infrared band data. 
Q 
i 2x39 
Ti =Z (x; x) / 9 eo. ® © (2) 
T) -Zz|px,x / 8 e€eoeo9 09909 (3) 
where x. is the value of the center pixel. 
T3 = 3] A, 1 / 4 ee 00000 (4) 
where i,j are adjacent pixels. 
After the application of principal component analysis to these 
parameters, the first principal conponent, of which proportion is 
92.5$, was used for classification as a feature vector(Figure 4). 
The second method is the employment of a non-parametric 
classifier. Although land use classes have large variance and 
non-Gaussian nature, these classes have very distinctive shape of 
histgram.  Consequentry, if the shape of histgrams can be utilized 
as a feature vector, it will increase the classification accuracy. 
From this point of view, we developed new classification algorithm 
called a histgram matching algorithm. 
In the histgram matching algorithm, shape of histgrams with 
neighbor points including a unknown pixel is compared with each 
shape of training data and the unknown pixel is classified to the 
class which has the nearest hidtgram shape. Since 5x5 pixels 
window was used in this study as neighbor points, histgrams of 
training data were normalized into cumulative frequency of 25.0. 
The distance between a unkown pixel and each classes is defined by 
eq. (5). 
4 
D.. = H. - ZH. — e090 0 0 
;5;7Xl8 09 Zu, (x-w)S(w)| (5) 
D, 3 distance between class i and pixel j 
HT: histgram of class i 
uj : histgram of neighbor points included a unknown pixel 
S : smoothing function 
r il (w=-1,0,1) 
Si) =. 0 (else) 
This algorithm time-consuming compared with maximum likelihood 
method, but has following advatages; 
1) training data need not indicated a Gausiun distribution, 
because it is a non-parametric classifier. 
2) Although training data have large variance, classification 
accuracy do not decrease. 
3) Since smoothing is performed automaticaly in this procedure, 
the results are not so match influenced by noises and uniform 
regions can be extracted. 
In this study, first and second principal components for 
spectral features (G,R,IR) and first pricipal component of texture 
features were used in land use classification. An example of a 
histgram of crop fields used in classification are shown in Figure 
24 
The third method is a utilization of decision tree classifier. 
A change pattern in land use has a regurarity. For example, 
forests can change to urbon areas, while urbon areas do not change 
to forests. A decision tree was constructed based on this 
680 
Se ARE TW ze 
e A a 
| 
ol