TELKOM
NIKA
, Vol.13, No
.3, Septembe
r 2015, pp. 1
037
~10
4
6
ISSN: 1693-6
930,
accredited
A
by DIKTI, De
cree No: 58/DIK
T
I/Kep/2013
DOI
:
10.12928/TELKOMNIKA.v13i3.1543
1037
Re
cei
v
ed Fe
brua
ry 10, 20
15; Re
vised
May 21, 20
15
; Accepte
d
Ju
ne 4, 2015
Burn Area Processing t
o
Generate False Alarm Data for
Hotspot Prediction Models
Imas Sukaes
ih Sitanggan
g
*
1
, Razali Yaakob
2
, Norw
ati Mustapha
3
, Ainuddin A. N.
4
1
Departme
n
t of Computer Sci
ence, F
a
cult
y o
f
Natural Scie
n
c
e and Math
e
m
atics,
Bogor Agr
i
cult
ural U
n
ivers
i
t
y
,
Indones
ia
2,3
F
a
culty
of Co
mputer Scie
nc
e and Inform
ati
on T
e
c
hnol
og
y, Universiti Putr
a Mala
ys
ia, Ma
la
ysia
4
Institute of
T
r
o
p
ical F
o
r
e
str
y
and F
o
rest Pro
ducts
(INT
ROP), Universiti Put
r
a Mala
ysi
a
, Mala
ysi
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: imas.sitang
g
ang
@ip
b
.ac.id
1
, razali
y@u
p
m.edu.m
y
2
,
nor
w
a
t
i
@u
pm.edu.m
y
3
, a
_ai
n
udd
in@
y
a
h
oo.
com
4
A
b
st
r
a
ct
Devel
o
p
i
ng
ho
tspot pred
ictio
n
mode
ls usi
n
g decis
io
n tree alg
o
rith
ms r
equ
ire targ
et classes t
o
w
h
ich ob
jects i
n
a d
a
taset are
classifie
d
. In mo
de
lin
g
hots
p
ots occurre
nce
,
target classes
are the tru
e
cl
ass
repres
entin
g
h
o
tspots occ
u
rr
ence
an
d th
e
false c
l
ass
in
d
i
catin
g
n
o
n
ho
tspots occurr
e
n
ce. T
h
is
pa
p
e
r
prese
n
ts the r
e
sults of s
a
tell
i
t
e imag
e pr
oce
ssing
in
order
to deter
mi
ne t
he ra
dius
of a
hotspot s
u
ch t
hat
rand
o
m
poi
nts are ge
nerat
ed
outsid
e
a
hots
pot buffer as false alar
m d
a
ta.
Clusterin
g an
d
maj
o
rity filterin
g
w
e
re perfor
m
e
d
o
n
th
e L
a
n
d
s
at TM i
m
a
ge
to extract
burn
scars
in th
e s
t
udy ar
ea
i.e.
Roka
n H
ilir,
Ri
a
u
Provinc
e
Indo
n
e
sia. Ca
lcul
ati
on on b
u
rn ar
e
a
s and F
I
RMS
MODIS fire/hotspots in
20
06
results the rad
i
us
of a
hotsp
ot 0
.
9073
7 k
m
. T
h
erefore,
no
n-h
o
tspots w
e
re
r
and
o
m
ly
ge
ner
ated
in
are
a
s
that are
l
o
cate
d
0.907
37 k
m
a
w
ay from a h
o
tspot. T
h
ree d
e
c
ision tre
e
al
g
o
rith
ms i.e. ID
3, C4.5 a
nd ex
tende
d spati
a
l
ID3
have bee
n
ap
p
lied on
a datas
et
contai
ni
n
g
2
35 o
b
jects that
have the tru
e
class an
d 32
6
obj
ects that ha
ve
the fals
e cl
ass.
T
he r
e
sults
ar
e d
e
cisi
on
tree
s for
mo
del
in
g
hotspots
occur
r
ence
w
h
ich
h
a
ve th
e
accura
cy
of 49.02
% for the ID3 dec
isi
on tree,
65.
24
% for the C4.
5
decis
io
n tree, and 7
1
.66
%
for the exten
d
e
d
spatia
l ID3 dec
ision tre
e
.
Ke
y
w
ords
: hot
spot, satellite i
m
a
ge pr
ocessi
ng, data
min
i
n
g
, decisi
on tree
Copy
right
©
2015 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
Predi
ctive m
odel
s for hot
spot
s o
c
currence a
r
e
e
ssential to d
e
velop
so th
at
damag
es
cau
s
e
d
by
f
o
rest
f
i
r
e
s
ca
n
be minimi
ze
d.
No
wa
d
a
ys, the la
rge
n
u
mbe
r
of forest fire
data
has
been tri
gge
re
d the devel
op
ment of
data
mining
syste
m
s to an
alyze influen
cing
factors for fo
rest
fires
and
thei
r relation
s [1
-5]. Data mi
ni
ng i
s
a
growi
ng a
r
ea
in
co
mputer scie
n
c
e th
at is
wid
e
ly
use
d
to extract intere
sting
and valid info
rmation
fro
m
large d
a
ta. One of data mining techniqu
es
namely
cla
s
sification al
go
rithms h
a
ve b
een a
pplie
d t
o
mod
e
l hot
spots
occu
rre
nce [
6
-8]. Th
e
t
a
sk of
cl
as
si
f
i
cat
i
on aim
s
t
o
discov
e
r cla
ssif
i
cation
rule
s on a collectio
n of object
s
whi
c
h
is
rep
r
e
s
ente
d
in a relatio
n
(a dataset). T
he rul
e
s d
e
te
rmine la
bel
cl
asse
s of any
obje
c
t (Y) fro
m
the value
s
of its attribute
s
(X). De
cisi
on
tree
is
one
of famous
meth
ods i
n
creatin
g cla
s
sificatio
n
model
s. A d
e
ci
sion tree i
s
a mo
del e
x
pressin
g
cl
a
ssifi
cation
rul
e
s
whi
c
h ha
s thre
e types of
node
s i.e.
a
root
nod
e, in
ternal,
and
l
eaf no
de
s. A
ro
ot no
de
o
r
a
n
inte
rnal
nod
e
cont
ains
attribute tes
t
c
o
nditions
to
s
e
parate
objec
ts
that
h
a
ve
different
cha
r
acte
ri
stics. L
eaf nod
es
ho
ld
the target classes
(true
cl
ass and false class) to
which
obj
ects
will be
classi
fied. In hotspot
s
occurre
n
ce
modelin
g, th
e cl
asse
s
are hot
spot
s o
c
curren
ce
(T
rue
cla
s
s) a
nd n
on
hotspots
occurre
n
ce (False
cl
ass).
