Estabilidade de Classificadores de Decisão em Árvore Binária para Dados Imagem em Alta Dimensão
This paper deals with the problem of classifying high-dimensional image data image data using a multiple stage classifier structured as a binary tree. The aim here consists in finding the optimal structure for the binary tree in the sense of achieving a stable accuracy. The advantage presented by a multiple stage classifier lies on the fact that only a sub-set of classes is considered at each stage, allowing a better selection of the features to be used at each node. The binary tree is a particular case of a tree structured classifier, on which only two classes are considered at each node. This peculiarity makes possible the direct use of statistical distances for feature reduction (selection or extraction). In this study the criterion used for feature reduction at each node consists in optimizing the Bhattacharyya distance separating both classes in the node. The optimization of Bhattacharyya distance was based on the covariance matrices. Once the final set of features is obtained at each particular node, the classification is performed using the Gaussian Maximum Likelihood decision rule. Tests were performed using high-dimensional image data collected by the sensor system AVIRIS covering a test area. The criteria to evaluate the performance of the classifiers are: the final accuracy yielded by the classifier, its stability, and the required computer time.