Applying class-based feature extraction approaches for supervised classification of hyperspectral imagery

Abstract

Global band selection or feature extraction methods have been applied to hyperspectral image classification to overcome the “curse of dimension”. We applied class-based feature extraction approaches and compressed the class data into different lower dimensional subspaces. Land cover classes in hyperspectral imagery could be roughly modelled as low-dimensional Gaussian clusters (i.e., “Gaussian pancakes”) floating in sparse hyperspace. Each pixel was labelled accordingly based on conventional classifiers. We evaluated and compared the class-based version of principal components analysis (PCA), probabilistic principal components analysis (PPCA), and probabilistic factor analysis (PFA) algorithms to find the lower dimensional class subspaces in the training stage, projected each pixel, and then assigned the class label according to the maximum likelihood decision rule. Results from simulations and the classification of a compact airborne spectrographic imager 2 (CASI 2) hyperspectral dataset were presented. The proposed class-based PCA (CPCA) algorithm provided a reasonable trade-off between classification accuracy and computational efficiency for hyperspectral image classification. It proved more efficient and provided the highest classification kappa coefficient (0.946) among all band selection and feature extraction classifiers in our study. CPCA is recommended as a useful class-based feature extraction method for classification of hyperspectral imagery. © 2007, Taylor & Francis Group, LLC. All rights reserved.