TELKOM
NIKA
, Vol.12, No
.2, June 20
14
, pp. 511~5
1
8
ISSN: 1693-6
930,
accredited
A
by DIKTI, De
cree No: 58/DIK
T
I/Kep/2013
DOI
:
10.12928/TELKOMNIKA.v12i2.2036
511
Re
cei
v
ed Fe
brua
ry 26, 20
14; Re
vised
Ap
ril 28, 201
4; Acce
pted
May 15, 20
14
A Decision Tree Based
on Spatial Relationships for
Predicting Hotspots in Peatlan
d
s
Imas Sukaes
ih Sitanggan
g
*
1
, Razali Yaakob
2
, Norw
ati Mustapha
3
, Ainuddin A. N.
4
1
Department o
f
Computer Sci
ence, F
a
cult
y o
f
Natu
ral Scie
n
c
e and Math
e
m
atics, Bogor
Agricult
ural
Univers
i
t
y
, Ind
ones
ia
2,3
F
a
cult
y
of C
o
mputer Sci
e
n
c
e and Inform
a
t
ion T
e
c
hnol
og
y, Univ
ersiti Pu
tra Mala
ysi
a
, Mala
ysi
a
4
Institute of
T
r
opic
a
l F
o
restr
y
and
F
o
rest Products (INT
ROP), Universiti P
u
tra Mala
ys
ia, 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@fsktm.upm.edu.m
y
2
,
nor
w
a
t
i
@fsktm.upm.edu.m
y
3
, ain
udd
in@f
orr.upm.ed
u
.m
y
4
A
b
st
r
a
ct
Predicti
ng h
o
tspot occurre
nc
e as an ind
i
cat
o
r of fo
rest and lan
d
fires is essenti
a
l in d
e
v
elo
p
in
g
an ear
ly warning system
for fire prev
ention. This work
applied
a spatial decision
tree algorithm
on spatial
data of forest fi
res. T
he alg
o
rit
h
m is th
e i
m
pr
ove
m
e
n
t of the conventi
o
n
a
l
decisi
on tre
e
a
l
gorit
hm
in w
h
i
c
h
the d
i
stance
a
nd to
pol
og
ical
relati
ons
hips
are i
n
cl
ude
d t
o
grow
u
p
s
p
atial
dec
isio
n t
r
ees. Sp
atial
d
a
ta
consiste
d of a target layer
an
d t
en exp
l
an
atory layers re
pr
esenti
ng p
h
ysi
cal, w
eather, socio-
econ
o
m
ic
an
d
peatl
a
n
d
ch
ara
c
teristics i
n
th
e
study
are
a
R
o
kan
Hil
ir
D
i
stri
ct, Indon
esia.
T
a
rget o
b
j
e
cts
w
e
re hots
pots
of
200
8 a
n
d
no
n-
hotspot
poi
nts. T
he r
e
su
lt w
a
s a
pru
n
e
d
s
patia
l d
e
cisi
on
tree w
i
th
122
leav
es
and
th
e
accuracy
of 7
1
.
66%. T
h
e sp
atial tr
ee
has
p
r
oduc
ed
hi
gher
accur
a
cy tha
n
the
non-s
pati
a
l trees th
at w
e
r
e
created
usin
g the ID3 an
d
C4.5 al
gorit
h
m
. T
he ID3
dec
ision tre
e
ha
d
accuracy of
49.02
% w
h
ile
the
accuracy of C4
.5 decisi
on tree
reache
d
65.2
4
%
.
Ke
y
w
ords
: sp
atial d
e
cisi
on tree al
gorith
m
, s
patia
l rela
ti
ons
hip, hotsp
ot, forest fires, peatl
and
1. Introduc
tion
Predi
cting ho
tspots
occu
rrence is
con
s
i
dere
d
a
s
one
of activities for fire p
r
eve
n
t
ion in
orde
r to red
u
ce d
a
mag
e
s
be
cau
s
e of
f
o
r
e
s
t
a
n
d
la
n
d
f
i
r
e
s
.
Hotspot
s (a
ctive fires) indicate
spatial di
strib
u
tion of forest and land fires.
Hotspots occurren
ce
model
s have
been devel
op
ed
in seve
ral stu
d
ies u
s
in
g ge
ogra
phi
cal inf
o
rmatio
n syst
ems a
nd re
m
o
te sen
s
in
g tech
nolo
g
ies.
In
addition,
data
mining
a
s
o
ne of
growin
g
areas in
co
mputer scie
n
c
e
ha
s b
een
applie
d to
sp
atial
forest fire
s da
tasets to obta
i
n cla
ssifi
cati
on model
s for hotspot
s occurren
ce.
De
cisi
on tree
is a famous
method for
cl
assi
ficatio
n
tasks a
nd it has be
en appl
ied to a
broa
d ra
nge
of application
s
. Some of deci
s
ion tre
e
a
l
gorithm
s are
ID3, C4.5 as a succe
s
so
r of
ID3, and CA
RT (Cla
ssifi
cation and Re
gre
ssi
on Tre
e
). The
s
e alg
o
rithm
s
are d
e
sig
ned for n
on-
spatial
data
s
ets. T
he
different
bet
wee
n
spatial
and
non
-spatial
deci
s
io
n d
a
ta
is th
at in
th
e
spatial d
a
ta, an obje
c
t ma
y have a sig
n
ificant influe
nce o
n
neigh
borin
g obje
c
t
s
. Therefo
r
e
,
improvem
ent
of the n
o
n
-
spatial
de
cisi
on tre
e
al
gorithm ha
s b
e
en d
one
by
involving spa
t
ial
relation
shi
p
s
betwe
en two
spatial o
b
je
cts.
Several
studi
es h
a
ve be
e
n
co
ndu
cted
on s
patial d
e
c
isi
on tree al
gorithm
s. Th
e sp
atial
deci
s
io
n tree
algorithm
was introdu
ce
d
in [1]
base
d
on the ID3
algorith
m
involving the sp
atial
relation
shi
p
Dista
n
ce. Th
e spatial
bina
ry tree al
g
o
rit
h
m wa
s p
r
op
ose
d
in [2] that works on t
he
dataset conta
i
ning p
o
int, li
ne, and
polyg
on featu
r
e
s
.
An extensi
o
n
of the
CART
method,
call
ed
the SCART (Spatial Cla
ssi
fication an
d Reg
r
e
ssi
on T
r
ee
s), was d
e
v
eloped in [3]
.
In the SCART,
topologi
cal a
nd
di
stan
ce relation
ship
s are used
to
t
e
st
whethe
r a
pre
d
ictive att
r
ibute
belo
n
g
s
to
the nei
ghbo
r
table. The
S
C
ART
was a
pplied
to an
a
l
yze traffic ri
sk u
s
in
g a
c
cid
ent informati
o
n
and the
m
ati
c
info
rmatio
n abo
ut ro
a
d
netwo
rks, popul
ation
cen
s
u
s
, buil
d
ing
s
, and
other
geog
rap
h
ic n
e
ighb
orh
ood
details [3].
