Clustering spatial data via mixture models with dynamic weights

Abstract

This article introduces a spatial mixture model for the modeling and clustering of georeferenced data. In this model, the spatial dependence of the data is taken into account through the mixture weights, which are modeled by logistic transformations of spatial coordinates. In this way, the observations are supposed to be independent but not identically distributed, their dependence being transferred to the parameters of these logistic functions. A specific EM algorithm is used for parameter estimation via the maximum likelihood method, which incorporates a Newton-Raphson algorithm for estimating the logistic functions coefficients. The experiments, carried out on synthetic images, give encouraging results in terms of segmentation accuracy.