16 
the expression of topography; some make sharp contrast 
between cover types. 
Another enhanced images were created by spatial 
filtering. In order to reduce all the sources of noise 
and propagation fluctuations, the moving average 
filter was employed for this effort. Eight original 
bands data were put into an IDIMS function called 
AVERAGE. The spectral responses for 3 X 3 pixels were 
averaged with an input mask of coefficients to be 
moved over the entire array of data. 
Besides the enhancement techniques mentioned in 
previous paragraphs, bilinear interpolation algorithm 
was applied to the 8 original bands data and several 
selected images. In general, resampling accomplishes 
two things: (1) It relocates the pixels to more 
accurate positions. (2) It can smooth out the blocky, 
mosaiclike pattern that results when individual 
pixels become visible. 
The analytical technique, that was used to process 
the airborne MSS data, was hybrid supervised/unsuper 
vised maximum likelihood pattern recognition approach. 
Experience has shown the desirability of using both 
supervised and unsupervised approaches together in a 
procedure (Fleming et al., 1977). 
2.3 Classification procedures 
Classification is the process of assigning each of 
the pixels of the airborne MSS data of the study area 
to a class based on the set of input statistics 
generated by training area and cluster formation. 
Eight original bands of airborne MSS data were sub 
jected to standard clustering and classification 
techiques. On grouping of training samples into 
meaningful classes, sixty training areas which were 
selected by their uniformity of light radiance using 
function TR, TA and IRRECV in IDIMS. The spectral 
variation is caused by different vegetation commu 
nities present and the presence of environmental 
parameters such as tree species, tree age, slope and 
aspect which are found to influence the reflectance 
values of vegetation cover types (table 2,3). 
Table 2. Age class of forest plantations. 
Class 
Age (year) 
1 
1 - 6 
2 
7-15 
3 
16 - 25 
4 
25 - 40 
5 
40 - 60 
6 
over 60 
Table 3. Slope class 
of forest stands 
Class 
Slope 
(%) 
1 
0 - 
5 
2 
6 - 
15 
3 
16 - 
30 
4 
31 - 
45 
5 
46 - 
60 
6 
61 - 
75 
7 
over 
75 
After the data of training areas were collected, 
they were then put into an IDIMS function entitled 
ISOCLA. The ISOCLA algorithm in IDIMS used a com 
bination of distance measurements and statistical 
parameters to distinguish differences between 
classes. This type of procedure was used to train 
the maximum likelihood classifier of IDIMS. The 
maximum likelihood classifier was used to determine 
the statistical parameter which characterizes the 
classes (Walsh, 1930). 
2.4 Best band combinations 
3 RESULTS 
The best band combinations were selected from 8 
original MSS data, ratio images, principal component 
transformations, spatial filtering data, resampling 
data and mixed bands. The symbols of above mentioned 
data are designated as follows: 
CnB: A band combination which is composed of selec 
ting n bands from the original 8 bands. n=l,2,...,8. 
It may represent one of the 8 bands or 2 or more bands 
combination. For instance, the symbol C6B represents 
the band combination that is composed of 6 bands out 
of 8 bands available. There are 28 possible combina 
tions of C6B. Among them, the band 5, 7, 8, 9, 10, 11 
combination is the best because it has the highest 
value of average divergence. The total band combina 
tion of CnB is 255. 
CnB'R: A resampling band combination of CnB. 
PmB: A band combination which is composed of selec 
ting m bands from 3 principal component transforma 
tions (PCj, PC 2 , PC3). m=l,2,3. 
PmB*R: A resampling band combination of PmB. 
RnB: A band combination which is composed of selec 
ting n ratio images from previously stated 8 ratio 
images n=l, 2 8. 
RnB-R: A resampling band combination of RnB. 
AnB: A mixed band combination which is selected from 
2 best band combinations from CnB; 3 principal com 
ponent transformations and 3 best band combinations 
from RnB. i.e., band 5, band 9, PC^, PC 2 , PC3, 10/8, 
7/9, and 5/11. n=1,2,...,8. 
AnB'R: A resampling band combination of AnB. 
AVnB: A spatial filtering band combination of CnB. 
AVnB’R: A resampling band combination of AVnB. 
In best band combination selection, average diver 
gence and minimum divergence are used as the ranking 
criteria. By the ranking of divergences, 47 best band 
combinations were selected from above mentioned 
various combinations. Means, variances and covariance 
matrices of these band combinations were input to a 
maximum likelihood classifier which was used to 
characterize the forest cover types. 
2.5 Accuracy assessment 
The most common way to represent the accuracy of a 
airborne MSS data classification is in the form of 
an error matrix. An error matrix is a square array 
of numbers set out in rows and columns which express 
the number of pixels assigned as a particular land- 
cover type relative to the actual land cover as 
verified in the field or from interpreted aerial 
photographs. The columns usually represent the 
corrected reference data and the rows indicate the 
computer assigned land-cover classes. 
Once the error matrix has been generated, both 
percent accuracy and mapping accuracy are computed. 
The percent accuracy is shown as: 
number of correctly 
pixels in a class 
total number of 
pixels in that class 
It is a measure of the number of pixels that were 
correctly classified within a particular surface 
cover type classification delineation. But it is 
somewhat inadequate for the expression of the error 
in class location on a map. Such error must combine 
the commission and omission in one summary measure 
since it must reflect the change in class pattern 
due to losses (omissions) to, as well as additions 
(commissions) from other classes (Kalensky et al., 
1975). Consequently, it is proposed the mapping 
accuracy as a new measure. 
N T 
percent accuracy (%) 
100 
mapping accuracy (%) = 
N I +E I 
X 100 
wljere 
Nj.: number of correctly classified pixels in 
class I 
Ej.: number of erroneous pixels in class I 
(sum of omissions and commissions) 
3. 1 Classifii 
The imageries 
approximate!} 
plantations, 
and miscellar 
According to 
ness, 29 cove 
study area. 1 
the ISOCLA fu 
the same cove 
distance were 
a transformed 
together. 29 
(table 4). Th 
were STDMAX:4 
ISTOP:10. 
Among the 2 
classes. Then 
in IDIMS were 
study area in 
3.2 Best band 
As previously 
Table 4. 23 c 
Class 
No. 
Code 
1 
A 
C: 
P- 
2 
3 
C 
D 
C] 
P- 
C1 
pj 
4 
E 
Ci 
pj 
5 
G 
Ne 
6 
H 
Me 
7 
I 
Na 
8 
J 
Ta 
Pi 
9 
K 
Mo 
10 
M 
Ta 
re 
11 
N 
Ta 
re 
12 
0 
Mo 
13 
P 
Ma 
14 
Q 
Na 
15 
R 
Nui 
16 
S 
Pai 
17 
T 
Mil 
pie 
18 
u 
Chi 
pie 
19 
V 
Taj 
cec 
20 
w 
Cor 
woe 
21 
X 
Mix 
22 
Y 
Con 
woo 
23 
Z 
Mix