A spatial
de
cision
tree
ba
sed
on
the I
D
3
algo
rithm
that
works on pol
ygon feature
s
wa
s intro
d
u
ce
d in [4].
The algo
rith
m was a
ppli
ed to cla
ssif
y
the
averag
e (p
er
farm) ma
rket value of sold
agri
c
ul
tu
ral produ
cts ba
se
d
on c
limate, the distri
butio
n
of the princip
a
l aquifers, crop
s cultivate
d
, and
the numbe
r of cattle and calve
s
per are
a
. The
spatial e
n
trop
y-based de
ci
sion tree met
hod was p
r
op
os
e
d
in [5] which u
s
e
s
the
spatial relati
on
Dista
n
ce to relate poi
nt an
d polygo
n
fea
t
ures. T
he al
gorithm
wa
s
use
d
to cl
assify gross valu
es
of agricultural
output [5] and the air pollu
tion index in main citie
s
in Chin
a [6]. A
new formula for
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 12, No. 2, June 20
14: 511 – 51
8
512
spatial i
n
form
ation gai
n wa
s p
r
op
ose
d
i
n
[7] by
incl
u
d
ing
spatial
a
u
toco
rrelation
(nei
ghbo
rh
o
o
d
split a
u
tocorrelation
ratio
)
.
The
alg
o
rith
m was ap
plie
d to th
e
ra
ste
r
format that
i
s
rep
r
e
s
ente
d
in
a set of pixels.
This work de
veloped a cl
assifier for predict
in
g hotspots occu
rre
n
ce u
s
ing th
e spatial
cla
ssifi
cation
algorith
m
n
a
mely the spatial de
ci
sio
n
tree
algo
rithm [8]. The
algorith
m
is an
extensio
n of t
he
covention
a
l ID3
algo
rithm [9]. The
n
e
w al
go
rithm
prop
osed i
n
[8] ca
n
work
on
spatial d
a
tasets co
ntainin
g
point, line a
nd polygo
n
feature
s
as
rep
r
esentation
s
of spatial obj
ects.
The form
ula
of entropy an
d informatio
n gain in t
he ID3 algorith
m
were mo
dified
by involving two
types of spati
a
l relation
shi
p
s nam
ely metric an
d t
opo
logical to rela
te two spatial
object
s
[8].
The
spatial
data
s
et used i
n
thi
s
work contai
ns fo
re
s
t
an
d la
n
d
fir
e
s
data
fo
r
the
s
t
ud
y a
r
ea
R
o
ka
n
Hilir
distri
ct i
n
Ria
u
Province
Indon
esi
a
. In
ad
dition to p
h
ysi
c
al, so
cio
-
e
c
o
nomic,
we
athe
r
cha
r
a
c
teri
stics of the stud
y area [8], th
is wo
rk
in
clu
des p
eatland
types and p
eatland de
pth to
predi
ct fire
s o
c
curren
ce in
peatlan
d
s.
A peatland fire is cla
ssifie
d
as a gro
u
n
d
fire because the fire burn peat soil in
side the
peatlan
d
s an
d we
can
onl
y see
smo
k
e
visible
on th
e
su
rfa
c
e.
Th
erefo
r
e, p
eatl
and fires a
r
e
not
easy to handl
e comp
ared to the fires in non-peatla
nd
s
[10,11]. A
study in [12] report
s
that Ri
au
is on
e of province
s in S
u
matra th
at has
high
d
e
f
o
re
station b
e
cau
s
e
of forest fire
s. Ri
au
provin
ce had
about 4.044
million
h
e
ct
are
s
(56.
19
%) of pe
atla
nd in
200
2
and it m
ade
the
provin
ce a
s
the larg
est area of peatlan
d
in Su
matera Island an
d Kalimantan Islan
d
. For that,
influen
cing fa
ctors for fire e
v
ents in peatl
and
s are
con
s
ide
r
ed in thi
s
study.
2. Rese
arch
Metho
d
2.1. Stud
y
Area and For
e
s
t Fires
Data
This
work d
e
v
eloped the
predi
ction m
o
del for
hot
sp
ots o
c
curren
ce based o
n
the fore
st
fires dataset for the
study
area
Roka
n
Hilir di
stri
ct in Riau P
r
ovince
in Indonesia.
Rokan Hilir
is
locate
d in the
area
betwee
n
100°
16'
- 1
01°2
1' Ea
st Longitud
e
and
1°14'
- 2°
30'
North
Latitud
e.
It covers a
n
a
r
ea of 8,881.
59 km
2
or a
b
out 10 perce
n
t
of Riau’s total land area [13].
The sp
atial fore
st fires d
a
ta includ
e p
h
ys
ical, soci
o
-
econo
mic, weather a
nd p
eatland
cha
r
a
c
teri
stics of th
e stu
d
y are
a
that m
a
y influenc
e fo
rest
and
lan
d
fire eve
n
ts.
The d
a
ta a
n
d
its
sou
r
ce are provided in Tab
l
e 1.
Table 1. Data
and its so
urce
Data
Source
Spread and coo
r
dinates of hotspo
t
s 2008 (for cr
eat
ing models
for hotspots occurrence p
r
ediction)
FIRMS M
O
DIS F
i
re/Hotspot, NAS
A
/University
of Mar
y
lan
d
Spread and coo
r
dinates of hotspo
t
s 2010 (for m
o
d
e
l
evaluation)
FIRMS M
O
DIS F
i
re/Hotspot, NAS
A
/University
of Mar
y
lan
d
Weather data
20
08 (in the Net
C
D
F
format
): ma
ximum daily
temperatu
r
e, daily rainfall, and sp
eed of
w
i
nd
Meteorological Climatological and
Geoph
ysical Agency (BMK
G)
Digital maps for r
oad, rivers,
cit
y
centers, land cover, and
administrative border
National Coordin
a
ting Agenc
y
fo
r
Surve
y
and
Mapping (BAKOSURTANAL
)
Digital maps for peatland depth a
nd peat
land t
y
pe
s
Wetland Internati
onal
Inhabitant’s income source
BPS-Statistics In
donesia
2.2. Spatial Relationship, Spatial En
trop
y
and Spatial Informati
on Gain
Spatial data
s
ets fo
r
cla
ssif
i
cation
tasks
ar
e
co
mpo
s
e
d
by
some
e
x
planatory l
a
yers and
one target la
yer. Each l
a
yer re
pre
s
e
n
ts a set
of
spatial o
b
je
cts whi
c
h i
s
chara
c
te
rized
by
several spati
a
l and
non
-spatial attribut
es. On
e of
n
on-spatial
attributes i
n
an
explanato
r
y layer
is the
explan
atory attrib
ute that ide
n
tifies
obje
c
ts
i
n
the laye
r.
