I
AE
S In
t
er
na
t
io
na
l J
o
urna
l o
f
Art
if
icia
l In
t
ellig
ence
(
I
J
-
AI)
Vo
l.
13
,
No
.
4
,
Dec
em
b
er
2
0
2
4
,
p
p
.
4
7
6
6
~
4
7
7
7
I
SS
N:
2
2
5
2
-
8
9
3
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijai.v
13
.i
4
.
p
p
4
7
6
6
-
4
7
7
7
4
7
6
6
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
ia
esco
r
e.
co
m
Lev
erag
ing
multimo
da
l deep
lea
rni
ng
f
o
r natura
l
disa
ster e
v
ent
cla
ss
ificatio
n and
its dam
a
g
e sev
erit
y
ana
ly
sis
t
hro
ug
h so
cia
l
media
pos
ts
Niv
edit
a
K
a
s
t
uri
1,
2
,
Sh
a
s
hik
um
a
r
G
urup
utr
a
T
o
t
a
d
2
,
G
o
ldi
na
G
ho
s
h
3
1
D
e
p
a
r
t
me
n
t
o
f
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
a
n
d
E
n
g
i
n
e
e
r
i
n
g
,
P
ES U
n
i
v
e
r
s
i
t
y
,
B
a
n
g
a
l
o
r
e
,
I
n
d
i
a
2
S
c
h
o
o
l
o
f
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
a
n
d
E
n
g
i
n
e
e
r
i
n
g
,
K
L
E
U
n
i
v
e
r
si
t
y
,
H
u
b
l
i
,
I
n
d
i
a
3
D
e
p
a
r
t
me
n
t
o
f
C
o
m
p
u
t
e
r
A
p
p
l
i
c
a
t
i
o
n
a
n
d
S
c
i
e
n
c
e
,
I
n
st
i
t
u
t
e
o
f
En
g
i
n
e
e
r
i
n
g
a
n
d
M
a
n
a
g
e
m
e
n
t
K
o
l
k
a
t
a
,
K
o
l
k
a
t
a
,
I
n
d
i
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Oct
17
,
2
0
2
3
R
ev
is
ed
Mar
26
,
2
0
2
4
Acc
ep
ted
Ap
r
18
,
2
0
2
4
Ac
c
u
ra
te
a
n
d
ti
m
e
ly
in
fo
rm
a
ti
o
n
is
c
rit
ica
l
to
e
ffe
c
ti
v
e
ly
c
o
o
rd
i
n
a
t
e
d
isa
ste
r
re
sp
o
n
se
.
Du
e
t
o
th
e
d
iv
e
rsity
a
n
d
c
o
m
p
lex
i
ty
o
f
d
a
ta
so
u
rc
e
s,
it
i
s
d
iffi
c
u
lt
fo
r
trad
i
ti
o
n
a
l
m
e
th
o
d
s
t
o
c
las
sify
d
isa
ste
r
e
v
e
n
ts
a
n
d
a
ss
e
ss
d
a
m
a
g
e
se
v
e
rit
y
.
P
re
v
io
u
s
stu
d
ies
h
a
v
e
m
a
in
ly
fo
c
u
se
d
o
n
s
p
e
c
ifi
c
tas
k
s,
su
c
h
a
s
in
fo
rm
a
ti
o
n
c
o
ll
e
c
ti
o
n
o
r
h
u
m
a
n
it
a
rian
a
ss
istan
c
e
,
b
u
t
h
a
v
e
n
o
t
a
d
e
q
u
a
tely
a
d
d
re
ss
e
d
th
e
a
ss
e
ss
m
e
n
t
o
f
d
isa
ste
r
l
o
ss
se
v
e
rit
y
.
Th
is
p
a
p
e
r
p
ro
p
o
se
s
a
h
y
b
ri
d
lea
rn
i
n
g
m
o
d
e
l
to
imp
ro
v
e
d
isa
ste
r
e
v
e
n
t
c
las
sifica
ti
o
n
a
n
d
d
a
m
a
g
e
se
v
e
rit
y
id
e
n
ti
fica
ti
o
n
.
T
h
e
m
o
d
e
l
c
o
m
b
i
n
e
s
ima
g
e
a
n
d
tex
t
d
a
ta,
u
sin
g
Re
sN
e
t5
0
to
e
x
trac
t
fe
a
tu
re
s
fro
m
ima
g
e
s,
a
n
d
u
si
n
g
lo
n
g
s
h
o
rt
-
term
m
e
m
o
ry
with
a
n
a
tt
e
n
ti
o
n
m
e
c
h
a
n
ism
t
o
lea
rn
se
q
u
e
n
c
e
s
fro
m
tex
t.
T
h
is
c
o
m
b
in
a
ti
o
n
a
ll
o
w
s
fo
r
a
m
o
re
c
o
n
t
e
x
tu
a
l
a
n
d
in
fo
r
m
a
ti
v
e
re
p
re
se
n
tatio
n
o
f
t
h
e
in
p
u
t
d
a
ta.
Th
e
e
x
p
e
rime
n
tal
re
su
lt
s
sh
o
ws
th
a
t
t
h
e
p
ro
p
o
se
d
m
u
lt
imo
d
a
l
a
p
p
r
o
a
c
h
a
c
h
iev
e
s
sig
n
ifi
c
a
n
t
ly
b
e
tt
e
r
re
su
lt
s
i
n
d
isa
s
ter
e
v
e
n
t
c
las
sifica
ti
o
n
a
tt
a
in
i
n
g
a
n
a
c
c
u
ra
c
y
o
f
9
0
.
3
1
%
,
c
o
m
p
a
re
d
to
e
x
is
ti
n
g
m
e
th
o
d
s.
F
u
rth
e
rm
o
re
,
th
e
m
o
d
e
l
d
e
m
o
n
stra
tes
p
r
o
m
isin
g
c
a
p
a
b
i
li
t
ies
in
a
ss
e
ss
in
g
d
a
m
a
g
e
se
v
e
rit
y
,
o
ffe
ri
n
g
sig
n
ifi
c
a
n
t
imp
r
o
v
e
m
e
n
ts
fo
r
d
isa
ste
r
m
a
n
a
g
e
m
e
n
t
a
n
d
re
sp
o
n
se
p
r
e
p
a
ra
ti
o
n
wh
e
re
a
c
c
u
ra
c
y
a
n
d
d
e
p
e
n
d
a
b
il
it
y
a
re
c
ru
c
ial.
K
ey
w
o
r
d
s
:
Dam
ag
e
s
ev
er
ity
id
en
tific
atio
n
Dis
aster
ev
en
t c
lass
if
icat
io
n
Hy
b
r
id
d
ee
p
lear
n
i
n
g
Mu
lti
-
m
o
d
ality
an
aly
tics
So
cial
m
ed
ia
p
o
s
ts
T
h
is i
s
a
n
o
p
e
n
a
c
c
e
ss
a
rticle
u
n
d
e
r th
e
CC B
Y
-
SA
li
c
e
n
se
.
C
o
r
r
e
s
p
o
nd
ing
A
uth
o
r
:
Niv
ed
ita
Kastu
r
i
Dep
ar
tm
en
t o
f
C
o
m
p
u
ter
Scie
n
ce
an
d
E
n
g
in
ee
r
in
g
,
PES U
n
i
v
er
s
ity
B
an
g
alo
r
e,
I
n
d
ia
E
m
ail: n
iv
ed
itak
@
p
es.e
d
u
1.
I
NT
RO
D
UCT
I
O
N
T
h
e
im
p
ac
ts
o
f
clim
ate
c
h
an
g
e
an
d
o
t
h
er
e
n
v
ir
o
n
m
en
tal
ch
a
n
g
es
ar
e
in
cr
ea
s
in
g
,
a
n
d
n
atu
r
al
d
is
aster
s
ar
e
in
cr
ea
s
in
g
i
n
f
r
e
q
u
en
c
y
an
d
s
ev
er
ity
,
wh
ic
h
p
o
s
e
a
m
ajo
r
th
r
ea
t
to
th
e
en
v
ir
o
n
m
e
n
t
an
d
h
u
m
an
life
[
1
]
,
[
2
]
.
I
n
o
r
d
er
to
co
n
tr
o
l
d
am
ag
e
a
n
d
f
ac
ilit
ate
th
e
tim
ely
r
e
co
v
er
y
p
r
o
ce
s
s
f
o
r
a
n
af
f
ec
ted
i
n
d
iv
id
u
al,
d
is
aster
p
r
ep
ar
e
d
n
ess
an
d
r
esp
o
n
s
e
s
t
r
ateg
ies
is
r
eq
u
ir
e
d
to
b
e
m
o
r
e
im
p
r
o
v
ed
an
d
en
h
an
ce
d
.
I
n
to
d
a
y
'
s
d
ig
itally
co
n
n
ec
ted
wo
r
ld
,
s
o
cial
n
etwo
r
k
in
g
p
latf
o
r
m
s
lik
e
T
witter
,
F
ac
eb
o
o
k
,
W
h
atsAp
p
,
an
d
I
n
s
ta
g
r
am
h
a
v
e
p
r
o
v
e
n
to
b
e
o
n
e
o
f
th
e
v
alu
ab
le
s
o
u
r
ce
s
f
o
r
co
m
m
u
n
icatio
n
,
co
n
n
ec
tio
n
,
in
f
o
r
m
atio
n
d
is
s
em
in
atio
n
[
3
]
.
Du
r
in
g
a
n
atu
r
al
cr
is
is
o
r
in
em
er
g
en
cie
s
,
s
o
cial
n
etwo
r
k
in
g
p
latf
o
r
m
s
h
elp
s
to
q
u
ick
ly
d
is
s
em
in
ate
in
f
o
r
m
atio
n
,
p
r
o
v
id
e
tim
ely
u
p
d
ates
to
f
ac
ilit
ate
s
i
tu
atio
n
al
awa
r
en
ess
,
an
d
co
n
tr
ib
u
tes
in
o
p
p
o
r
t
u
n
ities
to
s
e
ar
ch
f
o
r
s
u
p
p
o
r
t
o
r
ass
is
tan
ce
to
af
f
ec
ted
p
o
p
u
lati
o
n
s
[
4
]
.
I
n
a
d
d
itio
n
,
s
o
cial
m
ed
ia
p
latf
o
r
m
s
also
p
r
o
v
id
e
r
a
w
in
f
o
r
m
atio
n
th
at
ca
n
b
e
v
e
r
y
h
elp
f
u
l
f
o
r
em
er
g
e
n
cy
r
esp
o
n
d
e
r
s
as
th
e
p
e
o
p
le
o
n
th
e
s
ce
n
e
a
r
e
t
h
e
f
i
r
s
t
to
s
h
a
r
e
wh
at
is
g
o
in
g
o
n
,
in
clu
d
in
g
tex
t
u
p
d
ates,
p
h
o
to
s
,
an
d
ev
en
v
id
e
o
s
.
T
h
is
u
n
f
ilter
ed
in
f
o
r
m
atio
n
ca
n
h
elp
d
is
aster
r
esp
o
n
d
er
s
u
n
d
er
s
tan
d
h
o
w
b
ad
th
i
n
g
s
ar
e
an
d
g
et
aid
to
a
r
ea
s
th
at
n
e
ed
it
m
o
s
t
[
5
]
,
[
6
]
.
H
o
w
ev
e
r
,
t
h
e
r
e
ar
e
i
n
h
er
e
n
t
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2
2
5
2
-
8
9
3
8
Leve
r
a
g
in
g
mu
ltimo
d
a
l
d
ee
p
l
ea
r
n
in
g
fo
r
n
a
t
u
r
a
l d
is
a
s
ter ev
en
t c
la
s
s
ifica
tio
n
a
n
d
…
(
N
ived
ita
K
a
s
tu
r
i
)
4767
ch
all
en
g
es
i
n
u
s
i
n
g
s
o
c
ial
m
ed
ia
ef
f
e
cti
v
e
ly
t
o
r
es
p
o
n
d
t
o
n
a
t
u
r
al
ca
la
m
iti
es.
O
n
t
h
e
o
t
h
e
r
h
an
d
,
it
is
als
o
v
er
y
d
i
f
f
ic
u
lt
t
o
k
ee
p
tr
ac
k
o
f
ev
e
r
y
th
i
n
g
th
at
is
p
o
s
t
ed
,
es
p
ec
ial
ly
d
u
r
i
n
g
t
h
e
d
is
as
te
r
o
r
n
a
tu
r
al
ca
l
am
i
ties
.
Fi
n
d
in
g
th
e
m
o
s
t
i
m
p
o
r
t
a
n
t
in
f
o
r
m
a
ti
o
n
,
s
u
c
h
as
t
h
e
s
e
v
e
r
it
y
o
f
a
d
is
ast
er
a
n
d
t
h
e
l
ik
eli
h
o
o
d
o
f
i
m
m
in
e
n
t
d
a
n
g
er
,
i
n
s
o
c
ial
m
e
d
ia
p
o
s
ts
is
o
f
t
en
a
ch
all
en
g
i
n
g
t
as
k
.
I
n
t
h
is
r
e
g
a
r
d
,
e
f
f
e
cti
v
e
s
t
r
at
e
g
y
t
o
w
ar
d
s
d
e
v
el
o
p
i
n
g
b
e
tte
r
r
esp
o
n
s
e
m
ec
h
an
is
m
b
asic
all
y
d
ep
e
n
d
s
o
n
g
e
tti
n
g
t
h
e
r
i
g
h
t
i
n
f
o
r
m
at
io
n
an
d
in
ti
m
e
ly
m
a
n
n
er
[
7
]
.
T
h
e
c
o
n
v
e
n
ti
o
n
a
l
m
et
h
o
d
s
f
o
r
r
es
p
o
n
s
e
m
ec
h
a
n
is
m
o
f
t
en
i
n
v
o
l
v
es
p
e
o
p
le
m
a
n
u
all
y
an
al
y
z
e,
s
o
r
t,
an
d
an
n
o
ta
te
m
ill
io
n
s
o
f
in
f
o
r
m
at
io
n
,
w
h
i
ch
is
q
u
ite
ti
m
e
-
co
n
s
u
m
in
g
,
c
o
s
t
ly
,
a
n
d
o
f
te
n
s
u
b
je
cte
d
t
o
h
u
m
a
n
-
er
r
o
r
.
H
e
n
ce
,
t
h
e
co
n
v
e
n
t
io
n
a
l
ap
p
r
o
ac
h
es
a
r
e
n
o
t s
u
i
ta
b
le
f
o
r
s
it
u
a
ti
o
n
s
w
h
e
r
e
d
is
as
te
r
ev
e
n
ts
o
cc
u
r
s
o
r
a
p
i
d
l
y
an
d
u
n
ce
r
ta
in
l
y
[
8
]
.
