Semi-supervised hyperspectral image segmentation

This paper presents a new semi-supervised segmentation algorithm, suited to high dimensional data, of which hyperspectral images are an example. The algorithm implements two main steps: (a) semisupervised learning, used to infer the class distributions, followed by (b) segmentation, by inferring the labels from a posterior density built on the learned class distributions and on a Markov random field. The class distributions are modeled with a multinomial logistic regression, where the regressors are learned using both labeled and, through a graph-based technique, unlabeled samples. The prior on the labels is a multi-level logistic model. The maximum a posterior segmentation is computed by the α-Expansion min-cut based integer optimization algorithm. We give experimental evidence that the spatial prior greatly improves the segmentation performance, with respect to that of a semi-supervised classifier. The effectiveness of the proposed method is demonstrated with simulated and real data.