The ta
rget la
yer ha
s
a target
attribute that store
s
class l
abel
s of the target obj
ect
s
.
All obje
c
ts i
n
a laye
r
have
a p
a
rti
c
ula
r
geomet
ry type that m
a
y b
e
eithe
r
p
o
int
,
line o
r
polygon. The geometry type of object
s
is pre
s
e
n
te
d in a spatia
l attribute of
the layer. For
instan
ce, in t
h
is
study the
road l
a
yer re
pre
s
ent
s a ro
ad network in
which ea
ch
road segm
ent
has
the geometry
type of line.
Other layers in t
he dataset are the land cove
r layer and the ta
rget
layer. Spatial
obje
c
ts i
n
th
e land
cover l
a
yer
a
r
e
poly
gon featu
r
e
s
,
whe
r
e
a
s
obj
ects in the ta
rget
layer are poin
t
features indi
cating h
o
tsp
o
t
s and no
n-h
o
t
spots.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
A Deci
sion T
r
ee Based on
Spatial Relati
onships
for Predi
cting .... (Im
a
s Sukaesi
h
Sitanggang)
513
Relatio
n
bet
wee
n
spatial
obje
c
ts of t
w
o diffe
rent
layers i
s
essential in
spat
ial data
mining sy
ste
m
s. In our
study, for example, hot
sp
ots occu
rren
ce in the target layer ma
y be
influen
ced by
the existence of road
s b
e
ca
us
e roa
d
s open a
c
cess for human t
o
enter a forest
and th
eir
acti
vities may tri
gger forest fi
re ev
e
n
ts.
More
over, dif
f
erent la
nd
cover type
s m
a
y
provide
different risk level
s
of fire
s occurren
ce.
For
inst
an
ce,
f
i
re
s ar
e mo
re li
kely
t
a
ke
pla
c
e in
plantation a
r
e
a
s than tho
s
e
are in settlement are
a
s b
e
ca
use farme
r
s may u
s
e fires to ope
n n
e
w
plantation
s
.
Spatial relati
onship
s
allo
w us to in
cl
ude relation
s betwee
n
two spatial o
b
j
e
cts in a
dataset for a
cla
ssifi
catio
n
task. The
s
e
relation
shi
p
s
can
be to
pologi
cal
su
ch as
meet a
n
d
overlap, a
s
well as metri
c
, for example d
i
stan
ce.
In spatial databa
se
s, a layer is rep
r
e
s
ente
d
as
a rel
a
tion a
n
d
applyin
g
a
spatial
rel
a
tion bet
wee
n
two laye
rs
results
a ne
w relation. T
he
st
ru
ct
ur
e S
p
a
t
ial Join I
n
de
x
(S
JI
)
wa
s in
t
r
odu
ced
in
[1
4] to impleme
n
t spatial
rel
a
tionshi
ps i
n
the
relation
al d
a
taba
se frame
w
ork.
The
SJI is a
ne
w
rel
a
tion a
s
th
e
result
of join
in
dex bet
wee
n
two
relation
s that
consi
s
ts of i
ndices pai
rs each refe
ren
c
ing a tuple
of each rel
a
tion. The pairs of
indices refer t
o
obje
c
ts that
meet the join criteri
on.
The co
ncept of SJI was a
dopted in ou
r previou
s
wo
rk [8]. The work in [8] co
mputed
quantitative v
a
lue
s
resulte
d
from
topol
ogical a
nd m
e
tric rel
a
tion
ships.
A top
o
logical relatio
n
betwe
en two
spatial o
b
je
cts is
cal
c
ulat
ed by pe
rf
orming the ove
r
lap o
peratio
n. In additio
n
to
topologi
cal
re
lationship
s
, the al
gorith
m
involves
a
m
e
tric rel
a
tion
ship n
a
mely d
i
stan
ce f
r
om
a
spatial
obje
c
t
to an
other spatial o
b
ject.
Fo
r exampl
e
,
applying
th
e spatial
rela
tionshi
p
ov
er
la
p
on two p
o
lyg
ons result
s
a
n
ove
r
lappi
n
g
a
r
ea
with
a certai
n ext
ent. Mo
reov
er,
we m
a
y
also
cou
n
t how m
any hotsp
ot points in
a certain p
o
lygo
n or calculat
e distan
ce
b
e
twee
n hotspot
points to
a n
eare
s
t
river segment.
We
den
ote th
ese qu
antitative value
s
, i.e.
area,
count
a
n
d
distan
ce, a
s
spatial
mea
s
ure
s
of
spatia
l relatio
n
shi
p
s betwe
en t
w
o
obje
c
ts. In
stead of
usin
g t
h
e
SJI, our
work propo
se
s
what we
ca
lled
Spatial Joi
n
Relation
(S
JR),
as the
re
sult of
a
sp
atial
relation
b
e
tween t
w
o
laye
rs [8]. T
he
SJR contain
s
sp
atial
obje
c
ts fro
m
the
t
w
o l
a
yers
an
d its
asso
ciated
sp
atial measure
s
. The SJR o
f
a new layer
R is defin
ed a
s
follows [8]:
SJR
= {(p, Sp
atMes(r), q |
p in layer L
i
,
q in layer L
j
, and
r is a featu
r
e in R a
s
soci
ated to
p and q}.
(1)
The spatial measure
of a
layer R,
S
patMes(r)
, is
use
d
in the
spatial entrop
y
formula
whi
c
h repla
c
es the n
u
mb
er of tuple
s
i
n
a par
tition
in the non
-spatial entrop
y
formula. The
spatial
entro
py is d
e
fined
as follows [
8
]. Let t
he ta
rget attri
bute
C in
the ta
rget layer S h
a
s
l
dist
in
ct
cla
s
s
e
s (i.
e
.
c
1
, c
2
, …, c
l
), entro
py for S represe
n
ts the ex
pecte
d inform
ation nee
ded
to
determi
ne the
class of tuple
s
in the data
s
et and define
d
as
SpatMes(S)
)
SpatMes(S
log
SpatMes(S)
)
SpatMes(S
H(S)
i
i
c
l
1
i
2
c
(2)
SpatMes(S)
rep
r
e
s
ent
s the spatial m
e
asu
r
e of
lay
e
r S that may be either area, count
or
distan
ce.
The spatial d
e
ci
sion tre
e
a
l
gorithm p
a
rtitions
o
b
je
cts i
n
the target l
a
yer S based
on the
explanato
r
y (non-ta
rg
et) la
yer L. This
step re
sult
s a n
e
w layer
L(v
j
, S) for each possibl
e valu
e v
j
in L. Each new layer is a
s
sociate
d
to a new pa
rt
ition.
