aly 
ith 
out 
ever, 
ica- 
on. 
ica- 
(14 
eg 
ein 
en 
the 
er 
.vided 
Lied 
1Sing 
in the 
pixel 
incids 
    
  
    
mean and standard deviations of LANDSAT data acquired at 1979 to 
those acquired at 1972. The reson of this normalization is to 
remove ascertain changes such as detector responce differences and 
different atmospheric effects between the two images.  Lastry, a 
smoothing was applied to both images using 3x3 pixels window in | 
order to decrease misregistration effects between two images. Eg 
B. CHANGE DETECTION 
There are several algorithnsl'? for change detection such as 
subtraction, ratioing and so on. In this study, subtraction(eq.1) 
was used, because the purpose of this change detection is only to 
eliminate change regions for training area selection. 
= -— * | 
D : change region=1, no=0 fin 
U : logical sum operator Hi 
x..: pixel value of MSS band "j" 
MITIS UNO 26.1972) 
i=2 (Dec. 14,1979) i 
T : threshold value for MSS band "j" | 
I 
Results of change detection is shown in Figure 3. 
  
IV. NON-PARAMETRIC LAND USE CLASSIFI CATION 
WITH AERIAL PHOTOGRAPHS 
A. PREPROCESSING 
  
The three components (G,R,IR) of the infrared color photographs 
were digitized throuth rotating drum digitizer with the resolution 
of 100x100m“ which corresponds about 8x8m“ on the ground. A 
shading correction was first performed to the digitized data uging 
a low-pass filter". Next, the image was rectified with 10x10m Il 
ground resolutions corresponding to the grid cell of the NNLI Il 
using nearest neighbor resampling. | 
  
B. CLASSIFICATION 
  
Training areas for land use categories were selected from I! 
no-changed regions in NNLI. As feature vectors for classifica- | 
tion, spectral data(G,R,IR) extracted from the training areas were iii 
first examined. 
  
The separability (Mahalanobis' distance) among these training 
classes was insufficient for classification. The major reasons of il 
this bad separability are supposed as follows ; In 
1) Land use(not land cover) differences does not exactly | 
correspond to spectral feature differences. 
2) Training data do not exactly fit to a Gaussian distribution. 
(An example of crop fields is shown in Figure 1.) 
3) Small areas less than 10 pixels were usually neglected in the 
manual interpretation. 
In order to obtain sufficient classification accuracy, 
following three methods were employed. The first method is the 
utilization of texture features. Three parameters according to Î 
eq.(2),(3) and (4) were calculated with 3x3 pixels window for 
  
  
679 | 
  
st : 
tee Se OR ROI CERE D ND LUE UD M ES CRUEL Rs 8€ 
RU.