
syst
e
m
autom
a
ti
ca
ll
y
deco
m
pr
esses
the
file
into
BUFR
BIN,
s
el
ect
s
file
s
cor
r
esp
on
ding
to
the
are
a
of
interest
,
a
nd
decodes
the
m
us
ing
a
thir
d
pa
rty
li
br
ary
(BUF
Re
xtract)
to
the
corres
pondin
g
m
essages
an
d
tur
ns
them
into
two
CS
V
file
s
con
ta
ini
ng
al
l
of
the v
al
ues
of
interest
in
ne
ar
real
tim
e,
this
res
ul
ts
in
a
c
onsid
erab
le
re
du
ct
io
n
i
n
the
dim
e
ns
io
nalit
y
of
the
data
a
nd
th
e
sp
ace
it
no
r
m
al
l
y
occupies.
The
seco
nd
pa
rt
of
the
syst
e
m
fil
ls
the
m
issi
ng
va
lues
in
the
2
-
D
s
urface
of
inte
r
est
and
al
so
gen
e
rates new
d
at
a
points usi
ng
al
gorithm
ic
search
a
nd
a ne
ur
al
n
et
w
ork
arch
it
ect
ure
to g
et
a
nea
r
c
on
t
inuous
data stream
o
ut
pu
t t
hat is r
ead
y for e
xp
l
or
at
i
on, visuali
sat
io
n,
a
nd inte
rpret
at
ion
.
2.2.
Int
el
li
gen
t
In
t
erpola
tio
n
The
pre
dicti
on
of
m
issi
ng
values
is
based
on
th
ree
pre
-
trai
ned
Fee
d
-
F
orward
F
ully
Connect
e
d
Neural
net
wor
k
m
od
el
s
fit
to
fill
the
m
issi
n
g
values
i
n
the
2
-
D
s
urface
of
our
i
nterest
f
or
the
t
hr
ee
poll
utant
s
(CO
2,
C
H4, a
nd
N2O), a
nd th
e g
e
ner
a
l a
rc
hitec
ture of o
ur
ANNs
is
as
s
how
n
i
n
Fi
gure
2.
Figure
2. The
ANN Arc
hitec
ture
to pre
dict
m
issi
ng
v
al
ues
As
a
n
a
ct
ivati
on
f
unct
ion
f
or
our
m
od
el
,
we
c
hose
t
he
recti
fier
functi
on.
T
he
ge
neral
process
i
n
wh
ic
h
sel
ect
ed
m
issi
ng
po
i
nts
are pre
dicte
d
(
or not),
is
s
hown
as
Fi
gure
3
:
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N:
20
88
-
8708
Met
Op Satel
li
t
es Data
Pr
oces
sing f
or
Air P
ol
luti
on
M
onit
ori
ng
i
n
M
orocc
o
(
Mo
hame
d A
kram Z
aytar
)
4587
Figure
3. The
Fil
li
ng
m
issi
ng
points
Proced
ur
e
The
syst
em
p
re
dicts m
issi
ng
va
lues foll
owin
g
this
pro
ce
dur
e
:
1.
Ma
p
al
l m
issi
n
g values
w
it
h t
he nearest
100
neig
hbor
value
s.
2.
So
rt
points i
n f
un
ct
io
n o
f
the
nu
m
ber
of n
ei
ghbo
rin
g
m
issi
ng
po
i
nts a
nd th
e ave
rag
e
d
ist
a
nce, givi
ng a
new sc
or
e
f
or
e
ach m
issi
ng
po
int i
n
the
fo
rm
of :
=
∗
3.
If
t
he
T
op m
iss
ing
point’s
av
e
rag
e
d
ist
anc
e f
ro
m
all
n
ei
ghbor
points is
gre
at
er th
a
n 50 km
, o
r
if the
re is
no to
p
m
issi
ng
po
i
nt, brea
k
t
he
loop
a
nd
fini
sh
t
he pr
ocess.
4.
If
t
he
a
ver
a
ge dist
ance is less
than 5
0km
, p
r
edict
the m
issin
g p
oin
t
us
i
ng
the AN
Ns
m
odel
s and m
ark
th
e
po
i
nt as
done
a
nd lo
op b
ac
k
t
o
ste
p 2.
The
syst
em
autom
a
ti
cal
l
y
lo
op
s
over
thes
e
ste
ps
unti
l
al
l
m
issi
ng
val
ues
a
re
fill
ed
(for
possible
pre
dicti
on
s
),
the syst
em
r
epeats t
his wh
ole pro
ce
dure
f
or
the th
ree
po
ll
ut
ants
of
i
nterest.
3.
RESEA
R
CH MET
HO
D
3.1.
Data De
scri
pti
on
The
first
Data
set
us
ed
in
this
stud
y
was
colle
ct
ed
in
the
fo
rm
of
bu
lk
s
of
BUFR
m
essage
file
s
com
ing
directl
y
fr
om
two
sat
el
li
te
s,
Metop
-
A
a
nd
Me
to
p
-
B,
an
d
preci
s
el
y
fr
om
the
In
f
rar
e
d
at
m
os
ph
e
ric
so
un
ding
inte
r
ferom
et
er
(I
A
SI
)
sens
or,
w
hich
is
c
om
po
sed
of
a
Fouri
er
trans
f
or
m
sp
ect
r
om
et
er
a
nd
a
n
associat
ed
in
te
gr
at
e
d
Im
aging
Subsyst
em
(I
IS
).