The expe
cted
entropy value for splitting is
defined a
s
fol
l
ows:
S))
,
L(v
H(
SpatMes(S)
S))
,
v
SpatMes(L(
L)
|
H(S
j
q
1
j
j
, (3)
Spatial inform
ation gain for
the layer L
is
given by the followin
g
form
ula.
Gain
(
L
) =
H
(
S
)
H
(
S
|
L
) (
4
)
whe
r
e
H
(
S
) a
nd
H
(
S
|
L
) are given in Equation 2 an
d Equation
3 re
spe
c
tively. The layer L wi
th
the highe
st in
formation g
a
i
n
,
Gain
(
L
), is
sele
cted a
s
the splitting la
yer to partitio
n
the dataset.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 12, No. 2, June 20
14: 511 – 51
8
514
2.3. Spatial ID3 Algo
rith
m
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 an
d ea
rly 1980
s. This algo
rithm h
a
s the pri
n
ci
p
l
e that it builds the tree in
gree
dy mann
er
starting
fro
m
the root, an
d
sele
cting
mo
st info
rmat
ive
features at e
a
ch
st
e
p
[15].
Th
e al
gorith
m
use
s
inform
ation gain t
o
sele
ct the
best f
eature at each
step for
spli
tting a dataset.
Furthe
rmo
r
e,
the ID3 alg
o
r
ithm is
desi
gned fo
r n
o
n
-
sp
atial data
s
ets in
wh
ich
the input of the
algorith
m
i
s
a rel
a
tion
co
ntaining
so
m
e
obj
ect
s
of intere
st.
All obje
c
ts are
chara
c
te
rized by
several featu
r
es.
One
of t
he featu
r
e
s
i
s
a ta
rget
fe
ature
that
con
s
i
s
ts
of
class la
bels of o
b
je
cts,
whereas other features are expl
anatory features that will be us
ed to cl
assif
y
an object to a
certai
n cla
s
s label.
The ID3
algo
rithm has
bee
n improve
d
in
[4] such th
at the algo
rithm
can b
e
ap
plie
d on a
spatial d
a
taset containi
ng
polygon feat
ure
s
. On
th
e other h
and
, spatial data
s
ets m
a
y involve
not only poly
gon featu
r
e
s
but also poi
nt and lin
e feat
ure
s
. The
r
ef
ore in
ou
r previous
wo
rk [
8
],
we exten
ded
the ID3 alg
o
r
ithm ba
se
d on several
a
ppro
a
che
s
in
[4] so that the ne
w alg
o
rithm
can
wo
rk on
point, line
and p
o
lygon
feature
s
.
Our
pro
p
o
s
e
d
algo
rithm
use
s
the
sp
atial
informatio
n gain provide
d
in Equation 4 to se
le
ct the best splitting laye
r from a set
of
explanato
r
y layers.
Cre
a
ting
a sp
atial de
cisi
on
tree u
s
in
g th
e sp
atial de
ci
sion t
r
ee
alg
o
rithm [8] foll
ows the
basi
c
lea
r
ni
n
g
pro
c
e
s
s in t
he algo
rithm I
D
3 [9].
The
algorith
m
wo
rks
on spatial
data sto
r
ed i
n
a
spatial d
a
tab
a
se. Befo
re
the algo
rithm
is ex
ecuted
, the databa
se
contain
s
only a set o
f
explanato
r
y layers
and o
n
e
target laye
r. When t
he al
gorithm
wo
rks on the d
a
ta
base, som
e
n
e
w
layers
are
produ
ced a
s
th
e re
sult of sp
atial relatio
n
s betwe
en two
distin
ct layers. The
s
e n
e
w
layers
are
cre
a
ted from exi
s
ting expla
n
a
t
ory layers, a
nd the value
v
j
of predictiv
e attribute in
the
best splitting
layer.
The
value
v
j
i
s
a
selectio
n
crite
r
ion in
the
qu
ery to
rel
a
te
an expl
anato
r
y
layer a
nd the
best l
a
yer.
Each
ne
w lay
e
r i
s
a
s
soci
ated with
a
set
of tuples that
relate
obje
c
ts in
a layer to o
b
j
e
cts i
n
an
oth
e
r laye
r. Thi
s
work
con
s
i
ders
this set of
tuples as a
small
e
r sp
atial
dataset if on
e of two
relat
ed laye
rs is t
he ta
rget
layer. Eac
h
tuple in the datas
e
t has a
s
p
atial
measure whi
c
h is st
ore
d
in the Spatial Join
Rela
tion
(SJR). Inputs of t
he spati
a
l ID3 algo
rithm
are a spatial
dataset, a set of explanatory layers
, a
target layer and a SJR. Output of the
algorith
m
is a
spatial d
e
ci
si
on tree. Th
e tree ha
s
the
same st
ru
cture
as that of th
e cla
ssi
cal
on
e
in whi
c
h
the t
r
ee
co
nsi
s
ts
of a ro
ot no
d
e
, intern
al no
des
and
leav
e nod
es.
Th
e ro
ot nod
e a
nd
internal n
ode
s have the b
e
st splitting l
a
yers a
s
its l
abel
s. Mean
while, the lab
e
ls of leave n
ode
s
are ta
rget
cla
s
ses
of the target laye
r.
Ther
e a
r
e
so
me edg
es ou
tgoing fro
m
the ro
ot nod
e
and
internal n
ode
s. The label
of each e
dge
is one of
po
ssible valu
es i
n
the best spl
i
tting layer.
2.4. Tree Pru
n
ing
Overfitting i
s
one
of issu
es that may
be
encount
e
r
ed
whe
n
a
de
ci
sion tre
e
al
go
rithm is
applie
d on
real data
s
et
s. In this situ
ation,
as th
e de
cisio
n
tree b
e
come
too larg
e, the
gene
rali
zatio
n
error of de
cisi
on tree
starts to in
cre
a
s
e while its resu
bstitution
error continu
e
s to
d
e
c
r
e
as
e [1
6]. R
e
s
u
bs
titutio
n
er
r
o
r
s
are
mis
c
l
as
sification e
r
rors
on the
trai
nin
g
set, wh
ere
a
s
gene
rali
zatio
n
erro
rs are miscl
assification
e
rro
rs
on
the testing
se
t. Leaves i
n
large
tree
s m
a
y
reflect n
o
ises or outlie
rs th
at can in
crea
se ge
ne
raliza
t
ion errors when the tre
e
is appli
ed on t
h
e
testing
set. One of meth
ods to ove
r
come overfi
ttin
g
is po
st-p
ru
ning in
whi
c
h
the tree is f
u
lly
gro
w
n
at first, and th
en
al
l su
btree
s
of the tr
ee at
given n
ode
s
are
prune
d b
y
removin
g
it
s
bran
ch
es a
n
d
repla
c
ing it with a leaf [17]. The new
leaf is labele
d
with the ma
jority class in
the
subtree.
