Clustering-Based Ensemble of One-Class Classifiers for Hyperspectral Image Segmentation

In this paper, we propose a new ensemble for an effective segmentation of hyperspectral images. It uses one-class classifiers as base learners. We prove, that despite the multi-class nature of hyperspectral images using one-class approach can be beneficial. One need simply to decompose a multi-class set into a number of simpler one-class tasks. One-class classifiers can handle difficulties embedded in the nature of the hyperspectral data, such as a large number of classes, class imbalance and noisy pixels. For this task, we utilise our novel ensemble, based on soft clustering of the object space. On the basis of each cluster, a weighted one-class classifier is constructed. We show a fast method for calculating weights assigned to each object, and for an automatic calculation of preferred number of clusters. We propose to build such ensemble for each of the classes and then to reconstruct the original multi-class hyperspectral image using Error-Correcting Output Codes. Experimental analysis, carried on a set of benchmark data and backed-up with an extensive statistical analysis, proves that our one-class ensemble is an efficient tool for handling hyperspectral images and outperforms several state-of-the-art binary and multi-class classifiers.