TELKOM
NIKA
, Vol. 11, No. 12, Decem
ber 20
13, pp.
7174
~71
8
1
e-ISSN: 2087
-278X
7174
Re
cei
v
ed
Jun
e
29, 2013; Revi
sed
Jul
y
2
9
, 2013; Acce
pted Augu
st 29, 2013
Retrieving Atmospheric Precipitable Water Vapor
using Artificial Neural Network Approach
Wang Xin*, Deng Xiaobo
, Zhang She
nglan
Che
ngd
u Un
iv
ersit
y
of Inform
ation
T
e
chn
o
l
o
g
y
, Ch
en
gdu,
Chin
a, 61
022
5
Ke
y
Lab
orator
y of
Atmosphere
Sound
in
g, Chi
na Meteor
ol
ogi
cal
Admin
i
st
rati
on, Che
n
g
du, Chin
a, 61
022
5
*Corres
p
o
ndi
n
g
author, e-ma
i
l
:
w
x
w
e
ll
@12
6
.
c
om
A
b
st
r
a
ct
Discussin
g
of w
a
ter vapor an
d its variatio
n i
s
the importa
nt issue for syno
p
tic meteoro
l
o
g
y a
n
d
mete
oro
l
ogy. In physic
a
l Atmosph
e
ric, the mo
isture co
nte
n
t of the earth atmos
p
h
e
re is
one of the
mo
st
importa
nt par
a
m
eters;
it is h
a
rd to r
epres
e
n
t w
a
te
r vapor
beca
u
se
of it
s space-ti
me
variati
on. Hi
gh
-
spectral
reso
lu
tion At
mos
p
h
e
r
ic Infrare
d
So
und
er (A
IRS)
data c
a
n
be
u
s
ed to
retriev
e
the s
m
all
scal
e
vertical structu
r
e of air te
mp
er
ature, w
h
ich
provi
ded
a
mor
e
acc
u
rate
and
go
od i
n
i
t
ial fiel
d for th
e
nu
meric
a
l for
e
casting
and th
e larg
e-sca
le
w
eather an
aly
s
is. T
h
is pap
e
r
propos
es a
n
artificial
ne
ural
netw
o
rk to retrieve the cl
ear
sky atmos
pher
ic radi
at
ion
da
ta from AIRS and co
mpari
n
g
w
i
th the AIRS
Leve
l
-2 stan
da
rd prod
uct, and
gain a g
o
o
d
in
versio
n results.
Ke
y
w
ords
:
ne
ural n
e
tw
ork, AIRS, precipita
b
l
e
w
a
ter vapor, retrieval
Copy
right
©
2013 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
Wate
r vap
o
r i
s
n
o
t o
n
ly the
mo
st a
c
tive
compon
ent
an
d ab
und
ant g
r
een
hou
se
ga
se
s in
atmosp
he
re,
but also a
key factor to
affect clim
atic vari
ation. It plays a
n
im
portant
role
on
maintainin
g ecol
ogi
cal b
a
lan
c
e. So, water va
por is an e
s
se
ntial factor i
n
atmosphe
ric
gree
nho
use effect and the water
cycl
e of the earth
-ai
r
system,
it is the central element fo
r
climate forma
t
ion and clim
ate cha
nge.
Mean
while
, the wate
r vap
o
r content in
atmosp
here is
one
of mai
n
physi
cal
qua
ntities to i
m
p
a
ct
remo
te
sensi
ng
appli
c
ation [1]. Fo
r a lo
ng tim
e
,
the
influen
ce of
water vapo
r
in the
climat
e sy
st
em (su
c
h
as
green
hou
se effe
ct) ha
s not
be
en
discu
s
sed further due to la
cki
ng of accu
rate long
-term stability of the global
wa
ter vapor d
a
ta
record. As a result, dete
c
ting global atm
o
sp
heri
c
moi
s
ture in di
stri
bution and
ch
ange i
s
of gre
a
t
signifi
can
c
e t
o
weathe
r foreca
st, meteo
r
ologi
cal
sup
p
ort work, e
s
p
e
cially
wate
r
circulatio
n a
n
d
climate chan
ge re
sea
r
ch [2].
Half a century, the interna
t
ional co
nven
tional meteo
r
ologi
cal ra
dio
s
on
de net
wo
rk ha
s
been u
s
e
d
in detectin
g
water vap
o
r.
But, this kind of detecti
on mode o
n
l
y
provides t
he
distrib
u
tion of
water va
por over the fixe
d point
s, moreover, the gl
obal
radio so
undin
g
statio
n
network density far
cannot satisfy
the needs
of business and scientific
research
work. Satellit
e
is one of the
best choi
ce
s to obtain atm
o
sp
heri
c
info
rmation be
cau
s
e of its uniq
ue advanta
g
e
in time resol
u
tion and
sp
ace re
sol
u
tion, i
t
can m
a
ke
u
p
for
sou
ndin
g
data’
s in
suf
f
icient of va
st
oce
an, platea
u, dese
r
t and
polar
regi
ons
[3].