The attri
but
es of
obje
c
ts may incl
ude
som
e
supp
o
r
ting fa
ctors
for
hot
sp
ot
s o
c
c
u
rr
en
ce s
u
c
h
as phy
si
cal,
soci
o-e
c
o
n
o
m
ic, as
well as weathe
r d
a
ta. This stu
d
y
applie
d thre
e
deci
s
ion t
r
ee
algorithm
s i.
e. ID3,
C4.5
and extend
e
d
spatial I
D
3
[9] on the forest
fire dataset to develop mod
e
ls for
classifi
cation a
nd predictin
g hotspots o
c
curren
ce.
Hotspots
data are provided by s
e
veral ins
t
itut
ions
s
u
c
h
as
NASA/Univers
i
ty of Maryland
and
The ASEAN Speciali
sed Meteorol
ogical
Cent
re
(ASMC). In
addition to
hotspot
s as true
alarm
data, a
cla
s
sificatio
n
task in
mod
e
ling hot
spot
s
occurre
n
ce re
quire
s
non
-ho
t
spot p
o
ints a
s
false ala
r
m d
a
ta. This wo
rk aims to ge
nerate
n
on-h
o
tspot point
s near hot
spot
s to prep
are
the
target
cla
s
se
s fo
r m
odeli
ng h
o
tsp
o
ts occu
rren
ce
in Ro
kan
Hilir Di
stri
ct in
Ri
au
P
r
o
v
ince
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No. 3, September 20
15 : 1037 – 10
46
1038
Indone
sia. B
u
rn a
r
ea
pro
c
e
ssi
ng for t
he stu
d
y ar
e
a
wa
s pe
rformed to dete
r
mine the radi
us of
buffer fo
r a
h
o
tspot
and th
en out
side
of
the buffe
r, we gen
erated
non-hots
pot
points. T
h
e
r
e
are
two main
ste
p
s in bu
rn a
r
ea pro
c
e
s
sin
g
i.e.
image cla
ssifi
cation
and majo
rity filtering. Image
cla
ssifi
cation
identifies
cla
s
se
s on a
n
im
age ba
se
d o
n
its sp
ect
r
al
cha
r
a
c
teri
stics. In o
r
de
r
to
improve the
accuracy of image cl
as
sifi
cation, majo
ri
ty filtering is applie
d to remove very small
area
s resulte
d
from the im
age cl
assifica
tion.
Section 2 di
scu
s
ses mat
e
rials an
d meth
ods u
s
e
d
in o
u
r stu
d
y. The discussio
n
in
clud
es
the study a
r
e
a
and th
e dat
a utilize
d
in t
h
is
study
. In
addition, two
method
s in i
m
age p
r
o
c
e
s
sing
are outlin
ed in Section 2 n
a
mely cla
ssifi
cation an
d m
a
jority filtering. In Section 3, we pre
s
e
n
t the
results of b
u
r
n
are
a
p
r
o
c
essing
to g
e
nerate
fal
s
e
alarm
data.
The
stu
d
y i
s
summ
ari
z
e
d
in
Section 4.
2. Materials
and Method
s
2.1. Stud
y
Area and Data
The
study area is
Rokan
Hilir di
stri
ct i
n
Riau Provi
n
ce
i
n
Indonesia
(Figure
1). Rokan
Hilir
spans an area of 8,881.59 km
2
[10] or app
roxi
mately 10% of Riau’
s tota
l land area. The
site i
s
situate
d
in
the
are
a
between
10
0°16'
- 1
01°
2
1' Ea
st L
ongi
tude a
n
d
1°1
4' - 2
°
30'
North
Latitude [1
0]. Ro
ka
n Hili
r i
s
lo
cate
d in t
he
we
stern
p
a
rt of the
no
rt
h Sumate
ra, t
he
south
e
rn
part
of Beng
kali
s
distri
ct a
nd
Roka
n
Hulu
di
stri
ct, t
he e
a
stern
of Dumai
and
the
no
rthern
p
a
rt
of th
e
north Sumat
e
ra and Mal
a
cca
st
rait. Accordi
ng to [11], in
2002,
Rokan Hilir had
454,000
hecta
re
s (h
a) of p
eatland
s
or abo
ut 11.2
%
of the whole peatlan
d
s
i
n
Riau Provin
ce.
Figure 1. Study area
The data u
s
e
d
in burn are
a
pro
c
e
s
sing
are spread
and coo
r
dina
tes of FIRMS
MODIS
fire/hotsp
ots i
n
20
06, a
s
well
as the
L
and
sat TM
i
m
age
for
extractin
g
b
u
rn
area
s (Fig
ure
2)
(co
u
rte
s
y of t
he
U.S. Geol
ogical Survey
). The
a
c
qui
si
tion date
of the ima
ge i
s
2
4
July 200
6, the
resolution
of the ima
ge i
s
3
0
×3
0 m
2
an
d
the ban
d
com
b
ination
used
is 7, 4,
2. Thi
s
combi
natio
n
is u
s
e
d
in
th
e fire
man
a
g
e
ment a
ppli
c
ations
for po
st-fire
an
alysi
s
of burned
and non
bu
rned
foreste
d
area
s. In Figure 2,
the area
s co
vered by whit
e lines rep
r
e
s
ent burn a
r
e
a
s.
Figure 2. Lan
dsat TM satel
lite im
agery, Band co
mbin
ation 7, 4, 2
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Burn Are
a
Proce
s
sing to G
enerate Fal
s
e
Alarm
Data for Hot
s
pot…
(Im
a
s Sukae
s
ih Sitanggan
g
)
1039
2.2. Tools fo
r Data Proce
ssing
This
wo
rk utili
zed
ILWIS for pro
c
e
s
sing t
he satellite
i
m
age and Qu
antum
GIS
fo
r
spatial
data pro
c
e
ssing and d
a
ta
visualization
.
