Contextual Unmixing of Geospatial Data based on Gaussian Mixture Models and Markov Random Fields

In supervised and unsupervised image classification, it is known that contextual classification methods based on Markov random fields (MRF) improve non-contextual classifiers successfully. In this paper, we consider unsupervised unmixing problem by introduction of a new MRF. First, spectral vectors observed at mixels are assumed to follow Gaussian mixtures. Second, vectors representing fractions of categories are supposed to follow MRF over the observed area. Then, we derive an unsupervised unmixing method, which is also useful for unsupervised classification. The proposed method was evaluated through a synthetic data set and a benchmark data set for classification, and it showed an excellent performance.