Combining Weighted Mixture Model and Markov Random Field for Optical Remote Sensing Image Segmentation

Aiming at the problems of inaccurately building the statistic model and low accuracy and efficiency in remote sensing image segmentation, a remote sensing image segmentation algorithm combining weighted Gaussian mixture model and Markov random field is proposed in this paper. Firstly, considering the complicated distributions of spectral intensities in remote sensing image, the weighted Gaussian distributions is used as the component of mixture model, and weighted Gaussian mixture model is used to build the statistical model of image for overcoming the inaccurate modeling of traditional mixture model using a single probability distribution as its component. Secondly, the prior of component weight is built with the attribute probability of neighborhood pixels by Gibbs distribution. Its structure is simple and easy to implement. Finally, model parameters can be estimated by expectation maximization method. The structure of component weight distribution is conducive to derive its closed-form and reduce the amount of calculation. Cartosat and Worldview images are selected for experiments, and compared with the segmentation algorithms based on fuzzy C-means, Gaussian mixture model and student's-t mixture model. The results show that the proposed algorithm can more accurately model the complex distribution of remote sensing images, and the average segmentation accuracy is 16.44%, 16.00% and 16.17% higher than the comparison algorithm, respectively.