TELKOM
NIKA
, Vol.14, No
.2, June 20
16
, pp. 684~6
9
1
ISSN: 1693-6
930,
accredited
A
by DIKTI, De
cree No: 58/DIK
T
I/Kep/2013
DOI
:
10.12928/TELKOMNIKA.v14i1.3540
684
Re
cei
v
ed
Jan
uary 6, 2015;
Re
vised Ap
ril
20, 2016; Accepte
d
May 8
,
2016
Comparative Analysis of Spatial Decision Tree
Algorithms for Burned Area of Peatland in Rokan Hilir
Riau
Putri Thariq
a
1
, Imas Sukaesih Sitang
gang*
2
, Lailan Sy
aufina
3
1,2
Department of Computer S
c
i
enc
e, F
a
cult
y of Natural Sci
ence a
nd Math
ematics,
Bogor Agr
i
cult
ural U
n
ivers
i
t
y
,
Indones
ia
3
Departme
n
t of Silvicult
u
re, F
a
cult
y of F
o
res
t
r
y
, Bogor Agr
i
cultura
l
Univ
er
sit
y
, Indo
nesi
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: putri.thariq
a
@
gmai
l.com
1
, Imas.sitang
ga
n
g
@gma
il.com
2
,
sy
au
fi
na
l
a
i
l
an@
g
m
a
i
l
.
co
m
3
A
b
st
r
a
ct
Over one-y
ear
peri
od (Marc
h
201
3
-
March
20
14), 58
perce
nt
of all
d
e
tecte
d
hotspots i
n
In
don
esia
are foun
d in R
i
au Provi
n
ce.
Accord
i
ng to the data, Rok
a
n Hilir sh
ared
the greatest n
u
mber
of hots
pots,
abo
ut 75% h
o
tspots alert occ
u
r in pe
atlan
d
areas. T
h
is study app
lie
d spa
t
ial decis
io
n tree al
gorith
m
s t
o
classify class
e
s before b
u
rne
d
, burne
d a
nd
after burn
ed
from re
mote se
n
s
ed dat
a of pe
atlan
d
are
a
in
Kubu
and Pas
i
r Li
mau Kap
a
s sub
d
istrict, Rokan
Hilir, Riau.
The dec
isio
n tree al
gorith
m
base
d
on spa
t
ial
autocorr
e
lati
on
is ap
pli
ed
by i
n
volvi
ng
Nei
g
b
o
rho
od
S
p
lit A
u
tocorre
latio
n
Ratio
(N
SAR) t
o
the i
n
for
m
ati
o
n
gai
n of CA
RT
alg
o
rith
m. T
h
is
spatia
l d
e
cisi
o
n
tree
cl
assific
a
tion
metho
d
i
s
compar
ed to
the co
nventi
o
na
l
decisi
on tre
e
a
l
gorit
hms, n
a
m
ely,
Class
ificati
on an
d Re
gres
sion T
r
ees (CA
R
T
)
,
C5.0, and
C4.5 al
gorith
m
.
T
he ex
peri
m
en
tal res
u
lts s
h
o
w
ed that th
e
C5.0
alg
o
rith
m ge
ner
ate th
e
most
accur
a
te
classifier
w
i
th t
h
e
accuracy
of 9
9
.79%. T
h
e i
m
p
l
e
m
e
n
tatio
n
of sp
atial
de
cision
tree
al
g
o
rith
m succ
es
sfully i
m
prov
e
s
the
accuracy of CA
RT
algorit
h
m
.
Ke
y
w
ords
: cla
ssificatio
n
, deci
s
ion tree, pe
atl
and, spati
a
l a
u
tocorrel
a
tio
n
Copy
right
©
2016 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
Fire
s in pe
atland/fore
st a
r
e very difficul
t
to
be handl
e
d
than the fires that o
c
curred in the
area of
non-p
eat. Peat fires (g
rou
nd fire
) difficult
to detect be
cau
s
e it can spre
ad to the dee
pe
r
or
spread
to
more
di
stan
t location
s
without bein
g
see
n
from
th
e su
rfa
c
e [1]
.
Proce
s
sing
of
satellite
im
ages produced from
re
mote
sensing is able to provi
de convenience for
stakeholders
in monito
ring
the fire th
at has
hap
pen
e
d
, is h
app
eni
ng, and
e
s
timates th
e in
ciden
ce of fire
s in
the future. A
dditionally it
can
estim
a
te
the
a
r
ea burned
a
nd pre
d
icted
enviro
n
mental ch
an
ges
cau
s
e
d
by the fire for a ce
rtain pe
riod [2
].
One use of satellite image is
to m
a
ke t
he process
of clas
sification. There are
several
cla
ssif
i
cat
i
on
algorit
h
m
s s
u
ch as d
e
ci
sio
n
t
r
ees,
B
a
y
e
sia
n
Net
w
or
ks,
N
a
iv
e B
a
y
e
s,
Max
i
mum
Likeli
hood
an
d Minimu
m Distan
ce. Som
e
re
se
arche
s
on satellite i
m
age
cla
s
sification
have b
e
e
n
carrie
d out
usin
g de
ci
sio
n
tree
algo
ri
thms. Sha
r
m
a
et al.
[3]
conducted a satellite i
m
age
cla
ssifi
cation
usin
g the de
cision tre
e
alg
o
rithm an
d compa
r
ed
with
the ISODATA algorithm a
nd
maximum likelihoo
d. The
result sh
ows
that a de
ci
sio
n
t
r
ee h
a
s t
h
e be
st
ac
cu
r
a
cy
c
o
mpa
r
e
d
t
o
other alg
o
rith
ms. The de
ci
sion tre
e
ha
s proven to be
an efficient algorithm for t
he cla
s
sificati
on
of large data
s
ets.
Li and Cl
ara
m
unt [4] had built a deci
s
ion tre
e
by con
s
id
erin
g the sp
atial effect an
d
spatial a
u
tocorrelation a
s
pect
s
by integratin
g spa
t
ial entropy
into the ID3
deci
s
ion tree.
