Automatic selection of informative samples for SVM-based classification of hyperspectral data using limited training sets

In this work, we focus on how to select the most highly informative samples for effectively training support vector machine (SVM) classifiers in remotely sensed hyperspectral data classification. This issue is investigated by comparing different unsupervised algorithms which account for the spectral purity of training samples in the process of selecting those samples for classification purposes. Sample sets obtained using these algorithms are used to train an SVM architecture implemented using different kernels, with the ultimate goal of exploring the suitability of the aforementioned algorithms to reduce the number of training samples required by these architectures in the context of hyperspectral image classification. Experimental results are provided using the full version of a hyperspectral data set collected by NASA's Airborne Visible Infra-Red Imaging Spectrometer (AVIRIS) over the Indian Pines region in Northwestern Indiana.