At prese
n
t, busin
ess processi
ng AIRS
data mainly
use the
featu
r
e vecto
r
stat
istical
reg
r
e
ssi
on al
gorithm of th
e clea
r air
atmosp
he
re
bu
sine
ss retri
e
ving Intern
ation
a
l MODIS/AIRS
processi
ng
software package IMAPP
(I
nternational MODIS/AIRS
Pr
eprocessi
ng Package), t
h
is
statistic meth
od is sim
p
le,
so th
e retri
e
val accu
ra
cy is largely l
i
mited [4]. Artificial n
eura
l
netwo
rks ha
ve been pro
v
en as a rel
i
able tec
hniq
ue to diagno
se an
d have
good lea
r
ni
ng
cap
ability [5], with the d
e
v
elopment of
artificial n
e
u
ral n
e
two
r
k, neural net
work ret
r
ieving
atmosp
he
ric
comp
one
nts
has be
en
gre
a
tly applie
d
as
an
effecti
v
e metho
d
, the m
a
in
ben
efit
from it in the work is its high sp
eed
and the pot
e
n
tial for high
speed, a
s
well as its fa
ult
toleran
c
e
abi
lity. Neural
n
e
twork
ca
n d
i
rectly
d
edu
ce the
com
p
l
e
x and
un
cle
a
r
relation
shi
p
betwe
en i
n
p
u
t-output
fro
m
the t
r
aini
ng d
a
ta
without d
o
ing
any a
s
sumpt
i
on a
bout
d
a
ta
distrib
u
tion, so it is very suitable to be put in
to
h
y
p
e
r
s
p
ec
tra
l
r
e
mo
te
s
e
ns
ing r
e
tr
ie
va
l w
hos
e
data with the
cha
r
a
c
teri
stics
of high di
mensi
on, st
ro
ngly co
rrel
a
te
d, and noi
se
-sen
sitive[6]. In
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
Retrie
vin
g
Atm
o
sphe
ric Precipita
b
le Wa
ter Vapor
u
s
i
ng Artificial Neural
Network… (Wa
ng Xin)
7175
addition, a
simple topol
o
g
ical
stru
cture with
only
one hid
den l
a
yer can a
c
hieve the be
st
approximatio
n in the rang
e of allowabl
e
erro
r [7].
Resea
r
ch sh
ows that the method of retrievi
ng
pre
c
ipitabl
e
water vapo
r
by make u
s
e
of neu
ra
l net
work perfo
rm
ance
ha
s hig
her accu
ra
cy.
It
will play a
cruci
a
l rol
e
in storm
predi
c
tion,
num
eri
c
al
weather
forecast and other extreme
weath
e
r p
r
edi
ction by usi
n
g
the method o
f
neural
net
work
retrievin
g
pre
c
ipitabl
e water vapo
r.
2.
The Propos
e
d
Algorithm
Known
satelli
te inst
rume
nts in
different
spe
c
tral
b
and
of radiatio
n
observation,
then t
o
ascertai
n at
mosp
he
ric te
mperature
a
nd ab
so
rptio
n
ga
s (wate
r
vapor), this is the
retrie
val
probl
em. The retri
e
val met
hods
we
com
m
only use
are: GPS inversion,
satellite hyperspect
r
al
remote sensi
ng retrieval,
sate
llite i
n
frared remote sensing retr
ieval and MODIS near inf
r
ared
retrieval. In t
h
is p
ape
r, a
kind
of metho
d
ba
sed
on p
r
inci
pal
com
p
onent a
nalysi
s
(P
CA) n
eural
netwo
rk p
h
ysi
c
al stati
s
tical
retrieval i
s
ad
opted.
2.1. Data Sel
ection
AIRS is
one
of the
six mai
n
observation in
struments
of NASA Aqua satellite platform,
Infrare
d
ra
di
ation mate
ria
l
has th
e a
d
vantage
s of
more
ch
an
nels, mo
re i
n
formatio
n a
nd
narro
wer
sp
e
c
trum. It ha
s 2378 i
n
fra
r
e
d
sp
ec
trum
chann
els
sp
ectrum a
rray from 3.7
μ
m t
o
15.4
μ
m, ca
n
provide
atmosp
heri
c
info
rmation from
t
he groun
d to
the height of
40 km. Aqu
a
satellite
goe
s thoug
h
Chi
na a
bout B
e
i
jing time
1:3
0and
13:30
twice eve
r
yda
y
[8]. We
ca
n
download the requi
red data by visiting
http://disc.sci.gsfc.
nasa.gov/AIRS.
Usi
ng
satellit
e hyperspectral data
retri
e
val at
mospheric precipitat
ion i
s
hard t
o
meet
the training requireme
nt if only using ra
dial brig
ht
ne
ss from satellite actually, it i
s
ne
ce
ssary to
simulate AI
RS radial
brig
htness. And
AIRS radial
brightn
e
ss d
a
ta sim
u
latio
n
is the
ba
si
s on
usin
g AIRS o
b
se
rvation
of
hypersp
ect
r
al data
retri
e
val. Afte
r re
searchin
g a
n
d
discu
s
sing, t
h
is
pape
r choo
ses
University of Wisc
o
n
si
n-Ma
dison gl
obal
clea
r air atmosp
he
re
profile traini
ng
sampl
e
s
(Se
e
Bor Ve
rsi
o
n
5.0) an
d CRTM (Atmo
s
ph
eric
ra
diative transfe
r mo
d
e
l) to si
mulat
e
the observati
on radi
al brig
htness inform
ation fr
om at
mosp
he
ric inf
r
ared dete
c
to
r AIRS [9].
