Soft tree building in a binary hierarchical classifier framework

Access full-text files




Gupta, Ankur P.

Journal Title

Journal ISSN

Volume Title



Classification problems in remote sensing are often difficult because of high dimensionality of the input space, large number of class labels, overlapping class distributions, and limited quantity of training data. Various modular learning techniques have been developed that decompose the classification problem into simpler sub-problems in either the input space or the output space or both. The Binary Hierarchical Classifier (BHC) is a hierarchical modular learning framework that decomposes a ̀C' class problem in the output space. The BHC has the capability of automatically revealing the inherent taxonomy of classes according to the class separation determined by the linear discriminant. In this thesis, a new approach called the Binary Hierarchical Classifier-Soft Tree Build (BHC-STB) has been developed that provides the capability for each class to softly associate with more than one node. The algorithm seeks to partition the given set of classes into two overlapping sets in such a way that classes that are similar to each other, but dissimilar from the classes in the other set are assigned to one of the two sets, while classes that are similar to the members of both these groups (or dissimilar from the members of both the groups) are assigned to both sets with some probability. The process is recursively repeated until each set contains only one class. The soft tree build framework enhances the BHC framework for problems with mixed classes that share characteristics with both the child nodes or distinct classes that do not associate easily with one or the other child node. The proposed method shows improvement in classification accuracy for difficult problems with mixed classes and also reveals the enhanced domain knowledge compared to the BHC.


LCSH Subject Headings