Integration of
spatial entropy in
the cl
assificatio
n
p
r
ocess resu
lts a hierarchi
c
al structu
r
e
to
reflect the sp
atial distrib
u
tion of geog
ra
phic d
a
ta,
also produ
ce a
classification
that consi
d
e
r
s
spatial
aspe
ct. Deci
sion
tree i
s
a fam
o
u
s
meth
od
fo
r cla
ssifi
cation tasks and
it h
a
s been
ap
pli
ed
to a broa
d ra
nge of appli
c
ations. An ex
tensio
n
of the CART meth
o
d
, called the
SCART (S
pat
ial
Cla
ssifi
cation
and
Reg
r
e
s
sion T
r
e
e
s),
wa
s devel
op
ed in [5]. In the SCART, topologi
cal
and
distan
ce
rel
a
tionship
s
a
r
e
use
d
to te
st
whet
h
e
r
a p
r
edictive
attrib
ute bel
ong
s t
o
the
neig
h
b
o
r
table. The SCART
wa
s applied to ana
lyze traffic
risk usin
g accid
ent informati
on and them
ati
c
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Com
parative
Analysis of S
patial De
ci
sio
n
Tree
Algo
rit
h
m
s
for Burn
ed Area
… (P
utri Tha
r
iqa
)
685
informatio
n about road
netwo
rk
s,
popul
ation census, buil
d
i
ngs, an
d ot
her g
eog
rap
h
ic
neigh
borhoo
d
details [5].
The different
betwe
en
s
pat
ial and
non
-spatial de
ci
sio
n
data i
s
that
in
the spatial da
ta, an object may have a signifi
cant
influen
ce on nei
ghbo
ring o
b
j
e
cts. The
r
efo
r
e
,
improvem
ent
of the
non
-spatial
de
ci
si
on tree
al
go
rithm ha
s
be
en d
one
by
involving
spa
t
ial
relation
shi
p
s
betwe
en two
spatial o
b
je
cts [6].
The imp
o
rtan
ce of
spatial
and a
u
to
correlation
aspe
ct ma
ke
s Jia
ng
et al.
[7
] a
d
ded
autocorrelatio
n
a
s
pe
ct into
a de
ci
sion t
r
ee alg
o
rithm.
Implementati
on of
conve
n
tional de
ci
sio
n
tree algo
rith
ms
i
n
clu
d
ing
ID3, C4.5
and CART
i
n
the geo
graphi
cal cla
ssification
im
pli
c
itly
assume
s that
data item
s a
r
e ind
epe
nde
nt
and ig
nore
s
spatial a
u
to
correl
ation e
ff
ec
t. Thus
, the
cla
ssi
fi
cation
result co
ntain
s
salt
-n-pep
p
e
r noi
se. To
redu
ce th
e n
o
ise, auto
c
o
r
relation
aspe
cts
must b
e
con
s
ide
r
ed.
Jian
g
et al.
[7] condu
cted
a
classificatio
n
usin
g a
spati
a
l de
cisi
on tree
algorith
m
, wh
ich
combi
ned
spatial
auto
c
orrelation
as
a mea
s
u
r
e of
new fo
rmul
a
of informatio
n
gain. In the
new form
ula of informatio
n gain had the most important param
eter (
=
0
.26), this
para
m
ater i
s
also
appli
c
a
b
le fo
r the
ot
her area
s.
T
he study su
cce
ssfully re
d
u
ce
d
n
o
ise
a
nd
obtaine
d high
er accu
ra
cy than the C4.5
algor
ithm. B
u
t, Jiang research only compa
r
ed
spa
t
ial
deci
s
io
n tree
with C4.5
al
gorithm a
nd t
he auto
c
o
rrel
a
tion wa
s a
d
ded to info
rm
ation gain fro
m
C4.5 al
gorith
m
. Jiang
did
n
’t try to com
pare
with oth
e
r de
ci
sion t
r
ee alg
o
rithm
and di
dn’t try to
use
othe
r i
n
formatio
n g
a
in
. Variou
s alg
o
rithm
s
in
clu
d
ing I
D
3,
C4.
5
, CA
RT,
Ra
ndom
Fo
re
st
had
been
used t
o
cla
s
sify sa
tellite image
s, except th
e
C5.0
algo
rithm that i
s
still rarely
used
,
becau
se C5.0 algorith
m
is a new alg
o
rit
h
m as the de
velopment of C4.5 alg
o
rith
m.
This
work ap
plies the m
e
thod of de
cisi
on
tree b
a
se
d spatial a
u
toco
rrelation
namel
y
spatial d
e
ci
si
on tree
(SDT
) to cla
ssify pe
atland
befo
r
e
burn
ed, bu
rn
ed, and afte
r
burn
ed in Ku
bu
subdistri
c
t and Pasi
r Limau Kapas subdistri
c
t, Rokan Hilir, Riau Prov
ince. Parameter and
autocorrelatio
n
wa
s ad
de
d to CART
algorith
m
for this wo
rk, becau
se SDT have a si
mila
r
c
o
nc
ept with CART algorithm. This
work
tri
ed the
other pa
ram
e
ter for the
best result. The
results from the de
ci
sion t
r
ee b
a
sed
sp
atial aut
o
c
orrelation alg
o
rit
h
m are co
m
pare
d
with th
e
other de
ci
sio
n
tree algo
ri
thm like CA
RT algo
rith
m, C5.0 alg
o
rithm an
d C4.5 alg
o
rith
m.
Comp
ari
s
o
n
algorith
m
is d
one to determine wh
ethe
r the SDT is b
e
tter than tra
d
itional de
ci
sion
tree alg
o
rith
m. Acco
rdin
g
to the com
pari
s
on
of
the four al
gorit
hms, the b
e
s
t algo
rithm
for
cla
ssifying
pe
at fire
wa
s
a
nalyze
d
. Th
e
results we
re
expecte
d to
b
e
u
s
ed
to
cal
c
ulate
extent
of
the area in th
e cla
s
ses of b
e
fore bu
rn
ed,
burne
d, and
after burned.
