Hyperspectral Imbalanced Datasets Classification Using Filter-Based Forest Methods

Abstract

A large variety of algorithms have been proposed for hyperspectral data classification in the literature. Among them, classical forest methods, such as bagged tree, random forest, oblique random forest, and rotation forest, have received much attention, thanks to their efficiency. Recently, two filter-based forest methods, called balanced filter-based forest (BFF) and cost-sensitive filter-based forest (CFF), have been proposed for high-dimensional data classification. These methods also have the solutions for imbalanced data problem. In this paper, these two methods were examined and compared with the classical forest methods for the classification of several well-known imbalanced hyperspectral datasets. The results indicated higher efficiency and reliability of the filter-based forests as compared to the classical forests. Moreover, they were smaller and sparser ensembles than the competitors. Furthermore, these two forests (especially BFF) were generally more successful for the classification of minority classes in hyperspectral datasets.