T
he
F
ourie
r
tran
sf
or
m
sp
ect
ro
m
et
er
pr
ov
ides
inf
rar
e
d
s
pectr
a
with
high
re
so
l
ution bet
wee
n 645 a
nd
2760c
m
-
1
(3
.
6m
to
15.
5m
).
The
m
ai
n
goa
l
of
I
ASI
is
t
o
pro
vid
e
at
m
os
phe
ric
em
is
sion
sp
e
ct
ra
t
o
der
i
ve
te
m
per
at
ur
e
a
nd
hu
m
idit
y
pr
of
il
es
with
high
ve
rtic
al
reso
luti
on
a
nd
accu
rac
y.
Additi
on
al
ly
it
is
us
ed
f
or
t
he
dete
rm
inatio
n
of
trace
gas
es
s
uc
h
as
oz
one,
nitrous
oxide,
an
d
ca
rbo
n
dioxide,
a
s
well
as
l
and
an
d
sea
s
urface
te
m
per
at
ur
e
a
nd
e
m
issi
vity
an
d cl
oud pro
per
ti
e
s.
IA
S
I
m
eas
ur
es
in
the
infr
a
red
par
t
of
th
e
el
ect
ro
m
agn
et
ic
sp
ect
r
um
a
t
a
h
or
iz
on
ta
l
res
olu
ti
on
of
12
km
ov
er
a
sw
at
h
widt
h
of
about
2,
200km
.
W
it
h
14
orbits
in
a
su
n
-
syn
chro
nous
m
id
-
m
or
nin
g
orbit
(9:
30
Local
S
olar
Ti
m
e
equ
at
or
cr
ossi
ng,
de
scen
di
ng
node
)
globa
l
ob
se
r
vations
can
be
pr
ov
i
de
d
twic
e
a
day
(
ever
y
12
h
our
s),
the sate
ll
it
es
ta
ke
a
rou
nd
25
m
inu
te
s
to
scan
The
area
of
interes
t,
we
get
po
ll
ut
ion
data
f
ro
m
po
i
nts
appr
ox
im
at
ely
20
km
apar
t
fr
om
each
oth
e
r.We
co
ns
tr
ucted
the
sec
ond
dataset
from
a
lready
prep
ro
c
esse
d
data
points
i
n
t
he
go
al
o
f
trai
ni
ng
,
te
sti
ng,
a
nd v
al
idati
ng
ou
r
ne
ural
net
wor
k
m
od
el
s
an
d
s
olv
e
t
he
prob
le
m
of
fill
ing
m
issi
ng
data points
and
inter
po
la
ti
ng
ne
w po
i
nts in
th
e sele
ct
ed
a
rea
of interest
.
3.2.
Int
el
li
gent I
n
terp
ola
ti
on
We
ge
ne
rated
new
em
pty
po
i
nts
val
ues
in
w
hich
al
l
of
the
po
i
nts
in
the
ar
ea
of
i
nterest
a
re
dista
nce
d
from
each
othe
r by 5
km
, th
e s
yst
e
m
then
inte
ll
igently
interpolat
e all
em
pty
points.
3.2.1.
D
ata Col
le
ction
We
c
ollec
te
d
150
Giga
byte
s
of
prep
r
ocesse
d
data
or
the
e
qu
ivale
nt
of
ar
ou
nd
800
m
i
ll
ion
data
point
to
bu
il
d
an
int
el
li
gen
t
m
od
el
capab
le
of
pre
dicti
ng
m
issi
ng
poll
utant
val
ues.
Af
te
r
c
ollec
ti
ng
the
data
set
,
we
ran
a
gen
e
ral
sta
ti
sti
c
on
m
i
ssing
data
poi
nts
an
d
we
present
the
f
ollo
wing
res
ults
ba
sed
on
the
sa
m
pled
dataset
as s
ho
wn in T
able
2
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
201
8
:
4584
-
4592
4588
Table
2.
Mi
ssi
ng
Val
ues
i
n
th
e
Data Set
No
.
Po
llu
tan
t
% o
f
m
iss
in
g
valu
es
1
N2
O
0
.00
3
%
2
CO2
7
2
%
3
CH4
7
1
%
3.2.2.
Tr
ainin
g and tes
ting
da
t
a
The
data
wa
s
trans
form
ed
into
a
ta
ble
w
he
re
the
featu
res
are
the
50
ne
arest
points
a
nd
the
ta
rg
et
var
ia
ble
is
the
data
point
us
e
d
to
trai
n
the
art
ific
ia
l
neu
ral
ne
twork
,
the
dis
ta
nce
betwee
n
the
ta
rg
et
point
and
the
f
urt
hest
point
set
t
o
a
m
axi
m
u
m
and
t
he
sam
e
con
diti
on
s
we
a
pp
li
ed
w
hen
sel
ect
in
g
va
li
d
m
issi
ng
points
wer
e
ap
plied
w
hen
tra
nsf
or
m
ing
the
da
ta
.
