selected from 8
, principal component
ing data, resampling
Is of above mentioned
is composed of selec-
8 bands. n=l,2,...,8.
ands or 2 or more bands
symbol C6B represents
mposed of 6 bands out
28 possible combina-
and 5, 7, 8, 9, 10, 11
it has the highest
e total band combina-
ination of CnB.
is composed of selec-
omponent transforma-
ination of PmB.
is composed of selec-
usly stated 8 ratio
3 RESULTS
3.1 Classification results
The imageries selected from 60 training areas were
approximately grouped into 3 categories, cryptomeria
plantations, conifers and mixed conifers plantations
and miscellaneous group (hardwoods, bamboo etc.).
According to their statistical and spatial related
ness, 29 cover type classes were obtained in the
study area. The data of the 29 classes were put into
the ISOCLA function in IDIMS, cluster classes from
the same cover type whose interclass statistical
distance were less than the chosen threshold, i.e.,
a transformed divergence less then 1800, were grouped
together. 29 cover types were combined into 23.
(table 4). The parameters used in ISOCLA function
were STDMAX:4.5; DLMIN:3.2; MAXCLS:35; NMIN:30; and
ISTOP:10.
Among the 23 classes, there were several similar
classes. Then the MAP, LPMAP and TRANSFER functions
in IDIMS were used to combine the 512x312 pixels
study area into 12 classes.
ination of RnB.
which is selected from
nB; 3 principal com-
st band combinations
PCi, PC 2 , PC 3 , 10/8,
ination of AnB.
d combination of CnB.
aination of AVnB.
ction, average diver-
e used as the ranking
ergences, 47 best band
above mentioned
riances and covariance
ions were input to a
tiich was used to
ypes.
t the accuracy of a
i is in the form of
k is a square array
columns which express
3 a particular land-
al land cover as
aterpreted aerial
f represent the
a rows indicate the
asses.
a generated, both
auracy are computed.
c correctly
In a class , nn
aumber of
a that class
: pixels that were
articular surface
aation. But it is
ression of the error
a error must combine
me summary measure
a in class pattern
a well as additions
3 (Kalensky et al.,
ised the mapping
100
ed pixels in
n class I
¡sions)
3.2 Best band combination determination
As previously stated, the best band combination is
Table 4. 23 cover type classes
Class
No.
Code
Species
Age
class
Slope
class
Aspect
1
A
Cryptomeria
plantation
4
4
North
2
C
Cryptomeria
plantation
2
6
Northwest
3
D
China-fir
plantation
2
7
West
4
E
Cryptomeria
plantation
3
7
North
5
G
Natural hardwoods
4
Southeast
6
H
Moso bamboo
2
North
7
I
Natural hardwoods
6
Northwest
8
J
Taiwania
plantation
1
3
East
9
К
Moso bamboo
4
East
10
M
Taiwan
red cypress
5
4
West
11
N
Taiwan
red cypress
5
5
Northwest
12
0
Moso bamboo
2
Southeast
13
P
Ma bamboo
3
Northwest
14
Q
Natural hardwoods
6
Northwest
15
R
Nursery
3
Northwest
16
S
Parking lot
3
Northwest
17
T
Mixed conifer
plantation
2
7
Southeast
18
и
China-fir
plantation
2
2
Northeast
19
V
Taiwan incense-
cedar plantation
3
3
North
20
w
Conifers & hard
woods plantation
2
4
East
21
X
Mixed conifers
1
4
North
22
Y
Conifers & hard
woods plantation
1
3
North
23
Z
Mixed conifers
2
3
North
determined via divergences. All the band combinations,
i.e., CnB, PmB, RnB, AnB, AVnB and their resampling
data were introduced to the divergence module and
a request was made for the best band combination to
be ranked according to average divergence and minimum
divergence. The average divergence is the average
interclass divergence for all pairwise combinations
of cover type classes. The band combination offering
the highest average divergence is then ranked first.
On the other hand, minimum divergence determines the
lowest interclass divergence for each and every band
combination. That combination possessing the highest
minimum interclass divergence is then ranked first.
The 47 best band combinations are shown in table 5.
Table 5. 47 best band combinations
Code
Band combination
Remarks
ClB
C2B
C3B
C4B
C5B
C6B
C7B
C8B
6
5,9
6,8,9
5,8,9,11
5,7,8,9,11
5,7,8,9,10,11
5,6,7,8,9,10,11
4,5,6,7,8,9,10,11
original band
ClB-R
C2B-R
C3B-R
C4B-R
C5B-R
C6B-R
C7B-R
C3B-R
5
5,9
6,8,9
5,8,9,11
5,7,8,9,11
5,7,8,9,10,11
5,6,7,8,9,10,11
4,5,6,7,8,9,10,11
resampling
original band
PlB
P2B
P3B
PCl
PCi,PC 2
PC, ,PC?,PC 3
principal
component
transformation
P3B-R
PC,,PC?,PC,
resampling PC
RIB
R2B
R3B
R4B
R5B
R6B
R7B
R8B
10/8,7/9
10/8,7/9,5/11
10/8,7/9,9/11,4/7
10/8,7/9,9/11,4/7,5/11
10/8,7/9,10/5,9/11,4/7,5/11
10/8,10/6,9-6/9+6,10/5,9/11,4/7,5/11
10/8.10/6,7/9,9-6/946.10/5,9/11,4/7,5/11
ratio image
R8B-R
10/8,10/6,7/9,9-6/946,10/5,9/11,4/7,5/11
resampling
ratio image
A1B
A2B
A3B
A4B
A5B
A6B
A7B
A8B
PC!
PCi,7/9
PCi,PC 2 ,PC 3
PC 2 ,PC 3 ,7/9,9
PC 2 ,PC 3 ,10/8,7/9,9
PC 2 ,10/8,7/9,5/11,5,9
PC 2 ,PC 3 ,10/8,7/9,5/11,5,9
рсьРСр.рс,, 10/8,7/9,5/11,5,9
mixed band
A8B-R
PCi,PC 2 ,PC 3 ,10/8,7/9,5/11,5,9
resampling
' mixed band
AV1B
AV2B
AV3B
AV4B
AV5B
AV6B
AV7B
AV8B
6
5,9
6,8,9
5,8,9,11
5,7,8,9,11
5,7,8,9,10,11
5,6,7,8,9,10,11
4,5,6,7,8,9,10,11
spatial
filtering
band
AV8B-R
4,5,6,7,8,9,10,11
resampling
filtering
band
Table 6. Divergences of C2B and C5B
Code
Average
divergence
Minimum
divergence
Rank
Band
combination
C2B
1790
146
1
5,9
1789
114
2
5,10
1784
329
3
5,8
C5B
1965
1111
1
5,7,8,9,11
1965
1018
2
5,7,8,10,11
1963
1115
3
5,6,8,9,11
1963
1089
4
5,8,9,10,11
