Semi-supervised co-training and active learning framework for hyperspectral image classification

Hyperspectral imaging enables detailed ground cover classification with hundreds of spectral bands at each pixel. Rich spectral information can be a drawback since supervised classification of a hyperspectral image requires a balance between the number of training samples and its dimension. Achieving this balance requires a large number of training or ground truth samples, which is generally difficult, expensive and time-consuming. This led researchers to explore the use of semi-supervised learning techniques where new training samples (unlabeled) are obtained from a small set of available labeled samples without significant effort. In this paper, we propose a semi-supervised approach which adapts active learning to a co-training framework in which the algorithm automatically selects new training samples from abundant unlabeled pixels. Efficacy of the proposed approach is validated using a probabilistic support vector machine classifier. Our experimental results with an Indian Pines hyperspectral image collected by the National Aeronautics and Space Administration Jet Propulsion Laboratory's Airborne Visible-Infrared Imaging Spectrometer indicate that the use of this co-training based approach represents promising strategy in the context of hyperspectral image classification.