It is nece
s
sary to establish
global temp
er
atu
r
e humi
d
ity and atmosp
heri
c
pa
ramete
rs
vertical di
strib
u
tion datab
ase whe
n
we
re
trieve
atmosp
heri
c
pa
ramet
e
rs
by usin
g the metho
d
of
statistical inv
e
rsi
on m
e
tho
d
, statistics-p
h
ysical
retrie
val method
o
r
ne
ural
network metho
d
. In
addition, we
must kn
ow
th
e
surfa
c
e
te
mperature
a
n
d
nu
meri
cal
e
m
issivity d
ue
to the i
n
fluen
ce
of the su
rface. In orde
r to
meet
the ap
plicatio
n of re
mote se
nsi
n
g
retrieval
de
mand, University
of Wi
sconsin-Madi
son m
e
teorol
ogi
cal
satellite application res
earch i
n
stitut
e researchers
synthe
size th
e glo
bal
cle
a
r air atmo
sp
he
re
prof
ile
of t
he training
sa
mples (CIMS
S
, Coo
perative
Institute for
Meteorologi
cal Satellite Studies,
Un
iversity of Wi
sconsi
n
-M
adi
son) after
a great
deal
of re
sea
r
ch
analy
s
is experim
ental work, whic
h
contai
ns
a total of abo
ut 1
5704
gro
u
p
s
of
pre
c
isi
on of a
t
mosph
e
re profile lines, after pr
ocessing
, yields a total of 2,099,250 (155
50
×3
×3
×15
)
group
s
of atmosp
here profile lin
e
s
. At
last, we cho
o
se 23,
325 g
r
oup
s
of atmosp
here
profile lin
es
due to th
e computing
ne
eds
a lot
of
time and
so
me of the
co
mbination
is
not
necessa
rily reasona
ble, th
en we get
a t
o
tal of 23,32
5
AIRS gro
u
p
s
radial
bri
ghtn
e
ss (at-sen
s
o
r
radia
n
ce)
by usin
g gl
obal
atmosphe
ric model
a
fter cal
c
ulatin
g
every group
of atmosphe
re
profile lin
es
by CRT
M, the sele
cted
2
3
,325 g
r
ou
ps are the trai
ning sampl
e
data of neu
ral
network
data s
e
t in turn.
2.2. Data Pre
p
roces
sing
The
solution of the
problem that
sat
e
llit
e data retrieving atm
o
spheri
c parameters
(mainly
retrie
ving tempe
r
a
t
ure a
nd h
u
m
idity vertic
a
l
profile li
ne) is not th
e o
n
ly. On the
one
hand, the
in
formation
provided by
satellite dat
a
is in
suffici
e
n
t; on the
other
han
d,
the
information provided by
satellite data i
s
sometim
e
s unusefully
whil
e retrieving.
I
t
is necessary
to screen d
a
ta as well as redu
ce the ob
serv
atio
n dat
a dimen
s
ion
and elimin
ate
redun
dan
cy, so
it is very ne
cessary to
use the no
nline
a
r d
a
ta
processing
algo
rithm. In the
study of neu
ra
l
netwo
rk retri
e
ving
water
vapor, the
first qu
estio
n
required to
so
lve is
data
prepro
c
e
s
sing,
at
pre
s
ent, the
main metho
d
we u
s
e is PCA [10].
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN: 2
087-278X
TELKOM
NIKA
Vol. 11, No
. 12, Dece
mb
er 201
3: 717
4 – 7181
7176
3. Res
earc
h
Method
3.1. PCA (Pri
ncipal Comp
onent
Analy
s
is)
To re
set the
origin
al vari
able
s
to a n
e
w
set of fe
w varia
b
le
s
that each ot
her i
s
indep
ende
nt, at the same
time, accord
ing to a
c
tual
need
s, we can take out
a few le
ss to
tal
variable
s
to
reflect
the
original va
riabl
es i
n
fo
rm
atio
n
a
s
m
u
ch as po
ssi
ble, this statistica
l
approa
ch i
s
calle
d p
r
in
cip
a
l compo
nen
t analysi
s
, a
nd it is a
ki
nd of p
r
o
c
e
s
sing
dime
nsi
o
n
redu
ction m
e
thod in mathe
m
atics.
Its basi
c
i
dea
is to
re
set m
any ori
g
inal i
ndexe
s
which
have
certai
n
correl
ation (such as
the num
be
r i
s
P) to a
n
e
w set
of
comp
rehen
sive i
n
d
e
xes th
at ea
ch othe
r i
s
i
n
d
epen
dent, th
en
repla
c
e
the o
r
iginal
indexe
s
by the
ne
w indexe
s
. It is u
s
ually to
put the P in
d
e
xes into
line
a
r
combi
nation,
then take the
m
as n
e
w
co
mpre
hen
sive
indexe
s
. The
cla
ssi
cal a
p
p
r
oach is to u
s
e
the varia
n
ce
of F1
(the f
i
rst lin
ear co
mbinati
on,
n
a
mely the first comp
reh
e
n
sive in
dex)
to
expre
ss, the
bigge
r Var (F
1) is, the mo
re info
rmatio
n F1co
ntain
s
. The F1 shoul
d be the bigg
est
varian
ce in al
l of the linear combi
nation,
so F1 i
s
call
ed a
s
the first princi
pal co
mpone
nt. If the
first prin
cip
a
l comp
one
nt is not enoug
h to rep
r
e
s
ent the origi
nal inf
o
rmatio
n of the P indexes,
then
con
s
ide
r
sele
cting F
2
whi
c
h
is the
se
cond
li
near combin
ation, in o
r
d
e
r to
refle
c
t
the
origin
al information effecti
v
ely, the existing
informati
on of F1 will
not need to a
ppea
r agai
n in
the F2, if it requires that
Cov (F
1, F2
)=0,
we call F2
the se
co
nd
p
r
in
cipal comp
one
nt,
By
analogy, we
can
structure the third principal
com
ponent,the forth
prin
cipal com
ponent
and so
on.