2. Rese
arch
Metho
d
Study area i
n
this
re
sea
r
ch i
s
pe
atlan
d
in Kubu
subdi
strict a
n
d
Pasi
r Lima
u Kapa
s
sub
d
istri
c
t, Roka
n
Hilir,
Ri
au Province. This
st
udi
ed
use
d
remote
sen
s
in
g data,
peatla
nd m
a
p,
and h
o
tsp
o
ts.
Satellite ima
ges u
s
ed
we
re Lan
dsat 7
i
n
Rokan Hilir distri
ct,
Ria
u
Province
whi
c
h
wa
s taken fro
m
the
USGS
(Unite
d
State
s
G
eologi
cal
Survey). Th
er
e we
re fo
ur i
m
age
s u
s
e
d
that
are im
age
s a
c
uired in
May
,
July, August
,
and Novem
ber 2
002. Pe
atland ma
p 2
002 in
Riau
wa
s
use
d
to loca
te the peatland cover o
n
the sa
tellit
e image. Th
e map of p
eatland that
is
rep
r
e
s
ente
d
in polygon
s was obtain
ed from Wetland
s Internatio
n
a
l. The hotsp
ots in July 2
002
were obtai
ned from MODI
S Fire FIRM
S / Hotspo
t,
NASA / University of Maryland. Hot
s
pots
were used to determi
ne the
classe
s befo
r
e bu
rne
d
, bu
rned, an
d after bu
rne
d
.
2.1. Decision
Tree base
d Spatial Au
to
correla
tion
The de
ci
sion
tree metho
d
can a
u
tomati
cally sel
e
ct th
e app
rop
r
iate
supp
ortin
g
a
ttributes
that iterativel
y split the
gi
ven data
s
et i
n
to sma
lle
r g
r
oup
s acco
rd
ing to th
e dif
f
erent val
ues o
f
these
attrib
utes [8]. T
he
b
a
si
c
con
c
e
p
t
of the
deci
s
io
n tre
e
i
s
to
convert th
e d
a
t
a into a
tre
e
and
deci
s
io
n rul
e
s. The d
e
ci
si
on tree
co
nsi
s
ts of a
r
oot
node
at the top of the tre
e
,
the internal
node
whi
c
h is a b
r
a
n
ch of the tre
e
, and the lea
f
node whi
c
h i
s
the end of a
tree bra
n
ch.
The
spatial
d
e
ci
sion tree u
s
e
s
the n
e
igh
borh
ood
grap
h of traini
ng
pixels a
s
the
input an
d
build
s a spati
a
l de
cisio
n
tree mod
e
l. Th
e co
nvent
ion
a
l de
cisio
n
tree algo
rithm
use
s
info
rmat
ion
gain in the a
ttribute sele
ction; the pro
posed al
go
rit
h
m use
s
spa
t
ial information gain. Spa
t
ial
measure re
sulted from t
he spatial relation
ship
s
that may be
either to
pol
ogical or
m
e
tric
(dista
nce) i
s
use
d
in the formul
a of sp
atial in
formati
on gain in
ste
ad of numb
e
r of tuples in th
e
non-sp
atial i
n
formatio
n g
a
in [4]. Th
e
sp
atial d
e
ci
sion
tre
e
al
g
o
rithm
cal
c
ul
ates th
e
sp
a
t
ial
informatio
n gain by com
b
ining
conve
n
tional
an
d neigh
borhoo
d
split auto
c
o
rrel
a
tion ration
(NSAR). The
equatio
n use
d
to calculate
t
he value of NSAR is a
s
follows [7]:
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 14, No. 2, June 20
16 : 684 – 69
1
686
Г
′
Г
(
1
)
Whe
r
e
Г
i
an
d
Г
’
i
are local gam
ma
of sa
mple i
b
e
fore
and
af
ter
split resp
ectively. Spa
t
ial
autocorrelatio
n
is
often si
g
n
i
fi
cant at l
o
cal neig
hbo
rho
od level. Th
u
s
, we
ad
opt the lo
cal g
a
m
m
a
autocorrelatio
n
, formally de
fined as follo
ws [7]:
Г
∑
,
,
∑
,
,
(
2
)
W
h
er
e
i, j ar
e s
a
mp
le
in
d
i
ce
s
;
,
,
,
are spati
a
l simila
rity and cl
as
s
simil
a
rity, they are furthe
r
rep
r
e
s
ente
d
by W-matrix
,
and indi
cator
function
,
will
have a value of 1 if it has the sam
e
cla
ss a
nd i
s
0 if it has a different cl
ass;
is co
unt o
f
homoge
neo
us nei
ghb
ors.
NSAR value
s
use
d
in sp
atial informatio
n gain a
r
e NSAR value o
f
all sample
s. NSAR value of all sam
p
le
s
defined a
s
fol
l
ows [7]:
∑
(
3
)
Whe
r
e i is th
e index of a sampl
e
, varying from 1 to
m (m is the
numbe
r of sample
s). Fro
m
Equation (1), (2) and (3
)
spatial
informa
t
ion
gain
i
s
o
b
tained
as p
r
ese
n
ted at fol
l
owin
g equ
ation
[7]:
SIG
1
α
(
4
)
Whe
r
e
α
is a balan
cing pa
ra
meter.
2.2. Methodo
log
y
Analysis in
cl
ude
s four ma
jor step
s, na
mely
, (1) ima
ge pre
-
p
r
o
c
e
ssi
ng, (2
) determini
ng
cla
s
ses of im
age, (3)
dist
ri
bution of t
r
ain
i
ng data
an
d testing data
a
nd cla
ssification
p
r
o
c
e
ss, (4)
evaluation a
n
d
comp
arative analysi
s
of cla
ssifi
cation
results.
2.2.1. Image Pre-pro
ces
sing
The first pre
p
r
ocessin
g
sta
ge is g
eorefe
r
en
cing. G
eoreferen
cin
g
produ
ce
s ra
ster map
s
whi
c
h had th
e proje
c
ted
coordi
nate sy
stem UTM
Zo
ne 47 N with
WGS8
4, mea
n
ing that Ro
kan
Hilir is located at 47 N zone in
the UTM (Universal Transverse Me
rctator) proj
ect
i
on system
wi
th
geo
spatial referen
c
e
syste
m
of WGS84.