Wh
e
n
tr
ai
nin
g
the
m
od
el
to
predi
ct
ne
w
po
i
nt
val
ues
(Inter
po
la
ti
on)
,
the
syst
e
m
add
s
ne
w
points
(m
ark
ed
m
issin
g)
so
that
ev
ery
po
int
ha
s
a
po
int
at
le
ast
5
km
near
t
he
ne
xt
one,
a
fter
creati
ng
ne
w
gri
ds
of
2
-
D
points,
tr
ai
nin
g
set
s
we
r
e
sel
ect
ed
base
d
on
a
vaila
bili
ty
of
the n
ei
ghbori
ng
po
i
nts.
3.2.3.
Tr
ainin
g
6
M
od
el
s
co
nsi
sti
ng
of
3
f
ully
connecte
d
hid
de
n
la
ye
r
s
with
10
0,
50,
an
d
25
neurons
r
especti
vel
y
wer
e
us
e
d,
the
first
3
m
od
el
s
const
ru
ct
e
d
to
pr
e
dict
m
issi
ng
an
d
co
rru
pted
val
ues
an
d
the
la
st
3
we
re
trai
ne
d
to
inter
pola
te
new
po
i
nt v
al
ue
s,
the
traini
ng
detai
ls are
giv
e
n
as
:
a.
All
of
the
ne
uron
s
pa
ram
et
ers
were
ra
ndom
l
y
init
ia
li
zed
usi
ng
t
he
un
i
for
m
distribu
ti
on
betwee
n
−
0.1
a
nd
+0.1.
b.
The
Mi
ni
-
Ba
tc
h gr
a
dient
Des
cent wa
s
us
ed
to op
ti
m
iz
e the p
aram
et
ers.
c.
A
le
ar
ning
rate
of ε=
0.001 wa
s chose
n.
d.
Ba
tc
hes
of
1024 sam
ples and
200
e
poch
s
we
re traine
d.
3.2.4.
V
alida
tion
Fo
r
the
validat
ion
t
o
be
ef
fici
ent,
we
us
ed
10
-
fo
l
d
cr
os
s
va
li
dation
te
c
hn
i
qu
e
,
s
plit
ti
ng
t
he
data
set
into
m
ulti
ple training an
d
te
sti
ng sets to
v
e
rif
y t
he
ef
fici
enc
y of t
he
trai
ne
d m
od
el
s and to
avo
i
d ov
e
rf
it
ti
ng.
3.2.5.
I
nt
er
po
l
at
i
on
Metho
d
The
syst
em
us
es
three
pr
e
-
tr
ai
ned
neural
ne
twork
m
od
el
s
to
pr
e
dict
ne
wly
gen
e
rated
po
i
nts
an
d
interp
olate
the
whole
s
ur
face.
The
pr
ocess
is
sim
il
ar
to
the
proce
dure
of
predict
in
g
m
issin
g
val
ues,
ho
w
ever,
the
syst
e
m
do
esn’
t
set
a
thre
sh
ol
d
on
the
a
ver
a
ge
of
dista
nces
in
or
der
t
o
break
the
l
oop
of
pr
e
dicti
ons.
I
t
pr
e
dicts
an
d
f
il
ls
al
l
new
da
ta
po
i
nts
at
a
fixe
d
nei
ghbouri
ng
dista
nc
e
of
5km
,
t
he
f
ollo
wing
gr
a
ph
dem
on
strat
es t
he pr
ocess
as s
how
n
in
Fig
ure
4
.
Figure
4. I
nter
po
la
ti
on
by Fe
ed
-
Forwa
r
d Ne
ur
al
Netw
orks
The
syst
em
pr
edict
s
al
l
po
ints
and
up
dates
the
sorte
d
li
st
of
m
issi
ng
po
i
nts
as
it
go
es
unt
il
filling
al
l
of
t
he
m
issi
ng
values
,
the
on
l
y
diff
e
ren
ce
t
ha
t
this
m
od
el
ha
ve
wit
h
the
previo
us
on
e
is
t
hat
it
do
es
not
hav
e
a
crit
eria f
or wh
et
her
t
o pr
e
dict a m
issi
ng
poin
t or n
ot.
3.3.
Be
nchma
r
kin
g
To
m
easur
e
the
per
f
orm
ance
of
our
A
NN
-
ba
sed
inter
pola
ti
on
syst
em
,
we
be
nc
hm
ark
it
s
pr
e
dicti
on
s
against t
wo stat
e o
f
the a
rt alg
or
it
hm
ic
m
et
ho
ds
of sp
at
ia
l i
nt
erpolat
ion,
Kernel sm
oo
thin
g and K
rigi
ng.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N:
20
88
-
8708
Met
Op Satel
li
t
es Data
Pr
oces
sing f
or
Air P
ol
luti
on
M
onit
ori
ng
i
n
M
orocc
o
(
Mo
hame
d A
kram Z
aytar
)
4589
3.3.1.