The main
ste
p
s of prin
cip
a
l
compo
nent
analysi
s
a
s
follows:
(1)Index data standardi
z
ati
on (
SPSS sof
t
ware automatically);
(2)
Jud
g
ing th
e relation
shi
p
betwee
n
the index data;
(3)
Confirmin
g
the numbe
r of princip
a
l compon
ents
(m);
(4) O
b
tainin
g the expre
ssi
o
n
of the princi
pal com
pon
e
n
t Fi;
(5)
Namin
g
the prin
cipal
co
mpone
nt Fi
In recent years, as te
chnolo
g
y advanc
es a
nd
mature
s, PCA gets co
ntinuou
s
developm
ent, the princi
pal
compo
nent
s transfo
rm (P
CT) meth
od
and the proj
ected p
r
in
cipl
e
comp
one
nts (PPC) tran
sfo
r
m method ap
pears succe
s
sively.
Princi
pal com
pone
nt analy
s
is te
chni
que
is in co
ntinu
ous im
prove
m
ent and p
e
rfection
with tech
nica
l prog
re
ss, th
e proj
ectio
n
prin
cipal
com
pone
nt (proje
cted p
r
in
ciple
comp
one
nts,
PPC) tra
n
sfo
r
m propo
se
d
by Blackwell
has
been
proved to have
better dim
e
n
s
ion
red
u
ctio
n
perfo
rman
ce
[11]. Comp
ared with
the t
r
adition
al p
r
i
n
cip
a
l comp
o
nent, the p
r
o
j
ection
of the
prin
cipal
com
pone
nt ha
s b
e
tter pe
rform
ance in
retri
e
ving atmo
sph
e
ric tempe
r
at
ure
profile
lin
e.
Mean
while, B
l
ackwell indicates, the use
of 35
PPC can make ret
r
ieval error to
the minimum
value in the
o
r
y on water
vapor
retriev
ed by
u
s
ing
AIRS data [1
2]. So, the PPC tran
sform
method i
s
ad
opted i
n
this
pape
r. We fin
a
lly se
le
ct
4
1
wate
r
vap
o
r cha
nnel (4
1 PPC)
from
th
e
2378
ch
anne
ls after a
naly
z
ing lite
r
atu
r
e and
doi
ng
many expe
ri
ments, then
we g
e
t 24 P
P
C
though d
eali
ng with 41
PPC by usi
ng the meth
od of PCA, the AIRS ra
dial brig
htne
ss
dimen
s
ion i
s
red
u
ced fro
m
the o
r
igin
al 237
8 to 2
4
. This t
r
an
sformation
re
alize
s
the
da
ta
dimen
s
ion
re
ductio
n
a
s
well a
s
keep
s
the rel
a
ti
ve
more
informa
t
ion, co
nsequ
ently sh
orts the
neural network trainin
g
time and imp
r
ov
es t
he efficie
n
cy of the pro
posed alg
o
rit
h
m.
Known
satelli
te inst
rume
nts in
different
spe
c
tral
b
and
of radiatio
n
observation,
then t
o
ascertai
n at
mosp
he
ric te
mperature
a
nd ab
so
rptio
n
ga
s (wate
r
vapor), this is the
retrie
val
probl
em. The retri
e
val met
hods
we
com
m
only use
are: GPS inversion,
satellite hyperspect
r
al
remote sensi
ng retrieval,
sate
llite i
n
frared remote sensing retr
ieval and MODIS near inf
r
ared
retrieval. In t
h
is p
ape
r, a
kind
of metho
d
ba
sed
on p
r
inci
pal
com
p
onent a
nalysi
s
(P
CA) n
eural
netwo
rk p
h
ysi
c
al stati
s
tical
retrieval i
s
ad
opted.
3.2. Neural Net
w
o
r
k
BP neural n
e
t
work is the
most wi
dely u
s
ed
kind of
n
eural n
e
two
r
k in all artificial
neural
netwo
rk, it also call
ed error ba
ck p
r
op
agation ne
ural netwo
rk, it is a feed forwa
r
d net
wo
rk
comp
osed by
the nonlinea
r transfo
rmatio
n unit [13
].But this paper ul
timately choo
se
s multilayer
feed forward neural network radi
al ba
sis functi
on ne
ural netwo
rk
RBF (Radi
cal
Basis F
u
n
c
tio
n
)
whi
c
h
ha
s th
e be
st
app
ro
ximation pe
rf
orma
nce a
fte
r ma
ny te
st a
nd a
nalysi
s
, t
he
RBF
network
avoids trivial
lengthy
co
m
putation
of b
a
ck
p
r
o
pagat
ion b
e
twe
en
the inp
u
t lay
e
r
and
hid
d
e
n
layer, it makes learning
4
3
10
~
10
times fa
ster tha
n
usu
a
l BP ne
ural n
e
two
r
k [
14]. Figure 1
is the
neural network topology. As sho
w
n in figure, t
he inp
u
t layer has
24 neu
ron
s
, output layer ha
s
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
Retrie
vin
g
Atm
o
sphe
ric Precipita
b
le Wa
ter Vapor
u
s
i
ng Artificial Neural
Network… (Wa
ng Xin)
7177
one ne
uro
n
, two hid
den lay
e
r co
ntainin
g
respe
c
tively (18, 9) neu
ro
n
s
.