3. Results a
nd Discu
ssi
on
3.1. Spatial Decision Tr
e
e
for Ho
tsp
o
t
s Predic
tion
Applying the
spatial ID3 al
gorithm o
n
th
e fore
st fires
dataset re
su
lt
s a spatial d
e
ci
sion
tree
whi
c
h
ha
s 2
10 l
eave
s
. Accu
ra
cy of
the tre
e
o
n
th
e trai
ning
set
is 7
6
.51% m
e
aning
that 2
3
8
of 1013 targe
t
objects
are i
n
co
rrectly cla
ssifie
d
by the tree. Target o
b
ject
s are hot
spot
s and
no
n-
hotsp
ot poi
nts in
the
study
are
a
. Non
-
h
o
tspot
point
s
were g
ene
rat
ed o
u
tsid
e bu
ffers
of hot
sp
ots.
The ra
dius of
a buffer for a hotsp
ot is 0.9073
74
km
. It was defined by pro
c
e
ssi
ng bu
rn area
s
extracted
fro
m
the
Lan
dsat TM im
age.
Th
e first te
st laye
r of
th
e tre
e
i
s
in
co
me
sou
r
ce.
Thi
s
work
prepa
re
d a te
sting
set from th
e
spatial d
a
t
aba
se
by ap
plyin
g
several
sp
atial op
eratio
ns.
The te
sting
set co
nsi
s
ts of
561
obj
ect
s
(235
po
sitive
example
s
a
n
d
32
6 n
egativ
e exam
ples).
A
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
A Deci
sion T
r
ee Based on
Spatial Relati
onships
for Predi
cting .... (Im
a
s Sukaesi
h
Sitanggang)
515
positive
exa
m
ple i
s
a
n
o
b
j
ect
with the
true
cla
s
s,
wh
erea
s
a n
ega
tive example
is a
n
o
b
je
ct
with
the false
cla
s
s. Accuracy o
f
the tree on t
he test
in
g set
is 71.12% m
eanin
g
that 3
99 of 561 ta
rget
obje
c
ts are correctly cla
s
si
fied by the tree.
The
spatial
d
e
ci
sion
tree
a
s
a
p
r
edi
ction
model
for ho
tspots o
c
currence h
a
s the
si
ze
of
613. Size of
a tree i
s
nu
mber
of nod
e
s
in
cludi
ng a
root n
ode, int
e
rnal
and
lea
v
es no
de
s. T
he
numbe
r of cl
assificatio
n
rules g
ene
rate
d from the tre
e
is 134. A rule is obtai
ne
d from a tree
by
cre
a
ting a
pat
h from the
ro
ot to a leaf. In ord
e
r to
obt
ain a
simple
r
tree
with the
highe
r a
c
cura
cy,
the post
-
prun
ing method
was ap
plied to
the tree. In
this metho
d
, the tree i
s
fully grown
at first,
and then all subtree
s
at given node
s are prun
ed by
re
moving its branche
s and repla
c
ing it with a
leaf [17]. Thi
s
work im
ple
m
ented th
e p
o
st-p
ru
ning
method
up to
16 ite
r
ation
s
. The la
st p
r
u
ned
tree h
a
s th
e
accuracy
of 71.66% an
d
its si
ze i
s
4
8
5
. Starting from the
se
co
nd iteration, the
highe
st accu
racy of pru
n
e
d
tree
s for all
iterati
ons a
r
e the sam
e
i.e. 71.66%. Howeve
r, the size
of tree
de
cre
a
se
s f
r
om
59
9 in th
e
se
co
nd iteration t
o
48
5 in
the
16
th
iteratio
n. The
r
efore, t
he
numbe
r of
rul
e
s
gene
rate
d
from th
e tre
e
also de
clin
es. The
r
e a
r
e
1
08 rule
s g
ene
rated f
r
om th
e
s
i
mp
le
pr
un
ed
tr
e
e
.
Se
ve
ral rule
s are the followin
g
:
1.
IF income
_so
u
rce = Planta
t
ion
AND di
stance to the neare
s
t ro
ad (m)
≤
250
0 AND 1
500
<
distan
ce to th
e nearest rive
r (m)
≤
3000
THEN
Hotsp
o
t Occurren
ce = Tru
e
2.
IF income
_so
u
rce = Fo
re
st
ry
AND lan
d
_
c
over
= Bare
_land AND 1
≤
wind_
sp
ee
d (m/s) < 2
THEN Hotspot
Oc
cu
rr
ence =
Tr
ue
3.
IF income
_so
u
rce = Fo
re
st
ry AND lan
d
_
c
over
= Swa
m
p THEN
Hotspot O
c
currence =
TRUE
4.
IF income
_so
u
rce = Fo
re
st
ry
AND lan
d
_
c
over
= Bare
_land AND 0
≤
wind_
sp
ee
d (m/s) < 1
AND 29
7
≤
scre
en temp
erature (K
) < 2
98 AND p
eat
l
and_
depth
= D4 (Ve
r
y dee
p/Very thick
> 400
cm) T
H
EN Hot
s
pot
Occu
rre
nce = False
5.
IF income
_so
u
rce = Fo
re
st
ry A
ND lan
d
_
c
over
= Padd
y_field AND 0
≤
wind_
sp
ee
d (m/s) < 1
THEN
Hotspot Oc
currence =
Fals
e
6.
IF income
_so
u
rce = T
r
adin
g_re
s
tau
r
a
n
t THEN
Hotsp
o
t Occurren
ce = Fal
s
e
7.
IF income
_so
u
rce = Fo
re
st
ry A
ND lan
d
_
c
over
= Mix_
gard
en AND 0
≤
wind
_spe
ed (m/s)
≤
1 THEN
Hotspot Occu
rre
n
c
e = FALSE
8.
IF income
_so
u
rce = Fo
re
st
ry A
ND lan
d
_
c
over
= Plant
ation AND 0
≤
wind
_spee
d (m/s)
≤
1
AND pe
atlan
d_de
pth = Sh
allow/Thi
n
(5
0-
10
0 cm
)T
HEN Hot
s
pot
Occu
rre
nce = FALSE
9.
IF income
_so
u
rce = Fo
re
st
ry AND la
n
d
_
c
over
= Uni
rri
gated_
agri
_
field AND 2
≤
pre
c
ipitation
(mm/day)
≤
3 THEN
Hotspot Occu
rre
n
c
e = FALSE
10.