Kernel
Smoot
hing
A
kernel
sm
oo
ther
is
a
sta
ti
st
ic
al
te
chn
iqu
e
for
est
i
m
at
ing
a
real
valued
f
un
ct
io
n
f
(X)
(
X
∈
R
p)
by
us
in
g
it
s
no
isy
ob
s
er
vations,
wh
e
n
no
pa
ra
m
et
ric
m
od
el
fo
r
this
f
unct
io
n
is
known
.
Th
e
est
i
m
a
te
d
functi
on
is
sm
oo
th,
a
nd
th
e
le
vel
of
sm
oo
th
ness
is
set
by
a
sin
gle
pa
ra
m
et
er.
T
o
put
i
t
in
m
at
he
m
ati
cal
te
rm
s,
the
idea
of
the
nea
rest
nei
ghbor
sm
oo
the
r
is
the
f
ollo
wing.
F
or
eac
h
point
Xi
,
ta
ke
N
near
est
neig
hb
or
s
a
nd
est
im
a
t
e
the
value
of
F(Xi)
by
aver
a
ging
the
values
of
th
ese
neighb
or
s
.
This
ty
pe
of
in
te
rpolat
ion
is
m
os
t
app
r
opria
te
fo
r
low
-
dim
ensions
(p
<
3)
(the
dim
ension
al
it
y
cur
se
[
11
]
is
on
e
rea
son
f
or
tha
t).
Actuall
y,
the
ker
nel
s
m
oo
ther
represe
nts
the
set
of
ir
regula
r
data
po
i
nts
a
s
a
sm
oo
th
li
ne
or
s
urface,
i
n
our
case
(2
-
D
s
urface
)
thi
s
is
a
perfect
ly
reas
on
a
ble
s
olu
ti
on.
O
ne
way
to
fill
these
points
w
ou
l
d
be
to
us
e
Scip
y’s
[
12
]
(
pr
ec
ise
ly
sci
py.inter
po
la
te
.Rbf
)
im
ple
m
entat
ion
of
Ra
dial
Ba
sis
Functi
on
inter
po
la
ti
on
wh
ic
h
is
inte
nd
e
d
for
t
he
sm
oo
thing
/i
nte
rpolat
ion o
f
sc
at
te
red
data.
3.3.2.
Gaussi
an Pr
ocess Re
gr
ession
or Kri
ging
Kr
i
ging
or
Ga
us
sia
n
process
regressio
n
is
a
m
et
ho
d
of
inte
rpolat
ion
i
n
w
hich
the
inte
rpolat
ed
val
ue
s
are
m
od
el
le
d
by
a
Ga
us
sia
n
process
gove
r
ned
by
pri
or
c
ov
a
riances
,
as
oppose
d
to
a
pi
ecewise
-
poly
nom
ial
sp
li
ne
ch
os
e
n
to
opti
m
iz
e
s
m
oo
t
hn
e
ss
of
th
e
fitt
ed
values.
Under
s
uitabl
e
assum
ption
s
on
the
pr
i
or
s
,
kr
i
ging
giv
es
t
he
be
st
li
ne
ar
un
biase
d
pre
dicti
on
of
the
inte
rm
ediat
e
values.
In
t
erpolat
ing
m
et
hods
base
d
on
oth
e
r
crit
eria
su
c
h
as
sm
oo
thn
ess
m
ay
no
t
yi
el
d
th
e
m
os
t
li
kely
i
nterm
ediat
e
values.
T
he
m
et
ho
d
is
wi
dely
use
d
i
n
the
dom
ai
n
of
sp
at
ia
l
anal
ysi
s
and
c
ompu
te
r
ex
pe
rim
ents.
T
he
te
c
hn
i
qu
e
is
al
s
o
kn
own
as
W
ie
ne
r
Ko
lm
og
or
ov
predict
io
n.
We’
l
l
com
par
e
the
r
esults
of
K
rigi
ng
i
nter
po
la
ti
on
on
the
datase
t
us
in
g
the
G
a
us
sia
n
Pr
oc
ess Re
gr
es
sion i
m
ple
m
entat
ion
in
t
he Py
thon’s sci
kit
-
le
arn li
br
a
ry.
3.4.
Ha
rdw
are
A
Pyt
hon
im
pl
e
m
entat
ion
of
t
he
dee
p
ne
ur
al
networ
k
arch
it
ect
ur
e
with
hidden
la
ye
rs
of
100,
50,
25
nu
m
ber
of
ne
uro
ns
(r
es
pecti
ve
ly
),
G
oogle’s
Tens
orFlo
w
[
13]
li
br
ary
was
us
e
d
to
buil
d
a
nd
trai
n
t
he
m
od
el
.
An
N
V
IDIA
T
esl
a
K
80
sin
gl
e
GPU
de
vice,
with
49
92
C
UDA
c
ores,
24
GB
of
G
D
D
R5
m
e
m
or
y,
and
48
0
GB/s ag
gregat
e m
e
m
or
y bandw
i
dth
was
u
s
ed
to
train
the
neural
netw
ork
m
od
el
s.
4.