Figure 1. The
Neural Net
w
ork T
r
aini
ng
Topolo
g
y
This
pap
er ta
ke
s AIRS rad
i
al brig
htne
ss PPC
data
si
mulated
by CRTM a
s
th
e i
nput of
the neu
ral n
e
t
work; corre
s
pondi
ngly takes the
atmo
s
pheri
c
p
r
e
c
ipi
t
ation data from Unive
r
sity of
Wisco
n
si
n-M
adison
clea
r
sky p
r
ofile li
n
e
s d
a
ta
a
s
t
he outp
u
t of
the neu
ral
n
e
twork. At first,
sen
d
ing
the
input-o
utput
d
a
ta set to
mul
t
ilayer fe
e
d
fo
rwa
r
d
ne
ural
netwo
rk trai
ni
ng, choo
sin
g
a
optimize
d
ne
ural net
work
as a practi
ca
l applic
atio
n algorith
m
; Then the optimi
z
ation ne
ural
netwo
rk
who
s
e si
mulation
erro
r is
sma
ll enoug
h
is
applie
d to actual AIRS detection data,
at
last, we verify
the pre
c
isio
n
and stability of the algorith
m
.
After analy
s
is and
processing, the
24 P
P
C
ha
ndle
d
by
prin
cip
a
l compon
ent
a
n
a
lysis
techn
o
logy transfo
rmatio
n
are t
r
eate
d
as the
in
p
u
t of the ne
ural
netwo
rk, PWV a
s
the
o
n
ly
output of the neural network. The atmo
spheri
c
p
r
e
c
ipi
t
ation data which
corre
s
po
nd to the input
of the
neu
ral
network f
r
o
m
University
of Wi
sc
on
sin
-
Madi
so
n
cle
a
r
sky p
r
ofile
of the
trai
ning
sampl
e
s (Se
e
Bor Ve
rsion
5.0) i
s
tre
a
te
d a
s
the
ta
rg
et output of t
he n
eural net
work,
we
put
the
AIRS hyper spe
c
tral d
a
ta
simulated b
y
the
neural
network int
o
CRT
M, an
d con
s
id
er the
cal
c
ulate
d
re
sults
as
actu
a
l
output, we
calcul
at
e the root mean
sq
u
a
re e
r
ror b
e
t
w
ee
n the a
c
t
ual
output
a
nd
ta
rget output calcul
ated
i
s
0
.
063g/ cm2,
while
u
s
ing
st
atistical
ap
proach, a
s
sho
w
n
in Figure 2.
Figure 2. Co
mpari
s
o
n
bet
wee
n
the PWV Retrie
ved from Neu
r
al
Network ba
se
d
on Algorithm
and PWV fro
m
University of Wiscon
sin
-
Madison.
4. Resul
t
s
and
Analy
s
is
This p
ape
r se
lect typical re
gion of low
a
nd middl
e latitude area a
s
the re
sea
r
ch
obje
c
t,
middle l
a
titud
e
area’
s surf
ace type
is complex, in
clu
de: mou
n
tain
, stepp
e, de
sert, rivers
an
d
glaci
e
r, et
c. surface temp
e
r
atur
e emi
s
si
vity and atmo
sph
e
ri
c p
r
e
c
i
p
itation’s sp
a
t
ial
distri
bution
has a la
rge
chan
ge; the low-l
a
titud
e
area, mai
n
ly surrou
nd
ed by ocea
n
and islan
d
, its
temperature i
s
hig
h
, the water vapo
r
co
ntent is
rich, and
atmo
sp
h
e
ric prec
i
p
itation’s distri
buti
o
n
cha
nge
s very largely in time and
spa
c
e. We
ca
n
effectively verity the appli
c
ability of the
algorith
m
s while ta
king th
e two
area
s
as
an
exam
p
l
e. In orde
r t
o
validate
th
e multilaye
r f
eed
forward neural net
work i
n
version
al
gori
t
hm’s accuracy and feas
ibility(which based on
pri
n
ci
pal
comp
one
nt a
nalysi
s
),
we
select
a g
r
ou
p
of AIRS
L
1
B
infra
r
ed
ra
di
al bri
ghtne
ss
data, only
pa
rt
of the measu
r
ed data are listed to test
the result owin
g to thesis le
ngth.
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN: 2
087-278X
TELKOM
NIKA
Vol. 11, No
. 12, Dece
mb
er 201
3: 717
4 – 7181
7178
4.1. Lo
w
-
lati
tude Re
gion Wa
ter Vapor
Retriev
a
l Results and
An
aly
s
is
One AI
RS le
vel-1B im
age
cove
rin
g
(0N-22
N, 6
0
E-8
0
E
) on
May
9t
h, 201
2 at
08
:47:24-
08:53:24 (uni
versal time
) is sel
e
cte
d
for studying lo
w-latitude regi
o
n
.
In ord
e
r to
di
scuss the
infl
uen
ce of th
e
clou
ds,
we
d
on’t eliminate
pollution
clo
uds
pixel
in the inversi
on process, i
n
stea
d ret
r
iev
e
the whole
scen
e imag
es.
Becau
s
e
of o
b
taining a
wid
e
rang
e of atm
o
sp
heri
c
p
r
e
c
ipitation m
e
asu
r
ed
data
is difficult, we analy
z
e an
d com
p
a
r
e the
algorith
m
me
ntioned in thi
s
pap
er
with
AIRS
Level-2 stand
ard p
r
odu
cts
whil
e
doing reality
testing.