T
h
er
ef
o
r
e
,
d
e
v
elo
p
m
en
t
o
f
a
n
au
to
m
ated
ap
p
r
o
ac
h
is
r
eq
u
ir
ed
th
at
ca
n
p
r
ec
is
ely
class
if
y
th
e
n
atu
r
al
d
is
aster
f
r
o
m
th
e
s
o
cial
m
ed
i
a
p
o
s
ts
d
u
r
in
g
em
e
r
g
en
cies
a
n
d
o
r
e
v
en
ts
o
f
n
atu
r
al
ca
lam
ities
[
9
]
.
T
h
e
r
ec
en
t
ad
v
an
ce
m
e
n
ts
in
ar
tific
ial
in
tellig
en
ce
an
d
em
er
g
en
ce
o
f
d
e
ep
lear
n
in
g
a
p
p
r
o
ac
h
es
s
h
o
ws
a
p
r
o
m
is
in
g
s
co
p
e
to
au
to
m
ate
f
ea
tu
r
e
e
n
g
in
ee
r
i
n
g
task
s
an
d
in
s
ig
h
ts
f
r
o
m
m
u
ti
-
v
ar
iate
an
d
co
m
p
le
x
d
ata
lik
e
s
o
cial
m
ed
ia
p
o
s
ts
[
1
0
]
.
Ho
wev
er
,
d
ev
el
o
p
in
g
s
u
c
h
au
to
m
ate
d
s
ch
em
e
is
n
o
t
an
e
asy
task
esp
ec
ially
wh
en
d
ea
lin
g
with
b
o
th
te
x
tu
al
an
d
im
ag
e
d
ata
in
s
o
cial
m
ed
ia
p
o
s
ts
.
T
h
is
r
eq
u
ir
es
ca
r
e
f
u
l
an
aly
s
is
o
f
d
ata,
ad
o
p
tio
n
o
f
p
r
ep
r
o
ce
s
s
in
g
tech
n
iq
u
es
an
d
s
elec
tio
n
o
f
s
u
itab
le
d
ee
p
lea
r
n
in
g
m
o
d
e
ls
to
ex
tr
ac
t
m
ea
n
in
g
f
u
l
p
atter
n
s
an
d
p
er
f
o
r
m
class
if
icatio
n
o
f
n
atu
r
al
ca
lam
ities
.
I
n
th
e
ex
is
tin
g
liter
atu
r
e,
m
an
y
r
esear
ch
wo
r
k
s
h
av
e
b
ee
n
p
r
esen
ted
,
wh
er
e
f
ew
r
esear
ch
,
wo
r
k
s
ar
e
s
u
b
j
ec
ted
to
u
s
ag
e
o
f
o
n
ly
s
o
cial
m
ed
ia
im
ag
es
an
d
f
ew
ar
e
f
o
cu
s
ed
o
n
te
x
tu
al
in
f
o
r
m
atio
n
f
o
r
b
u
ild
in
g
d
is
aster
class
if
icatio
n
m
o
d
els
[
1
1
]
,
[
1
2
]
.
T
h
er
e
ar
e
f
ewe
r
wo
r
k
s
th
a
t
co
n
s
id
er
b
o
th
te
x
t
an
d
im
ag
e
f
ea
tu
r
es
f
o
r
n
atu
r
a
l
ca
lam
ity
class
if
icatio
n
an
d
d
is
aster
s
ev
er
ity
an
aly
s
is
b
ec
au
s
e
o
f
th
e
in
h
er
en
t
co
m
p
lex
ity
o
f
th
ese
ty
p
es o
f
d
ata.
So
cial
m
ed
ia
d
ata
in
th
e
f
o
r
m
o
f
tex
tu
al
c
o
m
m
u
n
icatio
n
o
f
ten
p
r
esen
ts
a
co
m
p
lex
d
ata
f
o
r
co
m
p
u
tatio
n
al
an
aly
tics
task
s
b
ec
au
s
e
u
s
er
s
o
f
ten
u
s
e
in
f
o
r
m
al
way
s
o
f
wr
itin
g
p
o
s
ts
th
at
in
clu
d
e
s
lan
g
,
em
o
tico
n
s
,
s
h
o
r
t
s
en
ten
ce
s
,
a
n
d
with
v
ar
io
u
s
la
n
g
u
a
g
e
[
1
3
]
.
I
m
ag
e
d
ata,
o
n
th
e
o
th
e
r
h
an
d
,
p
r
o
v
id
es
a
r
ich
er
co
n
tex
t,
b
u
t
d
is
aster
im
ag
es
a
r
e
co
m
p
le
x
d
u
e
to
t
h
eir
m
u
ltifa
ce
ted
n
atu
r
e
a
n
d
r
eq
u
i
r
e
ca
r
e
f
u
l
in
ter
p
r
etatio
n
an
d
u
n
d
er
s
tan
d
i
n
g
.
Fo
r
ex
a
m
p
le,
d
u
r
in
g
an
ea
r
th
q
u
a
k
e
ev
en
t,
a
s
in
g
le
im
ag
e
ca
n
d
ep
ict
d
am
ag
e
d
b
u
ild
i
n
g
s
,
in
ju
r
ed
p
eo
p
le,
d
eb
r
is
,
a
n
d
v
ar
io
u
s
o
th
er
ty
p
es o
f
d
am
ag
e
[
1
4
]
.
T
h
e
r
ef
o
r
e,
p
r
o
ce
s
s
in
g
im
ag
es
to
e
s
tab
lis
h
an
ef
f
ec
tiv
e
d
is
aster
r
esp
o
n
s
e
s
y
s
tem
p
r
o
v
es
to
b
e
e
x
tr
em
ely
ch
allen
g
in
g
d
u
e
t
o
th
e
wid
e
v
ar
iatio
n
in
im
ag
e
q
u
ality
,
m
u
ltifa
ce
ted
n
at
u
r
e
o
f
d
is
as
ter
s
ce
n
ar
io
s
,
an
d
d
iv
e
r
s
ity
o
f
c
o
n
ten
t.
I
n
th
e
cu
r
r
e
n
t
c
o
n
tex
t,
d
ep
lo
y
in
g
lear
n
i
n
g
m
o
d
els
ca
p
ab
le
o
f
h
an
d
lin
g
th
e
d
y
n
am
ic
n
at
u
r
e
o
f
d
is
aster
s
itu
atio
n
s
r
em
ain
s
an
o
n
g
o
in
g
ch
allen
g
e.
Dis
aster
r
esp
o
n
s
e
m
ec
h
an
is
m
s
o
p
er
at
e
u
n
d
er
tim
e
co
n
s
tr
ain
ts
.
I
t
is
cr
u
cia
l
to
en
s
u
r
e
th
eir
ac
cu
r
ac
y
an
d
h
ig
h
r
esp
o
n
s
iv
en
ess
in
r
ea
l
-
tim
e
s
it
u
atio
n
s
,
b
ec
au
s
e
th
e
ac
cu
r
ac
y
an
d
tim
elin
ess
o
f
in
f
o
r
m
atio
n
p
r
o
v
i
d
ed
b
y
s
u
ch
s
y
s
tem
s
will
s
er
io
u
s
ly
af
f
ec
t
d
is
aster
r
esp
o
n
s
e
ac
cu
r
ac
y
an
d
tim
elin
ess
.
Dis
tin
g
u
is
h
b
etwe
en
life
-
s
av
in
g
m
ea
s
u
r
e
s
an
d
p
o
ten
tial
lo
s
s
es.
I
n
th
e
r
ec
en
t
liter
atu
r
e,
m
a
n
y
wo
r
k
s
in
a
s
im
ilar
d
ir
ec
tio
n
h
av
e
b
ee
n
p
r
o
p
o
s
ed
b
y
d
i
f
f
er
en
t
r
esear
ch
er
s
,
m
o
s
t
o
f
wh
ich
p
er
f
o
r
m
a
s
in
g
le
task
in
th
e
c
o
n
tex
t
o
f
d
is
aster
r
esp
o
n
s
e
[
1
5
]
.
T
h
e
wo
r
k
o
f
An
b
ar
asan
et
a
l.
[
1
6
]
in
tr
o
d
u
ce
d
a
f
l
o
o
d
p
r
e
d
ictio
n
s
y
s
tem
b
ased
o
n
in
ter
n
et
o
f
th
in
g
s
an
d
co
n
v
o
l
u
tio
n
al
n
eu
r
al
n
etwo
r
k
(
C
NN)
to
d
ete
ct
an
d
p
r
e
d
ict
f
lo
o
d
ev
en
ts
.
T
h
is
wo
r
k
p
r
o
p
o
s
es
a
r
u
le
g
e
n
e
r
atio
n
m
eth
o
d
b
ased
o
n
p
r
ep
r
o
ce
s
s
ed
m
ater
ial
th
at
is
u
s
ed
as
in
p
u
t
to
a
c
o
n
v
o
lu
tio
n
al
d
ee
p
n
eu
r
al
n
etwo
r
k
(
C
DNN)
class
if
ier
to
class
if
y
th
e
o
cc
u
r
r
en
ce
a
n
d
n
o
n
-
o
cc
u
r
r
en
ce
o
f
f
lo
o
d
in
g
.
Z
h
a
n
g
et
a
l.
[
1
7
]
ad
d
r
ess
ed
th
e
ch
a
llen
g
e
o
f
ac
cu
r
ately
p
r
ed
ictin
g
la
n
d
s
lid
e
d
is
p
lace
m
en
ts
in
r
esp
o
n
s
e
to
r
ai
n
f
all
an
d
r
eser
v
o
ir
wate
r
lev
el
ch
a
n
g
es
in
C
h
in
a.
T
h
e
au
th
o
r
s
p
r
o
p
o
s
ed
a
d
y
n
am
ic
p
r
ed
ictio
n
m
o
d
el
b
ased
o
n
th
e
g
ated
r
ec
u
r
r
en
t
u
n
it
(
GR
U)
lear
n
in
g
alg
o
r
ith
m
,
f
o
llo
win
g
cu
m
u
lativ
e
d
is
p
lace
m
en
t d
ec
o
m
p
o
s
itio
n
an
d
tr
en
d
p
r
ed
ictio
n
.
Ge
et
a
l.
[
1
8
]
u
tili
ze
s
g
r
a
p
h
k
n
o
wled
g
e
c
o
n
ce
p
t
t
h
at
co
n
s
id
er
s
r
em
o
te
s
en
s
in
g
,
a
n
d
g
eo
g
r
ap
h
ic
d
ata
,
alo
n
g
with
in
teg
r
atio
n
o
f
d
o
m
ain
k
n
o
wled
g
e
to
b
u
ild
a
s
p
atio
-
tem
p
o
r
al
f
r
a
m
ewo
r
k
f
o
r
d
is
aster
p
r
ed
ictio
n
.
Sev
er
al
r
esear
ch
er
s
u
s
e
m
ac
h
in
e
lear
n
in
g
an
d
n
at
u
r
al
lan
g
u
a
g
e
p
r
o
ce
s
s
in
g
(
NL
P)
to
an
aly
ze
d
is
aster
s
en
tim
en
t
an
d
d
a
m
ag
e
v
ia
s
o
cial
m
ed
ia
an
aly
tics
[
1
9
]
,
[
2
0
]
.
Z
o
u
et
a
l
.
[
2
1
]
le
v
e
r
ag
es
t
h
e
p
o
te
n
ti
al
o
f
c
o
m
b
i
n
i
n
g
i
m
a
g
e
an
d
te
x
t
i
n
f
o
r
m
ati
o
n
b
y
em
p
l
o
y
i
n
g
a
d
ee
p
le
ar
n
i
n
g
ap
p
r
o
a
ch
f
o
r
v
is
u
a
l
f
e
at
u
r
e
ex
tr
ac
t
io
n
an
d
u
tili
zi
n
g
t
h
e
Fas
tT
ex
t
f
r
a
m
ew
o
r
k
f
o
r
t
e
x
t
u
a
l
f
e
at
u
r
e
e
x
tr
ac
ti
o
n
.
A
d
ata
f
u
s
io
n
m
o
d
el
is
d
e
v
e
lo
p
e
d
t
o
c
o
m
b
i
n
e
t
h
es
e
f
ea
t
u
r
es
f
o
r
cl
ass
i
f
y
in
g
r
el
ev
a
n
t
d
is
aste
r
im
ag
es.
Asi
f
et
a
l
.
[
2
2
]
in
tr
o
d
u
ce
d
a
d
is
aster
tax
o
n
o
m
y
o
f
em
er
g
en
cy
r
esp
o
n
s
e
s
y
s
tem
s
b
ased
o
n
d
ee
p
lear
n
in
g
class
if
icatio
n
m
o
d
els s
u
ch
a
s
VGG
-
1
6
an
d
YOL
O.
T
h
e
m
eth
o
d
u
s
ed
d
ec
is
io
n
tab
les
f
o
llo
wed
b
y
d
ata
a
n
aly
tics
o
p
er
atio
n
s
to
m
ap
im
a
g
e
o
u
tp
u
ts
to
th
e
d
is
aster
-
r
elate
d
in
f
o
r
m
ati
o
n
to
d
eter
m
in
e
s
u
itab
le
em
er
g
en
c
y
r
esp
o
n
s
es.
Z
h
an
g
et
a
l.
[
2
3
]
p
r
o
p
o
s
ed
a
m
u
ltimo
d
al
class
if
icatio
n
s
y
s
tem
u
s
in
g
laten
t
d
ir
ich
let
allo
ca
tio
n
(
L
DA)
f
o
r
clu
s
ter
in
g
an
d
a
b
i
d
ir
e
ctio
n
al
en
co
d
er
r
e
p
r
esen
tatio
n
with
a
p
r
e
-
tr
ain
ed
C
NN
m
o
d
el
to
a
n
aly
ze
tex
t
an
d
im
ag
e
d
ata
to
d
eter
m
i
n
e
th
e
m
atic
in
f
o
r
m
atio
n
d
u
r
in
g
a
d
is
aster
.
Hen
ce
,
it
ca
n
b
e
an
aly
ze
d
th
at
th
er
e
ar
e
cu
r
r
en
tly
m
an
y
s
tu
d
ies
o
n
n
atu
r
al
d
is
a
s
ter
clas
s
if
icatio
n
,
b
u
t
m
o
s
t
o
f
th
em
f
o
cu
s
o
n
tex
t
o
r
im
ag
e
d
ata.