RESU
LT
S
AND DI
SCUS
S
ION
The
res
ulti
ng
s
olu
ti
on
is
a
sys
tem
co
m
po
sed
of
th
ree
la
ye
rs
of
pr
ocesses,
t
he
first
la
ye
r
deco
m
pr
ess,
decode,
a
nd
no
rm
alizes
the
data.
the
s
eco
nd
l
ay
er
is
a
three
ANN
sta
ck
t
o
fill
in
the
m
issin
g
poll
utant
va
lues
,
and
la
stl
y
the
final
la
ye
r
w
hi
ch
is
com
po
se
d
of
a
no
t
her
sta
ck
of
neural
netw
ork
m
od
el
s
to
inter
pola
te
ne
w
data points
in o
ur area
of inte
r
est
.
4.1.
Data Pr
ocessi
ng
The
deco
m
pr
e
ssing
,
dec
odin
g,
m
erg
in
g,
cl
eanin
g
a
nd
no
rm
alizi
ng
of
Ra
w
BU
FR
dat
a
resu
lt
in
a
consi
der
a
ble
r
edu
ct
io
n
in
res
ources.
Si
nce
our
al
gorithm
runs
in
li
near
tim
e,
and
con
si
der
i
ng
the
vol
um
e
of
data
the
syst
em
pr
ocesses
at
each
ste
p,
a
sim
ple
co
m
pu
te
r
config
ur
at
io
n
(4
Gi
ga
byte
R
AM,
4
co
res
w
it
h
no
par
al
le
li
sm
)
resu
lt
in
t
he
f
ollo
wing
durati
ons
as s
how
n
in
Fi
gure
5.
Figure
5. A
verage
durati
on
of the
pr
e
process
ing
sta
ge
These
te
sts
we
re
co
nducte
d
m
ul
ti
ple
tim
es
for
each
vo
l
um
e
cat
ego
ry,
to
ens
ure
high
pr
eci
sio
n.
We
con
cl
ud
e
t
hat
the
syst
em
scales
pr
et
ty
well
and
ca
n
proces
s
la
rg
e
vo
l
um
e
s
of
data
(
up
to
te
ra
byte
s
per
hour
)
in r
e
la
ti
vely
s
hort
durati
on
of
tim
e.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
201
8
:
4584
-
4592
4590
4.2.
Int
el
li
gent In
t
erpola
tio
n
The
res
ulti
ng
s
urface
of
inter
est
is
a
18
87
by
1776
km
2
rectang
le
,
t
he
sy
stem
pr
edict
s
a
m
axi
m
u
m
nu
m
ber
of
12
3,568
points,
t
he
fo
ll
ow
i
ng
f
igures
showca
se
exam
ples
of
pr
e
dicti
ons
in
a
fixed
dat
e,
us
ing
kr
i
ging,
sm
oo
t
hi
ng a
nd ou
r n
eur
al
netw
ork m
od
el
as shown
in
Fig
ures
6,
7,
8.
Figure
6. I
nter
po
la
ti
on
visu
al
isa
ti
on
s
of N2
O
Figure
7. I
nter
po
la
ti
on
visu
al
isa
ti
on
s
of CH
4
Figure
8. I
nter
po
la
ti
on
visu
al
isa
ti
on
s
of CO
2
In
t
he
a
bove figures, t
he ro
un
ded m
ark
ers
re
pr
ese
nt a
know
n
sam
ple o
f p
ol
luti
on
data poi
nts,
a
nd the
interp
olate
d
surface
re
pr
ese
nt
the
the
res
ulti
ng
pre
dicti
on
s.
W
e
got
the
f
ol
lowing
trai
ning
r
esults
a
fter
cro
s
s
validat
in
g
the
m
od
el
s
as
sh
own
in
Fig
ur
es
9
an
d
10.
As
e
xp
ect
e
d,
the
s
yst
e
m
pr
oduce
s
bette
r
res
ults
wh
e
n
fill
ing
m
issi
ng
values
,
a
nd
ge
ner
al
ly
w
or
s
e
res
ults
w
hen
fill
ing
in
ne
w
data
po
i
nts.
but
w
he
n
com
pa
rin
g
interp
olate
d
da
ta
us
in
g
the
3
m
et
ho
ds,
we
fi
nd
i
nteresti
ng
resu
lt
s,
t
he
f
ollow
i
ng
grap
h
s
h
owcases
t
he
r
esults
of com
par
iso
ns.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N:
20
88
-
8708
Met
Op Satel
li
t
es Data
Pr
oces
sing f
or
Air P
ol
luti
on
M
onit
ori
ng
i
n
M
orocc
o
(
Mo
hame
d A
kram Z
aytar
)
4591
Figure
9. Com
par
i
ng the acc
uracy
of
diff
e
re
nt
m
od
el
s u
sin
g M
SE
Figure
10. Acc
ur
acy
of the
sugg
e
ste
d In
te
rpolat
ion
Me
thods
4.3.
