229 
3 
EXPERIMENTS 
Areas of landcover changes are usually a 
little in object images. If a landcover 
map of the object area exists before 
analyses of a set of multi temporal 
images, most of pixels included in each 
category on the map should correspond to 
correct training area of its category. 
Thus roughly correct training area can be 
decided by overlaying the the existing 
landcover map on a object image in a set 
of multi temporal images. 
Training data of each category correspond 
to all pixels data in its training area 
on the object image. The used existing 
landcover map and obtained training data 
is named as a reference map and initial 
training data, respectively. 
Since a few number of pixels included in 
the initial training data may be 
incorrect training data because of land- 
cover change between the reference map 
and the object image, it is necessary to 
remove these pixels from the initial 
training data. Those pixels are named as 
change pixels. Change pixels can be 
detected by searching pixels at a long 
distance from a centroid of each category 
in feature space. And classification 
classes, which change pixels should be 
assigned, can be generated by using 
clustering methods. Consequently, we may 
use the data combined generated classes 
by clustering for change pixels and 
remaining initial training data after 
removing of change pixels as final 
training data. 
As a result of above discussions, a 
following five stages procedure can be 
obtained(Fig. 1 ) . At the first stage, 
initial training data are extracted by 
overlaying a reference map on each object 
image. If a reference map can not be 
prepared, one object image in a set of 
multi temporal image data should be 
classified using a conventional method. 
And the classified result may be used as 
reference map. 
At the second stage, change pixels are 
extracted from the initial training data. 
In this study, Mahalanobis' distance is 
used to search change pixels. At the 
third stage, clustering is performed for 
change pixels and new classification 
classes are generated. At the fourth 
stage, final training data are construct 
ed by combining generated new classes and 
remaining of the initial training data. 
At the final stage, a object image is 
classified using obtained final training 
data. Maximum likelihood classifier are 
used in this study. Multi temporal land- 
cover maps can be obtained by applying 
this procedure to each image of a set of 
multi temporal images. 
In order to evaluate the proposed 
automatic classification procedure, 
following four temporal Landsat TM data 
were classified by using this method and 
a conventional method. 
[object area] 
Sagami River basin(12.8Kmxl2.0km) 
[observation date] 
Nov.4(1984), Jan.23(1985), 
Aug.6(1986), May 21(1987) 
[image size] 
512x480 pixels, 25mx25m/pixel 
[channels] 
1 , 2,3,4,5, and 7 
Classification results of conventional 
method are used to compare to classifi 
cation results of proposed method. And 
also, there are used as reference maps. 
The reason of preparing four reference 
maps is to evaluate dependences of a used 
reference map. 
Items in left hand side in Table 1 shows 
classification categories which were used 
in the conventional classification. The 
number of categories are fourteen. But 
the total number of classification 
classes are about sixty since each 
category has several cub-classes. 
Classification accuracies were estimated 
quantitatively by using digital test site 
data as shown in Figure 2. Test site data 
contain about 60 1 and-cover/use 
categories. As the categories used in 
land-cover/use test site data differ from 
those used in landcover classifications, 
fourteen landcover categories were merged 
to five major categories as shown in 
Table 1. Accuracy evaluations were 
performed based on these five major 
categories. 
Change pixels were extracted by searching 
pixel which exist at longer than 2.5 
Mahalanobis' distance from a centroid of 
each categories containing these pixels 
in the initial training data. Figure 2 
shows obtained change pixels of each 
temporal image. Change pixels had a area 
about 20%-30% for each temporal image. 
About 40 to 60 classes were generated by 
clustering for change pixels. Used 
clustering method is a hierarchical Ward 
method, namely, the number of classifi 
cation classes of final training data is 
about 100-120 classes. 
Table 2 shows estimated classification 
accuracies of classification results of 
the conventional methods and the proposed 
method for each temporal image. Mean 
accuracies shown in Table 2 were area 
weighted mean values for each category. 
The results of conventional method were 
derived by using training data which are 
extracted and modified by skilled 
operator. The skilled operator performed 
detailed ground truth before training 
area selection and during modification 
procedure of extracted training data.