Figure 3. The
Spatial Distri
bution of PWV
Retrieved from Neural Network
Figure 4. The
Spatial Distri
bution of PWV
Retrieve
d fro
m
AIRS Level-2 Standa
rd
Produ
ct
Figure 5. The
Absolute Error of PWVs
Retrieve
d by Neu
r
al Netwo
r
k an
d AIRS
Produ
ct
Figure 6. The
Cloud
Distri
b
u
tion from AIRS
Level-2 Stan
dard Produ
ct
The spatial d
i
stributio
n
of PWV
ret
r
ieve
d
by ne
ural
netwo
rk an
d
AIRS PWV p
r
odu
ct
are
sho
w
n
by Figure 3 a
n
d
Figu
re 4.
Compa
r
ed
to fi
gure
6, it can
be seen th
at blue a
r
ea fo
r
little clou
d a
r
ea, an
d
red
yellow area
for clo
udy a
r
e
a
.
Due
to the
se
lected
a
r
ea
is larger,
mari
n
e
climate type
i
s
spe
c
ially, t
he maxim
u
m
and
minim
u
m wate
r va
p
o
r
conte
n
t dif
f
eren
ce i
s
big
a
t
the sa
me ti
me. Figu
re 5
,
interpol
ated
throu
gh AIRS stand
ard
prod
uct
s
, ca
n analy
z
e th
eir
differen
c
e i
n
tuitively, the root
me
an
sq
uare
e
rro
r i
s
0.7552
g/cm
2
,
it can intuitiv
ely find out the
absolute erro
r distri
bution
bet
we
en PWV retrieved b
y
neural n
e
twork
and AIRS
PWV prod
uct.
As is shown in Figure 5, RMSE is sm
aller in li
ttle clou
d are
a
; similarly, RMS
E
is bigge
r in
clou
dy area,
the absolut
e error di
stri
bution
mat
c
h
e
s well with
the AIRS Level-2
stand
ard
prod
uct
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
Retrie
vin
g
Atm
o
sphe
ric Precipita
b
le Wa
ter Vapor
u
s
i
ng Artificial Neural
Network… (Wa
ng Xin)
7179
4.2. Middle-latitud
e
Re
gion Wa
ter Va
por
Re
triev
a
l Results and
Analy
s
is
One AIRS l
e
vel-1B ima
ge coverin
g
(32
N
-54
N
, 66E-9
7E) o
n
May 2th, 2012
at
20:47:24
-20:
53:24 (u
niversal time) i
s
se
lected for
stu
d
ying middle
-
latitude regi
o
n
.
Figure 7. The
Spatial Distri
bution of PWV
Retrieved from Neural Network
Figure 8. The
Spatial Distri
bution of PWV
Retrieve
d fro
m
AIRS Level-2 Standa
rd
Produ
ct
Figure 9. The
Absolute Error of PWVs
Retrieve
d by Neu
r
al Netwo
r
k an
d AIRS
Produ
ct
Figure 10. Th
e Clou
d Di
stri
bution from A
I
RS
Level-2 Stan
dard Produ
ct
The su
rfa
c
e
climate type sele
cted in t
h
is
pa
per i
s
rathe
r
com
p
l
e
x, include
mountain
land, la
ke
s,
rivers, etc. T
he
water va
por
ch
ang
es larg
ely in ti
me an
d
spa
c
e. Th
e
spat
ial
distrib
u
tion
of
PWV
retri
e
ved by
ne
ural
netwo
rk
a
nd AIRS
PWV produ
ct
a
r
e sh
o
w
n by
Fig
u
re 7
and Fig
u
re
8. Figure 9, interpol
ated
throug
h AIRS stan
dard
prod
uct
s
, can analy
z
e t
heir
differen
c
e i
n
tuitively, the root
mea
n
squ
a
re erro
r
is
0.
3359
g/cm
2
, whic
h more
c
l
os
e to and more
fit the AIRS standard prod
ucts. As
sh
o
w
n in Fig
u
re 9 and Fig
u
re 10, we
can fi
nd the ab
solu
te
error di
strib
u
tion match
e
s
well with the
clou
d di
stri
bu
tion of AIRS
Level-2
stand
ard inve
rsi
o
n
prod
uct
s
.
This pa
per
mainly resea
r
ch the atmo
sph
e
ri
c pre
c
i
p
itation wate
r vapor in cl
ear ai
r
con
d
ition, but
from the AIRS
Level-2 sta
ndard produ
ct we can
see
wheth
e
r lo
w latitude regi
on
or middl
e latitude re
gion
has a
ce
rtain
amount
of cloud, no do
u
b
t which will
cau
s
e certai
n
influen
ce to t
he inve
rsio
n
result. According to
the
a
nalysi
s
an
d compa
r
ison, we ca
n
see th
e
absolute erro
r of the low-latitude regi
o
n
is bi
gg
er th
an the middl
e latitude re
g
i
on’s. Th
e main
rea
s
on i
s
th
at the tempe
r
ature is hi
g
h
in low l
a
titude regio
n
, mean
while, t
he wat
e
r va
por
conte
n
t is ri
ch and it chan
ges ve
ry larg
ely in ti
me and space, so
the absol
ute
error i
s
slig
h
t
ly
bigge
r.
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN: 2
087-278X
TELKOM
NIKA
Vol. 11, No
. 12, Dece
mb
er 201
3: 717
4 – 7181
7180
In addition, this stu
d
y ha
s certai
n sig
n
i
f
ic
an
ce for p
opula
r
ization,
although
we
sele
ct
low an
d middl
e regio
n
s a
s
subj
ect to stu
d
y, it also applies to di
scu
ss hi
gh latitud
e
area.