IF income
_so
u
rce = Fo
re
st
ry A
ND lan
d
_
c
over
= Padd
y_field AND 0
≤
wind_
sp
ee
d (m/s)
≤
1 THEN Hotspot O
c
curre
n
ce
= FALSE
3.2. Compari
s
on bet
w
e
e
n
Spatial and Non
-
Spatial
Classifie
r
s
For compa
r
ison, the non-spatial de
ci
si
on
tree algo
rithms namely
C4.5 and ID3 have
been
applie
d
on the forest
fires d
a
taset [18]. Thes
e al
gorithm
s a
r
e
available in t
he data mi
nin
g
toolkit
We
ka
3.6.6. J48 i
s
a mod
u
le i
n
We
ka
as
Java imple
m
enta
t
ion of the
C4
.5 algo
rithm.
The
accuraci
es of
cla
s
sifiers
g
enerat
ed
by t
hese two al
g
o
rithm
s
we
re
determi
ned
u
s
ing
the
10
-folds
cro
s
s validati
on method. In addition to
non-sp
atial
d
e
ci
sion tre
e
a
l
gorithm
s, a logisti
c
reg
r
e
s
sion
model
wa
s
calcul
ated to
predi
ct h
o
tsp
o
ts o
c
curren
ce [18]. Hot
s
pots o
c
cu
rre
nce i
s
co
nsi
d
ered
as th
e de
pe
ndent va
riab
le and
dete
r
mina
nt
fact
ors (enviro
n
m
ental a
nd
human
facto
r
s)
influen
cing fire events are
the indepe
nd
ent variabl
e
s
.
Table 2 summari
ze
s the
accu
ra
cy of the
spatial a
nd n
on-spatial
cla
ssifie
r
s a
s
we
ll as
the num
ber of rul
e
s g
enerated fro
m
the trees.
Table 2. Accu
racy of the cl
assi
fiers and
numbe
r of ge
nerate
d
rul
e
s
Classifier
Accur
a
cy
Number of
generated
rules
Spatial decision t
r
ee
The Ext
ended S
patial ID3 Decision Tree
w
i
thout
pruning
71.12%
134
The Ext
ended S
patial ID3 Decision Tree
w
i
th pru
n
ing
71.66%
108
Non-spatial classifier
ID3 Decision Tre
e
49.02%
270
C4.5 Decision Tr
ee
65.24%
35
Logistic regression
68.63%
-
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 12, No. 2, June 20
14: 511 – 51
8
516
Table 2
sh
ows that the p
r
o
posed alg
o
rit
h
m nam
ely the sp
atial de
ci
sion tree al
go
rithm is
sup
e
rio
r
amo
ng other met
hod
s i.e. non-sp
atial de
cisi
on tree algo
ri
thms and lo
gistic re
gressio
n
.
The spatial
I
D
3 without
p
r
uning perfo
rms well
on
the testin
g se
t with the a
c
curacy
of 71.
12%
comp
ared to
the
cla
ssi
cal
ID3
(n
on-sp
atial ID3
)
wit
h
the
accu
ra
cy of
49.02
%. Furth
e
rmore
,
Table 2
sh
ows that the
spa
t
ial ID3 de
ci
sion tree
with p
r
unin
g
outp
e
rforms th
e C4.5 de
cisi
on tre
e
with 6.4
2
% o
f
accu
ra
cy hi
gher tha
n
the
C4.5
d
e
ci
sio
n
tre
e
.
Mo
reover, l
ogi
stic
reg
r
e
ssi
on
ha
s
been u
s
e
d
in seve
ral st
udie
s
to det
ermin
e
the relation bet
ween hot
spot
s occurren
ce
and
influen
cing fa
ctors of fi
re e
v
ents. Applyi
ng thi
s
me
thod to the forest fire datas
et
results
the bes
t
reg
r
e
ssi
on m
odel
with th
e
accuracy
of 6
8
.63%
wh
ich
is n
o
t bette
r t
han th
e
spati
a
l de
ci
sion
tree
algorith
m
tha
t
has the a
c
curacy g
r
eat
er than 7
1
%. According
to these re
sults, this
work
concludes that involving spatial relations in
the de
cisi
on tree al
gorithm p
r
od
uce
s
the bet
ter
cla
ssif
i
e
r
s f
o
r hot
sp
ot
s o
c
c
u
rr
en
ce.
The spatial I
D
3 al
gorithm
prod
uces m
o
re
simple t
r
ees
com
p
a
r
e
d
to the ID3
algorith
m
.
It can
be i
n
fe
rre
d fro
m
the
numb
e
r of rules ge
nerate
d
from
the tree a
s
sh
own
in Tabl
e 2.
T
he
spatial ID3 algorithm
witho
u
t prunin
g
gi
ves 134 ru
le
s whi
c
h is al
most a half o
f
the number of
rule
s g
ene
rat
ed by the
con
v
entional ID3
de
cisio
n
tr
ee
i.e. 270.
Ho
wever, i
n
term of the
num
ber
of rule
s gen
e
r
ated fro
m
the tree
s, the C4.5 algo
rithm
outperfo
rm
s the spatial ID3 algo
rithm with
pruni
ng
wh
ere the
C4.5
al
gorithm
re
sul
t
s only
35
rul
e
s
and
the
p
r
opo
se
d al
go
rithm p
r
od
uces
108 rul
e
s (Ta
b
le 2). The furthe
r study is req
u
ir
e
d
especi
a
lly in the tree pruni
ng
method in order
to obtain
mo
re si
mple
spat
ial de
ci
sion
trees.
O
n
the
other ha
nd, the C4.5 de
ci
sion tre
e
has
the
accuracy
of
65.24% that
i
s
slightly lo
wer th
an
th
e
spatial ID3 d
e
cisi
on tree
wi
th pruning
which
achi
eves th
e
accu
ra
cy of
71.66%
. Th
erefo
r
e, rega
rdle
ss the
si
ze
of tree
s,
the spatial I
D
3
algorith
m
wit
h
pruni
ng ha
s better perfo
rmance than the C4.5 al
gorithm.
3.3. Tree Ev
a
l
uation
The unp
run
e
d
and pru
n
e
d
trees were
applied to a new sp
atial dataset. The dataset
contai
ns the
same
expla
n
a
tory laye
rs
as th
ose for
cre
a
ting th
e t
r
ee
and
the
FIRMS MO
DIS
Fire/Hotsp
ots in 201
0. Th
e num
b
e
r
of hotsp
ots in
2
010 for Ro
ka
n Hilir
area i
s
77
4. As m
any
726 p
o
ints were
ran
domly
gene
rate
d n
ear
any hot
sp
ot
in 20
10. T
o
acco
mpli
sh
this ta
sk, b
u
ffers
with
the
radi
us of
0.90
73
74 km we
re cre
a
ted
for e
a
ch
hot
spot and
th
en ran
dom
p
o
ints were
gene
rated o
u
t
side the buff
e
rs. Th
ese random p
o
ints
are den
oted
as false ala
r
m data. Along
with hot
spot
s in 201
0 a
s
t
r
ue
alarm
dat
a, false
al
arm data
comp
ose ta
rg
et ob
jects i
n
the
n
e
w
target layer.
