Discriminative archetypal self-taught learning for multispectral landcover classification

Self-taught learning (STL) has become a promising paradigm to exploit unlabeled data for classification. The most commonly used approach to self-taught learning is sparse representation, in which it is assumed that each sample can be represented by a weighted linear combination of elements of a unlabeled dictionary. This paper proposes discriminative archetypal self-taught learning for the application of landcover classification, in which unlabeled discriminative archetypal samples are selected to build a powerful dictionary. Our main contribution is to present an approach which utilizes reversible jump Markov chain Monte Carlo method to jointly determine the best set of archetypes and the number of elements to build the dictionary. Experiments are conducted using synthetic data, a multi-spectral Landsat 7 image of a study area in the Ukraine and the Zurich benchmark data set comprising 20 multispectral Quickbird images. Our results confirm that the proposed approach can learn discriminative features for classification and show better classification results compared to self-taught learning with the original feature representation and compared to randomly initialized archetypal dictionaries.