No
n
e
o
f
th
ese
s
tu
d
ies
d
ea
lt
with
d
eter
m
in
in
g
th
e
s
ev
er
ity
o
f
d
is
aster
lo
s
s
es,
wh
ich
is
also
im
p
o
r
tan
t
f
o
r
ef
f
ec
tiv
e
d
is
aster
r
esp
o
n
s
e
s
y
s
tem
s
.
Acc
o
r
d
i
n
g
to
liter
atu
r
e
a
n
aly
s
is
,
p
r
e
d
ictin
g
d
is
aster
s
an
d
ass
es
s
in
g
th
eir
s
ev
er
ity
th
r
o
u
g
h
s
o
cial
m
ed
ia
p
o
s
ts
is
a
co
m
p
lex
an
d
m
u
ltifa
ce
ted
task
th
at
r
eq
u
ir
es
co
m
p
lex
d
ata
p
r
e
p
r
o
ce
s
s
in
g
a
n
d
f
ea
tu
r
e
en
g
in
ee
r
in
g
,
an
d
f
ac
es
t
h
e
ch
al
len
g
e
o
f
lar
g
e
a
n
d
i
m
b
ala
n
ce
d
d
ata
s
ets.
T
h
e
id
en
tifie
d
r
esear
ch
g
ap
s
ar
e
d
is
cu
s
s
ed
b
elo
w:
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell,
Vo
l.
13
,
No
.
4
,
Dec
em
b
er
2
0
2
4
:
4
7
6
6
-
4
7
7
7
4768
–
E
x
is
tin
g
d
is
aster
d
etec
tio
n
m
eth
o
d
s
d
o
n
o
t
e
f
f
ec
tiv
ely
e
x
tr
ac
t
th
e
im
p
o
r
tan
t
f
ea
tu
r
es
f
r
o
m
d
if
f
er
e
n
t
d
ata
ty
p
es
(
s
u
ch
as
tex
t
a
n
d
im
ag
es).
T
h
is
is
b
ec
au
s
e
f
ea
tu
r
e
en
g
in
ee
r
in
g
in
m
u
ltimo
d
al
d
is
aster
d
etec
tio
n
r
eq
u
ir
es
s
p
ec
if
ic
d
o
m
ain
k
n
o
w
led
g
e
f
r
o
m
a
d
ata
a
n
aly
s
is
p
er
s
p
ec
tiv
e,
wh
ich
is
n
o
t
well
r
ef
le
cted
in
c
u
r
r
e
n
t
r
esear
ch
.
–
Pre
v
io
u
s
ap
p
r
o
ac
h
es
to
d
is
ast
er
an
aly
s
is
h
av
e
f
o
cu
s
ed
o
n
s
p
ec
if
ic
task
s
,
s
u
ch
as
p
r
o
v
id
in
g
in
f
o
r
m
atio
n
o
r
h
u
m
an
itar
ian
ass
is
tan
ce
.
T
h
is
leav
es a
g
ap
in
a
n
aly
zin
g
t
h
e
s
ev
er
ity
o
f
d
am
ag
e
ca
u
s
ed
b
y
d
is
aster
s
.
–
Dep
lo
y
in
g
m
u
ltip
le
m
o
d
els
s
p
ec
if
ic
to
d
if
f
er
en
t
m
ater
ial
ty
p
es
an
d
task
s
ca
n
b
e
ch
allen
g
in
g
as
it
in
v
o
lv
es
m
o
d
el
s
elec
tio
n
,
o
p
tim
izatio
n
,
an
d
in
ter
p
r
etatio
n
.
Ho
we
v
er
,
th
e
ex
is
tin
g
liter
atu
r
e
d
o
es
n
o
t
ad
eq
u
ately
ad
d
r
ess
th
is
ch
allen
g
e.
–
Desig
n
in
g
an
d
d
e
v
elo
p
in
g
h
ig
h
ly
in
teg
r
ated
m
o
d
els
is
co
m
p
u
tatio
n
ally
ch
allen
g
in
g
task
,
r
eq
u
ir
es
u
n
iq
u
e
o
p
tim
izatio
n
tech
n
iq
u
es.
L
im
ited
atten
tio
n
h
as
b
ee
n
p
aid
to
m
an
a
g
in
g
m
o
d
el
co
m
p
le
x
ity
,
ad
d
r
ess
in
g
o
v
er
f
itti
n
g
is
s
u
es,
an
d
o
p
tim
iz
in
g
p
er
f
o
r
m
a
n
ce
.
–
T
h
e
liter
atu
r
e
lack
s
n
o
v
elty
a
n
d
d
o
es
n
o
t
ex
p
lo
r
e
ap
p
r
o
ac
h
es
to
cr
o
s
s
-
m
o
d
al
lear
n
i
n
g
,
w
h
er
e
m
o
d
els
ca
n
lear
n
f
r
o
m
tex
tu
al
an
d
im
ag
e
m
ater
ial
s
im
u
ltan
eo
u
s
ly
,
e
n
ab
lin
g
th
em
t
o
u
n
d
er
s
tan
d
th
e
r
e
latio
n
s
h
ip
s
an
d
d
ep
en
d
e
n
cies b
etwe
en
d
if
f
e
r
e
n
t m
o
d
alities
.
T
h
er
ef
o
r
e,
th
is
p
ap
er
s
u
g
g
ests
a
h
ig
h
ly
s
y
n
ch
r
o
n
ized
an
d
in
t
eg
r
ated
co
m
p
u
tatio
n
al
f
r
am
ew
o
r
k
b
ased
o
n
NL
P
m
ec
h
an
is
m
an
d
h
y
b
r
i
d
d
ee
p
lear
n
in
g
ap
p
r
o
ac
h
.
T
h
e
p
r
im
e
aim
o
f
th
e
p
r
o
p
o
s
ed
s
y
s
tem
is
to
o
f
f
er
a
r
eliab
le
au
to
m
ated
s
y
s
tem
u
s
i
n
g
s
o
cial
m
ed
ia
p
o
s
ts
th
at
ca
n
h
elp
em
er
g
e
n
cy
r
esp
o
n
d
e
r
s
m
ak
e
b
etter
d
ec
is
io
n
s
ab
o
u
t
p
r
o
v
id
in
g
r
elief
to
af
f
ec
ted
co
m
m
u
n
ities
.
T
h
is
p
r
o
p
o
s
ed
wo
r
k
is
a
n
o
v
el
m
u
lti
-
m
o
d
al
m
u
lti
-
task
h
y
b
r
id
lear
n
in
g
m
eth
o
d
th
at
au
to
m
ati
ca
lly
co
m
b
in
es
tex
tu
al
an
d
v
is
u
al
p
o
s
ts
in
s
o
cial
n
etwo
r
k
in
g
to
cl
ass
if
y
d
is
as
ter
ev
en
ts
an
d
au
to
m
atica
lly
ass
ess
th
eir
d
am
ag
e
s
ev
er
ity
.
I
n
t
h
e
h
y
b
r
id
lear
n
in
g
ar
ch
itectu
r
e,
lo
n
g
s
h
o
r
t
-
te
r
m
m
em
o
r
y
(
L
STM
)
is
u
s
ed
to
c
ap
tu
r
e
th
e
l
o
n
g
-
ter
m
d
e
p
en
d
e
n
ce
o
f
t
h
e
tex
t
c
o
n
ten
t
o
f
th
e
p
o
s
t
an
d
e
x
tr
ac
t
k
ey
in
f
o
r
m
atio
n
r
elate
d
to
d
is
aster
ev
en
ts
.
At
th
e
s
am
e
tim
e,
it
ad
o
p
ts
th
e
R
esNet
ar
ch
itectu
r
e
as
an
ad
v
an
ce
d
f
ea
tu
r
e
lear
n
er
f
o
r
C
NN
m
o
d
e
ls
to
p
r
o
ce
s
s
v
is
u
al
co
n
ten
t.
Sin
ce
th
e
d
ataset
u
s
ed
to
tr
ain
th
e
p
r
o
p
o
s
ed
h
y
b
r
id
lear
n
in
g
m
o
d
el
is
a
co
m
p
lex
m
u
lti
-
m
o
d
al
task
an
d
is
ass
o
ci
ated
wi
th
th
e
clas
s
im
b
alan
ce
p
r
o
b
lem
.
T
h
er
ef
o
r
e,
th
e
p
r
o
p
o
s
ed
lear
n
in
g
m
o
d
el
i
n
co
r
p
o
r
ates
an
atten
tio
n
m
ec
h
an
is
m
to
m
ak
e
th
e
s
y
s
tem
p
a
y
m
o
r
e
atten
tio
n
t
o
th
e
f
ea
tu
r
es o
f
a
f
ew
ca
teg
o
r
ie
s
.
T
h
e
s
ch
em
atic
ar
ch
itectu
r
e
o
f
th
e
p
r
o
p
o
s
ed
s
y
s
tem
is
s
h
o
wn
in
Fig
u
r
e
1
.
Fig
u
r
e
1
.
Hig
h
lev
el
ar
c
h
itectu
r
e
o
f
th
e
p
r
o
p
o
s
ed
s
y
s
tem
As
it
ca
n
b
e
s
ee
n
f
r
o
m
Fig
u
r
e
1
,
th
e
p
r
o
p
o
s
ed
s
y
s
tem
h
as
f
o
u
r
m
ain
lay
er
s
:
in
p
u
t
lay
e
r
,
v
i
s
u
al
lay
er
,
s
eq
u
en
ce
lear
n
er
o
f
tex
t
f
ea
tu
r
es
,
an
d
p
r
o
jectio
n
lay
er
.
T
h
e
in
p
u
t
lay
er
is
r
esp
o
n
s
ib
le
f
o
r
r
ec
eiv
in
g
in
itial
tex
t
an
d
v
is
u
al
in
f
o
r
m
atio
n
.
T
h
e
v
is
u
al
lay
er
u
s
es
C
NN
to
p
er
f
o
r
m
tr
an
s
f
er
lear
n
in
g
ac
co
r
d
i
n
g
to
th
e
p
r
e
-
tr
ain
e
d
d
ee
p
lear
n
in
g
m
o
d
el.
T
h
e
s
tu
d
y
em
p
lo
y
ed
R
esNet,
a
p
r
e
-
tr
ain
ed
C
NN
a
r
ch
itectu
r
e
th
at
s
er
v
es
as
a
h
ig
h
-
lev
el
f
ea
tu
r
e
ex
tr
ac
t
o
r
f
o
r
im
a
g
e
cla
s
s
if
icatio
n
task
s
.
T
h
e
s
eq
u
en
c
e
lear
n
in
g
m
o
d
u
le
o
f
th
is
s
y
s
tem
u
s
es
L
STM
an
d
atten
tio
n
m
ec
h
an
is
m
to
p
r
o
ce
s
s
tex
t
co
n
ten
t,
wh
ile
th
e
p
r
o
jec
tio
n
lay
er
is
r
esp
o
n
s
ib
le
f
o
r
f
e
atu
r
e
co
n
ca
ten
a
tio
n
an
d
class
if
icatio
n
o
f
d
is
aster
ev
en
ts
,
an
d
u
s
es
th
e
f
u
lly
co
n
n
ec
ted
(
FC
)
lay
er
af
ter
So
f
tMa
x
(
SF
)
f
o
r
d
a
m
ag
e
s
ev
er
ity
an
aly
s
is
ac
tiv
atio
n
f
u
n
ctio
n
.
T
h
e
s
y
s
tem
aim
s
to
im
p
r
o
v
e
th
e
ac
cu
r
ac
y
an
d
c
o
m
p
r
eh
en
s
iv
e
n
ess
o
f
d
is
aster
ev
en
t
class
if
icatio
n
an
d
d
am
a
g
e
s
ev
er
ity
a
n
aly
s
is
b
y
in
teg
r
atin
g
in
f
o
r
m
atio
n
f
r
o
m
m
u
ltip
le
d
ata
m
o
d
es.
T
h
is
n
o
v
el
ap
p
r
o
ac
h
b
ased
o
n
d
u
al
d
ata
p
atter
n
s
an
d
o
p
tim
ized
d
ee
p
lear
n
i
n
g
m
o
d
els
ca
n
b
etter
h
elp
d
is
aster
r
esp
o
n
d
er
s
,
r
esear
ch
e
r
s
,
an
d
p
o
licy
m
ak
er
s
d
ev
elo
p
m
o
r
e
ef
f
ec
tiv
e
an
d
tar
g
e
ted
d
is
aster
r
esp
o
n
s
e
s
tr
ateg
ies
f
o
r
em
er
g
en
cies,
s
u
ch
as
r
ep
o
r
ts
,
r
eq
u
ests
f
o
r
ass
is
tan
ce
,
an
d
u
p
d
ates
o
n
c
h
an
g
i
n
g
d
is
aster
co
n
d
itio
n
s
f
r
o
m
less
r
elev
an
t
o
r
n
o
n
-
ac
tio
n
a
b
le
co
n
ten
t.
B
y
au
to
m
atin
g
th
is
p
r
o
c
ess
,
th
e
r
esear
ch
aim
s
to
s
p
ee
d
u
p
t
h
e
d
eliv
er
y
o
f
cr
itical
in
f
o
r
m
atio
n
to
em
e
r
g
en
cy
r
esp
o
n
d
er
s
,
r
elief
o
r
g
an
izatio
n
s
an
d
af
f
ec
ted
c
o
m
m
u
n
ities
,
th
er
eb
y
im
p
r
o
v
in
g
o
v
er
all
d
is
aster
r
esp
o
n
s
e
ef
f
icien
c
y
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2
2
5
2
-
8
9
3
8
Leve
r
a
g
in
g
mu
ltimo
d
a
l
d
ee
p
l
ea
r
n
in
g
fo
r
n
a
t
u
r
a
l d
is
a
s
ter ev
en
t c
la
s
s
ifica
tio
n
a
n
d
…
(
N
ived
ita
K
a
s
tu
r
i
)
4769
2.
M
E
T
H
O
D
T
h
is
s
ec
tio
n
d
etails
th
e
p
r
o
p
o
s
ed
s
y
s
tem
an
d
d
etails
th
e
im
p
lem
en
tatio
n
p
r
o
ce
d
u
r
es
ad
o
p
ted
f
o
r
th
e
p
r
o
p
o
s
ed
d
is
aster
r
esp
o
n
s
e
s
y
s
tem
.
First,
th
e
d
ata
s
et
u
s
ed
i
s
b
r
ief
ly
d
escr
ib
ed
,
f
o
llo
wed
b
y
d
ata
an
aly
s
is
an
d
m
etad
ata
co
n
s
tr
u
ctio
n
.