A new data
s
et co
ntain
s
707 obje
c
ts (2
7
7
po
sitive example
s
and 4
30 negative
example
s
).
Applying the
spatial de
ci
sion tre
e
s
algorithm o
n
the ne
w dat
aset results
the
accuracy
of
60.06% for the tre
e
with
out pruni
n
g
and 6
1
.89%
for the
tre
e
with
pruni
ng.
More
over, th
e tree is u
n
a
b
le to cla
ssif
y
some obj
ects in the ne
w dataset. The
r
e are
51 of
707
(7.21%)
obje
c
ts that
can
not be cl
assified by
the tree with
ou
t prunin
g
. The nu
mbe
r
of
uncl
a
ssified
obje
c
ts de
cre
a
se
s to 30
of 707 (4.24%) when th
e tre
e
with p
r
unin
g
wa
s exe
c
ut
ed
on the
ne
w
d
a
taset.
Tabl
e 3
gives cha
r
acte
ri
stics of
un
cla
ssifie
d
objetcs
ba
se
d on
lan
d
co
ver,
peatlan
d
type, peatland de
pth and inco
me sou
r
ce. Mo
st of uncl
a
ssified obje
c
ts are locat
ed in
non-peatla
nd
s in which in
come
so
urce
s of peo
ple li
ving in these area
s a
r
e mo
stly forest
ry and
agri
c
ultu
re.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
A Deci
sion T
r
ee Based on
Spatial Relati
onships
for Predi
cting .... (Im
a
s Sukaesi
h
Sitanggang)
517
Table 3. Ch
aracteri
stics of
uncl
a
ssified o
b
ject
s
Explanator
y att
r
ibute
True class
False class
Total
Land cover
Plantation 2
6
8
Dr
y
l
and
_forest
2
7
9
Bare_land
0
1
1
Shrubs 2
2
4
Padd
y
_
field
0
1
1
Sw
am
p
3
0
3
Mix_gard
e
n
1
3
4
Peatland_t
y
p
e
Hemists/Saprists(60/40),
M
oder
ate
2
4
6
Hemists/Saprists(60/40),V
er
y_de
ep 2
2
4
Non_peatland
6
7
13
Saprists/min(90/10),Mode
rate
0
1
1
Saprists/min(50/50),Shallo
w 0
6
6
Peatland depth
D1 (Shallo
w
/
Thin
50-100 cm)
0
9
9
D2 (Mode
rate 10
0-200 cm)
2
2
4
D3 (Dee
p/Thick 200-400 cm)
2
2
4
Non_peatland
6
7
13
income_source
Other
_agricultur
e
2
1
3
Forestr
y
0
10
10
Agriculture 8
9
17
4. Conclusio
n
This
wo
rk
a
pplied th
e spatial ID3
al
gorithm
on t
he spatial fo
rest fires
dat
aset. Th
e
dataset con
s
i
s
ts of phy
sical,
weathe
r, so
cio-econo
mic an
d peat
land cha
r
act
e
risti
cs th
at may
influen
ce fire
s occu
rren
ce
in the
study a
r
ea
Ro
kan
Hil
i
r Di
strict, Ind
one
sia. The
result is
a sp
atial
deci
s
io
n
tree for
p
r
edi
cting
hotsp
ots occurren
ce with
the a
c
cu
ra
cy
of 76.51%
on
the trainin
g
set
and 71.1
2
% on the testing
set. Size of the tree
is 6
1
3
and the nu
mber of rule
s generated from
the tree is
13
4. To simplif
y the tree, the pos
t
-
prunin
g
method h
a
s been impl
em
ented. Applying
this m
e
thod
on the
spati
a
l de
ci
sion
tree p
r
od
uc
es a p
r
u
ned
tree
whi
c
h i
s
simple
r th
an
the
unprune
d tre
e
. Th
e p
r
un
e
d
tre
e
h
a
s the
accu
ra
cy of
71.66%
with i
n
com
e
sou
r
ce a
s
the
first test
layer. The si
ze of the tree
decrea
s
e
s
to 485 an
d
the n
u
mbe
r
of gen
erated
rule
s d
e
clin
es to 10
8.
In compa
r
i
s
o
n
with the spatial ID3 al
gorithm
, this work also a
pplied the n
on-spatial
deci
s
io
n tre
e
algorith
m
s i.e
.
ID3 an
d C4.5 on th
e fore
st fire
s data
s
et. The exp
e
r
imental
re
sul
t
s
sho
w
that th
e
pro
p
o
s
ed
alg
o
rithm h
a
s be
tter perfo
rm
a
n
ce i
n
term of
accu
ra
cy tha
n
the two no
n-
spatial al
gorit
hms. The a
c
cura
cy of ID3 de
ci
sion
tree is 49.
02
% and the accura
cy of C4.5
deci
s
io
n tree is 65.24%. Moreove
r
, the spatial ID
3 al
gorithm outp
e
r
form
s t
he lo
gistic regressi
on
model that h
a
s the a
c
curacy of 68.63
%. The s
patial ID3 algo
rithm ha
s been
tested to cla
ssif
y
obje
c
ts in th
e new fo
rest
fires data
s
et
. The result
s sh
ow that there a
r
e
3
0
of 707 or a
b
out
4.24% obje
c
t
s
which cann
ot be cl
as
sifie
d
by the prun
ed tree. T
h
e
s
e un
cla
ssified
obje
c
ts mo
st
ly
take
pla
c
e i
n
non-peatla
nd
s in
which in
come
sou
r
ce
s
of peo
ple livi
ng in
the
s
e
areas are fo
re
stry
and a
g
ri
cultu
r
e. Mo
reove
r
, most of
un
cla
ssifie
d
obj
ects
are lo
ca
ted in pla
n
tation an
d dryla
n
d
fores
t.
This work co
nclu
de
s that involving dista
n
ce a
nd topol
ogical relatio
n
s bet
wee
n
o
b
ject
s in
the spatial
cla
ssifi
cation ta
sk re
sult
s the spatial d
e
ci
si
on tree a
s
a
model for p
r
e
d
icting h
o
tsp
o
ts
occurre
n
ce with the high a
c
cura
cy.
Ackn
o
w
l
e
dg
ment
This
wo
rk was
sup
p
o
r
te
d by Indo
ne
sia
Di
recto
r
a
t
e Gene
ral
of High
er E
ducation
(IDG
HE), Mi
nistry of Na
tional Edu
c
a
t
i
on unde
r
Grant [nu
m
b
e
r 17
24.2/D4.4/2008]; a
nd
Southeast Asian Regio
nal
Center for Graduate Study and
Re
search in Agri
cult
ure
(SEARCA)
unde
r Grant [Ref. No. GCS
10-2
129].