Discussio
n
As
we
can
se
e
from
the
resu
lt
s,
the
optim
al
interp
olati
on
te
ch
nique
is
generall
y
bette
r
tha
n
our
trai
ned
ne
ur
al
netw
ork
m
odel
s,
howe
ve
r,
in
the
case
of
N2O
a
nd
C
H4
we
ca
n
sa
y
that
our
m
od
el
is
com
petit
ive
with
the
oth
e
r
tw
o
cl
assic
al
2
-
D
interp
olati
on
al
gorithm
s,
and
since
we
ha
d
70%
m
issi
ng
data,
that
op
e
ns
the
po
s
sibil
it
y
of
be
tt
er
per
f
orm
a
nce
with
gr
eat
er
volum
es
of
data,
if
trai
ne
d
on
la
r
ger
vo
l
um
es
of
data,
our
syst
e
m
can
m
ake
bette
r
pr
e
dicti
ons
an
d
the
refor
e
introd
uce
an
optim
al
so
luti
on
an
d
a
c
om
petito
r
t
o
the krigi
ng or s
m
oo
thing
i
nter
po
la
ti
on al
gorithm
s.
5.
CONCL
US
I
O
N
At
the
prese
nt
tim
e,
t
he
siz
e,
var
ie
ty
an
d
co
m
plexit
y
of
ra
w
data
is
hu
ge
and
c
onti
nu
es
to
increas
e
ever
y
day.
T
he
us
e
of
data
processin
g
syst
em
s
to
store,
process,
an
d
a
na
ly
ze
data
strea
m
s
has
change
d
ho
w
we
disco
ver
a
nd
vis
ualise
bi
g
data
in
ge
ne
ral.
I
n
t
his
pa
pe
r,
we
pr
ese
nt
ed
a
s
oft
war
e
so
luti
on
c
om
po
se
d
of
m
ul
ti
ple
sta
cked
la
ye
rs
of
s
ubsyst
e
m
s
that
tr
ansfo
rm
and
process
c
on
si
derable
volum
es
of
ra
w
poll
utio
n
dat
a
in
near
real
ti
m
e,
ta
king
the
data
from
it
s
native
com
pr
essed
form
at
to
a
structu
re
d,
cl
ea
ned,
norm
al
ized
,
an
d
con
ti
nu
ous
data stream
that is l
igh
t a
nd easy
to expe
rim
ent
with.
In
the
f
uture,
sign
ific
a
nt
cha
ll
eng
es
an
d
pr
ob
le
m
s
con
cer
ning
Bi
g
En
vironm
ental
Dat
a
m
us
t
be
addresse
d
by
the
industry
a
nd
aca
dem
ia
,
current
work
on
t
op
ic
s
rangin
g
f
ro
m
util
iz
ing
AI
f
or
plant
m
on
it
or
ing
[14
]
,
work
i
ng
on
so
ci
al
awar
e
ne
ss
co
ncernin
g
cl
i
m
at
e
chan
ge
[15,
16
]
,
a
nd
the
use
of
bi
ologica
l
m
et
ho
ds
[
17
]
to
fi
gh
t
cl
im
at
e
change
is
i
m
po
rta
nt.
B
ut
ne
w
c
halle
ng
e
s
to
ta
c
kle
are
in
the
fi
el
d
of
env
i
ronm
ental
data
sci
ence,
fu
tu
re
w
ork
fo
c
us
e
d
on
how
to
bu
il
d
new
e
nv
i
ron
m
ental
data
l
earn
i
ng
par
a
dig
m
s,
sci
ent
ific
com
p
uting
e
nv
i
ron
m
ents,
and
a
n
al
l
aro
un
d
bette
r
infr
as
tructu
re
f
or
poll
utio
n
m
on
it
or
ing i
s a
n
ecessi
ty
for a
ll
o
f us.
ACKN
OWLE
DGE
MENTS
The
aut
hors
a
re
than
kful
to
the
Mi
nistry
of
Higher
E
ducat
ion
an
d
Sci
entifi
c
Re
sear
ch,
an
d
the
Nati
on
al
Ce
ntr
e
for
Scie
ntific
and
Tech
nic
al
Re
search
(CNRST)
f
or
fun
ding
this
st
ud
y
,
unde
r
the
pr
oj
ect
:
PPR/
2015/7
.
REFERE
NCE
S
[1]
Le
lieve
ld,
J.
,
E
vans,
J.
S.,
Fnai
s,
M.,
Gianna
d
a
ki,
D.,
&
Poz
zer,
A
.
“
The
contributi
on
of
outd
oor
ai
r
poll
u
ti
o
n
source
s to
p
rema
ture
m
ortality
on
a
g
lobal
sca
le
”
.
Nature
,
2015
;
52
5(7569):
367
-
37
1.
[2]
Agus
ci
k
,
A.,
Ik
ob,
R.
,
&
Putra
,
S.
A.
“
The
Le
vel
of
Malond
ia
ld
eh
y
de
in
Pe
ople
Exposed
t
o
Air
Pollut
ion
”
.
Inte
rnational
Jo
urnal
of Publ
i
c Healt
h
S
ci
en
ce (
IJP
HS)
,
2017
;
6(
1)
:
99
-
103.
[3]
Haja
t
,
A.
,
Alli
so
n,
M.,
Diez
-
Rou
x,
A.