Combi
nation
red
g
r
ee
n
blue
(RGB)
of a fe
w b
a
nds cau
s
e
s
image
s h
ad
different
informatio
n. In this
study, t
he combi
nati
on of t
he i
m
a
ge involved i
s
band
7, ba
n
d
4, an
d ba
n
d
2.
The b
and
7
is represente
d
in red, the
band
4
is
repre
s
e
n
ted i
n
green, a
n
d
the ban
d 2
is
rep
r
e
s
ente
d
i
n
blue. In thi
s
ba
nd
com
b
ination,
veget
ation a
r
ea i
s
sho
w
n
by th
e green
col
o
r,
becau
se the
band 4
whi
c
h had hi
gh reflectan
c
e of
the vegetation re
pre
s
e
n
ted by the gree
n
colo
r. Band
7 is
sen
s
itive to radi
atio
n thus
it all
o
ws dete
c
tin
g
a heat
so
urce. Moreo
v
er,
according to
Wagte
ndo
nk
et al.
[9] use
s
of b
and
4
and b
and
7
of Land
sat E
T
M+ i
s
valid
to
detect the bu
rn scars.
This
study used
satellite i
m
ages of rokan Hili
r that
have peatland
cover. Overlay and
cro
p
satellite
image
s with
maps of pe
at are
ne
ce
ssa
ry to get the i
m
age th
at ha
s pe
atland
co
ver.
Overlaying
was
cond
ucte
d
to determin
e
the area
s of
peatlan
d
cov
e
r, and
cropp
ing wa
s
carri
ed
out to ta
ke
p
eatland
area
only. Satellite image
s
with
peatla
nd
co
ver still
have
a lot
of cl
ou
ds;
therefo
r
e it was n
e
cessa
r
y to sele
ct a
subset of
ima
ge with
cle
a
n
of the clo
ud.
The results
of
sub
s
et imag
e
s
incl
ude in K
ubu subdi
stri
ct and Pa
sir L
i
mau Kapa
s subdi
strict, Ro
kan
Hilir, Ria
u
.
2.2.2. Dete
r
m
ine Class
of Image
At this stage the overlay
pr
ocess
of hotspot
s with
satellite image aims to
obtain the
requi
re
d
cla
s
se
s. Burned
cla
s
s de
rived
from
hot
spo
t
s was ove
r
l
a
yed
with th
e ima
ge i
n
Juli.
Before bu
rne
d
cla
ss d
e
riv
ed from hot
spots was
ove
r
layed with th
e image in M
e
i. After burn
ed
cla
ss d
e
rived
from hotsp
ots wa
s overl
a
yed with
the image in Novembe
r
. Hotsp
o
ts used in this
study we
re ta
ken fro
m
July
2 to July 5, 2002.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Com
parative
Analysis of S
patial De
ci
sio
n
Tree
Algo
rit
h
m
s
for Burn
ed Area
… (P
utri Tha
r
iqa
)
687
2.2.3. Distrib
u
tion of Tr
aining Data an
d
Testing Da
ta
and Class
ification
Pro
cess
Before the cl
assificatio
n
st
age, the data
were
divid
e
d
into several
experim
ental
grou
ps
usin
g K-fold
cro
s
s validati
on with
k of 10. In
whi
c
h
9/10 data
we
re u
s
ed
as training d
a
ta a
n
d
1/10 data were use
d
as te
sting data. Cla
ssifi
ca
tion p
r
o
c
e
ss
wa
s co
n
ducte
d usi
ng
R software.
The de
ci
sion
tree ba
se
d spatial auto
c
o
r
relation
algo
ri
thm usin
g
sp
atial inform
ation gai
n
in Equation
(4). Thi
s
stu
d
y use
s
several
balan
ce p
a
rameters (alp
ha) in
equ
ation (0.1, 0.1
4
and
0.26)
of 0.1 b
e
ca
use it re
sult t
he be
st a
c
cura
cy of th
e cla
s
siffier
. Instea
d of u
s
i
ng the
entro
p
y
of
C4.5 al
gorith
m
in the
cal
c
ulation of info
rmation
gain
as in
Jia
ng
et
al.
[7], this st
udy used the
gini
gain value in
CART al
gorit
hm. Here is the equ
ation u
s
ed to calcula
t
e the gini gai
n [10]:
1
∑
(
5
)
,
∑
|
|
|
|
(
6
)
Whe
r
e
is th
e probability of
in t
he cla
s
s
,
is the partition of S in
duced by the value o
f
attribute A.
2.2.4. Ev
alua
tion and Co
mparativ
e Analy
s
is of Classifica
tion
Resul
t
s
Evaluation
was
con
d
u
c
ted
on a
c
curacy
, size of tre
e
s
, an
d num
b
e
r of
rule
s
of the fou
r
algorith
m
s. A
c
cura
cy was calculated u
s
ing
conf
u
s
io
n matrix. Ne
xt, comparati
v
e analysi
s
of
cla
ssifi
cation
outcom
e
s u
s
ed spatial
de
cisi
on tree
b
a
se
d spatial
autocorrelatio
n
algo
rithm, t
h
e
CART al
gorit
hm, the C4.5
algorith
m
and
the C5.0 alg
o
rithm.
3. Results a
nd Analy
s
is
3.1. Dete
rmine Class o
f
Image
Buffers with radiu
s
1 km were create
d
for
hotspots d
a
ta. This radi
us of 1 km was u
s
e
d
becau
se the
area fo
r o
ne
hotsp
ot in av
erag
e is
2.
58
657
km, therefore the
ra
di
us of the
circl
e
is
0.9073
7
km.
This value
is
con
s
id
ere
d
a
s
the
radiu
s
o
f
a b
u
ffer fo
r
a hot
spot.
Ou
tside th
e
buffers,
rand
om p
o
int
s
a
r
e
gen
era
t
ed a
s
false
alarm
data [
11]. Overlayi
ng bet
wee
n
buffer
zon
e
of
hotspots and satellite im
age was usef
ul to get informatio
n about the cla
ss before burned,
burn
ed, a
nd
after bu
rne
d
.