T
h
e
s
u
b
s
eq
u
en
t
s
ec
tio
n
s
th
en
d
e
s
cr
ib
e
th
e
co
m
p
u
tatio
n
al
o
p
er
atio
n
s
ad
o
p
ted
in
p
r
ep
ar
in
g
d
ataset
a
n
d
f
ea
tu
r
e
e
n
g
in
ee
r
in
g
to
en
h
a
n
ce
th
e
p
r
e
d
ictiv
e
ca
p
ab
ilit
y
a
n
d
f
ea
tu
r
e
g
e
n
er
aliza
tio
n
ab
ilit
y
o
f
th
e
p
r
o
p
o
s
ed
h
y
b
r
id
lear
n
i
n
g
m
o
d
el.
Fu
r
th
e
r
m
o
r
e,
an
i
m
p
lem
en
tatio
n
s
tr
ateg
y
is
o
u
t
l
in
ed
f
o
r
th
e
f
ea
tu
r
e
ex
tr
ac
tio
n
an
d
its
co
n
ca
ten
atio
n
in
th
e
p
r
o
jectio
n
la
y
er
o
f
h
y
b
r
id
d
ee
p
lear
n
in
g
m
o
d
el.
2
.
1
.
Da
t
a
s
et
T
h
e
d
ataset
ad
o
p
ted
in
th
is
r
esear
ch
wo
r
k
is
cr
is
is
v
i
s
io
n
b
en
ch
m
ar
k
(
C
VB
)
[
2
4
]
in
tr
o
d
u
ce
d
f
o
r
b
u
ild
in
g
a
d
is
aster
r
esp
o
n
s
e
an
d
p
r
e
p
ar
ed
n
ess
s
y
s
tem
.
T
h
is
d
ataset
was
d
ev
elo
p
ed
b
y
r
esear
ch
er
s
f
r
o
m
Go
o
g
le
AI
,
Stan
f
o
r
d
Un
i
v
er
s
ity
,
an
d
t
h
e
Un
iv
er
s
ity
o
f
W
ash
in
g
to
n
.
I
t
co
n
s
is
ts
o
f
v
ar
iety
o
f
d
is
aster
ev
en
ts
lik
e
f
lo
o
d
s
,
wild
f
ir
e
s
,
ea
r
th
q
u
a
k
es,
h
u
r
r
ic
an
es,
an
d
m
an
y
m
o
r
e
in
th
e
f
o
r
m
o
f
b
o
th
im
a
g
es
an
d
tex
t
s
f
r
o
m
a
v
ar
iety
o
f
s
o
u
r
ce
s
s
u
ch
as
s
o
cial
m
ed
ia,
s
atellite
im
ag
er
y
an
d
d
r
o
n
es.
T
h
e
d
is
tr
ib
u
tio
n
o
f
d
ata
s
am
p
l
e
in
C
V
B
d
ataset
i
s
s
h
o
wn
in
Fig
u
r
e
2
f
o
r
a
v
a
r
iety
o
f
d
is
aster
r
esp
o
n
s
e
an
d
p
r
e
p
ar
ed
n
ess
task
s
as f
o
llo
ws:
–
Dis
aster
d
etec
tio
n
task
f
o
r
id
e
n
tify
in
g
th
e
p
r
esen
ce
o
f
a
d
is
aster
in
an
im
ag
e.
–
Dam
ag
e
ass
ess
m
en
t f
o
r
an
aly
zin
g
th
e
s
ev
er
ity
o
f
d
a
m
ag
e
ca
u
s
ed
b
y
a
d
is
aster
ev
en
ts
.
–
Ob
ject
d
etec
tio
n
task
f
o
cu
s
es
o
n
s
p
ec
if
ic
o
b
jects
id
en
tific
atio
n
in
d
is
aster
im
ag
es,
s
u
ch
as
p
e
o
p
le,
b
u
ild
i
n
g
s
,
an
d
v
e
h
icles.
–
Scen
e
class
if
icat
io
n
task
is
s
u
b
jecte
d
to
p
r
ed
ictin
g
o
r
class
if
y
in
g
th
e
d
is
aster
ev
en
ts
o
r
n
atu
r
al
ca
lam
ities
,
s
u
ch
as f
lo
o
d
s
,
wild
f
ir
es,
a
n
d
ea
r
th
q
u
ak
es.
Fig
u
r
e
2
.
T
h
e
d
is
tr
ib
u
tio
n
o
f
d
ata
s
am
p
les in
th
e
d
ataset
As
s
h
o
wn
in
Fig
u
r
e
2
,
th
e
d
is
aster
ca
teg
o
r
y
d
am
ag
ed
_
in
f
r
astr
u
ctu
r
e
h
as
th
e
lar
g
est
n
u
m
b
er
o
f
s
am
p
les,
f
o
llo
wed
b
y
d
am
a
g
ed
_
n
atu
r
al
,
f
ir
e,
an
d
f
lo
o
d
.
T
h
e
h
u
m
an
_
d
am
ag
e
ca
teg
o
r
y
h
as
th
e
f
ewe
s
t
s
am
p
les,
an
d
th
e
n
o
n
_
d
am
ag
e
ca
te
g
o
r
y
h
as
th
e
m
o
s
t
s
am
p
les,
r
esu
ltin
g
in
a
s
er
io
u
s
d
ata
im
b
ala
n
ce
.
Un
e
v
en
d
is
tr
ib
u
tio
n
o
f
d
ata,
esp
ec
ially
lar
g
e
d
if
f
er
e
n
ce
s
b
etwe
en
n
o
n
-
d
am
a
g
e
ca
te
g
o
r
i
es
an
d
o
th
er
d
is
aster
ca
teg
o
r
i
es,
m
ay
b
ias
d
ee
p
lear
n
in
g
m
o
d
els.
T
r
ad
itio
n
al
m
o
d
els
m
ay
b
e
b
iased
to
war
d
m
o
r
e
co
m
m
o
n
ca
teg
o
r
ies,
wh
ich
m
ay
l
ea
d
to
m
is
class
if
icatio
n
o
f
im
p
o
r
tan
t
d
is
aster
s
.
T
h
ese
m
is
cla
s
s
if
ic
atio
n
s
ca
n
h
av
e
s
er
io
u
s
co
n
s
eq
u
en
ce
s
in
d
is
aster
m
an
ag
em
en
t sit
u
atio
n
s
,
wh
e
r
e
tim
ely
an
d
ac
c
u
r
ate
in
f
o
r
m
ati
o
n
is
cr
u
cial.
2
.
2
.
Da
t
a
prepa
ra
t
i
o
n
T
h
e
d
ata
s
et
ad
o
p
ted
p
r
esen
ts
s
o
m
e
s
tr
u
ctu
r
al
an
d
o
r
g
an
izati
o
n
al
ch
allen
g
es
th
at
m
ak
e
it
d
if
f
icu
lt
to
d
ir
ec
tly
u
s
e
it
f
o
r
d
is
aster
ev
en
t
class
if
icatio
n
an
d
d
am
ag
e
s
ev
er
ity
an
aly
s
is
.
T
h
ese
is
s
u
es
in
clu
d
e
in
co
n
s
is
ten
cies
in
d
ata
f
o
r
m
attin
g
.
T
h
e
ar
r
a
n
g
em
en
t
a
n
d
s
tr
u
ctu
r
izatio
n
o
f
d
ata
ar
e
m
o
r
e
task
-
ce
n
tr
ic,
wh
ich
s
u
b
jects it to
in
h
er
en
t c
o
m
p
lex
ities
th
at
n
ee
d
to
b
e
a
d
d
r
ess
ed
f
o
r
ef
f
ec
tiv
e
u
tili
za
tio
n
in
tr
ai
n
in
g
th
e
m
o
d
el.
2
.
2
.
1
.
Da
t
a
s
et
s
t
ruct
ure
T
h
e
d
ataset
u
s
ed
f
o
r
d
is
aster
ev
en
t
class
if
icatio
n
an
d
d
am
ag
e
s
ev
er
ity
an
aly
s
is
is
o
r
g
an
ized
in
a
m
an
n
er
th
at
r
e
f
lects
th
e
co
m
p
l
ex
ity
o
f
th
e
task
.
I
t
c
o
m
b
in
es
v
is
u
al
d
ata
with
tex
tu
al
in
f
o
r
m
a
tio
n
to
en
h
a
n
ce
th
e
u
n
d
er
s
tan
d
i
n
g
o
f
ea
ch
d
is
aster
s
ce
n
ar
io
.
T
h
is
k
i
n
d
o
f
d
ata
s
tr
u
ctu
r
izatio
n
will e
n
s
u
r
es th
at
t
h
e
m
o
d
el
ca
n
lear
n
ef
f
ec
tiv
ely
f
r
o
m
th
e
d
iv
er
s
e
ty
p
es o
f
d
ata
it c
o
n
tain
s
.
T
h
e
s
tr
u
ctu
r
e
o
f
d
ataset
is
s
u
m
m
ar
ized
as f
o
llo
ws:
–
T
h
e
p
r
im
ar
y
d
ata
u
n
it
is
an
im
ag
e
r
e
p
r
esen
tin
g
ea
ch
o
f
th
e
s
ix
d
if
f
er
en
t
d
is
aster
class
es.
T
h
ese
im
ag
es
ar
e
o
r
g
an
ized
with
in
d
is
tin
ct
s
u
b
-
d
ir
ec
to
r
ies co
r
r
esp
o
n
d
in
g
to
e
ac
h
class
.
–
W
ith
in
ea
ch
s
u
b
d
ir
ec
to
r
y
,
tex
t
d
ata
is
p
r
o
v
id
ed
f
o
r
m
o
s
t im
ag
es,
i.e
.
,
twee
ts
p
o
s
ted
o
n
th
e
s
p
ec
if
ic
d
is
aster
ev
en
t o
r
ca
te
g
o
r
y
to
wh
ich
th
e
im
ag
e
b
elo
n
g
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell,
Vo
l.
13
,
No
.
4
,
Dec
em
b
er
2
0
2
4
:
4
7
6
6
-
4
7
7
7
4770
–
A
s
ep
ar
ate
f
o
ld
er
d
escr
ib
es
in
f
o
r
m
atio
n
o
n
d
am
ag
e
s
ev
er
ity
a
n
aly
s
is
.
T
h
is
f
o
ld
er
is
o
r
g
an
ize
d
s
y
s
tem
atica
lly
in
to
tr
ain
in
g
,
test
in
g
,
an
d
v
alid
atio
n
s
u
b
s
ets.
2
.
2
.
2
.
I
s
s
ues
w
it
h t
he
da
t
a
s
et
T
h
e
d
ataset
p
r
esen
ts
s
ev
er
al
c
h
allen
g
es
wh
ich
r
eq
u
ir
ed
to
b
e
ad
d
r
ess
ed
to
en
s
u
r
e
ef
f
ec
tiv
e
tr
ain
in
g
o
f
m
o
d
el
an
d
ac
cu
r
ate
p
r
ed
icti
v
e
o
u
tco
m
es.
T
h
ese
is
s
u
es
ar
e
m
o
s
tly
r
elate
d
to
th
e
d
a
ta's
im
b
alan
ce
an
d
ar
r
an
g
em
e
n
t,
wh
ich
ca
n
h
a
v
e
a
b
ig
ef
f
ec
t
o
n
h
o
w
well
m
ac
h
in
e
lear
n
in
g
alg
o
r
ith
m
s
wo
r
k
.
T
h
er
ef
o
r
e,
it
v
e
r
y
im
p
o
r
tan
t
to
m
itig
ate
th
ese
is
s
u
es
in
o
r
d
er
to
cr
ea
te
a
r
o
b
u
s
t
an
d
tr
u
s
two
r
th
y
p
r
ed
ictiv
e
m
o
d
el
f
o
r
d
is
aster
class
if
icatio
n
an
d
d
am
ag
e
s
ev
er
ity
an
aly
s
is
.
T
h
e
k
e
y
is
s
u
es a
s
s
o
ciate
d
with
d
ataset
is
h
ig
h
lig
h
ted
as f
o
llo
ws:
–
A
k
ey
is
s
u
e
is
th
e
in
h
er
e
n
t
im
b
alan
ce
o
f
th
e
d
ataset
in
ter
m
s
o
f
th
e
n
u
m
b
er
o
f
im
a
g
es
an
d
ass
o
ciate
d
tex
t
an
n
o
tatio
n
s
o
r
twee
ts
.
–
Data
o
n
d
am
a
g
e
s
ev
er
ity
is
p
r
e
-
p
ar
titi
o
n
ed
i
n
to
tr
ain
in
g
,
tes
tin
g
an
d
v
alid
atio
n
s
u
b
s
ets,
ad
d
in
g
ad
d
itio
n
al
co
m
p
lex
ity
.
2
.
2
.
3
.
Da
t
a
prepa
ra
t
io
n a
pp
ro
a
ch
T
h
e
s
tu
d
y
ad
o
p
ts
an
ef
f
ec
tiv
e
m
eth
o
d
o
l
o
g
y
t
o
ad
d
r
ess
is
s
u
es
with
th
e
d
ataset.
A
a
s
p
ec
if
ic
alg
o
r
ith
m
is
d
esig
n
ed
to
h
ar
m
o
n
ize
an
d
i
n
teg
r
ate
d
ata
f
r
o
m
tex
t
an
d
im
ag
e
d
atasets
.
T
h
e
f
o
llo
win
g
ar
e
th
e
m
ain
h
ig
h
lig
h
ts
o
f
th
e
alg
o
r
ith
m
f
o
r
d
ata
p
r
ep
a
r
atio
n
:
–
Sin
ce
d
am
ag
e
s
ev
er
ity
d
ata
is
m
o
s
tly
in
tex
t
f
o
r
m
at,
th
e
im
a
g
e
d
ata
n
ee
d
s
to
b
e
p
r
ep
a
r
ed
a
cc
o
r
d
in
g
to
th
e
tex
tu
al
d
ataset.
T
h
is
o
p
er
ati
o
n
is
p
er
f
o
r
m
ed
b
y
r
eo
r
g
a
n
izin
g
th
e
im
ag
e
d
ata
b
ased
o
n
its
id
e
n
tity
an
d
p
at
h
s
,
en
s
u
r
in
g
co
n
s
is
ten
t a
lig
n
m
en
t
with
th
e
co
r
r
esp
o
n
d
in
g
tex
tu
al
d
ata.
–
T
h
e
n
ex
t o
p
er
atio
n
is
ca
r
r
ied
o
u
t to
war
d
s
cr
ea
tin
g
a
m
etad
ata
in
.
C
SV
f
ile
f
o
r
m
at.
