Integration of association-rule and decision tree for high resolution image classification

Association rule is one of the most important rules in nature. Each type of object in a remotely sensed image relates to special association rules, thus association rules are important features for image classification, and the mining and rational selection of the effective rules is the key issues for accurate classification. In this paper, an approach that integrates association rules analysis and decision tree is presented and applied to object-oriented high resolution image classification. The association rules analysis is adopted for mining strong rules from an image, and the decision tree is for finding the optimal rules for classification. A Geoeye-1 image is used for experimental data. Firstly, the Geoeye-1 image is segmented, then spatial, spectral, textural, color space and band ration features are selected. The association rules in a training set are mined, and a decision tree is designed with consideration of confidence, support of mined rules, as well spectral complexity and the generation sequence of rules. The visual comparison with the results of K-nearest neighbors and accuracy estimation validate the effect of the proposed approach.