sat MSS data 
d by the pro- 
e similar fea- 
'e in the near 
. In this case, 
'as appeared. 
mage Is near 
field, so clas- 
1 not be exe- 
od nodes and 
rameter p be- 
increases but 
ion accuracy. 
performance. 
‘y, by exclud- 
)de. Present 
sample num- 
' distribution 
tegories with 
ned 
ACY 
ch 
5.3 Summary of the results 
Tree structures which support the assumption and 
sufficient classification accuracy were obtained in the 
tests. Proportion of sizes of determined terminal node 
and undetermined terminal nodes was dependent on 
classification accuracy about training samples and 
specified tolerance parameter p. 
Proposed triplet tree classifier was demonstrated 
that it not only has advantages of general tree classi- 
fiers, but also enables to treat uncertainty of data. 
Samples were effectively classified by the decision 
tree. Moreover, both relations between categories and 
the uncertainty were shown in the hierarchical tree 
structure. 
6. CONCLUSION 
Undetermined nodes are introduced as an extension 
of binary decision trees. Data with indistinct feature 
at binary division are classified to the undetermined 
node at each node division. 
Proposed triplet tree classifiers can be widely 
adapted even when adjacent pair of category exists 
or representabilities of training samples is relatively 
poor. 
Proposed triplet tree was shown to enable classi- 
fication with flexible and effective boundaries. Pro- 
posed design method ensures classification reliability 
in determined nodes about training samples. Classi- 
fication accuracy for determined nodes is almost the 
same to the Bayesian classifiers. 
It was also demonstrated that the hierarchical 
structure can represent the relations of categories and 
uncertainly-classified data part. 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
References 
[1] Wang Ru-Ye.1986a. “An approach to tree- 
classifier design based on a splitting algorithm,” 
Int. J. Remote Sensing, vol.7,89-104. 
[2] S. Tanaka,1992. “A Comparison and Rating 
of Conditioned Bayesian Discriminant Classi- 
fiers by Quantitative Term of Training Repre- 
sentability,” J. Remote Sensing Society of Japan, 
v0l.12,3-21(In Japanese). 
[3] Wang Ru-Ye,1986b. “An approach to tree- 
classifier design based on a splitting algorithm," 
Int. J. Remote Sensing, vol.7,89-104. 
[4] B. Kim and D. A. Landgrebe,1991. “Hierarchi- 
cal Classifier Design in High-Dimensional, Nu- 
merous Class Cases," IEEE Trans. on Geoscience 
and Remote Sensing,vol.29,518-529. 
[5] R. Chin, P. Beaudet, and P. Argentiero.,1980. 
*An Automated Approach to the Design of of 
Decision Tree Classifiers," Proceedings of the 5th 
International Conference in Pattern Recognition, 
660-665. 
993