T
h
is
d
ata
th
en
s
er
v
es a
s
a
ce
n
tr
al
r
ep
o
s
ito
r
y
o
f
in
f
o
r
m
a
tio
n
,
in
clu
d
i
n
g
im
ag
e
p
ath
s
,
r
e
lated
twee
ts
,
d
is
aster
o
cc
u
r
r
en
ce
an
d
s
ev
er
ity
.
–
Ad
d
itio
n
ally
,
twee
ts
u
s
ed
as
tex
t
an
n
o
tatio
n
s
ar
e
s
u
b
s
eq
u
en
tly
ad
d
ed
to
th
e
m
eta
d
ata.
T
w
ee
ts
ar
e
m
atch
ed
to
im
ag
es
b
ased
o
n
th
e
im
a
g
e
-
id
co
lu
m
n
,
s
o
th
e
v
is
u
al
c
o
n
ten
t
o
f
ea
c
h
im
ag
e
is
alwa
y
s
p
air
ed
with
co
r
r
esp
o
n
d
in
g
te
x
tu
al
m
ater
ial
.
–
T
ex
t
d
ata
f
r
o
m
twee
ts
ar
e
p
r
e
p
r
o
ce
s
s
ed
to
clea
n
a
n
d
s
im
p
lif
y
,
r
em
o
v
in
g
ir
r
elev
an
t
in
f
o
r
m
atio
n
to
im
p
r
o
v
e
th
e
ef
f
ec
tiv
en
ess
o
f
s
u
b
s
eq
u
en
t a
n
aly
s
is
.
T
h
e
s
u
cc
ess
f
u
l e
x
ec
u
tio
n
o
f
th
e
p
r
o
p
o
s
ed
d
ata
p
r
ep
ar
atio
n
p
r
o
ce
s
s
r
etu
r
n
s
a
s
tr
u
ctu
r
ed
d
ata
f
r
am
e,
as
h
ig
h
lig
h
ted
in
Fig
u
r
e
3
.
T
h
e
d
ata
f
r
am
e
i
n
clu
d
es
th
e
p
ath
o
f
ea
ch
im
ag
e,
its
ass
o
ciate
d
twee
ts
,
ass
o
ciate
d
d
is
aster
ev
en
ts
,
an
d
s
ev
er
ity
.
Af
ter
p
r
ep
r
o
ce
s
s
in
g
,
a
n
ew
co
l
u
m
n
f
o
r
clea
n
ed
tex
t
d
ata
is
ad
d
ed
,
wh
ich
is
cr
u
cial
f
o
r
ex
tr
ac
tin
g
tex
tu
al
f
ea
tu
r
es
an
d
co
n
d
u
ctin
g
p
r
e
d
ictiv
e
an
a
ly
tics
.
T
h
is
r
ef
in
ed
d
ata
f
r
am
e
is
th
en
u
tili
ze
d
a
s
th
e
f
in
al
tr
ain
in
g
d
ataset
f
o
r
th
e
s
tu
d
y
.
Fig
u
r
e
3
.
A
s
am
p
le
v
is
u
aliza
tio
n
o
f
th
e
m
etad
ata
a
f
ter
th
e
d
a
ta
p
r
ep
ar
atio
n
o
p
e
r
atio
n
T
ab
le
1
s
h
o
ws
th
e
d
is
tr
ib
u
tio
n
s
tatis
t
ics
f
o
r
th
e
d
is
aster
ev
e
n
t
class
if
icatio
n
task
o
b
tain
ed
af
t
er
th
e
d
ata
p
r
ep
ar
atio
n
task
,
h
ig
h
lig
h
tin
g
s
ig
n
if
ican
t
d
if
f
er
en
ce
s
b
etwe
e
n
ca
teg
o
r
ies.
No
n
-
d
a
m
ag
e
ca
te
g
o
r
ies
d
o
m
in
ate
th
e
d
ataset,
wh
ile
ev
en
ts
s
u
ch
as
f
ir
e
an
d
h
u
m
an
d
am
a
g
e
ar
e
u
n
d
er
r
e
p
r
esen
ted
.
T
ab
le
2
f
o
cu
s
es
o
n
th
e
d
is
aster
s
ev
er
ity
id
en
tific
atio
n
task
,
wh
er
e
th
e
n
u
m
b
e
r
o
f
litt
le
o
r
n
o
n
e
s
ev
er
ity
lev
els
is
m
u
ch
g
r
ea
ter
th
an
o
th
er
lev
els.
B
o
th
tab
les h
ig
h
lig
h
t th
e
c
h
allen
g
e
o
f
class
im
b
alan
ce
an
d
th
e
n
ee
d
f
o
r
a
s
tr
ateg
ic
m
o
d
el
tr
ain
in
g
ap
p
r
o
ac
h
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2
2
5
2
-
8
9
3
8
Leve
r
a
g
in
g
mu
ltimo
d
a
l
d
ee
p
l
ea
r
n
in
g
fo
r
n
a
t
u
r
a
l d
is
a
s
ter ev
en
t c
la
s
s
ifica
tio
n
a
n
d
…
(
N
ived
ita
K
a
s
tu
r
i
)
4771
T
ab
le
1
.
Data
d
is
tr
ib
u
tio
n
co
n
ce
r
n
in
g
d
is
aster
ev
en
t c
lass
if
ic
atio
n
task
S
I
.
N
O
D
i
sast
e
r
e
v
e
n
t
s
Tr
a
i
n
d
a
t
a
Te
st
d
a
t
a
1
D
a
mag
e
d
_
i
n
f
r
a
s
t
r
u
c
t
u
r
e
1
0
0
9
2
5
8
2
D
a
mag
e
_
n
a
t
u
r
e
3
4
7
1
1
4
3
F
i
r
e
s
2
9
4
56
4
F
l
o
o
d
2
6
1
50
5
H
u
ma
n
_
d
a
ma
g
e
1
8
0
42
6
N
o
n
_
d
a
ma
g
e
2
1
7
7
5
1
3
T
ab
le
2
.
Data
d
is
tr
ib
u
tio
n
co
n
ce
r
n
in
g
d
is
aster
s
ev
er
ity
id
en
tific
atio
n
task
S
I
.
N
O
D
i
sast
e
r
e
v
e
n
t
s
Tr
a
i
n
d
a
t
a
Te
st
d
a
t
a
1
l
i
t
t
l
e
_
o
r
_
n
o
n
e
2
9
0
0
7
0
0
2
sev
e
r
e
8
5
3
2
0
9
3
mi
l
d
5
1
5
1
2
1
2
.
3
.
Sequ
ence
lea
rner
T
h
is
s
ec
tio
n
d
is
cu
s
s
es th
e
ap
p
r
o
ac
h
tak
en
to
d
esig
n
a
d
ee
p
s
eq
u
en
ce
lear
n
er
s
y
s
tem
f
o
r
cla
s
s
if
icatio
n
o
f
d
is
aster
ev
e
n
ts
an
d
id
e
n
tific
atio
n
o
f
th
eir
d
am
a
g
e
s
ev
e
r
ity
.
T
h
is
p
h
ase
o
f
th
e
r
esear
ch
im
p
lem
en
ts
th
e
L
STM
m
o
d
el
as a
s
eq
u
en
ce
lear
n
er
b
ec
au
s
e
it is
b
est s
u
ited
to
ca
p
tu
r
e
lo
n
g
-
ter
m
d
ep
en
d
en
cies o
n
th
e
tex
t (
twee
ts
)
o
f
ea
ch
d
ata
s
am
p
le.
Fu
r
th
er
m
o
r
e,
th
e
p
r
o
p
o
s
ed
s
eq
u
en
ce
lear
n
er
co
m
b
in
ed
with
t
h
e
a
tten
tio
n
m
ec
h
an
is
m
allo
ws
th
e
p
r
o
p
o
s
ed
s
eq
u
e
n
ce
lear
n
e
r
m
o
d
el
t
o
f
o
c
u
s
o
n
t
h
e
m
o
s
t
im
p
o
r
tan
t
p
ar
ts
o
f
t
h
e
tex
t
s
eq
u
en
ce
,
w
h
ich
ca
n
im
p
r
o
v
e
th
e
ac
c
u
r
ac
y
o
f
th
e
class
if
icatio
n
task
.
Du
e
to
d
ata
im
b
alan
ce
,
esp
ec
ially
with
r
eg
ar
d
to
d
a
m
ag
e
s
ev
er
ity
task
s
,
in
te
g
r
atin
g
atten
tio
n
m
ec
h
an
is
m
s
ca
n
h
elp
r
e
d
u
ce
o
v
er
f
itti
n
g
b
y
p
r
e
v
en
tin
g
th
e
L
STM
m
o
d
el
f
r
o
m
m
em
o
r
izin
g
t
h
e
tr
ain
in
g
d
ata
to
o
clo
s
ely
,
t
h
er
eb
y
p
r
o
m
o
tin
g
b
etter
g
en
er
aliza
tio
n
to
u
n
s
ee
n
d
ata.
Ho
wev
er
,
b
ef
o
r
e
tr
ain
in
g
o
u
r
s
eq
u
en
ce
lear
n
in
g
m
o
d
el,
th
e
tex
t
d
ata
n
ee
d
s
to
b
e
p
r
ep
r
o
ce
s
s
ed
f
o
r
f
ea
tu
r
e
ex
tr
ac
tio
n
i
n
th
e
em
b
e
d
d
in
g
la
y
er
.
T
h
e
s
tu
d
y
f
ir
s
t
co
n
v
er
ts
th
e
tex
tu
al
p
h
r
ases
in
to
n
u
m
er
ic
al
v
ec
to
r
s
,
a
n
d
th
en
ap
p
lies
a
wo
r
d
em
b
ed
d
in
g
m
o
d
el
to
th
em
to
o
b
tain
t
h
e
s
em
an
tic
an
d
tem
p
o
r
al
r
ep
r
esen
tatio
n
o
f
th
e
tex
tu
al
p
h
r
ases
.
T
h
e
o
b
tain
e
d
em
b
ed
d
in
g
v
ec
to
r
s
ar
e
th
en
u
s
ed
t
o
tr
ain
a
s
eq
u
e
n
ce
lear
n
er
m
o
d
e
l,
wh
ich
lear
n
s
th
e
co
n
tex
t
-
lev
el
p
r
o
p
e
r
ties
o
f
ea
ch
p
h
r
ase.
T
h
en
,
e
x
p
lo
it
th
e
p
o
ten
tial
o
f
th
e
atten
tio
n
m
ec
h
an
is
m
to
r
etain
th
e
m
o
s
t im
p
o
r
tan
t f
ea
t
u
r
es f
r
o
m
s
eq
u
en
tial
lay
er
s
2
.
3
.
1
.
T
ex
t
v
ec
t
o
r
co
ns
t
ruct
io
n
T
h
e
in
itial
s
tep
o
f
tex
t
v
ec
t
o
r
izatio
n
p
r
ep
r
o
ce
s
s
es
th
e
te
x
tu
al
d
ata
to
r
em
o
v
e
e
m
o
jis
,
s
y
m
b
o
ls
,
p
u
n
ctu
atio
n
s
,
web
UR
L
s
,
an
d
ex
ce
s
s
s
p
ac
es.
B
y
em
p
lo
y
in
g
Py
th
o
n
lib
r
ar
ies
s
u
ch
as
B
ea
u
t
if
u
lSo
u
p
an
d
em
o
ji,
th
e
s
tu
d
y
en
s
u
r
es a
clea
n
an
d
s
tan
d
ar
d
f
o
r
m
at
f
o
r
all
te
x
t d
ata
.
On
ce
clea
n
ed
,
th
e
tex
t
u
al
d
at
a
is
co
n
v
er
ted
in
to
its
r
esp
ec
tiv
e
s
tr
in
g
d
ata
ty
p
e
f
o
r
co
n
s
is
ten
cy
.
T
h
e
d
is
tr
ib
u
ti
o
n
o
f
tex
t
le
n
g
th
f
o
r
ea
ch
d
ata
s
am
p
le
is
an
aly
ze
d
as d
ep
icted
in
Fig
u
r
e
4
.
Fig
u
r
e
4
.
Dis
tr
ib
u
tio
n
o
f
te
x
t le
n
g
th
co
n
ce
r
n
in
g
ea
ch
d
is
tin
ct
d
is
aster
class
B
ased
o
n
v
is
u
al
in
s
p
ec
tio
n
o
f
Fig
u
r
e
4
,
it
ca
n
b
e
f
o
u
n
d
th
at
m
o
s
t
tex
ts
ar
e
ar
o
u
n
d
1
5
0
wo
r
d
s
in
le
n
g
th
.
T
h
er
ef
o
r
e,
a
len
g
th
o
f
1
5
0
is
ch
o
s
en
as
th
e
p
a
d
d
in
g
cr
iter
i
o
n
to
en
s
u
r
e
th
at
all
s
eq
u
en
c
es
ar
e
o
f
co
n
s
is
ten
t
len
g
th
.
T
h
e
to
k
en
izer
is
im
p
lem
en
ted
u
s
in
g
a
d
ef
i
n
ed
v
o
ca
b
u
lar
y
o
f
5
0
,
0
0
0
wo
r
d
s
to
cr
ea
te
an
in
itial
v
ec
to
r
r
ep
r
esen
tatio
n
o
f
th
e
twee
t.
T
h
is
p
r
o
ce
d
u
r
e
en
s
u
r
es
th
at
ce
r
tain
s
y
m
b
o
ls
,
p
u
n
ctu
atio
n
m
ar
k
s
,
an
d
o
th
er
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell,
Vo
l.
13
,
No
.
4
,
Dec
em
b
er
2
0
2
4
:
4
7
6
6
-
4
7
7
7
4772
ch
ar
ac
ter
s
ar
e
f
ilter
ed
o
u
t
d
u
r
in
g
th
e
to
k
e
n
izatio
n
p
r
o
ce
s
s
.
I
f
th
e
to
k
e
n
izer
en
co
u
n
ter
s
a
wo
r
d
o
u
ts
id
e
o
f
its
v
o
ca
b
u
lar
y
,
it
r
ep
r
esen
ts
it
u
s
in
g
th
e
to
k
en
"<
o
o
v
>"
.
T
h
e
t
o
k
en
izer
co
n
v
er
ts
ea
ch
twee
t
in
to
a
s
eq
u
en
ce
o
f
s
in
g
le
-
wo
r
d
in
d
e
x
es.
Ho
wev
er
,
s
in
ce
th
ese
s
eq
u
en
ce
s
m
ay
v
ar
y
in
len
g
th
,
th
e
y
ar
e
p
ad
d
ed
to
a
f
ix
ed
len
g
th
o
f
1
5
0
.
