ISPRS, Vol.34, Part 2W2, "Dynamic and Multi-Dimensional GIS”, Bangkok, May 23-25, 2001 
139 
and with different precision. 
The difference between the training samples and the export 
from the net produces error, which can be gradually decreased 
by means of repeating iteration and finally calculating the 
solution in the training process. The error indicates the coming 
classification accuracy of a NN classifier at least indirectly. 
Obviously, the other accuracy indexes can be easily calculated 
after every time of iteration. Besides, the accuracy of the net that 
is underway or has been trained can be evaluated through 
another group of independent checking samples. They may be 
fuzzy and the evaluation of accuracy may be nearly real-time. 
From the above, a fully fuzzy supervised process may be 
realized by means of the NN method. It can theoretically 
describe any nonlinear relation and have the capability of 
learning and generalization. It can also get used to the dynamic 
change of the learning environment by renewing to learn. 
Therefore, it is a good choice to realizing the fully fuzzy 
classification by means of the NN method. 
TEST 
In order to prove the property of the study, we take a small 
scene of Landsat TM image as the test data. Its pixel size is 30 
meters. It is located in the Delta of the Yellow River in China. It is 
suitable for the test field of fuzzy classification for its land cover 
types interpolating mutually. According to the practical conditions, 
the system of image classification in the test is classified into 
seven types, namely arbor tree, bush, grassland, reed, dry land, 
paddy field and water. After visual interpretation, The bands 3, 4 
and 5 are selected as the original data for practical classification 
in order to simplify the latter management. The rectification of 
the image is performed according to the coordinates of a group 
of easily identified ground features from a map, such as the 
crossing of roads, corners of huge building structure. 
Table 1. The Classification results by Non-fuzzy (N) and Fully Fuzzy (F) Methods 
r ' ' 1 11 1 IU ' , - 1 - 1 i,J - ■ 
Training 
Classification 
Accuracy evaluation 
Classification 
method 
Mode 
Pixel 
number 
Iteration 
number 
Pixel 
number 
mode 
Pixel 
number 
Classification 
accuracy(%) 
Cross 
-entropy ! 
Maximum probability 
N 
416 
1 
1566 
N 
1566 
38 
- 
: j 
NN 
N 
143 
5567 
1709 
N 
1566 
51 
: 
Fully fuzzy NN 
F 
80 
5186 
1709 
F 
1629/96 
45.6/57.3 
0.456 
(Note: The data under the diagonal is acquired from the crispy pixels) 
For the purpose of comparison, the routine maximum 
probability method is taken for image classification at first. The 
appropriately distributing and representative pixels are used for 
its training. The sample pixel numbers of arbor tree, bush, 
grassland, reed, dry land, paddy field and water are respectively 
55, 60, 53, 63, 45,58,56 and 26. The result of classification is 
listed in table 1. 
After repeating experiments, the NN classifier designed by 
this paper is a net structure including 3 nodes in import layer, 10 
nodes in hide layer and 7 nodes in export layer. Its selected step 
length of learning is 0.1 and its momentum parameter is 0.1. 
Using the traditional method, the NN training is performed by 
means of equably distributing and more representative crispy 
pixels as samples. They include 25 arbor tree classes, 30 bush 
classes, 15 grassland classes, 22 reed classes, 18 dry land 
classes, 28 paddy field classes and 5 water classes. Next, the 
classification is performed by the fully fuzzy NN method. The 
training of the net is based on a fuzzy sample having 80 pixels 
totally. The sample pixel numbers of arbor tree, bush, grassland, 
reed, dry land, paddy field and water are respectively 15, 15, 13, 
13, 12,10 and 2. They are all selected from the fuzzy reference 
data layer by computer at random, which have different affiliation 
degrees between 0 and 1. The fuzzy reference data layer can be 
acquired from the aerial photo covering the same area by means 
of geostatistical method. Except for the training samples, all the 
pixels are used to evaluate the classification accuracy. From the 
table 1, it can be seen that the training samples of the maximum 
probability method are biggest and its accuracy is lowest. 
Generally speaking, the accuracy of the NN classification is 
higher than the accuracy of the fully fuzzy NN classification. But 
the former one need more training samples, which are 
comparably crispy. It is a over-strict request in the fuzzy 
phenomenon. As a whole, the result of the fully fuzzy NN 
classification is best. 
CONCLUSIONS 
From the table 1, we can see that the whole accuracy of 
classification is not good enough. It is the fuzziness of the land