A Method for Automatic Clustering of Remote Sensing Data

The paper presents a method for estimating parameters of the mixture of Gaussian distributions that can be used for constructing likelihood functions at automatic clustering of remote sensing data. The idea of the method consists in successive improvement of sets of estimated parameters for the mathematical model represented by a mixture of basis functions. To improve the estimates the information on the errors occurring at approximation of the signature empiric distribution at the previous step of the iterative procedure is used. As an example the paper provides the results of automatic generation of the list of classes with their probabilistic descriptions. It also presents the results of further clustering of a satellite image upon the maximum likelihood criterion. The paper proves the efficiency of the developed method. It shows that the method can be used for describing real remote sensing data in the tasks of images recognition and cluster analysis, as it allows defining the number of components in the mixture of Gaussian functions (that corresponds to the number of observed classes of objects) and estimating parameters of those components.