Vis
u
al
in
s
p
ec
tio
n
o
f
th
e
e
n
co
d
in
g
an
d
p
ad
d
i
n
g
s
eq
u
en
ce
s
r
ev
ea
ls
th
e
tr
an
s
f
o
r
m
atio
n
p
r
o
ce
s
s
.
Fo
r
in
s
tan
ce
,
th
e
tex
t:
"a
Sri
L
an
k
an
m
an
r
em
o
v
es
b
elo
n
g
in
g
s
f
r
o
m
h
is
s
u
b
m
er
g
ed
h
o
u
s
e
at
W
eh
an
g
alla
Villag
e
in
th
e
Kalu
tar
a
Dis
tr
ict.
.
.
"
is
co
n
v
er
t
ed
in
to
a
s
eq
u
en
ce
lik
e
[
5
,
7
5
0
3
,
1
2
6
1
2
,
5
2
8
,
.
.
.
]
an
d
s
u
b
s
e
q
u
en
tly
,
it
is
p
ad
d
e
d
to
en
s
u
r
e
a
f
ix
ed
len
g
t
h
o
f
1
5
0
.
I
n
t
h
e
p
r
o
v
id
e
d
d
ataset,
th
e
tr
ain
in
g
s
et
co
m
p
r
is
es
4
2
6
8
s
eq
u
en
ce
s
,
an
d
th
e
test
s
et
in
clu
d
es
1
0
3
3
s
eq
u
en
ce
s
.
An
im
p
o
r
tan
t
o
b
s
er
v
atio
n
is
t
h
e
ac
tu
al
v
o
ca
b
u
lar
y
s
ize,
wh
ich
tu
r
n
e
d
o
u
t
to
b
e
3
6
,
1
9
5
,
s
m
aller
th
an
th
e
in
itially
d
ef
i
n
ed
5
0
,
0
0
0
,
wh
ich
in
d
icate
s
th
at
th
e
d
ataset
co
n
tain
s
3
6
,
1
9
5
u
n
iq
u
e
wo
r
d
s
.
2
.
3
.
2
.
T
ex
t
em
bedd
ing
W
h
en
p
r
o
c
ess
in
g
tex
t
d
ata,
tex
t
em
b
ed
d
in
g
is
th
e
b
asis
o
f
d
ee
p
lear
n
i
n
g
m
o
d
els.
As
t
h
e
n
a
m
e
s
u
g
g
ests
,
th
ey
em
b
ed
wo
r
d
s
in
to
a
co
n
ti
n
u
o
u
s
m
ath
em
atica
l
s
p
ac
e
s
o
th
at
lear
n
in
g
m
o
d
els
ca
n
u
n
d
e
r
s
tan
d
an
d
p
r
o
ce
s
s
th
em
.
T
h
e
em
b
e
d
d
in
g
p
r
o
ce
s
s
co
n
v
e
r
ts
d
is
cr
ete
s
y
m
b
o
ls
(
s
u
ch
as
s
in
g
le
wo
r
d
s
)
in
t
o
f
i
x
ed
-
s
ize
d
en
s
e
v
ec
to
r
s
.
I
n
th
e
co
n
tex
t
o
f
th
e
p
r
o
p
o
s
ed
p
r
ed
ictiv
e
lear
n
in
g
task
,
th
e
e
m
b
ed
d
in
g
lay
er
co
n
v
er
ts
in
teg
er
in
d
ices
(
o
b
tain
ed
f
r
o
m
t
o
k
en
izatio
n
)
in
to
d
e
n
s
e
v
ec
to
r
s
o
f
f
ix
ed
s
ize
1
5
0
.
T
h
i
s
lay
er
is
cr
u
cial
b
ec
au
s
e
it
p
r
o
jects
wo
r
d
s
in
to
a
co
n
tin
u
o
u
s
v
ec
to
r
s
p
ac
e,
wh
e
r
e
s
em
an
tically
s
im
ilar
wo
r
d
s
ar
e
m
ap
p
ed
t
o
n
ea
r
b
y
p
o
in
t
s
.
T
h
e
s
ize
o
f
th
e
em
b
ed
d
in
g
s
p
ac
e
is
d
eter
m
in
ed
b
y
th
e
n
u
m
b
er
o
f
wo
r
d
s
in
th
e
v
o
ca
b
u
la
r
y
,
wh
ich
is
3
6
,
1
9
5
in
th
is
d
ataset.
Go
o
d
em
b
e
d
d
in
g
s
ca
n
c
ap
tu
r
e
th
e
s
em
an
tic
r
elatio
n
s
h
ip
s
b
et
wee
n
wo
r
d
s
.
T
h
e
e
m
b
ed
d
in
g
l
ay
er
ac
ts
as
a
lo
o
k
u
p
tab
le
to
r
etr
iev
e
a
d
en
s
e
v
ec
to
r
r
ep
r
esen
tatio
n
o
f
ea
ch
wo
r
d
d
u
r
in
g
tr
ain
in
g
.
T
h
ese
v
ec
to
r
s
ar
e
th
en
ad
ju
s
ted
d
u
r
in
g
tr
ain
in
g
to
r
ed
u
ce
th
e
lo
s
s
,
m
ak
in
g
t
h
e
em
b
ed
d
in
g
s
m
o
r
e
ac
c
u
r
ate
an
d
s
u
itab
le
f
o
r
class
if
icatio
n
task
s
.
T
h
er
ef
o
r
e,
wo
r
d
em
b
ed
d
in
g
s
ca
p
tu
r
e
a
lar
g
e
am
o
u
n
t
o
f
in
f
o
r
m
atio
n
ab
o
u
t
th
e
d
ataset
an
d
s
er
v
e
as
co
m
p
ac
t
an
d
d
en
s
e
r
ep
r
esen
tatio
n
s
o
f
wo
r
d
s
,
ca
p
tu
r
in
g
t
h
eir
m
ea
n
in
g
s
an
d
r
elatio
n
s
h
ip
s
.
2
.
3
.
3
.
L
o
ng
s
ho
rt
-
t
er
m m
e
mo
ry
la
y
er
s
T
h
is
s
tu
d
y
em
p
lo
y
s
a
b
id
ir
ec
tio
n
al
lo
n
g
s
h
o
r
t
-
ter
m
m
em
o
r
y
(
B
i
-
L
STM
)
m
o
d
el,
wh
ich
is
f
u
r
th
er
in
teg
r
ated
with
an
atten
ti
o
n
m
ec
h
an
is
m
to
g
en
er
ate
co
n
te
x
tu
al
r
ep
r
esen
tatio
n
s
o
f
t
h
e
in
p
u
t
tex
t
f
r
o
m
b
o
th
th
e
f
o
r
war
d
an
d
b
ac
k
war
d
d
ir
ec
tio
n
s
[
2
5
]
.
T
h
e
s
eq
u
en
ce
in
p
u
t
g
o
es
th
r
o
u
g
h
th
e
em
b
e
d
d
in
g
lay
e
r
an
d
th
en
in
to
t
h
e
Bi
-
L
STM
lay
er
.
T
h
e
m
o
d
el
co
n
s
is
ts
o
f
two
B
i
-
L
STM
lay
er
s
,
wh
er
e
th
e
f
ir
s
t
lay
er
h
as
1
2
8
u
n
its
an
d
th
e
s
ec
o
n
d
lay
er
h
as
6
4
u
n
its
.
T
h
e
s
ec
o
n
d
lay
er
r
etu
r
n
s
s
eq
u
en
ce
s
an
d
s
t
ates.
T
h
e
f
o
r
war
d
an
d
b
ac
k
wa
r
d
h
i
d
d
e
n
s
tates
o
f
th
e
s
ec
o
n
d
lay
e
r
ar
e
co
n
n
ec
ted
as
(
1
)
:
ℋ
=
ℋ
⃗
⃗
⃗
⃗
⃗
;
ℋ
⃖
⃗
⃗
⃗
⃗
⃗
(
1
)
W
h
er
e
ℋ
⃗
⃗
⃗
⃗
⃗
an
d
ℋ
⃖
⃗
⃗
⃗
⃗
⃗
,
ar
e
th
e
h
id
d
en
s
tates
a
t
tim
e
t
f
r
o
m
th
e
f
o
r
war
d
an
d
b
ac
k
war
d
L
STM
s
r
esp
ec
tiv
ely
.
B
asically
,
th
e
ad
o
p
tio
n
o
f
Bi
-
L
STM
m
itig
ates
th
e
v
an
is
h
in
g
g
r
ad
ie
n
t
p
r
o
b
lem
ass
o
ciate
d
with
lo
n
g
s
eq
u
en
tial
d
ata.
T
h
is
m
o
d
el
allo
ws
th
e
le
ar
n
in
g
to
ac
c
ess
b
o
th
p
ast
(
ℋ
⃖
⃗
⃗
⃗
⃗
⃗
)
an
d
f
u
t
u
r
e
(
ℋ
⃗
⃗
⃗
⃗
⃗
)
tem
p
o
r
al
i
n
f
o
r
m
atio
n
,
wh
ich
is
ess
en
tial
f
o
r
u
n
d
er
s
tan
d
in
g
th
e
s
em
an
tics
o
f
a
g
iv
en
co
n
tex
t
in
th
e
i
n
p
u
t
te
x
t
p
h
r
ases
.
Ho
wev
er
,
n
o
t
all
wo
r
d
s
in
a
tex
t
twee
t
co
n
tr
i
b
u
te
e
q
u
ally
to
its
class
if
icatio
n
.
T
h
er
ef
o
r
e,
t
h
e
s
tu
d
y
em
p
lo
y
ed
a
weig
h
ted
atten
tio
n
m
ec
h
an
is
m
to
h
i
g
h
lig
h
t
t
h
e
m
o
s
t
im
p
o
r
tan
t
wo
r
d
s
d
u
r
i
n
g
tex
t
class
if
icatio
n
.
T
h
e
atten
tio
n
m
ec
h
an
is
m
is
ap
p
lie
d
to
th
e
co
n
ca
ten
ate
d
h
id
d
e
n
s
tates
f
r
o
m
th
e
B
i
-
L
STM
lay
er
in
th
e
p
r
o
p
o
s
ed
s
eq
u
e
n
ce
lear
n
er
m
o
d
el.
Fo
r
ea
c
h
h
id
d
en
s
tate
ℋ
,
an
atten
tio
n
s
co
r
e
is
co
m
p
u
ted
,
in
d
icatin
g
th
e
im
p
o
r
tan
ce
o
f
t
h
e
co
r
r
esp
o
n
d
in
g
wo
r
d
in
th
e
s
eq
u
en
ce
.
T
h
e
co
m
p
u
tatio
n
o
f
th
e
atten
tio
n
s
co
r
e
is
g
iv
e
n
as
(
2
)
:
(
ℋ
)
=
ℎ
(
1
ℋ
+
2
ℎ
̅
)
(
2
)
wh
er
e
1
an
d
2
ar
e
lear
n
ab
le
we
ig
h
t
m
atr
ices,
an
d
ℎ
̅
is
th
e
co
n
ca
ten
ated
f
o
r
war
d
an
d
b
ac
k
w
ar
d
s
h
id
d
en
s
tates.
T
h
e
atten
tio
n
m
ec
h
an
is
m
ass
ig
n
s
a
we
ig
h
t
to
ea
ch
wo
r
d
f
ea
tu
r
e
f
r
o
m
th
e
s
eq
u
e
n
tial
lay
er
,
f
o
cu
s
in
g
o
n
th
e
o
u
tp
u
t la
b
els g
iv
en
as (
3
)
:
=
(
(
ℋ
)
)
∑
(
(
ℋ
)
)
′
=
1
(
3
)
wh
er
e
d
en
o
tes
atten
tio
n
weig
h
t
f
o
r
h
id
d
en
s
tate
ℋ
an
d
T
is
th
e
s
eq
u
en
ce
len
g
th
.
A
weig
h
ted
s
u
m
o
p
er
atio
n
g
en
er
ates
an
atten
tiv
e
f
ea
tu
r
e
v
ec
to
r
(
c
o
n
s
is
ts
a
s
u
m
m
ar
ized
r
ep
r
esen
tatio
n
o
f
i
n
p
u
t
s
eq
u
e
n
ce
s
)
f
o
r
ea
ch
tex
t,
n
u
m
er
ically
g
i
v
en
as
(
4
)
:
=
∑
=
1
(
4
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2
2
5
2
-
8
9
3
8
Leve
r
a
g
in
g
mu
ltimo
d
a
l
d
ee
p
l
ea
r
n
in
g
fo
r
n
a
t
u
r
a
l d
is
a
s
ter ev
en
t c
la
s
s
ifica
tio
n
a
n
d
…
(
N
ived
ita
K
a
s
tu
r
i
)
4773
T
h
e
co
n
tex
t
v
ec
to
r
C
is
th
en
c
alcu
lated
as
th
e
weig
h
ted
s
u
m
o
f
all
h
id
d
e
n
s
tates
b
ased
o
n
th
e
atten
tio
n
weig
h
ts
,
en
ab
lin
g
th
e
m
o
d
el
t
o
f
o
cu
s
m
o
r
e
o
n
th
e
r
elev
an
t
p
ar
ts
o
f
th
e
in
p
u
t
s
eq
u
en
ce
.
T
h
e
co
n
tex
t
v
ec
to
r
,
wh
ich
en
ca
p
s
u
lates
th
e
m
o
s
t
e
s
s
en
tial
in
f
o
r
m
a
tio
n
f
r
o
m
th
e
s
eq
u
en
ce
,
is
th
en
p
ass
ed
th
r
o
u
g
h
a
d
en
s
e
lay
er
with
r
ec
tifie
d
lin
ea
r
u
n
it
(
R
eL
U)
a
ctiv
atio
n
.
T
h
e
ad
v
an
tag
e
o
f
t
h
e
atten
tio
n
m
ec
h
a
n
is
m
is
its
ab
ilit
y
to
f
o
cu
s
o
n
s
p
ec
if
ic
p
ar
ts
o
f
th
e
in
p
u
t
s
eq
u
en
ce
,
ess
en
tially
weig
h
in
g
th
em
b
ased
o
n
th
eir
r
e
lev
an
ce
to
th
e
cu
r
r
e
n
t
p
r
o
ce
s
s
in
g
s
tep
.
2
.
4
.