Referen
ces
[1]
Ester M, Kriegel HP, San
der
J.
Spatial Dat
a
Mini
ng: A Databas
e Appr
o
a
ch
. Procee
di
n
g
s of the 5th
Internatio
na
l Symp
osi
u
m on L
a
rge Sp
at
ia
l Databas
es. Berli
n
. 1997: 4
7
-66.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 12, No. 2, June 20
14: 511 – 51
8
518
[2]
Kopersk
i K, Han J, Stefanov
ic N.
An Efficient T
w
o-step
Method for Cl
assificati
on of Spatia
l Data
.
Procee
din
g
s of
the Internatio
n
a
l S
y
m
pos
ium
on
Spati
a
l D
a
ta Han
d
li
ng. Va
ncouv
er. 199
8: 45-54.
[3]
Chel
gh
oum
N,
Z
e
itoun
i K, B
o
ulmak
oul
A.
A
Decisi
on
Tree
for Mu
lti-lay
e
re
d Sp
atia
l D
a
ta
.
Proce
edi
ngs
of the Joint International S
y
m
posium on
Geospatial T
heor
y
,
Pr
oc
essing and Applic
ations. Otta
w
a
.
200
2: 8-12.
[4]
Rinziv
ill
o S, Fr
anco
T
.
Classif
i
catio
n
i
n
Ge
og
raph
ical
Infor
m
ation
Syste
m
s
. In: Boulicaut JF, Esposit
o
F
,
Giannotti F
,
Pedresc
h
i D. Editors.
Artificial Intelli
genc
e
. Ne
w
Y
o
rk. Sprin
ger-Verl
ag. 20
04: 374
–3
85.
[5]
Li
X, Clar
a
m
unt C. A Sp
atial E
n
trop
y-
base
d
Dec
i
sio
n
T
r
ee for Classicati
on
of Geogra
phic
a
l
Information.
T
r
ansacti
ons i
n
GIS
. 2006; 10(
3): 451-4
67.
[6]
Z
hao M, Li
X.
An Applic
ati
on of Sp
atial
Decisi
on T
r
ee
for Classific
a
tion of Air Po
llutio
n
Ind
e
x
.
Procee
din
g
s of
the 19th Intern
ation
a
l Co
nfer
e
n
ce on Ge
oinfo
rmatics. Shang
hai. 20
11: 1-6.
[7]
Jian
g Z
,
Shekhar S, Mohan
P, Knight J, Corcoran J.
Lear
nin
g
Spatia
l Decis
i
on T
r
ee for Geogr
aph
ic
a
l
Classific
a
tio
n
: A Summary of
Results
. Proce
edi
ngs of the
2
0
th Inte
rnati
o
n
a
l Co
nfere
n
ce
on Adv
ance
s
in Geogr
ap
hic Informatio
n
S
y
s
t
ems. Californ
i
a. 2012: 3
90-3
93.
[8]
Sitang
ga
ng IS,
Yaak
ob
R, M
u
staph
a N,
Ain
udd
in A
N
. Cl
a
ssificatio
n
Mo
d
e
l for
Hotsp
o
t
Occurrences
usin
g Spati
a
l D
e
cisio
n
T
r
ee Algorithm.
Jour
n
a
l of Co
mp
uter
Science
. 20
13
; 9(2): 244-25
1
.
[9]
Quinl
an JR. Inductio
n
of Deci
sion T
r
ees.
Machin
e Le
arni
ng
. 1986; 1(1): 81
–10
6.
[10] S
y
a
u
fin
a
L,
Nu
rudd
in AA. F
o
r
e
st F
i
re i
n
P
e
a
t
F
o
rest: An O
v
ervie
w
.
M
a
n
a
j
e
men
Huta
n T
r
opik
a
. 20
00;
6(1): 75–
83.
[11]
Adin
ugro
ho W
C
, Sur
y
a
d
i
putr
a
INN, Saharj
o
BH, Siboro L.
Manu
al for the Contro
l of F
i
re in Peatl
a
n
d
s
and P
eatla
nd
Forest. Climate Ch
ang
e, Fore
sts and
Peatla
nds i
n
Indo
nesi
a
Proj
ect
. Wetlands
Internatio
na
l, Indo
nesi
a
Progr
amme
an
d W
i
l
d
life H
abitat C
ana
da. 20
05.
[12]
W
a
h
y
unto, S
u
r
y
a
d
i
putra IN
N.
Peatla
nd
Distri
butio
n i
n
Su
ma
tra an
d
Ka
li
ma
ntan-ex
pla
nati
on
of its D
a
ta
Sets Inclu
d
i
ng
Source
of Infor
m
ati
on,
Acc
u
ra
cy, Data
Co
nstraints
and
Gap
s
. W
e
tlands
Int
e
rnati
ona
l
,
Indon
esi
a
Prog
ramme. 200
8.
[13]
Rokan Hilir Dis
t
rict: Overview
of district [Inte
r
net]. 2009. Riau: Rokan Hilir
District; [cited 2012 May
30].
Avail
abl
e from: http://
w
w
w
.
r
o
h
ilkab.
go.id/
?
ta
mpil=
li
nka
ndac
t=
profilan
d
id=
4
.
[14]
Z
e
itoun
i K, Ye
h L, Aufa
ure
MA.
Join In
di
ces as a
T
ool
for Spatia
l D
a
ta Min
i
ng
. Pr
oc
eed
ings
of th
e
Internatio
na
l W
o
rksho
p
on T
e
mporal, Sp
atial
and S
pati
o
T
e
mporal D
a
ta Mi
nin
g
. L
y
on. 20
00: 102-
11
4.
[15] Marsland
S.
Machi
ne L
earn
i
n
g
: An Algorith
m
ic Pers
pectiv
e
. Boca Rato
n. CRC Press. 2009: 13
3-1
39.
[16]
T
an P, Steinba
ch M, Kumar
V.
Introductio
n
to Data
Min
i
n
g
. Boston. P
e
a
r
son A
ddis
on
W
e
sle
y
.
2
006
:
172-
176.
[17]
Han J, Kam
b
e
r
M.
Data Min
i
ng: Co
nce
p
ts and T
e
c
h
n
i
qu
e
s
. Second
Edit
ion. Sa
n F
r
anc
isco. Morg
a
n
Kaufman
n
. 200
6: 304-3
0
6
[18]
Sitang
ga
ng IS, Yaako
b
R,
Mustaph
a N, Ai
nud
din A
N
. Predictiv
e Mod
e
l
s
for Hotspots
Occurrenc
e
usin
g Dec
i
si
on
T
r
ee Algorith
m
s and
Lo
gisti
c
Regr
essio
n
.
Journ
a
l
of App
lied
Scie
nces
. 201
3;
13(
2):
252-
261.
Evaluation Warning : The document was created with Spire.PDF for Python.