The Integra
t
ed Land a
n
d
Wate
r Information Syst
em
(ILWIS) i
s
o
pen source
softwa
r
e for
remote
sen
s
i
ng and g
eog
raphi
cal info
rmation sy
ste
m
s
develop
ed by
the Fa
culty
of Geo
-
Information S
c
ie
n
c
e
and E
a
rth
Ob
servatio
n, of the
Unive
r
sity
of Twe
n
te (
h
ttp://www.ilwi
s
.org/).
Quant
um GIS (
QGI
S) is
a fr
ee a
nd op
en
so
ur
ce
Geo
g
ra
phi
c
Information
S
y
stem. Seve
ral m
a
in fe
ature
s
p
r
ovid
ed
by QGIS i
n
cl
ude vi
suali
z
at
ion, man
agin
g
,
editing, analy
s
ing
spatial d
a
ta, and com
posi
ng pri
n
ta
ble map
s
(htt
p://www.q
g
is.
o
rg
).
2.3. Digital Image Proce
ssing
Digital imag
e
processin
g
refers to a proc
e
s
s that is condu
cted t
o
improve an
image.
The pu
rpo
s
e
of this pro
c
e
s
s is to a
ssi
st the ex
tractio
n
of informatio
n about obj
ects in a satellit
e
image. Imag
es in digital i
m
age p
r
o
c
e
s
sing a
r
e d
a
ta
acqui
red
by remote sen
s
ors
on satelli
te
,
aerial,
or
gro
und pl
atform
s. The
imag
e
s
a
r
e avail
abl
e in the
digit
a
l format
with
spe
c
ific spat
ial,
radio
m
etri
c, and sp
ectral chara
c
te
risti
c
s.
A digital image is rep
r
e
s
ented by a matrix in whi
c
h ea
ch ele
m
ent in the matrix is
denote
d
a
s
a
pixel (pi
c
tu
re
eleme
n
t). A
pixel is
asso
ciated
with
a
Digital
Numb
er (DN), a
s
well
as ro
ws and
colum
n
s
whi
c
h determin
e
the coo
r
di
nat
e
of the image. Referen
c
e [12] states th
at
Digital Nu
mb
ers
(DNs) re
pre
s
ent a di
screte me
asure of the radian
ce (L)
detecte
d by the
sen
s
o
r
s a
nd
measured i
n
Watts pe
r sq
uare
met
r
e p
e
r steradia
n
(W·m
–2
· sr
–1
).
Actual
phy
si
cal
measures of the radiatio
n are
continu
o
u
sly acq
u
ired
and then t
he analogi
cal/
digital conve
r
ters
will alter the
s
e mea
s
u
r
e
s
i
n
to discrete l
e
vel [12].
In addition to
DN value
s
a
n
d
the coordinat
e of
the imag
e, the spe
c
tral
resolution i
s
anothe
r
e
s
sential
ch
ara
c
teri
stic
of a
satellite im
age.
Acco
rdi
ng to [12], the spe
c
tral resolutio
n
is the wav
e
length inte
rval (
λ
) to whi
c
h the ra
dian
ce
rep
r
e
s
ente
d
by its Di
gital
Numb
er
refe
rs. Several i
m
age
s
can
b
e
availabl
e fo
r the
sam
e
scen
e
to co
mpo
s
e
a
multispe
ctral
imag
e. Ea
ch ima
ge
re
fe
rs the
radia
n
ce record
ed i
n
definite
sp
ectral
rang
es [12].
2.4. Image Classifica
tion
Image
cla
ssifi
cation
is
a p
r
oce
s
s to
re
co
gnize
classe
s on a
n
ima
ge
based o
n
its
spe
c
tral
cha
r
a
c
teri
stics [12]. Classification tasks ca
n be divided into two grou
ps: un
sup
e
rvised a
nd
sup
e
rvised. I
n
un
su
pe
rvised
cla
ssifi
cati
on, pixel
s
in
a data
s
et
are
clu
s
te
red
ba
sed
on
statist
i
cs
only and the
con
c
ept of
distan
ce
(for
example,
Eu
clide
an),
with
out any user-define
d
train
i
ng
cla
s
ses.
This app
roa
c
h
d
oes not
req
u
i
re exte
rnal
i
n
formatio
n fo
r a
s
signin
g
t
he pixel
s
to
the
different
cla
s
se
s. K-Me
an
s
clu
s
terin
g
is th
e
co
mmonly u
s
e
d
alg
o
rithm
in un
su
pervi
sed
cla
ssifi
cation.
In su
pervi
se
d cla
s
sificatio
n
, a pr
i
o
ri kn
owle
dge abo
ut
the
cl
asse
s
for a suffici
ent
numbe
r
of pi
xels (t
raini
n
g
set
s
) is ne
e
ded [1
2]. Th
e traini
ng
set
s
a
r
e
prepa
red by
an
an
alyst
based on hi
s/her perso
nal
experien
c
e, previou
s
kno
w
led
ge abo
ut thematic ma
ps, and in-fie
ld
s
u
r
v
ey. Pixels
in the super
v
is
ed c
l
assific
a
ti
on meth
od are divid
ed into two
sets n
a
mely
the
training
set and the test set. The training set is use
d
to determin
e
a classifica
tion model. The
model is th
en
utilized to cl
assify object
s
in the
test se
t. The succe
s
sful su
pe
rvised cla
s
sificati
on
depe
nd
s on the definition
of classe
s to whi
c
h t
he pixels should b
e
assi
gned. S
o
me tech
niqu
es
applie
d in the
supe
rvise
d
cl
assificatio
n
in
clud
e Ne
ural
Network and
Suppo
rt Vector Ma
chin
es.
2.5. Majorit
y
Filtering
Majority filtering i
s
a post-classifi
cation
meth
od to
improve the
accu
ra
cy of
imag
e
cla
ssifi
cation.
This m
e
thod
can
red
u
ce
the “sal
t-a
nd-pape
r” re
sult
ed
from per-pixel
cla
s
sifie
r
s.
