Particle swarm optimization algorithm based Gaussian mixture models for remote-sensing image recognition

The model identification of the surface feature spectra is the basis of remote sensing image recognition. In this paper, Akaike´s Information Criterion (AIC) is first used to determine the number of Gaussian components in a Gaussian mixture model (GMM), then the estimation of GMM parameters is performed by using particle swarm optimization (PSO). A new parameterization procedure is applied to ensure the positive definiteness and symmetry of an arbitrary covariance matrix. The results of the GMM identification of a simulated dataset are compared between expectation-maximization (EM) and PSO algorithm based model parameter estimation methods, which showed the effectiveness of PSO based GMM. Furthermore, PSO is used to estimate parameters of the GMMs of the Statlog dataset collected by NASA, USA. The PSO-GMM based classification method is demonstrated to outperform the EM-based GMM method.