Cro
ppin
g
p
r
o
c
e
s
s was
pe
rformed
to get
a buffe
r a
r
e
a
in the
satell
ite
image.
The buffer zo
ne wa
s co
nverted into a d
i
gital numbe
r. The image in August that
would
be u
s
ed
for
cla
ssifi
cation
also
converte
d into a
digit
a
l num
ber. I
n
on
e pixel t
here
were
th
re
e
digital numbe
rs. A digital numbe
r wa
s o
b
tained from
band 7, ban
d
4, and band
2 becau
se o
f
comp
osite
p
r
oce
s
s. Di
gital
value
s
d
e
riv
ed from the
buffer
zo
ne
were m
a
tch
e
d
with
the
di
gital
values of the
image
in A
u
gust. If tho
s
e
three
digital
values of the
buffer zone
equal
to di
gital
numbe
r of im
age in Aug
u
st, then the pixel has
a buffe
r zo
ne a
s
the
matche
d cl
a
ss. If in imag
es
acq
u
ire
d
in Augu
st there a
r
e pixels that
doe
s not
hav
e a cla
ss o
r
n
o
t equal with
the pixels of the
existing buffe
r zon
e
, then the pixels a
r
e
cla
ssifie
d
as
a non pe
atlan
d
.
3.2. Ev
aluation and Com
p
arativ
e An
a
l
y
s
is of Classification
Re
sults
Comp
arative analysi
s
of the deci
s
io
n tree alg
o
ri
thms was p
e
rform
ed in
term of
accuracy,
size of tree
s, a
nd num
be
r o
f
rule
s.
A
c
c
u
racy
wa
s
cal
c
ulat
e
d
u
s
ing
t
he c
r
o
ss f
o
lds
validation wit
h
k=10. Th
e
accuracy
re
sulted
from t
he de
ci
sion t
r
ee al
go
rithm
based
spati
a
l
autocorrelatio
n
(SDT
), the
CART al
gori
t
hm, t
he C4.5 algo
rithm
and the
C5.0
algorithm
s a
r
e
pre
s
ente
d
in
Table 1. Th
e
results
were
averag
e of a
c
cura
cie
s
fro
m
10-fold. T
h
e C4.5 a
nd t
h
e
C5.0 al
gorith
m
had a
n
a
c
curacy of 3.3
%
greate
r
th
an the oth
e
r t
w
o al
gorith
m
s, be
cau
s
e th
ese
algorith
m
s ha
d a l
a
rge
nu
mber of
rule
s and
the l
a
rg
e si
ze
of tree.
The
C4.5 a
n
d
C5.0 al
gorit
hms
are u
s
u
a
lly used to p
e
rfo
r
m cla
ssifi
cati
on with
ca
teg
o
rical data a
n
d create a tre
e
with multi-split.
When there
were data
wit
h
continuo
us
attribute, that
algorithms
will
create a tree to binary split.
The used of binary split on an algorith
m
would ma
ke existing attributes ap
pea
r several time
s in
the tree.
Rep
e
tition of these attribute
s
could b
e
si
mpl
i
fied whil
e ch
angin
g
a d
e
ci
sion t
r
ee i
n
to
a
set of rule
s.
Neverth
e
less, that repeti
t
ion made la
rge
r
tree
s a
nd more co
mplex. The C5.
0
algorith
m
h
a
s an
accu
ra
cy
of 0.9% g
r
eat
er th
an
C4
.5
algorithm. Yet, result
from
C5.0 ha
s
la
rg
er
rule
s a
nd la
rg
er tree
s than
C4.5. Th
at is
becau
se the
C5.0
algo
rith
m ha
s b
o
o
s
ting a
nd
winn
o
w
s
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 14, No. 2, June 20
16 : 684 – 69
1
688
whi
c
h
wo
uld
improve th
e
accu
ra
cy a
nd di
sca
r
d t
he attri
bute
s
that h
a
s le
ss
contri
buti
on o
r
irrel
e
vant [12].
Table 1
.
Cla
s
sif
i
cat
i
o
n
com
pari
s
on
re
sult
Algorithm
Accuracy
(%
)
Number of
rules
Size of tree
C4.5
98.89
1681
3362
C5.0
99.79
595
1603
CART
95.67
8
15
Spatial Decision
Tree
(SDT
)
96.39
11
20
CART a
nd th
e SDT algo
rithm create tre
e
s by
usi
ng b
i
nary split. Binary split
sim
p
lify the
distrib
u
tion
criteria by co
n
s
ide
r
ing
all p
o
ssible
divid
ed attribute
s
,
then ch
oo
si
ng the be
st one.
This
situation
cau
s
e
s
o
n
n
u
mbe
r
of rul
e
s an
d
si
ze
of tree in
CART
and SDT are
lowe
r than
C4.5
and
C5.0. T
h
e sim
p
lify rul
e
s th
at re
sult
ed fro
m
bi
na
ry split
ha
s t
he lo
we
r a
ccura
cy comp
a
r
ed
with multi
s
plit
rule
s. Altho
u
gh, both
CA
RT and
SD
T u
s
ed
gini
inde
x, SDT ha
s
a
c
cura
cy of
0.
7%
better than
CART algo
rith
m, beca
u
se the SDT alg
o
rithm includ
es spatial auto
c
orrelation a
s
p
e
ct
that co
nsi
ders n
e
igh
bou
rhood
value
from eve
r
y
pixels i
n
si
de
inform
ation
gain
comp
uting
p
r
oc
es
s
.
The compa
r
i
s
on of the f
our alg
o
rithm
s
sh
ows that
the algorith
m
C5.0 ha
d the best
accuracy in multi-split crit
eria an
d algo
rithms
SDT had the best a
c
cura
cy in binary split criteria
.
