Information fusion in tree classifiers

The classification performance of a decision-tree classifier is strongly influenced by the classification strategy employed at each decision node and the presence of noise. In this study, multispectral data classification was attempted with maximum likelihood and backpropagation neural networks, as well as their combination. Three methods of fusing their information were studied in detail. The motivation for the information fusion was to enhance the interpretation of a particular pixel under study with the classifier that has a minimum uncertainty in assigning the pixel to one of desired classes. The classification performance with the fusion methods was found to be better than that of the individual classifiers. In addition, their recognizing ability in the presence of additive noise was found to be better, especially with the fusion using the fuzzy integral.