抄録
Characteristics of the category classification methods for remotely sensed multispectral image data has been mainly evaluated by correct rates of classification, which are given with diagonal term of the confusion matrix. This evaluation method cannot sufficiently considered the information, which off diagonal term of this matrix gives, and the evaluation is in aspect like. In this paper, the characteristics and the tendency of the classification methods are asserted in such a way that it should be analyzed using the whole of this matrix.
Next, an evaluation method using dendrogram is presented as an example. The merit of proposed evaluation method is able to evaluate the relationship among the hierarchies of categories ; here each classification method has been defined. Similarly, recognition and resemblance tendency of categories of a human sense can be also analyzed. It is meant that if the result of the trend analysis among classification methods is compared, it cannot to prove only the classification accuracy but evaluate in the relation with a human sense for various categories.
This example points out the existence of various problems to be discussed in future, that is, the relations between the utilization purpose of a user and statistical category classification method based on spectral characteristics. The evaluation method proposed in this paper is the influential gambit method for the sake. Moreover, this evaluation method can be utilized for the selection of classification method with the classification characteristics in proportion to the change of the hope, if the classification tendency, which the user hopes, changes.
