CLASSIFICATION 
The classification process can be described by the order of 
the numerical computations used in partitioning the signal 
space. The use of a clustering algorithm for classification 
is performed in a manner similar to LARSYS,i.e., the maximum 
likelihood classification, putting the emphasis on the super 
vised one. The partition of the color space is performed by 
the use of the discriminant G.(x) for multivariate Gaussian 
1 
density functions of the data vector x 
G ± (x) =logP (w ± ) -nlog (2tt) /2-log | K . | /2- (X-M.) T K i _1 (X-M ± ) /2 , (9) 
where P(w.) is the apriori probability for the pattern class 
i, M. is the mean value vector of the pattern class i, and 
K. is the covariance matrix of the pattern class i. The fol 
lowing inequality for all i^j 
G i (x)>G^(x) (j=l, 2, ,n) (10) 
gives such criterion that the data vector x belongs to the 
class w.. 
l 
The particulate method being used determines the form of the 
quadratic classifier. The unknown parameters in these dis- 
crimants are determined in a preliminary process call and 
training, via the supervised algorithm, when one is supplied 
with a set of training sample patterns of known classifica 
tion. In other words, these samples are used to develope dis 
criminants, which may then be used to classify ^unknown sam 
ples. Then, so far as the training samples are truely repre 
sentative of the classes and the discriminant is appropriat 
ely computed, the classification will be reasonably reliable. 
Thus, a crucial aspect of the classification procedure is the 
election of training samples. It is accomplished by visual 
inspection of the imagery, coupled with additional sources 
of such informations as the topographic maps, aircraft pho- 
tographes, and personal knowledge of the area. In the pre 
sent stu^y, training data are selected from Kanazawa area 
in 300km , which includes a portion of Japan sea, Kanazawa 
harbor,Kahoku lagoon, Asano- and Sai-rivers, Kanazawa city, 
and its outskirts. The line and sample coordinates of the 
selected area are determined from a computer line printer 
diaplay ,because of the readily identifiable pixels. If the 
training site boundaries indicated on the classification 
map are not adequate because of misclassification, the trai 
ning site selection may be iterated until the satisfactory 
data are obtained. The classes are given in Table 3 and Fig.3. 
No 
1 
2 
3 
4 
5 
Table 3 List o 
Classes Locations 
Sea Uchinada offshore 
Lagoon Kahoku lagoon 
Beach Uchinada seabeach 
Bared soil Race track 
and grass 
Residential Uchinada 
area town 
f land use classes 
No classes Locations 
6 Urban area Musashi quar- 
Harvested 
ter in city 
field Outskirts of town 
8 Farm Outskirts of town 
9 Coniferous Mountain on 
trees city 
10 Broad- Mountain on 
leaved trees town 
11 River,bank Kanakusare-kawa