V.
,
Jenn
y
,
N.
S.,
J
orge
nsen
,
N.
W
.
,
Szpiro
,
A.
A.,
&
Kaufm
an,
J.
D
.
.
“
Long
-
te
rm
exposure
to
ai
r
poll
uti
on
and
m
ark
ers
of
infl
amm
at
ion,
coa
gul
at
ion
,
and
endot
helial
a
ct
i
vat
ion:
a
r
epe
a
t
-
m
ea
sures
ana
l
y
s
is
in
the
Multi
-
Et
hnic
Stud
y
o
f
Ather
oscl
ero
sis
(MESA
)
”
.
Epi
demiol
ogy
(
Cambr
idge
,
Mass
.
),
26(3),
310
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
201
8
:
4584
-
4592
4592
[4]
Bee
l
en,
R.
,
Raas
chou
-
Niel
sen,
O.,
Stafoggi
a
,
M.,
Andersen,
Z
.
J.,
W
ei
nm
a
y
r
,
G.,
Hoffm
ann,
B.
,
.
.
.
&
Vine
is,
P.
“
Eff
ec
ts
of
long
-
te
rm
exposure
to
ai
r
pol
lut
ion
on
nat
ura
l
-
ca
us
e
m
orta
li
t
y
:
an
a
naly
s
is
of
22
E
urope
an
cohor
ts
withi
n
th
e
m
ulti
ce
ntr
e ESCA
PE
proje
c
t
”
.
The
La
nce
t
,
2014
;
383(
9919)
:
785
-
795.
[5]
Pio,
D.
V.,
Engler,
R.
,
L
inde
r
,
H.
P.,
Monadj
em, A
.
,
Cott
eri
l
l,
F. P
.
,
Tay
l
or,
P.
J.
,
.
.
.
&
Sal
amin,
N
.
“
Cli
m
at
e
cha
ng
e
eff
ects
on
ani
m
a
l
and
pla
n
t
ph
y
l
ogene
t
ic
div
ersity
in
southern
Af
ric
a
”
.
Global
Ch
ange
Bi
olog
y,
2
014;
20(5),
1538
-
1549.
[6]
EUMETSAT,
ht
tp:
//
ww
w.e
um
ets
at
.
in
t/
websit
e/
h
om
e/
AboutUs
/i
ndex.
html,
13
05
2017.
[7]
Moham
ed
Akram
Zay
t
ar,
Chak
er
El
Am
ran
i,
“
Sequenc
e
to
Seq
uenc
e
W
eathe
r
Forec
asti
ng
with
Long
Short
-
Te
rm
Mem
ory
Re
cur
r
ent
Neur
al
Net
works
.
”
Int
ernati
onal
Journal
of
Computer
A
ppli
cations
,
201
6
;
Volum
e
143
-
No.11
.
[8]
J
y
othi,
M.
N.,
D
ina
kar
,
V.,
T
ej
a,
N.
S.
R.
,
&
Kishore,
K.
N
.
“
NA
RX
Based
Short
Te
rm
W
ind
Po
wer
Forec
asti
ng
Model
”
.
Indone
s
ian
Journal
of
Elec
tri
cal
Engi
n
eer
ing
and
Compu
te
r Sc
ie
nc
e
,
201
5
;
15(1):
20
-
25.
[9]
Khale
l
,
S.
I
.
,
R
a
hm
at
,
M.
F.
,
&
Mus
ta
fa,
M.
W
.
B
.
“
Sensoring
Le
ak
age
Cur
ren
t
to
Pred
ict
Pollut
ion
L
evels
to
Im
prove
Tra
nsm
ission
Li
ne
Model
vi
a
AN
N
”
.
Inte
rnationa
l
J
ournal
of
Elec
tr
ic
al
and
Compu
te
r
Engi
n
ee
ring
,
2017
;
7(1):
68.
[10]
BUF
R
File
Support
Software
,
h
ttp://
ww
w.e
ln
at
h
.
org.
uk/, 13
05
2
017.
[11]
Bac
h,
F.
“
Breaking
the
cur
se
of
dimensional
ity
with
conv
ex
neur
al
ne
tworks
”
.
Journal
of
M
achi
ne
Lear
nin
g
Re
search
,
2014
;
18(19)
:
1
-
53
.
[12]
Jones,
Eric, T
r
av
is Ol
ipha
n
t, a
nd
Pear
u
Pet
erson.
“
SciPy
:
op
en
so
urc
e
sc
ie
n
ti
fi
c
t
o
ols f
or
P
y
thon.”, (2004).
[13]
Abadi,
Martn
,
e
t
al.
“
Te
nsorflo
w:
La
rge
-
sc
ale
m
ac
hine
l
ea
rn
in
g
on
het
ero
g
eneous
distri
bute
d
s
y
stems
.
”,
arXi
v
pre
print a
rXiv
:1
603.
04467
(201
6).
[14]
Ila
m
at
hi
,
P.
,
Sel
l
adur
ai,
V.
,
&
Ba
la
m
uruga
n,
K
.
