543
Table 3. confusion matrix of classification results
(2)
'• s.)
•ix of clusters
(3)
r K
persion matrix
(4)
1 cluster
different
the use of the
nd feature en-
s are shown in
expression (1 )
age set is the
e basic image
cessing.
iliary height
lassification
tion into clas-
that the deci
se he mem can
Cj pure classes mixture classes
0)
p ij
w±
—
' G
-F -
-c
"(c)-
-F -
c -
W -
(C)
(CJ
c
F
W
+
+
+
+
+
+
+
+
+
(F)
(G)
(F)|(F) (C)
(w)
(c)
(G,F )(S,F ,W )
cultiv. land
72.«
il .8
1 (j. 3
60.0
45.7
43.2
£#0
•H
CD
j Forest
,2.2
72. 1
5.8
24. 1
42.7
6.8
water
5.0
0.3
70.9
3.6
0.7
40.1
grass
3.!
10.0
0
3.4
6.1
0
-p
orchards
2. 0
2.3
1.0
6.1
3.7
0.5
3
0
settlem. Ian
3..,
»
0
2.,
0.4
0.5
43
-P
bear land
"
0.,
0.7
0
?
area
3822
4042
104
6403
3440
220
'S
cultiv. land
80.5
8.6
0.9
0
,2.5
5b. G
50.1
26.0
62.3
29.0
30.2
61 3
•H
Forest
7.0
78.0
0
34.7
47.1
26.8
41.9
64.7
6.2
0
27.6
10.9
-p
j water
4.3
0.2
90. 1
0
0.2
2.8
0.7
0.5
23.7
66.4
0.8
ine
bO
l grass
2.8
7.0
^“5
02.7
32.6
3.2
1.9
2.0
3.4
0.9
29.1
10.3
•H
a>
i orchards
2.7
2. 1
n
0
4.5
6.4
3.3
2.8
0.6
0
5.5
2.8
43
settlem. Ian
1.7
0
0
0
0
2. 1
0.8
0.2
3.9
3.7
0
0
43
-p
bear land
0
3.5
0
2.7
3.1
°-'
1.4
2.9
0
0
0.8
0
area
4075
230,
9,
1275
0739
855
3772
355
107
254
604
&i (x) = + - 2 Inpc^«) (5)
Where, the P(W{) represents a priori probability of
a feature class (W*), which is usually estimated by
the area percentage of class (Wi) in whole study area
(see the figures in last line of table 1). However,
the a priori probabilities of ground feature in dif
ferent hieght range are different in practice (see
table 1). So the better classification result can
only be got when P (Wt) is estimated in certain
ground height range and the classification is per
formed within the image area corresponding to the
same height ra.nge .
In our experiment, the study area was first digitals
iy segemented into different height range area by
taking- the "density (height)" sliced DTM image as
masks. Then the classification was performed se-
perately in different height range areas. Finally,
the results from them were digitally mosaiced each
other and forming- resulting 1 classification image.
Table 3 shows the confusion matrix of classification
seperately without (upper block) and with (lower
block) introducing ground height information. From
the table we can find that the classification ac
curacy was improved by 8% for cultivated land when
height information was introduced.
(2) Itsuhito Ohnuki (1981 ): Terrain Effect Nor-
marization Method of landsat Data and its Efficiency
of Forest Type Glassification, Forestry and Forest
Production Institute, P.0. Box 16, Tsukuba Norin-
kenkyu-danchi, Ibaraki 305» Japen.
(3) J.A. Richards, D .A. landgrebe, P.H. Swain (1982)
A Means for Utilizing Auxiliary Information in Mul-
tispectral Classification. R.S. of Eavironment, 12,
463 - 477.
(4) Yang Kai, Lin Kaiyu, Chen Jun, Lu Jian (1985) s
A Classification Scheme of landsat Multitemporal
Feature Images with the Use of Auxiliary DTM Data.
Acta Geodetica et Cartographies Sinica, Vol. 14»
Ho. 3 China.
5* Conclusions
Based on above classification processing, the areas
of defined ground feature classes in the testing-
county was calculated and compared with the existing
findings. The results shown that about 30% of cul
tivated land area in the county was not taken into
account in the existing findings. So that, we can
concluded that by elaborate classification scheme,
particularly by introducing ground height information
into classification procedure, the LANDSAT MSS images,
although whose geometric resalution is origionally
not high enough, can satisfyingly be used to verify
the lack fidelities of existing findings to cultiva
ted land in county level.
References
(1 ) J.T. Tou, Gonzalez (1 974): Pattern Recognition
Principles, Addison-Wesley Publishing Company.