Automatic clustering of multispectral data using a non-Gaussian statistical model

This paper proposes an unsupervised clustering algorithm for multispectral images, which automatically determines the number of statistically distinct clusters in the image. It uses the multivariate student-t distribution as a more flexible underlying statistical model, with the Gaussian as only a special case. The algorithm shows better data modeling flexibility than the Gaussian case. Excellent and reproducible clustering results are observed for both simulated data and real data from Worldview-2 multispectral sensor.