5. Conclu
sion
Information reso
urce
s integratio
n in coll
abo
rative logisti
cs n
e
twork m
a
kes u
s
e o
f
sup
e
rio
r
re
so
urces
in network, achieve
s
the
pu
rp
ose of
sh
arin
g
n
e
twork re
sou
r
ces by
tech
nical
mean
s. It m
a
inly virtual
i
n
tegrate
s
the
re
so
urce
s whi
c
h have comp
etitive advantag
es and
geo
spatial di
screte distri
b
u
tion
in
n
e
twork, stre
ngth
ens
co
he
sive
relation
s, im
prove
s
resou
r
ce
respon
se
spe
ed and redu
ces re
so
urce redun
dant wa
ste [15].
Neu
r
al n
e
twork
nonli
nea
r retri
e
val a
l
gorithm n
o
t only sho
w
s the sta
b
ili
ty and
effectivene
ss of the statisti
cal reg
r
e
ssi
o
n
met
hod, bu
t
also sho
w
s the
ac
cu
ra
cy
of the physi
cal
inversi
on m
e
thod. Th
e alg
o
rithm i
s
e
s
p
e
cially
suitabl
e for atm
o
sp
heri
c
remote
sen
s
in
g retrie
val
probl
em ba
sed on the
ch
ara
c
teri
stics
of the
neu
ral
netwo
rk a
n
d
the nonlin
ea
r, non
-Gau
ssi
an
relation
shi
p
o
f
atmosph
e
ri
c remote
se
nsi
ng data.
With
the rapi
d de
velopment of
netwo
rk, o
p
e
n
and re
co
nfig
urabl
e netwo
rk facility ca
n
well achi
ev
e
the goal for multi-net
work integration,
so
that the facility for NGN (n
ext generatio
n netwo
rk) wil
l
be more a
n
d
more po
pula
r
[16].
AIRS infra
r
ed
dete
c
tion d
a
t
a ca
n di
spla
y the su
btle atmosp
he
ric stru
cture, it contain
s
more info
rma
t
ion several
hund
red time
s than p
r
evio
us dete
c
tion
instru
ment, the method
o
f
neural net
wo
rk is
used in t
h
is p
ape
r to
handl
e AI
RS
data, then inv
e
rt atmo
sph
e
r
ic p
r
e
c
ipita
b
le
water va
po
r, finally we compa
r
e the
PWVs re
triev
ed by neu
ral
netwo
rk
an
d AIRS Leve
l
-2
stand
ard p
r
o
duct, the re
sult sho
w
s th
at Neural
ne
twork ret
r
ieval algorith
m
is simpl
e
an
d
feas
ible, what
’s
more, its
error is
rather small.
In additio
n
, it is
wo
rth
our attention i
s
t
hat
the retri
e
val
spatial resol
u
tion of neural
netwo
rk
algo
rithm is 13.5
k
m [17], the spatial re
sol
u
tion of AIRS st
anda
rd p
r
o
d
u
c
t is 4
5
km, b
y
comp
ari
s
o
n
, this alg
o
rithm
improve
s
the
spatial re
sol
u
tion 2 times higher tha
n
AIRS standa
rd
prod
uct, T
h
e
r
efore, the
neural n
e
two
r
k algo
ri
thm
produ
cts can b
e
tter
re
flect the
act
u
a
l
distrib
u
tion of
wate
r vapo
r, com
pared
wi
th AIRS
stan
dard
produ
ct
s, the fine
r le
vels have
be
en
obviou
s
ly improved. It ma
ke
s g
r
eat
se
nse to
exactl
y descri
p
t an
d analy
z
e th
e exch
ang
e
and
transpo
rt of water vapo
r, and discu
s
s sma
ll ran
ge of atmosp
he
ric
state activitie
s
.
Ackn
o
w
l
e
dg
ement
T
h
e
da
ta
us
ed
in
th
is
s
t
udy a
r
e
ac
qu
ired
from Unive
r
sity
of Wisconsi
n
-M
adi
sion,
and
the auth
o
rs than
k the
Key Labo
rat
o
ry of At
mo
sph
e
re
Sou
n
d
ing,
Chin
a
Meteo
r
olo
g
i
c
al
Adminis
t
ration for this
work
s
u
pport.
Referen
ces
[1]
HL Sen
i
or. Cli
mate and W
a
te
r. Englan
d:
Nat
u
re
. 200
2; 419
(690
3).
[2]
E Smith, G Asr
a
r, Y F
u
ru
ham
a et
al. Inter
nat
io
n
a
l Gl
ob
al Pr
ecipit
ation
Me
a
s
ureme
n
t (GPM) Progr
am
and Miss
ion: A
n
Overvie
w
. In Meas
uri
ng Pre
c
ipitati
on F
r
om
Space.
USA: Springer
. 20
07
; 611-65
3.
[3]
W
ang Yin
g
, Hua
ng Yo
ng,
Si
yuan
Hu
an
g. A Pr
elimi
n
ar
y
Stu
d
y
of
the Retriev
a
l
Methods for
Atmospher
e an
d Humi
dit
y
Pr
o
f
iles.
Re
mote S
ensi
ng for La
n
d
& Resourc
e
s
.
2008; 3(1).
[4]
Xu
eh
ui Z
h
a
ng,
Guan Li, Z
h
e
nhu
i W
ang, H
an Ji
n
g
. Retri
e
vin
g
Atmosp
heric T
e
mpera
t
ure Profiles
Using Artifici
al
Neur
al Net
w
o
r
k Approac
h. 20
09; 11(3
5
).
[5]
Li Guop
in
g, Z
hang Qin
g
w
e
i, Ma Xi
ao. Com
b
in
at
ion
of F
ault
T
r
ee and N
e
ural N
e
t
w
orks i
n
Exc
a
vato
r
Diag
nos
is.