Acco
rdi
ng to
[13], the majority filter is d
e
te
rmin
ed by
identifying a
nei
gh
borhoo
d
stru
cture
an
d
a
threshold
val
ue. Thi
s
m
e
thod
applie
s
a moving
wi
ndo
w in
whi
c
h the m
a
jorit
y
cla
s
s of pi
xels
within the
windo
w is a
s
signed to the
central pixel [1
4]. The majority class
of pixels is the m
o
st
freque
ntly occurrin
g value of a pixel and its neighbo
rs in the windo
w. A
standard majority filter
whi
c
h works in a 3×3 environment
whi
c
h con
s
id
ers 9
pixels in the
input map (I
LWIS
(3.5
)
h
e
lp
2008
). The
predomin
ant va
lue, i.e. mostl
y
freque
ntly
occurring val
u
e, or
cla
ss
na
me is
assig
n
ed
to the cente
r
pixel in the o
u
tput map. F
o
r ex
am
ple, 9
pixel values
encounte
r
ed i
n
the input m
ap
is sh
own in Table 1 [15].
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No. 3, September 20
15 : 1037 – 10
46
1040
Table 1. Nin
e
pixel values i
n
the input map
9 3
9
11 5
7
7 7
13
The p
r
ed
omi
nant value i
s
7. Therefore, the va
lue fo
r
the output pix
e
l is 7. T
he v
a
lue o
r
cla
ss n
a
me that is en
cou
n
t
ered first will
be assign
ed
to the center pixel as outp
u
t if there is no
pred
omin
ant value ca
n be
found in the 9
pixel values.
2.6. Decision
Tree Algorithms
De
cisi
on tree
is one
of
wi
dely u
s
ed
cl
a
ssifi
cation
m
e
thod
s in
dat
a minin
g
. A
deci
s
io
n
tree al
gorith
m
gen
erate
s
a tre
e
mod
e
l to cl
assify obje
c
ts to
their
cla
s
se
s based
on t
h
e
cha
r
a
c
t
e
ri
st
ic
s of
t
h
e o
b
je
ct
s.
A
d
e
ci
si
on t
r
e
e
h
a
s
three type
s
of nod
es:
1) a root n
ode
, 2)
internal n
ode
s, and 3
)
leav
es o
r
termin
al
node
s. The root node a
n
d
internal n
ode
s hold attri
but
e
test con
d
ition
s
to partition
records that
have diffe
rent
characte
ri
stics. Leave
s
n
ode
s (termi
n
a
ls)
store
cla
s
s la
bels of obj
ect
s
. Trave
r
si
ng
a deci
s
ion tree from the root node to t
he leave
s
no
des
results a
set
of classification rul
e
s. The rules
are utilized to descri
be char
acteri
stics of objects
and to pre
d
ict
unkn
o
wn cla
ss la
bel
s of object
s
.
The ID3 de
ci
sion tre
e
alg
o
rithm was d
e
velope
d by J. Ro
ss Q
u
i
n
lan du
ring t
he late
1970
s a
nd
e
a
rly 19
80
s.
The al
gorith
m
ha
s the
p
r
inci
ple,
whe
r
e it build
s th
e tree
in
gre
edy
manne
r sta
r
ting
from
th
e root,
and
sele
cting mo
st
in
formative fea
t
ures at
ea
ch
step
[16].
In
orde
r to sel
e
ct the best
feature for splitting
the set of objects, the algo
rithm cal
c
ula
t
es
informatio
n g
a
in. A feature
with the high
est info
rm
atio
n gain is
sele
cted a
s
a split
ting feature.
The
C4.5
de
cisi
on tree
al
gorithm
is a
su
ccesso
r of
ID3. Th
e
C4.
5
alg
o
rithm
u
s
e
s
al
so
Information
Gain to sele
ct optimal sp
litting attr
ibutes. This al
go
rithm uses a
different method
calle
d rule p
o
s
t-p
r
uni
ng. There a
r
e thre
e main tasks
in C4.5: 1) ge
nerate the tre
e
usin
g the ID3
algorith
m
, 2) conve
r
t the tree to a set of if-t
hen rul
e
s, and
3) p
r
une
ea
ch ru
le by removi
ng
pre
c
on
dition
s if the accu
ra
cy of t
he rule
increa
se
s wit
hout it [16].
Both ID3 an
d C4.5 u
s
e i
n
formatio
n g
a
in as
a me
asu
r
e fo
r attribute
sele
cti
on. The
formula
of inf
o
rmatio
n gai
n is
cal
c
ulate
d
as follows.
Let
p
i
b
e
th
e probability
that an a
r
bitrary
tuple in D b
e
l
ong
s to cla
ss C
i
, es
timated by |C
i
, D
|/|D| [17]. The entropy is a me
a
s
ure of expe
cted
informatio
n for cla
ssifying a
tuple in D. T
he formul
a of entropy is a
s
follows [17]:
)
(
log
)
(
2
1
i
m
i
i
p
p
D
Info
(
1
)
Whe
r
e
p
i
is t
he pro
bability
that an arbitrary tuple in
D belon
gs to
class C
i
a
nd i
s
e
s
ti
ma
te
d b
y
|C
i,D
|/|D|. The formul
a to
calcul
ate information ne
ede
d after u
s
in
g
A to split D i
n
to v partition
s to
cla
ssify
D [17
]
:
)
(
|
|
|
|
)
(
1
j
v
j
j
A
D
I
D
D
D
Info
(
2
)
Information
gain is
defi
ned a
s
the
differen
ce
betwe
en the
origin
al informatio
n
requi
rem
ent
(i.e. based o
n
just the p
r
oportio
n
of classe
s) a
nd
the new
req
u
irem
ent (i.e.,
obtaine
d after partitioning o
n
A) [17].
)
(
)
(
)
(
D
Info
D
Info
A
Gain
A
(
3
)
The exten
d
e
d
ID3
alg
o
rit
h
m is an
im
provem
ent of
the ID3 al
g
o
rithm
su
ch
that the
algorith
m
ca
n be dire
ctly applied on
a spatial dat
aset containi
ng a set of layers [9]. The
algorith
m
use
s
spatial information gai
n to sele
ct
the
best laye
r for splitting the
spatial d
a
taset.
The form
ula
of spatial inf
o
rmatio
n gai
n is defin
ed
as follo
ws [9]. Let a target
attribute C i
n
a
target layer
S has
l
di
stin
ct cla
s
se
s (i.
e
. c
1
, c
2
, …, c
l
), entropy for S re
pre
s
e
n
ts the expe
cted
informatio
n n
eede
d to dete
r
mine the
cla
ss of
tuple
s
in
the dataset a
nd define
d
as:
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Burn Are
a
Proce
s
sing to G
enerate Fal
s
e
Alarm
Data for Hot
s
pot…
(Im
a
s Sukae
s
ih Sitanggan
g
)
1041
)
(
)
(
log
)
(
)
(
)
(
1
2
S
SpatMes
S
SpatMes
S
SpatMes
S
SpatMes
S
H
i
i
c
l
i
c
(
4
)
Sp
a
t
Me
s(
S)
re
p
r
es
en
ts
s
p
a
t
i
a
l
me
as
ur
e o
f
l
a
ye
r
S th
at can
be
are
a
of interse
c
ti
on polyg
on
s
or
distan
ce b
e
tween two
spati
a
l feature
s
.