Although the
C5.0 al
go
rithm ha
s 3.4%
better a
c
cu
racy than th
e
algorith
m
s S
D
T, but it is
not
efficient. The
efficien
cy of
an al
gorith
m
is p
e
rfo
r
m
ed in te
rm o
f
the sp
eed,
scalability,
and
interp
retation
[13]. The
sp
eed of
an al
g
o
rithm
wa
s o
b
se
rved
wh
e
n
the mo
del i
s
u
s
ed
to cl
a
ssify
a ne
w
data.
The
rule
s
are
gen
erat
ed from the
de
ci
si
on tree
s. The
num
ber of
rules ge
nerated
from C5.0 alg
o
rithm is
gre
a
ter than S
D
T algorith
m
with 595 rul
e
s
and SDT
alg
o
rithm only h
a
ve
1
1
ru
les
.
T
h
e imp
l
e
m
en
ta
tio
n
o
f
c
l
as
s
i
fic
a
tio
n
n
e
w
da
ta
us
in
g
C
5
.0
a
l
g
o
r
i
th
m,
ma
y ta
k
e
lo
nger
time process.
SDT algorithm will
require shorter time
because t
he
number of rules generated is
less
than C5.0 algorithm.
But, the acc
u
rac
y
res
u
lt
ed
fr
o
m
SDT
a
l
go
r
i
th
m is
s
m
alle
r
th
an
th
ose
o
f
C5.0 algorithm. Both of these al
gorithm
s
satisfy the criteria of
scal
ability because they were able
to build a
mo
del that ha
d a
fairly goo
d a
c
cura
cy
with
a larg
e nu
mb
er of d
a
ta. C5
.0 algorith
m
wa
s
more
diffic
u
lt
to interpret,
becaus
e
it has
a co
mple
x rule
s an
d t
r
ee
s. It differs fro
m
the S
D
T
algorith
m
that was ea
sy to unde
rsta
nd
becau
se
of
the simpler rules a
nd trees. The tim
e
compl
e
xity of the C4.5 alg
o
rithm an
d C5.0 algorith
m
is
О
(m
n
2
)
,
where m i
s
the
size of data
s
ets
and n is the
numbe
r of attributes [14]. The time
co
mplexity of the algo
rithm CART a
nd SDT
whi
c
h a
pplie
d
the con
c
ept
of a bin
a
ry tree is
О
(N l
o
g
N)
[15], wh
ere N i
s
a
num
ber
of attribut
es.
CART a
nd SDT algo
rithm
had sim
p
le
r complexity than C4.5 an
d C5.0 algorith
m
.
Li
et al.
[16]
also
com
pare
d
deci
s
io
n tre
e
algo
rithm in
remote
sen
s
i
ng. The a
c
cu
racy of
C4.5
algo
rith
m wa
s 0.866
and
a
c
cura
cy of CA
RT al
gorithm
was
0.857.
C4.5
algorith
m
h
a
d
a
good a
c
curacy, although CART algo
rith
m has mo
re trainin
g
sam
p
l
e
s than C4.5. This sh
ows the
C4.5
algo
rtih
m is the
be
st algo
rithm
ho
wever t
he
co
ndition
of the
data. But th
e C4.5 al
gorit
hm
wa
s lo
st th
an
the
C5.0
alg
o
rithm, it
ca
n
be p
r
ove
d
on th
e
r
e
s
u
lts o
f
th
is s
t
ud
y wer
e
d
i
sc
uss
e
d
in
the para
g
rap
h
above.
Table 2
.
Co
nfusio
n matrix of classifier from C5.0 alg
o
r
ithm
Actual
Prediction
Non-Peat
Before
Burn
ed
After
Burne
d
Burned
Non-Peat
327 0
3
0
Before Burn
ed
0 81716
35
0
After Burne
d
2 5
9674
46
Burned
0 0
29
9393
Table 3
.
Co
nfusio
n matrix of classifier from SDT algo
rithm
Actual
Prediction
Non-Peat
Before
Burn
ed
After
Burne
d
Burned
Non-Peat
210 56
25
37
Before Burn
ed
3 81284
434
0
After Burne
d
0 1059
7216
1466
Burned
0 15
469
8955
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Com
parative
Analysis of S
patial De
ci
sio
n
Tree
Algo
rit
h
m
s
for Burn
ed Area
… (P
utri Tha
r
iqa
)
689
Table 2 an
d
Table 3 a
r
e confusi
on matrix der
ived fro
m
one expe
ri
ment fold. Accuracy of
the cl
assifier
from
C5.0
al
gorithm
in T
a
ble 2
is
99.8
8
%, and t
he
accuracy
of t
he
cla
ssifie
r
f
r
om
SDT alg
o
rith
m in Ta
ble 3
is 9
6
.47%. Confu
s
io
n m
a
trix obtain f
r
om the
cla
s
sification p
r
o
c
ess
us
ing SDT algorithm
s
h
owed that
there are s
i
milarities
between a
fter burned c
l
as
s
with burned
cla
ss,
and
b
e
twee
n after burned
cl
ass
with b
e
fore bu
rne
d
cla
ss. Simil
a
ritie
s
b
e
twe
en
a
fter
burn
ed
cla
ss
with befo
r
e b
u
rne
d
cl
ass o
c
curred
be
ca
use
co
ndition
of the land
a
fter burned
h
a
s
turn b
a
ck i
n
to
peatla
nd
s an
d the
col
o
r
b
e
cam
e
g
r
e
e
n
agai
n. The
g
r
een
colo
r
sh
ows veg
e
tation.
Mean
while th
e simila
rities betwee
n
after bu
rn
e
d
cl
ass with b
u
rned
class o
c
curre
d
be
ca
u
s
e
burn
ed a
r
ea
still in the red
color,
red col
o
r sh
ows bu
rned area.
