GENERALIZED WISHART MIXTURES FOR UNSUPERVISED CLASSIFICATION OF POLSAR DATA

摘要

This paper presents an unsupervised clusteringrnalgorithm based upon the expectation maximizationrn(EM) algorithm for finite mixture modelling, using therncomplex wishart probability density function (PDF)rnfor the probabilities. The mixture model enables tornconsider heterogeneous thematic classes which couldrnnot be better fitted by the unimodal wishartrndistribution. In order to make it fast and robust torncalculate, we use the recently proposed generalizedrngamma distribution (GΓD) for the single polarizationrnintensity data to make the initial partition. Then we usernthe wishart probability density function for therncorresponding sample covariance matrix to calculaternthe posterior class probabilities for each pixel. Thernposterior class probabilities are used for the priorrnprobability estimates of each class and weights for allrnclass parameter updates. The proposed method isrnevaluated and compared with the wishart H-Alpha-Arnclassification. Preliminary results show that thernproposed method has better performance.