Let an expl
a
natory attrib
u
t
e V in an e
x
planatory
(n
on-ta
rget) layer L
ha
s
q
distin
ct
values (i.e. v
1
, v
2
, …, v
q
).
We
pa
rtition t
he o
b
je
cts in
target l
a
yer S acco
rdi
ng to
the laye
r L
th
en
we have a se
t of layers L(v
i
, S) for each possibl
e value v
i
i
n
L
.
I
n
o
u
r
w
o
r
k
,
w
e
a
s
s
u
m
e
t
h
a
t
t
h
e
layer L cove
rs all are
a
s in
the layer S.
The
expe
cted
entropy valu
e for splitting
is given by:
))
,
(
(
)
(
))
,
(
(
)
|
(
1
S
v
L
H
S
SpatMes
S
v
L
SpatMes
L
S
H
j
q
j
j
(
5
)
The sp
atial in
formation g
a
i
n
for layer L i
s
given by:
Gain(L) =
H(S)
H
(
S
|
L
)
(
6
)
Gain(L)
den
o
t
es ho
w mu
ch inform
ation
woul
d be
ga
ined by b
r
an
chin
g on th
e
layer L.
The l
a
yer L
with the
highe
st informatio
n
gain,
(Gai
n
(
L))
,
i
s
ch
os
e
n
as
th
e
s
p
l
i
t
ti
n
g
l
a
ye
r
a
t
a n
o
de
N in a sp
atial deci
s
io
n tree.
3. Results a
nd Discu
ssi
on
3.1. Cluste
ring and Majo
rit
y
Filtering
The main p
u
r
po
se of bu
rn are
a
pro
c
e
ssi
ng is to d
e
fine the ra
d
i
us of a buff
e
r for a
hotsp
ot such
that ra
ndom
points a
s
non
-hotspots
will
be g
ene
rated
outsi
de
the
b
u
ffer. Th
ere
a
r
e
two mai
n
ste
p
s in
imag
e
pro
c
e
ssi
ng:
clusteri
ng, o
r
unsupe
rvise
d
cla
ssifi
cation
, to grou
p pix
e
ls
and m
a
jority
filtering to re
move very
small area
s.
These two ta
sks
we
re
con
ducte
d u
s
ing
the
tool Ilwis 3.7.
To p
e
rfo
r
m
clu
s
terin
g
a
n
d
majo
rity filtering,
we
det
ermin
ed the
map
sub
s
et f
o
r
each ban
d (b
and 7, 4, 2).
The coordina
tes u
s
ed
to
create a sub
s
et
of the map a
r
e (6314
78.2
3
,
1662
90.54
) a
nd (747
008.0
3
, 8744
9.25
). Clu
s
teri
ng
was a
pplie
d on
the sub
s
et of
image
with t
h
e
numbe
r of clu
s
ter i
s
15. Fig
u
re 3 sho
w
s t
he re
sult of cl
usteri
ng on th
e sub
s
et of image.
Figure 3. Clu
s
terin
g
on the
sub
s
et of image, numb
e
r
of cluste
r is 1
5
Furthe
rmo
r
e,
majority
filte
r
to r
e
mov
e
v
e
ry sm
all areas wa
s
app
lied fou
r
time
s in th
e
clu
s
tere
d ima
ge. Th
e resu
lts are p
r
ovid
ed in
Figu
re
4. The im
age
s resulted fro
m
the 1
st
and
the
2
nd
majority fi
ltering contai
n small area
s as sh
o
w
n i
n
the recta
n
g
u
lar re
gion.
The sm
all areas
were re
duced
after we appl
ied the 3
rd
and the 4
th
majority filtering. The imag
es
before a
nd af
ter
applying maj
o
rity filter are given in Figure 5. The
u
s
e of majority
filtering four times re
sults
the
smooth
e
r ima
ge com
p
a
r
ed
to those befo
r
e applying m
a
jority filterin
g as sho
w
n in
Figure 5.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No. 3, September 20
15 : 1037 – 10
46
1042
(a) 1
st
maj
o
rit
y
filtering
(b) 2
nd
majority filtering
(c
) 3
rd
majorit
y
filtering
(d) 4
th
majo
rit
y
filtering
Figure 4. Applying majority filtering on th
e image
(a) B
e
fo
re ma
jority
filtering
(
b) After maj
o
rity filtering
Figure 5. Image before (a
) and after
(b)
applying maj
o
rity filtering
False
alarms were gen
era
t
ed outsid
e
b
u
ffers of h
o
tspots a
s
tru
e
alarm d
a
ta u
s
ing the
tool Qua
n
tum
GIS 1.7.2. T
he buffe
r op
e
r
ation th
at
is
available i
n
Quantum
GI
S 1.7.2 is
ap
plied
to point featu
r
es (ve
c
tor fo
rmat).
Th
erefore, the im
ag
e in the
ra
ste
r
form
at (tiff file) resulted from
majority
filteri
ng wa
s conve
r
ted
to
the ve
ctor form
at (p
olygon).
Fig
u
r
e
6
sho
w
s p
o
lygon
s o
n
ly fo
r
bare
d
land
s (clu
ster 8
)
, burn area
s (clu
st
er 13
), and n
e
w bu
rn a
r
ea
s (clust
er 15
).
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Burn Are
a
Proce
s
sing to G
enerate Fal
s
e
Alarm
Data for Hot
s
pot…
(Im
a
s Sukae
s
ih Sitanggan
g
)
1043
Figure 6. Burn area
s an
d b
a
red la
nd
s
In orde
r to ge
nerate n
on h
o
tspot poi
nts,
this
work inv
o
lved only ne
w burn areas (clu
ster
15). Th
ese b
u
rn
scars
were ove
r
lai
d
with hot
spot
s that occu
rre
d in two
we
eks befo
r
e t
he
acq
u
isitio
n d
a
te for im
age
(24
July
200
6) (Figu
r
e
7).
