Table 1. a priori probability p (i) of different classes in different
height ranges
classes
r V'
2
3
4
—
8
?
P(h)
height
probability p(i;
(c)
(0)
(F)
(B)
(w)
(S)
(G)
(1)
0m~20 ,n
86
■ ! ■
1
11
1
-
11
(2)
21 ni—50i>i
67
4
18
4
5
3
34
(3)
51 m~IOOni
«
6
45
2
2
2
1
28
(4)
101 ■—200>ii
21
1
6»
4
•
4
17
(5)
20l">~900»>
13
5
51
26
8
10
entire
county P( W-^)
46
4
36
6
3 2
1
100
Table 2. calculation results for feature image selection
sequen
feature images
deraen-
t ion
X',
X,
X,
nr
J ■
/.
1
BI089
2
1.14
0.51
-
-
1.65
2
RT89
2
1.67
0.11
-
1.08
3
KL89
4
1.77
1.10
0.73
0.09
1.86
1.83
4
B79
4
1.14
1.08
0.85
0.59
1.73 j
1.94
5
B?8
4
1.62
1.13
0.93
0.44
2.06 |
2.06
distinguished in advance. The a priori probabilities
of different classes (cultivated land (C), orchard
land (0), forest land (f), grass land (G), water
area (w), settlement area (S) and bear area (B) are
listed in table 1 .
The class attributes of each sampling pixels were
encoded by class numbers and imput to computer
system for forming ground truth sampling images.
3. Selection of feature image-s
Before classification action, an important prepro
cessing is so-called feature selection. To this end,
some image transformations in multispectral domain
should be performed, and certain criterion for
feature selection should be defined.
In our experiment following alternative feature
images had been prepared by digital image transfor
mation techniques.
(a) B78 origional four bands' LANDSAT MSS
image set corresponding to the year 1978;
(b) B79 origional four bands' LANDSAT MSS
image set corresponding to the year 1979;
(c) KL89 two-temporal princple component
image set, with four bands images seperately cor
responding to the 1st and 2nd principle components
of origional B78 as well as B79 image sets;
(d) RT89 two-temporal ratio image set, with
two bands' images seperately corresponding to the
ratios of band 7 to band 5 of origional B78 as well
as B79 images;
(e) BI089 two-temporal biomass index image
set, with two bands' images seperately corresponding
to the Biomass Index transformations of origional
B78 as well as B79 images.
In order to selecte optimum one from above alter
native feature images, the "dispersion matrix
criterion" was used in our practice:
if J(i) = min { J(B79),..., J(BI089) } . (1 s
then (i) is the optimum feature image ' '
J = tr (Si' • S, ) = Z Aj (2 )
Xj — eigenvalue of matrix ’ S,)
S, internal dispersion matrix of clusters
S » = ]r ^ x'sojn ^ ) ( X —Mk Y J ( 3 )
k cluster number
X intensity vector
Mk expected vector of cluster K
Si general (or external) dispersion matrix
of clusters
Si= ? (Xj-AKXt-s.y (4)
1 feature point number
Jt, expected vector of general cluster
Each J (i) can be calculated based on different
alternative feature images, and with the use of the
image part corresponding to real ground feature en
coding images. The calculation results are shown in
table 2. According to the criterion (expression (1 )
we can find that the BI089 feature image set is the
optimum one. So it was selected as the basic image
data for following classification processing.
4.-Classification processing with auxiliary height
information
An efficient strategy for improving classification
is to introduce ground height information into clas
sification process. It is well known that the deci
sion function of a maximum likelyhood schemem can
be expressed as
whe re
