18
Table 7. Ranks of 47 best band combinations
Percent
Mapping
KHAT
CPU
Code
Band combination
accuracy
accuracy
value
time
Rank
(%)
(%)
(min. )
AV8B-R
4,5,6,7,8,9,10,11
98.64
97.45
.982865
160
2
C8B-R
4,5,6,7,8,9,10,11
98.46
97.09
.980647
155
1
C7B-R
5,6,7,8,9,10,11
97.13
94.92
.963907
129
3
C6B-R
5,7,8,9,10,11
97.00
94.87
.962316
109
4
AVÖB
4,5,6,7,8,9,10,11
96.92
94.52
.961431
155
6
AV6B
5,7,8,9,10,11
96.22
93.58
.952531
109
4
AV7B
5,6,7,8,9,10,11
95.54
92.69
.943983
129
7
R8B-R
10/8,10/6,7/9,
9-6/946,10/5,
9/11,4/7,5/11
95.24
92.24
.940128
158
8
A8B-R
PC,,PC 2 ,PC 3 ,
94.10
91.42
.925931
186
11
10/8,7/9,5/11,
5,9
AV5B
5,7,8,9,11
94.05
90.78
.925385
90
9
C5B-R
5,7,8,9,11
93.92
90.44
.923883
90
9
C4B-R
5,8,9,11
92.89
89.03
.911200
70
12
AV4B
5,8,9,11
91.53
87.35
.894271
70
13
C8B
4,5,6,7,8,9,10,11
89.08
82.87
.865814
150
16
C7B
5,6,7,8,9,10,11
88.35
81.99
.857194
125
15
A3B
PC!,PC,,PC,,10/8,
7/9,5/11,5,9
88.10
82.70
.852243
181
17
C6B
5,7,8,9,10,11
86.99
80.35
.840499
105
14
C5B
5,7,8,9,11
85.56
78.72
.822961
85
18
R3B
10/8,10/6,7/9,
9-6/946,10/5,
9/11,4/7,5/11
84.82
77.50
.812077
153
19
A7B
PC,,PC,,10/8,7/9
5/11,5,9
84.22
78.26
.804332
157
20
P3B-R
PC,,PC,,PC,
5,8,9,il
83.87
81.76
.798504
62
21
C4B
83.21
76.57
.795936
65
22
A6B
PC 2 ,10/8,7/9,
5/11,5,9
82.28
77.46
.780216
136
23
A5B
PC,,PC,,10/8,•
7/3,9
81.27
75.36
.766500
116
24
C3B-R
6,8,9
80.39
74.55
.758890
47
25
P3B.A3B
PC,,PC,,PC 3
10/8,10/6,
79.76
74.69
.752586
60
26
R7B
79.25
74.09
.743437
128
28
9-6/946,10/5,
9/11,4/7,5/11
R6B
10/8,7/9,10/5,
9/11,4/7,5/11
78.22
72.04
.731961
107
27
A4B
PC,,PC,,7/9,9
76.58
74.63
.711602
96
31
AV3B
6,8,9
75.70
69.41
.703845
47
29
R5B
10/8,7/9,9/11,
4/7,5/11
75.45
70.61
.697990
88
30
R4B
10/8,7/9,9/11,
4/7
74.31
69.96
.684153
67
34
P2B
PCj.PC,
10/8,7/9,5/11
73.66
66.56
.678652
60
33
R3B
73.61
70.53
.675170
46
32
AV2B
5,9
68.94
63.93
.624756
31
35
A2B
PC,,7/9
67.61
67.95
.607137
60
37
C3B
6,8,9
65.82
57.01
.599942
45
36
C2B-R
5,9
64.18
63.06
.572122
31
38
C2B
5,9
54.05
48.29
.473751
30
39
R2B
10/8,7/9
50.16
43.28
.434277
31
40
AV1B
6
48.42
51.82
.373982
15
41
C1B-R
6
46.38
54.41
.353796
16
42
PIBjAlB
PC!
44.57
41.42
.357885
30
43
C1B
6
41.59
42.82
.321694
15
44
RIB
7/9
29.22
29.98
.228814
15
45
Their ranks were determined by average divergence
first and then minimum divergence. For example, among
the C2B's 28 possible band combinations, the combina
tion of band 5 and band 9 is the best because it
possesses the highest average divergence, 1790 (table
&). The C5B represents a band combination with 5
bands. Among their 56 possible band combinations, the
combination of band 5, 7, 8, 9, 11 and combination of
band of 5, 7, 8, 10, 11 possess the same average
divergence. According to their minimum divergences,
we ranked the former first, the latter second.