There
were
298 h
o
tsp
o
ts
in non
-p
eatla
nds
and one h
o
tspot in peatl
and
s found i
n
the perio
d 10 – 24 July
2006 in whi
c
h 24
3 hotsp
ots
occurre
d
in the burn
scars.
Figure 7. Ne
w burn area
s and hot
spot
s for the peri
o
d
10 – 24 July
2006
To avoid sin
g
l
e pixels label
ing from the image,
we
co
nsid
er only th
e burn
scars with the
area
at le
ast
1 h
a
that i
s
equivalent
to
aro
und
3
×
3
Land
sat T
M
pixels. Th
erefore, b
u
rn
sca
r
s
with the
are
a
less tha
n
1
ha were re
m
o
ved. Thi
s
a
ppro
a
ch i
s
al
so a
dopte
d
i
n
the
work
of [18].
Table 2 p
r
ovi
des the
sum
m
ary of hotspots for t
he p
e
riod 1
0
-24 July 2006 in n
e
w bu
rn a
r
ea
s at
least
1
ha wh
ere
the count of
hotspot
s
(a
ssoci
a
ted
wit
h
burn scars) is
243.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No. 3, September 20
15 : 1037 – 10
46
1044
Table 2. De
n
s
ity of hotspo
t
s and a
r
ea fo
r one hot
spot
Area in km
2
Densit
y
(num
ber
of hotspot
per km
2
)
Area for o
ne hots
pot (Area in km
2
/
c
ount of
hotspot), in km
2
Max 44.75715
51.54616
11.18929
Average
9.78013
4.14282
2.58657
Min 0.01940
0.08937
0.01940
Sum 557.46733
For si
mplicity
,
it is assume
d that the are
a
for a
hotsp
ot is a circle
becau
se a bu
ffer of a hotspot
is
rep
r
e
s
ente
d
in
a
circle.
Th
e
radi
us of the
ci
rcle
is given
by
/
km
in
hotspot
one
for
area
2
whe
r
e
= 3.1415
9. As show
n in Tabl
e 2, the area
for one h
o
tsp
o
t in average
is 2.5865
7 km
2
,
therefo
r
e the
radiu
s
of the circle i
s
0.9073
7 km. T
h
is value i
s
con
s
id
ere
d
a
s
the ra
diu
s
of a
buffer for a h
o
tspot. Outsi
de the buffers, random
p
o
in
ts are g
ene
ra
ted as false a
l
arm data.
3.2. Genera
ti
ng Targe
t
O
b
jects
for Ho
tspo
t Prediction Models
As many 517
hotsp
ots w
e
re found in Ro
kan
H
ilir in 20
08. These hot
spot
s we
re a
c
qui
red
by the MO
DIS satellite
se
nso
r
. Buffers with the
radi
us of
0.9073
7 km
we
re
created
for e
a
c
h
hotsp
ot usin
g
Quantum GI
S 1.7.2. Furthermo
re,
as m
any 513 no
n hotsp
ot point
s we
re rand
o
m
ly
gene
rated o
u
t
side buffers. Therefore a n
on hotsp
ot point is locate
d at least 0.90737
4 km aw
ay
from a hot
sp
ot (Figu
r
e 8
)
. We
con
s
ide
r
these p
o
ints
as fal
s
e ala
r
m data w
h
ich
are
combi
n
e
d
to
obtain targ
et obje
c
ts for
th
e cla
ssifi
catio
n
task.
Figure 8. Tru
e
and false al
arm data a
s
target obj
ect
s
3.3. Predictiv
e
Models fo
r Hots
pots O
ccurre
nce
The de
cisi
on
tree algo
rith
ms nam
ely ID3 an
d C4.5
have bee
n a
pplied o
n
the
forest fire
dataset. Furt
her
discu
s
sio
n
re
garding
these alg
o
rith
ms
can
be fo
und in
[19] a
nd [16]. Th
e
s
e
algorith
m
s
are availabl
e in
the data mi
n
i
ng tool
kit W
e
ka
3.6.6. In
addition,
we
cre
a
ted
a mo
del
for p
r
edi
cting
hotsp
ots
occurren
ce u
s
in
g
our
propo
se
d metho
d
n
a
m
ely the exte
nded
sp
atial I
D
3
algorith
m
[9]. The al
gorith
m
is a
n
imp
r
ovement
of t
he existin
g
spatial ID3
alg
o
rithm int
r
od
uce
d
by [20]. Instead of ru
nnin
g
on the
non
-sp
a
tial data
s
et, our p
r
op
o
s
ed
algo
rith
m wo
rks o
n
the
spatial d
a
taset which con
t
ains several
explanato
r
y layers a
nd
one target la
yer. In a sp
atial
databa
se, layers
stores sp
atial object
s
that can be
re
pre
s
ente
d
either in point, line, or polygo
n
.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Burn Are
a
Proce
s
sing to G
enerate Fal
s
e
Alarm
Data for Hot
s
pot…
(Im
a
s Sukae
s
ih Sitanggan
g
)
1045
Explanatory l
a
yers in
clud
e
layers
of su
pportin
g
fa
ct
ors for forest
fire
s
whe
r
ea
s the
target l
a
yer
con
s
i
s
ts of ho
tspots a
s
tru
e
alarm data a
nd non
-hot
sp
ot points a
s
false al
arm d
a
t
a.
The
sp
atial d
a
taset fo
r
mo
deling
hot
spo
t
s o
c
curren
ce
ha
s
1030
o
b
j
e
cts,
one
targ
et layer
and ten
expl
anatory l
a
yers (dista
nce to
nea
re
st city
cente
r
(dist_
city), distan
ce
to nea
re
st riv
e
r
(dist_
river), d
i
stan
ce to
ne
are
s
t ro
ad
(d
ist_ro
ad
), income
sou
r
ce, land
cove
r, peatlan
d
type,
peatlan
d
dep
th, pre
c
ipitation in mm/d
a
y
, screen te
mperature i
n
K, 10m win
d
spe
ed in m/
s. In
orde
r to appl
y the ID3 and C4.5 algo
ri
thm, we con
ducte
d seve
ral step
s to prepare a data
s
et
from a sp
atial
dataset on fo
rest fire
s. Th
ese
step
s are
as follow
s
:
1.