“
Predic
ti
v
e
m
odelling
and
opti
m
izati
on
of
nit
rog
en
oxide
s e
m
ission
in
coa
l
pow
er
p
la
nt
using
Artif
i
ci
a
l
Neura
l
Ne
t
work
and
Sim
ula
te
d
Anne
al
ing
”
.
IAE
S
In
te
rnati
onal
Journal
of
Ar
ti
ficial Int
el
l
ig
enc
e
,
2012
;
1(1)
:
11.
[15]
Sulist
y
awa
ti
,
S.,
&
Nisa,
I.
“
Cli
m
at
e
Change
an
d
Hea
lt
h
Te
en
a
ger
’s
Perc
eption
s
as
a
Basis
for
Inte
rve
n
ti
ons
”
.
Inte
rnational
Jo
urnal
of Publ
i
c Healt
h
S
ci
en
ce (
IJP
HS)
,
2016
;
5(
3)
:
267
-
273.
[16]
Montana
ro,
T.,
Corno,
F.,
Migl
iore
,
C
.
,
&
Cast
rogiova
nni
,
P.
“
Sm
art
Bike
:
an
I
oT
Crowd
Sensing
Plat
form
for
Monitori
ng
Cit
y
Air
Pollut
ion
”
.
Inte
rnational
J
ournal
of
El
ec
tr
ic
al
and
Computer
Engi
ne
ering
(
IJE
CE)
,
2017
;
7(6)
:
3602
-
3612
.
[17]
Ghane
m
i,
A.
,
&
Boubert
akh
,
B
.
“
Biol
ogical
tool
s to
de
al
wi
th
po
llution:
se
lecte
d
a
dvanc
es
and
nov
el
p
erspe
ctives
”
.
Inte
rnational
Jo
urnal
of Publ
i
c Healt
h
S
ci
en
ce (
IJP
HS)
,
2014
;
3(
1)
:
57
-
62.
BIOGR
AP
HI
ES OF
A
UTH
ORS
Moham
ed
Akram
Zay
ta
r
Obt
ai
n
ed
his
Bac
h
el
or
of
Applie
d
Ma
th
emati
cs
and
Co
m
pute
r
Scie
nc
e
degr
ee
from
The Fac
ulty
of
Scie
n
ce
and
T
ec
hnolo
g
y
,
Ta
ngi
er
in
2
013.
He
recei
ved
his maste
r,
in
Com
pute
r
S
y
ste
m
s
and
Networki
ng
from
the
FS
TT
,
in
2016.
Curre
ntly
,
a
PhD
student
a
t
th
e
FS
TT
.
His
pri
m
ar
y
rese
arc
h
i
nte
rest
are
in
Data
Sci
ence,
Artifi
c
ia
l
In
te
l
ligence,
Mac
h
ine
Le
arn
ing, Cl
oud
Com
puti
ng,
and
Big
Dat
a.
Dr.
Chake
r
El
Am
ran
i
is
Doctor
in
Mathe
m
atic
al
Modell
ing
and
Num
eri
ca
l
Sim
u
la
ti
on
from
the
Univer
sit
y
of
L
i
ege
,
Bel
g
ium
(2001).
He
jo
ine
d
Abdelmale
k
Essaa
di
Univer
si
t
y,
Morocc
o
in
2003.
He
is
cur
r
ent
l
y
Cha
ir
of
th
e
Com
pute
r
Eng
ine
er
ing
Depa
rt
m
ent
at
th
e
Fac
ulty
of
Scie
n
ce
and
T
ec
hnolog
y,
T
angi
er
.
He
lectur
es
d
istri
but
e
d
s
y
stems
and
i
s
prom
oti
ng
High
Perform
anc
e
Com
puti
ng
educat
ion
in
th
e
Uni
ver
sit
y
.
His
rese
a
rch
int
er
ests
in
cl
ude
Cloud
Co
m
puti
ng,
Big
Data
Mining
a
nd
Envi
ronm
ental
Inform
at
ion
S
y
stems
.
Dr.
E
l
Am
ran
is
rese
arc
h
has
bee
n
supported
b
y
n
at
ion
al
and
in
ternat
ion
al
org
anis
m
s.
Dr
El
Am
ran
i
h
as
serve
d
as
an
activ
e
volunt
e
er
in
I
E
EE
Morocc
o
Se
ct
ion
.
He
is
cur
ren
tly
Vi
ce
Ch
a
ir
of
IE
EE
Com
m
unic
at
ion
and
Com
pute
r
Societie
s
Morocc
o
C
hapt
er
,
and
adv
isor
of
the
IE
E
E
Com
pute
r
Societ
y
Student
Branc
h
Ch
apt
er
at
Abdelmal
ek
Essaa
di
Univer
sit
y
.
He
is
th
e
NA
TO
Partne
r
Countr
y
Proje
ct
Dire
ct
or
of
a
real
-
ti
m
e
remote
se
nsi
ng
ini
ti
a
ti
ve
f
or
ea
rl
y
w
arn
ing
and
m
it
iga
ti
on
of
disaste
rs
and
epi
demics in Mo
roc
co.
Evaluation Warning : The document was created with Spire.PDF for Python.