Vis
ua
l f
ea
t
ure
e
x
t
ra
ct
o
r
T
h
e
p
r
o
p
o
s
ed
s
tu
d
y
ad
o
p
ted
a
co
n
ce
p
t
o
f
tr
a
n
s
f
er
lear
n
i
n
g
b
y
im
p
lem
en
tin
g
t
h
e
p
r
e
-
tr
ai
n
ed
C
NN
m
o
d
el
R
esNet5
0
[
2
6
]
as
a
h
i
g
h
er
-
lev
el
f
ea
t
u
r
e
ex
tr
ac
t
o
r
th
at
lev
er
ag
es
r
esid
u
al
lear
n
in
g
to
tr
ain
v
er
y
d
ee
p
n
etwo
r
k
s
.
T
h
e
k
ey
id
ea
b
eh
i
n
d
R
esNet
is
in
tr
o
d
u
cin
g
a
s
k
ip
co
n
n
ec
tio
n
th
at
b
y
p
ass
es o
n
e
o
r
m
o
r
e
lay
er
s
.
T
h
e
o
p
er
atio
n
i
n
a
r
esid
u
al
b
lo
ck
c
an
b
e
r
e
p
r
esen
ted
m
ath
e
m
atica
lly
as
(
5
)
:
=
(
)
+
(
5
)
W
h
er
e
d
en
o
tes
to
th
e
o
u
tp
u
t,
(
∙
)
is
a
m
ap
p
in
g
f
u
n
ctio
n
t
h
at
tr
a
n
s
f
o
r
m
s
th
e
in
p
u
t
d
ata
(
)
f
ea
tu
r
es
to
th
e
o
u
tp
u
t
class
f
o
llo
win
g
th
e
co
n
v
o
lu
tiv
e
lay
er
s
in
ea
ch
b
lo
c
k
.
T
h
e
R
esNet5
0
m
o
d
el
h
as
l
ea
r
n
ed
a
v
ar
iety
o
f
f
ea
tu
r
es
f
r
o
m
a
d
iv
e
r
s
e
d
ataset
co
n
tain
in
g
m
illi
o
n
s
o
f
im
a
g
e
s
f
r
o
m
th
o
u
s
an
d
s
o
f
class
es.
T
h
e
m
o
d
el
weig
h
ts
ar
e
in
itialized
with
th
e
v
alu
es
o
b
tain
ed
d
u
r
in
g
t
h
e
p
r
e
-
tr
ai
n
in
g
.
T
h
e
m
ath
e
m
atica
l
r
ep
r
es
en
tatio
n
o
f
tr
ain
i
n
g
in
v
o
lv
es
m
in
im
izin
g
a
lo
s
s
f
u
n
ctio
n
L
u
s
in
g
o
p
tim
izatio
n
al
g
o
r
ith
m
s
lik
e
s
to
ch
asti
c
g
r
a
d
i
en
t
d
escen
t
(
SGD)
,
wh
ich
u
p
d
ates th
e
weig
h
ts
W
o
f
th
e
tr
ain
a
b
le
lay
er
s
b
ased
o
n
th
e
g
r
a
d
ien
t o
f
th
e
lo
s
s
co
n
c
er
n
in
g
th
e
weig
h
ts
:
=
−
(
6
)
W
h
er
e
is
th
e
lear
n
in
g
r
ate
an
d
L
W
⁄
is
th
e
g
r
ad
ien
t
o
f
th
e
lo
s
s
w
it
h
r
esp
ec
t
to
th
e
weig
h
ts
.
Af
ter
R
esNet5
0
is
im
p
lem
en
ted
,
a
g
lo
b
al
av
er
ag
e
p
o
o
lin
g
o
p
er
atio
n
lay
e
r
is
ap
p
lied
to
r
ed
u
ce
th
e
s
p
atial
d
im
en
s
io
n
s
o
f
th
e
f
ea
tu
r
e
m
ap
s
,
n
u
m
er
ically
g
iv
en
as (
7
)
:
ℱ
,
=
1
×
ℋ
∑
∑
ℱ
,
,
ℎ
ℋ
ℎ
=
1
=
1
(
7
)
W
h
er
e
an
d
ℋ
ar
e
t
h
e
wid
th
an
d
h
eig
h
t
o
f
th
e
f
ea
tu
r
e
m
a
p
ℱ
an
d
ℱ
,
,
ℎ
r
ef
er
s
to
th
e
v
alu
e
at
s
p
atial
p
o
s
itio
n
(
,
ℎ
)
in
th
e
f
ea
tu
r
e
m
ap
ℱ
.
T
h
e
p
o
o
led
o
u
tp
u
t
is
th
en
f
latten
ed
a
n
d
f
e
d
in
t
o
a
d
en
s
e
la
y
er
,
wh
i
ch
p
er
f
o
r
m
s
a
lin
ea
r
o
p
er
atio
n
f
o
llo
we
d
b
y
a
n
o
n
-
lin
ea
r
ac
tiv
atio
n
f
u
n
ctio
n
g
iv
en
as (
8
)
:
=
(
⋅
ℱ
+
)
(
8
)
W
h
er
e
d
en
o
tes
th
e
weig
h
ts
,
r
ef
er
s
to
b
ias,
a
n
d
is
th
e
n
o
n
-
li
n
ea
r
ac
tiv
atio
n
f
u
n
ctio
n
.
T
h
e
ex
tr
ac
ted
f
ea
tu
r
es
ar
e
th
en
r
ea
d
y
to
b
e
c
o
n
ca
ten
ated
with
f
ea
tu
r
es
f
r
o
m
o
th
er
m
o
d
alities
(
e.
g
.
,
tex
t)
f
o
r
f
u
r
th
e
r
p
r
o
ce
s
s
in
g
an
d
f
in
al
p
r
ed
ictio
n
s
.
2
.
5
.
P
r
o
j
ec
t
io
n la
y
er
I
n
th
e
p
r
o
p
o
s
ed
m
o
d
el,
th
e
p
r
o
jectio
n
lay
er
is
r
esp
o
n
s
ib
le
f
o
r
in
teg
r
atin
g
t
h
e
f
ea
tu
r
es
ex
tr
ac
ted
f
r
o
m
b
o
th
v
is
u
al
f
ea
tu
r
e
ex
tr
ac
to
r
a
n
d
s
eq
u
en
tial le
ar
n
e
r
m
o
d
els
to
f
o
r
m
a
c
o
h
esiv
e
an
d
u
n
if
ie
d
r
ep
r
esen
tatio
n
to
co
n
s
tr
u
ct
a
f
u
s
ed
r
e
p
r
esen
t
atio
n
,
en
s
u
r
i
n
g
b
alan
ce
d
co
n
tr
ib
u
tio
n
s
f
r
o
m
v
is
u
al
an
d
tex
tu
al
in
f
o
r
m
ati
o
n
.
T
h
e
f
u
s
ed
f
ea
tu
r
e
v
ec
to
r
,
ℱ
,
is
n
u
m
er
ically
ex
p
r
ess
ed
as
(
9
)
:
ℱ
=
⨁
(
9
)
W
h
er
e
⨁
d
en
o
tes th
e
co
n
ca
ten
at
io
n
o
p
er
atio
n
.
T
h
is
u
n
if
ied
f
ea
tu
r
e
v
ec
to
r
th
en
p
r
o
g
r
ess
es th
r
o
u
g
h
an
e
n
s
u
in
g
h
id
d
en
la
y
er
e
n
co
m
p
ass
in
g
n
n
eu
r
o
n
s
an
d
s
u
b
s
eq
u
en
tly
r
ea
c
h
es
a
So
f
tMa
x
lay
er
to
c
o
n
d
u
c
t
class
if
icatio
n
s
.
T
o
p
r
ev
en
t
o
v
er
f
itti
n
g
,
a
d
r
o
p
o
u
t
lay
er
is
s
tr
ateg
ically
p
lace
d
b
ef
o
r
e
th
e
h
id
d
en
lay
er
.
T
h
e
tr
an
s
m
u
tatio
n
o
f
th
e
f
u
s
ed
f
ea
tu
r
e
v
ec
to
r
t
h
r
o
u
g
h
t
h
e
h
id
d
e
n
lay
er
ca
n
b
e
m
ath
e
m
atica
lly
r
ep
r
esen
ted
as
(
1
0
)
:
ℋ
=
(
⋅
ℱ
+
)
(
1
0
)
W
h
er
e
ℋ
s
y
m
b
o
lizes
th
e
o
u
tp
u
t
f
r
o
m
th
e
h
id
d
e
n
lay
e
r
an
d
is
o
f
th
e
d
im
en
s
io
n
s
an
d
ar
e
t
h
e
weig
h
t
m
atr
ix
an
d
b
ias
v
ec
to
r
af
f
iliated
with
t
h
is
h
id
d
e
n
la
y
er
.
Fin
ally
,
th
e
o
u
tp
u
t
f
r
o
m
th
e
h
id
d
en
lay
er
n
a
v
ig
ates
th
r
o
u
g
h
th
e
So
f
tMa
x
lay
er
t
o
d
eter
m
in
e
th
e
p
r
o
b
ab
ilit
y
d
is
tr
ib
u
tio
n
o
v
er
th
e
class
es.
T
h
e
n
ex
t
s
ec
tio
n
d
is
cu
s
s
es
th
e
p
er
f
o
r
m
a
n
ce
a
n
aly
s
is
o
f
t
h
e
p
r
o
p
o
s
ed
s
y
s
tem
f
o
r
d
is
aster
ev
en
t
class
if
icatio
n
an
d
its
d
am
ag
e
s
ev
er
ity
an
aly
s
is
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell,
Vo
l.
13
,
No
.
4
,
Dec
em
b
er
2
0
2
4
:
4
7
6
6
-
4
7
7
7
4774
3.
RE
SU
L
T
AND
DI
SCUS
SI
O
N
T
h
e
s
y
s
tem
was
d
e
s
ig
n
ed
an
d
d
ev
elo
p
e
d
o
n
th
e
Py
th
o
n
p
r
o
g
r
a
m
m
in
g
lan
g
u
ag
e
r
u
n
n
i
n
g
in
th
e
An
ac
o
n
d
a
d
is
tr
ib
u
tio
n
.
Fo
r
a
b
etter
ex
p
er
ien
ce
,
m
o
d
el
ex
ec
u
tio
n
n
ee
d
s
to
s
u
p
p
o
r
t
s
p
ec
if
ic
co
m
p
u
te
s
tack
s
o
f
Nv
id
ia
an
d
C
UDA
p
ac
k
a
g
es.
E
x
p
er
im
en
tal
an
aly
s
is
an
d
r
es
ea
r
ch
r
esu
lts
wer
e
o
b
tain
ed
r
u
n
n
in
g
th
e
p
r
o
p
o
s
ed
m
o
d
el
o
n
W
in
d
o
ws
1
0
with
1
6
GB
R
AM
an
d
GPU
GT
X
1
6
6
0
T
i.
T
h
is
s
ec
tio
n
p
r
esen
ts
th
e
r
esu
lts
o
b
tain
ed
b
y
th
e
p
r
o
p
o
s
ed
s
y
s
tem
,
wh
ich
is
ev
alu
ated
b
ased
o
n
wid
ely
ad
o
p
ted
class
if
icatio
n
m
etr
ics
s
u
ch
as
ac
cu
r
ac
y
,
p
r
ec
i
s
io
n
,
r
ec
all,
a
n
d
f
1
-
s
co
r
e.
T
h
e
r
esu
lts
o
f
t
h
e
p
r
o
p
o
s
ed
m
o
d
el
a
r
e
d
is
cu
s
s
ed
f
o
r
b
o
th
u
n
im
o
d
al
a
n
d
m
u
ltimo
d
al
d
ata
in
ter
m
s
o
f
b
o
th
v
is
u
al
an
d
n
u
m
e
r
ical
o
u
tc
o
m
e
d
is
cu
s
s
io
n
.
Fig
u
r
e
5
s
h
o
ws
a
co
n
f
u
s
io
n
m
atr
ix
an
aly
s
is
f
o
r
a
u
n
im
o
d
al
tex
t
-
b
ased
ap
p
r
o
ac
h
i
n
d
is
aster
ev
en
t
d
etec
tio
n
an
d
d
am
a
g
e
s
ev
er
it
y
ass
ess
m
en
t.
An
aly
s
is
o
f
th
e
co
n
f
u
s
io
n
m
atr
ix
in
Fig
u
r
e
5
(
a)
s
h
o
ws
th
at
th
e
m
o
d
el
p
e
r
f
o
r
m
s
well
in
p
r
ed
ic
tin
g
n
o
d
am
a
g
e
(
ND)
e
v
en
ts
.
Ho
wev
er
,
th
e
m
o
d
el
h
as
s
o
m
e
co
n
f
u
s
io
n
b
etwe
en
d
am
ag
ed
_
in
f
r
astru
ctu
r
e
(
DI
)
,
an
d
d
am
a
g
ed
_
n
atu
r
e
(
DN)
cla
s
s
es
a
s
th
er
e
ar
e
m
o
r
e
m
is
class
if
ied
in
s
tan
ce
s
.
T
h
e
class
if
icatio
n
o
f
f
i
r
e
-
r
elate
d
d
i
s
aster
s
an
d
f
lo
o
d
s
is
g
en
er
ally
ac
cu
r
ate
b
u
t
th
er
e
is
a
litt
le
c
o
n
f
u
s
io
n
with
o
th
er
d
is
aster
ty
p
es.
Als
o
,
h
u
m
an
d
am
ag
e
(
HD)
s
h
o
ws
s
o
m
e
r
o
o
m
f
o
r
im
p
r
o
v
em
en
t
in
th
e
m
o
d
el
as
th
er
e
ar
e
1
3
m
is
class
if
ied
in
s
tan
ce
s
o
u
t
o
f
a
to
tal
o
f
4
2
in
s
tan
ce
s
.
On
th
e
o
th
er
h
an
d
,
an
aly
s
is
o
f
th
e
co
n
f
u
s
io
n
m
atr
i
x
in
Fig
u
r
e
5
(
b
)
d
em
o
n
s
tr
ates
th
e
e
f
f
ec
tiv
en
ess
o
f
th
e
m
o
d
el
i
n
id
en
tify
in
g
co
r
r
ec
t
in
s
tan
ce
s
with
lo
w
d
am
ag
e,
b
u
t
it
s
tr
u
g
g
les
m
o
r
e
with
m
ild
d
am
ag
e
with
ze
r
o
co
r
r
ec
t
class
if
icatio
n
s
,
an
d
m
is
class
if
ie
s
it
as
lo
w
an
d
s
ev
er
e.
Ov
er
all,
th
e
an
aly
s
is
h
ig
h
lig
h
ts
b
o
th
s
tr
o
n
g
p
er
f
o
r
m
an
ce
a
r
ea
s
an
d
o
p
p
o
r
tu
n
ities
to
r
ef
i
n
e
th
e
tex
t
-
b
ased
d
is
aster
class
if
icatio
n
.