Cal
c
ulating di
stan
ce from t
a
rget obj
ec
t
s
to neare
s
t cit
y
center, rive
r, and road
2.
Relatin
g
laye
rs that
contai
n explanato
r
y
at
tributes
an
d the targ
et layer that con
s
ist
s
of target cla
s
se
s.
3.
Integrating all
layers in by matchin
g
ide
n
tifiers of obj
ects to create
a dataset for the
cla
ssifi
cation task.
4.
Remove d
upli
c
ate obj
ect
s
in the dataset
Applying the
s
e ste
p
s on
th
e spatial
data
s
et o
n
fo
re
st
fires re
sult
s 5
61 o
b
je
cts
(2
35 tru
e
cla
s
ses a
nd 3
26 false
cla
s
ses). Th
e expe
rimental
resul
t
s sho
w
that the accu
ra
cy of ID3 deci
s
io
n
tree is 4
9
.02
%
and the accuracy of C
4
.5 deci
s
io
n tre
e
is 65.24%
.
Furthe
rmo
r
e,
in term num
b
e
r
of rules g
ene
rated from th
e trees, the C4.5 algo
rith
m outperfo
rm
s the ID3 alg
o
rithm. The ID3
algorith
m
h
a
s
2
7
0
leave
s
with
peatl
and type
a
s
the first te
st attribute
w
here
a
s the
C4.5
algorith
m
pro
duces
only 3
5
rule
s a
nd th
e first test
attribute of the tree is p
eatlan
d
type. The C
4
.5
deci
s
io
n tree
has
several test attribute
s
to classify
the object
s
to the target cla
sses, i.e. peatland
type, distan
ce to nea
re
st road, di
stan
ce
to neares
t
ci
ty center,
screen temp
erature, di
stan
ce
to
nearest rive
r, and incom
e
sou
r
ce. For compa
r
ison, our pro
p
o
s
ed
algorith
m
(Sitangg
ang et a
l
.
2011
) gen
era
t
ed a spatial deci
s
io
n tree
with 134 lea
v
es and the first test layer of the tree is
incom
e
source. The spatia
l deci
s
ion t
r
e
e
ha
s hi
gh
er
accuracy tha
n
the ID3 a
n
d
C4.5 d
e
ci
si
on
trees i.e. 71.1
2
%
.
After pruning, the sp
atial deci
s
ion
tree be
come
s smaller
with 122 leave
s
a
nd
its accu
racy i
s
71.66%
.
4. Summar
y
This
work p
r
oce
s
sed b
u
rn are
a
s in th
e study area
to generate non hot
spot
points a
s
false ala
r
m d
a
ta in modeli
ng hotsp
ot occurre
n
ce mo
dels. Pro
c
e
ssing on the La
ndsat TM image
and FI
RMS
MODIS fire/h
otspot
s in
20
06 sho
w
s th
at the a
r
ea f
o
r o
ne h
o
tsp
o
t in averag
e is
2.5865
623
89
km
2
. Theref
ore
with the assumptio
n
that the are
a
for a hotspot
is a circle, the
radiu
s
of a b
u
ffer is 0.90
7
374 km. Exp
e
rime
nts on
t
he fore
st fire
s data
s
et re
sult three de
ci
sion
tree mo
del
s
for hotspots
occurre
n
ce p
r
edi
ction.
Th
e data
s
et co
ntains i
n
fluen
cing fa
ctors f
o
r
forest fi
re
s, h
o
tspot
s a
s
tru
e
ala
r
m d
a
ta
and n
o
n
-
hot
spots
as fal
s
e
alarm
data. T
he three m
o
d
e
ls
are the ID3 d
e
ci
sion tre
e
with the accu
racy of
49.02
%
,
the C4.5 deci
s
io
n tree
with the accu
racy
of 65.24%
an
d the spatial
deci
s
io
n tree
with the accu
racy of 71.66
%
.
Ackn
o
w
l
e
dg
ements
The auth
o
rs
woul
d like to
thank In
don
esia
Directo
r
ate Gen
e
ral
of High
er Ed
ucatio
n
(IDG
HE), Mi
nistry of
Nati
onal Ed
ucation, Indo
ne
sia for
su
ppo
rting Ph
D Sch
o
larship
(Con
tract
No. 172
4.2/D
4
.4/2008
) an
d Southea
st
Asian
Regi
on
al Cente
r
for
Grad
uate Stu
d
y and R
e
se
arch
in Agriculture
(SEARCA
) for partially sup
porting the
re
sea
r
ch.
Referen
ces
[1]
T
a
y
SC, Hsu
W, Lim KH, Yap LC.
Spati
a
l
d
a
ta
mi
nin
g
: clu
s
tering
of h
o
t s
pots a
nd
patter
n
rec
ogn
itio
n
.
Paper
prese
n
t
ed at the I
EEE Internati
ona
l G
eosci
e
n
ce a
nd R
e
mote Sens
ing
S
y
mp
osium
(IGARSS'
03). T
oulouse. 20
0
3
.
[2]
Yu L, Bian F
.
An Incrementa
l
Data Mini
ng M
e
thod for Spati
a
l Associati
on
Rule i
n
GIS Based F
i
repro
o
f
System
. Pap
e
r
presente
d
at the Internati
o
n
a
l
Conf
er
ence o
n
W
i
reless Co
mmunicati
ons,
Net
w
o
r
kin
g
and Mo
bil
e
Co
mputin
g, (W
iCom 200
7). Sha
ngh
ai, Chi
na. 2
007.
[3]
Prasad
KSN,
Ramakris
hn
a
S. An Aut
ono
mous F
o
r
e
st F
i
re D
e
tectio
n S
y
stem
bas
ed
o
n
Sp
atial
D
a
t
a
Minin
g
and
F
u
zz
y
Lo
gic.
Int
e
rnati
ona
l J
o
u
r
nal
of C
o
mpu
t
er Scie
nce
a
nd
Netw
ork S
e
curity.
2
008;
8(12): 49-
55.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No. 3, September 20
15 : 1037 – 10
46
1046
[4]
Hu
L, Zhou G
,
Qiu Y.
Ap
pl
icatio
n of
Apri
ori A
l
gor
ith
m
t
o
the
D
a
ta Mi
nin
g
of
the
W
i
ldfire
. Paper
prese
n
ted at t
he 6th
intern
atio
n
a
l co
nfere
n
c
e on F
u
zz
y s
y
stems a
nd k
n
o
w
le
dg
e disc
o
v
er
y
.