Figure 1. Image before cla
ssifi
cation
(a) C5.0
Algo
rithm
(b) C4.5
Algo
rithm
(c
) SDT Algo
r
i
thm
(d)
CART Alg
o
rithm
Non Peat
Burned
Noise
Before burne
d
After burned
Figure 2. Image cla
s
sificati
on re
sults
we
re
co
ntanin
g
noise aro
und
non pe
at cla
ss
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 14, No. 2, June 20
16 : 684 – 69
1
690
(a) C5.0
Algo
rithm
(b) C4.5
Algo
rithm
(c
) SDT Algo
r
i
thm
(d)
CART Alg
o
rithm
Non Peat
Burned
Noise
Before burned
After Burned
Figure 3. Image cla
s
sificati
on re
sults
we
re co
ntanin
g
noise aro
und
burn
ed cl
ass
and after bu
rned cl
ass
The re
sulte
d
of image cl
assificatio
n
will
had salt-n
-pe
pper
noi
se. Noise
salt-n-p
e
pper i
s
white d
o
ts o
r
black
contai
n
ed in the i
m
a
ge cl
assi
fication result
s. Noise
ari
s
e
s
b
e
ca
use there
is a
cla
s
s of mi
scl
assified. Fig
u
r
e 1
is imag
e
before
cl
a
ssi
fication pro
c
e
ss.
In
Figu
re 1
after burne
d
cla
ss lo
oks li
ke noi
se b
e
cause after b
u
rne
d
area a
ppea
r aroun
d
burn
ed area
and they ha
ve
small me
asure. Image that has
the m
o
st noi
se is i
m
age resulte
d
from CART
algorithm, a
n
d
SDT al
gorith
m
is
able to
redu
ce
that
noise.
The
m
o
st n
o
ise resulted a
r
ou
nd
non
peat
cl
ass,
burn
ed
cla
ss,
and after b
u
r
ned
cla
s
s. Figure
2 are image resulte
d
from C5.0, C4.5, SDT, a
n
d
CART
alg
o
rit
h
m which h
a
s
n
o
ise a
r
ou
nd n
on p
eat
cla
ss. Fi
gu
re
3 a
r
e i
m
age
re
sulted
fro
m
4
algorith
m
s which
ha
s
noi
se in
after bu
rned
cla
s
s a
n
d
bu
rn
ed
cla
s
s. Th
e
C5.0
a
l
gorithm
ha
d l
e
ss
noise, becau
se this al
gorit
hm had the b
e
st accu
ra
cy.
Rule
s f
r
om S
D
T
algo
rithm
indicated th
at befo
r
e
burne
d cl
ass
ha
s b
and
4 valu
e
great
e
r
than the ban
d 7 value, a burn
ed cl
ass has ban
d 7 value gre
a
ter than the band 4 value, after
burn
ed cl
ass was in the
middle value
of t
he band, and the non
-peat
class h
a
s ba
nd 2 va
lu
e
greate
r
than
any other ba
n
d
. Here
we
re
11 rule
s resul
t
ed from the SDT algo
rith
m:
1.
IF Band4 > 5
4
AND Ban
d
7
> 13 AND Band7
≤
51 THEN Before Burned
2.
IF Band7 > 5
1
AND Ban
d
4
> 70 THE
N
Before Burne
d
3.
IF Band4 > 5
4
AND Ban
d
7
≤
13 THE
N
Before Burned
4. IF
Band4
≤
54
AN
D
Ba
nd
7
≤
41 AND Ba
nd2
≤
49 THEN Before Burned
5.
IF Band4
> 5
4
AND Band
7 > 5
1
AND
Band4
≤
70
AND Ba
nd7
≤
79 T
H
E
N
After
Bu
r
n
ed
6.
IF Band7 > 4
1
AND Ban
d
4
≤
43 AND Ba
nd7
≤
54 T
H
EN After Burned
7. IF
Band4
≤
5
4
AND Ba
nd
7 > 41
AND
Band7
≤
66 AND
B
and
4 >
4
3
T
H
EN After
Bu
r
n
ed
8. IF
Band4
≤
54
AN
D
Ba
nd
7 >
6
6
T
H
EN
Bu
r
n
ed
9.
IF Band4 > 5
4
AND Band
4
≤
70 AND Band7
> 79 THEN Burne
d
10. IF
Band4
≤
54
AN
D
Ba
nd
7
≤
66 AND Ba
nd7 > 5
4
THEN Burne
d
11. IF
Band4
≤
54
AN
D
Ba
nd
7
≤
41 AND Ba
nd2 > 4
9
THEN Non
-
Peat
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Com
parative
Analysis of S
patial De
ci
sio
n
Tree
Algo
rit
h
m
s
for Burn
ed Area
… (P
utri Tha
r
iqa
)
691
4. Conclusio
n
The de
ci
sio
n
tree alg
o
rithm based
on s
patial
autoco
r
relat
i
on wa
s su
ccessfully
impleme
n
ted
by involving NSAR (Neig
borh
ood Aut
o
co
rrelation
Split Ratio) to the inform
ation
gain of the
CART alg
o
rith
m. That algo
rithm is a
b
le t
o
improve th
e accu
ra
cy of
CART
algo
rit
h
m.
Although C5.0 and C4.5 a
l
gorithm h
ad
high a
c
cu
ra
cy, but the number of rul
e
s generated from
the tre
e
a
nd
the si
ze
of t
r
ee
wa
s ve
ry
large
a
nd
t
h
e
cla
s
sif
i
er
w
a
s q
u
it
e
com
p
lex
,
so
it
c
o
u
l
d
redu
ce th
e ef
ficien
cy in the used of the
cla
ssifie
r
to
cla
ssify ne
w
data. In additi
on, the re
sult
s of
c
l
as
s
i
fic
a
tio
n
u
s
ing
SD
T
a
l
g
o
r
ith
m
s
h
o
w
s
th
a
t
th
e
r
e
is
s
i
mila
r
i
ty o
f
p
i
xe
ls
b
e
t
w
e
e
n
a
fte
r
bu
r
ned
cla
ss
with bu
rned
cla
s
s, and after b
u
rn
ed cla
s
s wi
th
before
burn
ed cla
s
s. Thi
s
is b
e
cau
s
e
the
land after bu
rned ha
s be
g
un to cha
nge
back be
cam
e
peat or ha
s not chan
ged
. The most n
o
ise
resulted a
r
ou
nd non p
eat cla
ss, bu
rne
d
class, and
a
fter burn
ed cl
ass. The C5.
0
algorith
m
h
a
d
less noi
se, be
cau
s
e thi
s
alg
o
rithm ha
d the best a
c
cura
cy.