TEL
K
OMNIKA
. 2013; 11(4): 1
787
-179
6.
[6]
Hornik
K, Stin
chcomb
e M,
White H. M
u
ltil
a
y
er
fe
edfor
w
a
rd n
e
t
w
orks
ar
e u
n
ivers
a
l
ap
pro
x
imat
ors.
Neur
al Netw
or
ks
. 1989; 2(5): 359-
366.
[7]
A Galla
nt, H
W
h
ite. T
here exists a
neur
a
l
net
w
o
rk that
does
not m
a
ke avo
i
d
abl
e
mistakes.
In
Procee
din
g
s of
the Internatio
n
a
l Joi
n
t Confer
ence o
n
Ne
ural
Netw
orks
. 1988; 657–
66
4.
[8]
W
enhu
a
Gao
，
F
engsh
e
n
g
Z
hao
，
C
han
gs
ong Ga
i. Val
i
d
a
tion
of Airs
Retriev
a
l T
e
mperatur
e a
n
d
Moisture Pr
od
ucts an
d the
i
r
Appl
icatio
n i
n
numeric
al M
o
d
e
ls.
Acta Met
e
orol
ogic
a
Si
nic
a
.
20
06; 6
4
(3): 2712
28
0.
[9]
Shen
gla
n
Z
h
a
ng. A Stu
d
y
of Preci
p
itab
le
W
a
ter Va
por
Retriev
a
l
Usi
ng Infrar
ed
H
y
p
e
rsp
e
ctral
Soun
din
g
s (de
g
ree p
a
p
e
r) Ch
eng
du; Ch
eng
du Un
iversit
y
o
f
Information T
e
chn
o
lo
g
y
. 201
1.
[1
0
]
JA So
b
r
i
n
o
,
ZL
L
i
, MP Sto
l
l
,
e
t
a
l
. Mu
l
t
i
-
ch
ann
e
l
an
d
mul
t
i
-
an
g
l
e
a
l
g
o
r
i
t
h
m
s fo
r e
s
ti
ma
ti
ng
se
a
and
lan
d
surface te
mperatur
e
w
i
th
AT
SK data.
Internati
ona
l Jour
nal of Re
mote Sensi
n
g
. 19
96;
17: 2089-
211
4
.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
Retrie
vin
g
Atm
o
sphe
ric Precipita
b
le Wa
ter Vapor
u
s
i
ng Artificial Neural
Network… (Wa
ng Xin)
7181
[11]
W
J
Black
w
el
l. A Neur
al-N
etw
o
rk T
e
chniq
u
e
for the
Retr
ieval
of Atmos
pher
ic T
e
mperature a
n
d
Moisture Pr
ofil
es from Hi
gh
Spectral
Reso
l
u
tion S
o
u
ndi
n
g
Data.
IEEE
Trans.
Geosci. Rem
o
te
Sensi
n
g
. 20
05;
43(11): 25
35-
254
6.
[12]
William J Bl
ac
k
w
e
l
l
，
F
r
e
deri
ck W
Chen. Neura
l
Net
w
ork
s
in Atmosph
e
r
ic Remote Se
nsin
g. MIT
.
200
9.
[13]
T
i
e W
ang, Chao W
a
ng. B
o
ra En
gin
e
C
H
Em
issio
n
s
Diag
nos
is Bas
ed o
n
Ne
ural
Net
w
ork.
TEL
K
OMNIKA
. 2012; 1
0
(8): 2
343-
235
0.
[14]
Lia
ngj
un Z
h
an
g, Cao J
i
n
g
, Shizh
ong
Jia
ng.
Neur
al N
e
t
w
or
k Practical T
u
torial.
Chi
na M
a
chin
e Press.
200
8; 2.
[15]
Xi
aofe
ng Xu, Jinlo
u
Z
h
ao, Yirui Den
g
.
In
tegr
ated
Opti
mizatio
n
a
nd
Depl
o
y
me
nt M
e
cha
n
ism
of
Information
R
e
sourc
e
s in
Compl
e
x Ma
n
u
facturin
g C
o
ll
abor
ative
Log
i
s
tics Net
w
ork.
Journ
a
l of
T
heoretic
al an
d Appl
ied Infor
m
ati
on T
e
ch
no
logy
. 20
12; 44(
1): 153 – 1
60.
[16]
Z
hou Mi
nhu
i, W
ang W
e
imi
n
g,
Z
hou Ji
ngj
in
g. Routi
ng Opt
i
miza for F
o
rce
s
Based
on T
r
affic Matrix.
Journ
a
l of T
h
e
o
retica
l and A
p
plie
d Infor
m
ati
on T
e
chn
o
l
ogy
. 2012; 44(
1); 007 - 011
[17]
Shen
gla
n
Z
h
a
ng, L
i
sh
eng
Xu, Jil
i
e
Din
g,
Ha
il
ei
Li
u,
Xi
a
obo
De
ng. Pr
ecipit
abl
e W
a
t
e
r Va
por
Retriev
a
l Usi
n
g Neur
al N
e
t
w
ork from Infrared H
y
p
e
rsp
e
ctral Sou
n
d
i
ngs.
Satellit
e Rem
o
te Sens
in
g
Lab C
o
ll
eg
e of Atmospher
ic Soun
din
g
, Che
n
gdu
U
n
ivers
i
t
y
of
Information T
e
chnolog
y.
2
010.
Evaluation Warning : The document was created with Spire.PDF for Python.