T
i
anji
n
,
Chin
a. 20
09.
[5]
Anga
ya
rkka
ni
K, Radh
akrish
nan. Efficie
n
t F
o
rest
F
i
re De
tection S
y
stem
: A Spatial Dat
a
Mini
ng a
n
d
Image Proc
ess
i
ng Bas
ed A
p
p
r
oach.
Intern
ati
ona
l Jour
nal
of Co
mputer Sc
i
ence a
nd N
e
tw
ork Security.
200
9; 9(3): 100
- 107.
[6]
Stojan
ova D,
Panov P, Kob
l
er
A, Džerosk
i
S,
T
a
škova K.
Learni
ng to
Predict F
o
res
t
F
i
res w
i
th
Different Dat
a
Minin
g
T
e
chn
i
qu
es
. Paper
prese
n
ted at t
he co
nferenc
e
on Data M
i
ni
ng a
nd D
a
ta
W
a
reho
uses.
L
j
ubl
ja
na, Slov
e
n
ia. 20
09.
[7]
Sitanggang IS, Ismail MH. Classifi
cati
on mo
del for
hotsp
ot
occurre
nces
us
ing
a d
e
cisi
on t
r
ee meth
od.
Geomatics, Na
tural Ha
z
a
rds a
nd Risk
. 20
11;
2(2): 111-
12
1.
[8]
Sitang
ga
ng IS, Yaako
b
R, Mustaph
a N, Nu
rudd
in AN
. Ap
plicati
on
of cla
ssificatio
n
al
go
rithms in dat
a
minin
g
for
hot
spots occ
u
rren
c
e pr
edicti
on
i
n
Ri
au
Provi
n
ce Ind
o
n
e
sia.
Journ
a
l of
T
h
eoretic
al an
d
Appl
ied Infor
m
ation T
e
ch
no
lo
gy
. 2012; 4
3
(2)
:
214-221.
[9]
Sitang
ga
ng IS,
Yaak
ob
R, M
u
staph
a N,
Nu
rudd
in A
N
.
An e
x
te
nd
ed
ID
3 d
e
c
i
s
i
o
n
tree
a
l
go
ri
thm
for
spatia
l d
a
ta
.
Paper
pres
ent
ed
at the
IEEE Internati
o
n
a
l
Co
nferenc
e
o
n
Sp
atial
D
a
ta
Mini
ng
an
d
Geogra
phic
a
l
Kno
w
l
e
d
ge Ser
v
ices (ICSDM). F
u
zhou, Chi
n
a
.
2011.
[10]
Roka
n Hi
lir
District. Overvie
w
of D
i
stri
ct. last mod
i
fied
200
9. Accessed M
a
y 3
0
, 20
12.
http://
w
w
w
.
ro
hi
lkab.g
o
.id/?tam
pil=
li
nka
ndact=
profil
and
id=
4
.
[11] W
a
h
y
unto,
Rit
ung
S,
S
u
p
a
rto, Sub
agj
o H.
Peatla
nd
distri
butio
n a
nd
its carbo
n
co
ntent
in S
u
mater
a
and Ka
li
ma
nta
n
.
Project of Climate Ch
an
ge
, Forests and P
eatla
nds in I
ndo
nesi
a
. Bog
o
r: W
e
tlands
Internatio
na
l – Indon
esi
a
Prog
ramme an
d W
i
l
d
life H
abitat. C
ana
da. 20
05.
[12] Gomarasca
M.
Basics of Geomatics
. Ne
w
Y
o
rk: Spring
er. 200
9.
[13]
KIM KE. 1996
. Adaptiv
e Maj
o
rit
y
F
ilteri
ng f
o
r Co
nte
x
tua
l
Classific
a
tio
n
of Remote
Se
nsin
g Dat
a
.
Internatio
na
l Journ
a
l of Re
mote Sensi
ng.
1
996; (17): 1
083
-108
7.
[14]
Cant
y Morton
J. Im
age Anal
ysis, Class
ifi
c
ation, an
d C
han
ge
D
e
tecti
on in R
e
mote
Sensin
g:
w
i
t
h
Algorit
hms for ENVI/IDL. Boca Ra
ton: CRC Press.
2010: 237-238.
[15]
ILWIS 3.4 Open. ILWIS
(3.4
) Help. last
modified
2008. Acce
ssed
July
13, 2011,
http://spatial-
analy
st.net/ILWIS/help.html.
[16]
Marslan
d
S. Machi
ne L
earn
i
n
g
: An Algor
ith
m
ic Perspectiv
e
. Chapm
an &
Hall/C
RC mac
h
in
e lear
ni
ng &
pattern reco
gn
i
t
ion seri
es. Boca Raton: CR
C
Press. 2009.
[17]
Han J, Kambe
r
M. Data Mi
ning: Conc
epts and T
e
chni
que
s. Sec
ond e
d
iti
on. T
he Morgan Kaufma
n
n
series i
n
data
mana
geme
n
t systems. San F
r
ancisc
o
: Morga
n
Kaufman
n
. 2
006.
[18]
T
anse
y
K, Bes
t
on J, Hosc
ilo
A, Page SE, P
a
red
e
s
Hern
án
dez C
U
. Re
lati
onsh
i
p
bet
w
e
e
n
Mod
i
s F
i
r
e
Hot Spot C
oun
t and Bur
ned
Area i
n
a D
egr
ade
d T
r
opi
cal
Peat S
w
am
p F
o
rest in C
entra
l Kalim
anta
n
,
Indon
esi
a
.
Jou
r
nal of Geop
hy
sical Res
earc
h
. 2008; 11
3 (D
23): 1-8.
[19]
Quinl
an JR. Inductio
n
of deci
s
ion trees.
Mac
h
in
e Le
arni
ng.
198
6; 1(1): 81-
106.
[20]
Rinziv
ill
o S, F
r
anco T
.
Classif
i
catio
n
i
n
Geo
g
r
aph
ical I
n
for
m
ation
Syste
m
s
. Pap
e
r pres
ent
ed at th
e
8th
Europ
e
a
n
Con
f
erence o
n
Pri
n
cipl
es an
d Practice of
Kno
w
led
ge Disc
o
ver
y
in
Databases.
Pisa, Italy
.
200
4.
Evaluation Warning : The document was created with Spire.PDF for Python.