Referen
ces
[1]
Adin
ugro
h
o
W
C
, Sur
y
a
d
ip
utr
a
INN, S
a
h
a
rjo
BH, Si
bor
o L.
Pan
d
u
an
Pen
gen
dal
ia
n K
e
b
a
kara
n H
u
ta
n
dan
L
aha
n Ga
mbut. Pro
y
ek
Climate
C
han
g
e
, Forest
a
n
d
Peatla
nds
in
Indo
nesi
a
. Bo
g
o
r: Wetlan
ds
Internatio
na
l-IP; 2005.
[2]
Hadi M. Pem
o
del
an sp
asia
l k
e
ra
w
a
na
n ke
b
a
kara
n di l
a
h
a
n
gam
but: stud
i kasus ka
bu
pa
ten Ben
g
kal
i
s,
provi
n
si Ri
au. Und
e
rgra
du
ate
T
hesis.
Bogor:
Institut Pert
anian Bog
o
r; 200
6.
[3]
Sharma
R, Gh
osh A, J
o
shi
P
K
. Decisi
on
tre
e
a
ppro
a
ch for
classific
a
tion
of
remotely
sensed satellite
data usi
ng o
p
e
n
source su
pp
ort.
T
hesis. India: T
E
RI Univer
sit
y
Ne
w
Del
h
i;
2013.
[4]
Li
X, Cl
aramu
n
t
C. A apati
a
l e
n
trop
y-b
a
se
d d
e
cisi
o
n
tree f
o
r
classific
a
tion
of geo
gra
phic
a
l inform
ation.
IEE Transaction in GIS
. 2006;
10(3): 451-
46
7.
[5]
Rinziv
ill
o S, Fr
anco
T
.
Classifica
tio
n
i
n
Ge
og
raph
ical
Inform
ation
Sy
stems. In: Boulicaut JF, Esposito
F
,
Giannotti F
,
Pedresc
h
i D.
Editors
. Artificial
Intelli
genc
e. Ne
w
Y
o
rk: Sprin
ger-Verl
ag. 20
04: 374-
38
5.
[6]
Sitang
ga
ng IS, R Yaakob, N Mustaph
a, AN Ainu
ddi
n. A Decision T
r
ee Ba
sed on Sp
atial
Relati
ons
hip
s
for
Pred
icting
Hotspots in
Pe
atlan
d
s.
T
E
LK
OMNIKA T
e
lec
o
mmunic
a
tio
n
Co
mp
uting
El
e
c
tronics
and
Contro
l
. 201
4; 12(2): 51
1-5
1
8
.
[7]
Jian
g Z
,
Sh
ek
har S, M
oha
n
P, Knig
ht J, C
o
rcora
n
J.
Le
arn
i
ng
spa
t
i
a
l de
ci
si
on
tree
fo
r g
e
o
g
r
ap
hi
cal
classification:
a summary of results
. 20
th
Internation
a
l
Confere
n
ce
Advanc
ec in
Geogra
phic
Information S
ystems. Ne
w
Y
o
rk. 2012; 12: 3
90-3
93.
[8]
Sitang
ga
ng IS,
R Ya
ako
b
, N
Mustaph
a, AN
Ainu
dd
in. Classification
Model for
Hotspot 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: 244-25
1.
[9]
W
agtend
onk J
W
, Root RR, Ke
y
C
H
. Compa
r
ison
of AVIRIS and La
nds
at ET
M+
detection cap
abi
liti
e
s
for burn severity
.
R
e
m
o
te
Se
nsi
n
g
o
f
En
vi
ro
nm
en
t
. 2004; 9
2
:
397
−
408.
[10]
Breima
n L, F
r
i
edma
n
JH, Ol
shen
RA, Sto
ne JC. C
l
assifi
cation
an
d Re
gressi
on T
r
ees. Ne
w
York:
Cha
p
man a
nd
Hall/C
RC. 19
8
4
.
[11]
Patil
N, Lath
i
R,
da
n C
h
itre
V. Comp
ariso
n
of C5.
0
&
CA
RT
classificati
on
alg
o
rithms
usin
g pr
uni
ng
techni
qu
e.
Internatio
nal Jo
urn
a
l of Engi
ner
in
g Rese
arch a
n
d
T
e
chno
lo
gy.
201
2.
[12]
Sitang
ga
ng IS,
R Ya
akob,
N
Mu
staph
a, AN
Ainu
ddi
n. Burn
Area Pr
ocessi
ng to Ge
ner
ate
F
a
lse A
l
arm
Data for Hotspot Prediction
Models.
T
E
LKOMNIKA T
e
lecommunic
a
tio
n
Co
mputi
ng E
l
ectronics
and
Contro
l
. 201
5; 13(3): 10
37-
10
46.
[13]
Han J, Kamber
M, Pei J. Data Minin
g
Conc
e
p
t and T
e
chniq
ue. Unite
d
State: Elsevier Inc.
2012.
[14]
Su J, Z
han
g H
.
A fast decis
i
on tree
le
arn
i
n
g
al
gorit
hm
. Pr
ocee
din
g
s of
2
1
st
nati
ona
l co
nferenc
e o
n
Artificial Intel
e
g
ence. 20
06; 1: 500-
505.
[15]
T
an PN, Steinbach M, Kumar V. Introduction
to Da
ta Mini
ng.
Boston: Pears
on Add
i
so
n W
e
sle
y
.
200
5.
[16]
Li C,
W
ang
J,
W
ang
L, Hu
L,
Gong
P.
C
o
mparis
on
of cl
as
sificatio
n
a
l
g
o
ri
thms
an
d trai
n
n
in
g sa
mple
si
z
e
s
in
urb
a
n
la
nd
class
i
fica
tion w
i
th
la
nds
at the
m
atic
mapp
er i
m
ag
ery
. MDPI (Mo
lec
u
lar
Div
ersit
y
Preservati
on In
ternatio
nal). R
e
motese
ns. 20
14: 964-
98
3.
Evaluation Warning : The document was created with Spire.PDF for Python.