As previously mentioned, on land cover types classi
fication accuracy assessment, error matrix was used.
By the error matrix of each band combination, the
percent accuracy, mapping accuracy and KHAT value
were computed (table 7). The KHAT value is a measure
of agreement which represents how well the classifi
cation data agree with the reference data (Congalton
et al., 1983).
Table 8 shows the CPU time of various operations.
Generally speaking, the computer time required for
the classification increases approximately as double
of the number of band used. On the best band combina-
Table 8. CPU time of various operations
Operation
CPU time
(min.)
Spatial filtering
5
Resampling
5
Ratio images
3
Principal component transformation
15
Classification with 8 bands
150
Classification with 7 bands
125
Classification with 6 bands
105
Classification with 5 bands
85
Classification with 4 bands
65
Classification with 3 bands
45
Classification with 2 bands
30
Classification with 1 bands
15
tion selection, computer time needed should be the
top consideration. For instance, according to the
testing of significance, there is no siginficant
difference between AV8B-R and C8B-R. But AV8B-R spent
more CPU time on data processing, therefore C8B*R
ranked first, and AV8B-R second.
4 CONCLUSIONS
Bearing in mind the objective to find the optimal
band combinations of airborne MSS data for forest
cover types classification, the following conclusions
are drawn. (1) The result has shown that the past
experience, maximum accuracy in cover type classifi
cation can be obtained by using all wavelength bands
available, is true. From table 7, one can find that
the best band combination contains more bands.
(2) Resampling and spatial filtering techniques
increase 7-10% classification accuracy because both
of them have the ability in sharpening the images to
approach a better resolution. (3) Principal component
transformations seem no effect on improving the
accuracy of forest cover types classification because
they are less effective on contrast enhancement for
vegetations. (4) Ratio images are also not good enough
for forest cover types classification because the
ratio images tend to suppress the difference in
albedo. Thus dissimilar materials with different
albedos may be inseparable on ratio images. (5) Six
out of eight components of mixed bands are unable to
improve classification accuracy. The mixed bands
perform naturally insufficiently.
ACKNOWLEDGMENTS
The authors would like to express their appreciation
to the Remote Sensing Planning Committee, Council of
Agriculture, Republic of China, for providing the
financial support. Thanks are also due to the Center
for Space and Remote Sensing Research, National
Central University, for data processing services.
REFERENCES
Congalton, R.G.; R.G. Oderwald; R.A. Mead 1983.
Assessing Landsat classification accuracy using
discrete multivariate analysis statistical tech
niques. Photogrammetric Engineering and Remote
Sensing, Vol. 49, No.12, Dec. 1983, pp.1671-1678.
Fleming, M.D.; R.M. Hoffer 1977. Computer-aided
analysis techniques for an operational system to
map forest lands utilizing Landsat MSS data, LARS
technical report 112277, Purdue University, West
Lafayette, Indiana, U.S.A. pp.1-231.
Kalensky, Z.; L.R. Scherk 1975. Accuracy of forest
mapping from Landsat Computer Compatible tape.
Proceedings of the 10th international symposium on
Remote Sensing of environment. PP.1155-1163.
Walsh, S.J. 1980. Coniferous tree species mapping
using Landsat data. Remote sensing of Environment
9. PP.11.26.
Symposiur
Method
for auto
R.I.Elman
All-Union ’Les 1
ABSTRACT: /
sidered. Th
matlon degt
fects in cc
ur networks
into the ve
code thread
lay process
1 CONTOUR L
In the fore
use of info
space photo
on contour 1
contour bas
operations i
characterisi
complicacy <
of contour :
In the wo:
input of th<
help of digJ
put is effe<
not too comj
segments» T1
in case of i
ous contour
tion of whi<
thod is slo*
a lot of mar
A higher t
tion of the
can be achic
cal-mechanic
case means c
urs on photc
units to be
having a cl«
al and space
ated automat
threshold pr
detail in li
Sobel, Kirsc
cultural tas
burns, cutov
part of the
unextracted.
rarely, in p
A more com
thod of the
parating dif
other is bas
with a furth
segments. In
Potapov, Svi
Pamorozski 1'
re features i
on. For exam]
tapov, Sviri<