T
h
e
n
ex
t a
n
aly
s
is
is
p
r
esen
ted
f
o
r
th
e
u
n
im
o
d
al
im
ag
e
-
b
ased
ap
p
r
o
ac
h
.
(
a)
(
b
)
Fig
u
r
e
5
.
C
o
n
f
u
s
io
n
m
atr
ices f
o
r
u
n
im
o
d
al
tex
t c
lass
if
icatio
n
p
er
f
o
r
m
an
ce
: (
a
)
d
is
aster
ev
e
n
t d
etec
tio
n
an
d
(
b
)
d
a
m
ag
e
s
ev
er
ity
ass
ess
m
en
t
Fig
u
r
e
6
p
r
esen
ts
th
e
co
n
f
u
s
i
o
n
m
atr
ix
an
al
y
s
is
f
o
r
u
n
im
o
d
al
im
ag
e
class
if
icatio
n
p
er
f
o
r
m
an
ce
in
d
is
aster
ev
en
t
d
etec
tio
n
an
d
d
am
ag
e
s
ev
er
ity
ass
es
s
m
en
t.
I
n
Fig
u
r
e
6
(
a)
,
th
e
im
ag
e
class
if
ica
tio
n
m
o
d
el
ex
h
ib
its
s
lig
h
tly
b
etter
r
esu
lts
th
an
th
e
tex
t
-
b
ased
m
o
d
el
in
m
o
s
t
class
e
s
,
ex
ce
p
t
f
o
r
DN
an
d
f
lo
o
d
,
wh
er
e
th
e
p
er
f
o
r
m
an
ce
s
lig
h
tly
d
ec
r
ea
s
es
.
Fig
u
r
e
6
(
b
)
in
d
icate
s
th
at
th
e
m
o
d
el
p
er
f
o
r
m
e
d
well
in
id
en
t
if
y
in
g
lo
w
d
a
m
ag
e,
b
u
t
f
ac
e
d
d
if
f
icu
lty
in
d
is
tin
g
u
is
h
in
g
b
etwe
en
m
ild
an
d
s
ev
er
e
d
a
m
ag
e,
alt
h
o
u
g
h
it
s
ti
ll
o
u
tp
er
f
o
r
m
s
th
e
tex
t
-
b
ased
u
n
im
o
d
al
ap
p
r
o
ac
h
.
T
h
e
n
e
x
t
an
aly
s
is
f
o
cu
s
es
o
n
th
e
m
u
lti
-
m
o
d
al
ap
p
r
o
ac
h
c
o
m
b
in
in
g
im
ag
e
an
d
tex
t d
ata
u
s
in
g
a
h
y
b
r
id
lear
n
i
n
g
m
o
d
el.
(
a)
(
b
)
Fig
u
r
e
6
.
C
o
n
f
u
s
io
n
m
atr
ices f
o
r
u
n
im
o
d
al
im
ag
e
class
if
icati
o
n
p
er
f
o
r
m
a
n
ce
: (
a)
d
is
aster
ev
en
t d
etec
tio
n
a
n
d
(
b
)
d
a
m
ag
e
s
ev
er
ity
ass
ess
m
en
t
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2
2
5
2
-
8
9
3
8
Leve
r
a
g
in
g
mu
ltimo
d
a
l
d
ee
p
l
ea
r
n
in
g
fo
r
n
a
t
u
r
a
l d
is
a
s
ter ev
en
t c
la
s
s
ifica
tio
n
a
n
d
…
(
N
ived
ita
K
a
s
tu
r
i
)
4775
Fig
u
r
e
7
s
h
o
ws
th
e
co
n
f
u
s
io
n
m
atr
ix
an
al
y
s
is
o
f
th
e
m
u
ltimo
d
al
im
ag
e
-
te
x
t
class
if
icatio
n
p
er
f
o
r
m
an
ce
.
T
h
e
a
n
aly
s
is
o
f
F
ig
u
r
e
7
(
a)
s
h
o
ws
th
at
th
e
m
u
ltimo
d
al
m
o
d
el
co
m
b
in
in
g
im
ag
es
an
d
tex
t
p
er
f
o
r
m
s
r
o
b
u
s
tly
in
id
e
n
tify
in
g
n
o
d
is
aster
s
(
ND)
,
an
d
also
f
o
r
m
o
s
t
d
is
aster
ty
p
es,
its
r
esu
lt
s
ar
e
im
p
r
o
v
e
d
o
v
e
r
th
e
u
n
im
o
d
al
ap
p
r
o
ac
h
.
Fig
u
r
e
7
(
b
)
s
h
o
ws
th
at
th
e
m
o
d
el
f
u
r
th
er
o
u
tp
er
f
o
r
m
s
th
e
u
n
im
o
d
al
b
ased
ap
p
r
o
ac
h
in
ter
m
s
o
f
d
am
ag
e
s
ev
er
ity
,
an
d
h
as
h
ig
h
er
ac
cu
r
ac
y
in
d
ete
ctin
g
s
m
all
d
am
a
g
es
as
well
as
m
in
o
r
an
d
s
ev
er
e
ca
teg
o
r
ies,
co
m
p
a
r
ed
to
th
e
te
x
t
an
d
im
ag
e
u
n
im
o
d
al.
T
h
e
o
v
er
all
an
aly
s
is
s
h
o
ws
th
at
th
e
m
u
ltimo
d
al
ap
p
r
o
ac
h
h
as
a
s
ig
n
if
ican
t
im
p
r
o
v
em
en
t
in
im
p
r
o
v
in
g
th
e
class
if
icatio
n
ac
cu
r
ac
y
.
T
h
e
n
ex
t
an
aly
s
is
i
s
p
r
esen
ted
in
T
ab
le
3
with
co
n
s
o
lid
ated
o
u
tco
m
e
s
tatis
tics
f
o
r
ea
ch
class
if
icatio
n
task
with
th
e
d
if
f
e
r
en
t
lear
n
in
g
m
o
d
els.
(
a)
(
b
)
Fig
u
r
e
7
.
C
o
n
f
u
s
io
n
m
atr
ices f
o
r
m
u
lti
-
m
o
d
al
im
ag
e
-
tex
t c
la
s
s
if
icatio
n
p
er
f
o
r
m
a
n
ce
: (
a)
d
is
aster
ev
en
t
d
etec
tio
n
an
d
(
b
)
d
am
ag
e
s
ev
er
ity
ass
ess
m
en
t
T
ab
le
3
.
C
o
n
s
o
lid
ated
o
u
tco
m
e
s
tatis
tic
s
f
o
r
u
n
im
o
d
al
an
d
m
u
lti
-
m
o
d
al
class
if
icatio
n
m
o
d
els
M
o
d
e
l
Ty
p
e
C
l
a
s
si
f
i
c
a
t
i
o
n
t
a
sk
A
c
c
u
r
a
c
y
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
F1
-
S
c
o
r
e
U
n
i
m
o
d
a
l
(
t
e
x
t
d
a
t
a
)
D
i
sast
e
r
e
v
e
n
t
8
8
.
0
8
8
.
4
8
8
.
0
8
7
.
8
5
D
a
mag
e
s
e
v
e
r
i
t
y
7
7
.
5
4
7
2
.
4
7
7
7
.
5
4
7
4
.
0
8
U
n
i
m
o
d
a
l
(
i
ma
g
e
d
a
t
a
)
D
i
sast
e
r
e
v
e
n
t
8
3
.
6
4
8
3
.
5
3
8
3
.
6
4
8
3
.
4
4
D
a
mag
e
s
e
v
e
r
i
t
y
7
8
.
6
1
7
5
.
7
4
7
8
.
6
1
7
6
.
0
6
M
u
l
t
i
-
mo
d
a
l
(
t
e
x
t
+
i
ma
g
e
)
D
i
sast
e
r
e
v
e
n
t
9
0
.
3
1
9
0
.
2
0
9
0
.
3
1
9
0
.
1
5
D
a
mag
e
s
e
v
e
r
i
t
y
8
0
.
2
5
7
6
.
6
0
8
0
.
2
5
7
8
.
4
1
T
h
e
tex
t
-
b
ased
u
n
im
o
d
al
d
is
aster
ev
en
t
class
if
icatio
n
m
eth
o
d
ex
h
ib
ited
a
h
ig
h
ac
cu
r
ac
y
o
f
8
8
.
0
%
an
d
p
r
ec
is
io
n
o
f
8
8
.
4
%,
in
d
icatin
g
th
at
m
o
s
t
d
is
aster
ev
en
ts
class
if
ied
b
y
th
e
m
o
d
el
wer
e
co
r
r
ec
t
.
T
h
e
r
ec
all
a
n
d
F1
s
co
r
es
wer
e
8
8
.
0
%
an
d
8
7
.
8
5
%,
r
esp
ec
tiv
ely
,
in
d
icatin
g
th
a
t
th
e
m
o
d
el
was
u
n
ab
le
to
ca
p
tu
r
e
m
o
s
t
in
s
tan
ce
s
co
r
r
ec
tly
ac
r
o
s
s
all
ca
teg
o
r
i
es.
A
s
im
ilar
an
aly
s
is
ca
n
b
e
p
er
f
o
r
m
e
d
i
n
th
e
ca
s
e
o
f
d
am
a
g
e
s
ev
er
ity
id
en
tific
atio
n
.
T
h
e
m
o
d
el
ac
h
iev
ed
an
ac
cu
r
ac
y
o
f
7
7
.
5
4
%
an
d
an
F1
s
co
r
e
o
f
7
4
.
0
8
%.
On
th
e
o
th
er
h
a
n
d
,
u
n
im
o
d
al
im
ag
e
-
b
ased
m
eth
o
d
s
s
h
o
w
s
lig
h
tly
b
etter
p
er
f
o
r
m
a
n
ce
o
n
b
o
th
class
if
icatio
n
p
r
o
b
lem
s
.
I
m
ag
e
-
b
ased
un
im
o
d
al
u
s
in
g
a
C
NN
m
o
d
el
ac
h
iev
es
8
3
.
6
4
%
ac
c
u
r
ac
y
an
d
8
3
.
4
4
%
F1
s
co
r
e
in
class
if
y
in
g
d
is
aster
ev
en
ts
.
I
n
th
e
d
a
m
ag
e
s
ev
er
ity
r
ec
o
g
n
itio
n
task
,
it
ef
f
ec
tiv
ely
o
u
t
p
er
f
o
r
m
s
th
e
tex
t
-
b
ased
u
n
i
m
o
d
al
m
eth
o
d
with
7
8
.
6
1
%
ac
cu
r
ac
y
an
d
7
6
.
0
6
% F1
s
co
r
e.
T
h
e
p
r
o
p
o
s
ed
m
u
lti
m
o
d
al
h
y
b
r
i
d
lear
n
in
g
m
o
d
el
i
n
teg
r
atin
g
tex
t
a
n
d
im
ag
e
d
ata
was f
o
u
n
d
to
b
e
th
e
m
o
s
t e
f
f
ec
tiv
e
am
o
n
g
th
e
two
s
in
g
le
m
o
d
alities
.
I
t a
ch
iev
ed
a
h
ig
h
er
ac
cu
r
ac
y
o
f
9
0
.
3
1
%
an
d
an
F1
s
co
r
e
o
f
9
0
.
1
5
%
in
d
is
aster
ev
en
t
class
if
icatio
n
.
Mo
r
e
o
v
er
,
i
n
th
e
p
e
r
f
o
r
m
an
ce
e
v
alu
atio
n
o
f
d
am
ag
e
s
ev
er
ity
id
e
n
tific
atio
n
,
th
e
p
r
o
p
o
s
ed
h
y
b
r
id
m
o
d
el
o
u
tp
e
r
f
o
r
m
ed
th
e
s
in
g
le
m
o
d
al
ap
p
r
o
ac
h
,
ac
h
iev
in
g
an
ac
cu
r
ac
y
o
f
8
0
.
2
5
%
an
d
an
F1
s
co
r
e
o
f
7
8
.
4
1
%
,
in
d
icatin
g
th
at
th
e
p
r
o
p
o
s
ed
m
o
d
el
is
v
er
y
r
o
b
u
s
t
in
ass
es
s
in
g
d
am
ag
e
s
ev
er
ity
,
wh
ich
co
m
b
in
es
th
e
f
ea
tu
r
es
o
f
tex
t
an
d
im
ag
e
d
ata.
T
h
e
p
r
o
p
o
s
ed
class
if
icatio
n
m
o
d
els we
r
e
b
en
c
h
m
ar
k
e
d
an
d
co
m
p
a
r
ed
an
d
an
aly
ze
d
,
as sh
o
wn
in
T
a
b
le
4
.
T
ab
le
4
.
T
h
e
co
m
p
ar
ativ
e
an
al
y
s
is
f
o
r
d
is
aster
ev
en
t c
lass
if
icatio
n
M
o
d
a
l
i
t
y
A
c
c
u
r
a
c
y
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
F1
-
sc
o
r
e
O
f
l
i
e
t
a
l
.
[
2
7
]
Te
x
t
8
0
.
8
8
1
.
0
8
1
.
0
8
0
.
9
I
mag
e
8
3
.
3
8
3
.
1
8
3
.
3
8
3
.
2
M
u
l
t
i
8
4
.
4
8
4
.
1
8
4
.
0
8
4
.
2
P
r
o
p
o
se
d
Te
x
t
8
8
.
0
8
8
.
4
8
8
.
0
8
7
.
8
5
I
mag
e
8
3
.
6
4
8
3
.
5
3
8
3
.
6
4
8
3
.
4
4
M
u
l
t
i
9
0
.
3
1
9
0
.
2
0
9
0
.
3
1
9
0
.
1
5
T
ab
le
4
p
r
o
v
id
es
a
co
m
p
ar
ativ
e
an
aly
s
is
wh
er
e
th
e
r
esu
lts
o
f
th
e
p
r
o
p
o
s
ed
d
is
aster
ev
en
t
cla
s
s
if
icatio
n
s
y
s
tem
ar
e
co
m
p
ar
ed
with
s
i
m
ilar
ex
is
tin
g
wo
r
k
b
y
Of
li
et
a
l
.
[
2
7
]
.
A
co
m
p
ar
ativ
e
a
n
al
y
s
is
o
f
d
if
f
er
en
t
test
ca
s
es c
o
n
s
id
er
in
g
s
in
g
le
m
o
d
a
lity
(
tex
t a
n
d
im
ag
e
in
d
e
p
en
d
e
n
t)
an
d
m
u
ltim
o
d
ality
(
tex
t a
n
d
im
ag
e
m
o
d
alities
Evaluation Warning : The document was created with Spire.PDF for Python.