Hierarchical Mixture Model Based High-resolution Remote Sensing Image Segmentation Method

The features of high-resolution remote sensing images present that the heterogeneity in the same region and the homogeneity in different regions have been increased,which have resulted in significant challenges in designing image segmentation method.In statistics,the statistical distributions of spectral intensities in each region mainly appear asymmetric,heavy-tailed and multimodal characteristics.The fact that one effective way of obtaining accurate segmentation results is accurately modeling the distribution of spectral intensities is well-known.While traditional mixture models are difficultly to satisfy the requirements of modeling the complicated distributions of spectral intensities in high-resolution remote sensing images.In order to resolve this problem,this thesis proposes a hierarchical mixture model to characterize the complicated distributions and a high-resolution remote sensing image segmentation method based on hierarchical mixture model.The main researches are presented as follow.(1)The proposed hierarchical mixture model is a mixture model with the hierarchical structure,which is used to resolve the problem of modeling the complicated distribution of spectral intensities.The proposed hierarchical mixture model is defined by the weighted sums of component distributions and includes two layers.Its first layer consists of its components,which are defined by the weighted sums of element distributions to model the complicated distributions of spectral intensities in object regions.Its second layer consists of the elements,which are defined by the known probability distributions to model the statistical distribution of spectral intensities in sub-regions of object regions.The proposed hierarchical mixture model can effectively utilize the spectral information of image by accurately modeling the statistical distribution of spectral intensities.(2)Based on the proposed hierarchical mixture model,a high-resolution remote sensing image segmentation method is proposed.The proposed hierarchical mixture model is used to characterize the statistical properties of image.The prior distribution of component weight is modeled by Gaussian-Markov random field to effectively utilize the spatial location information of pixels.Following Bayesian theory,the posterior distribution of model parameters is modeled as segmentation model.Finally,Expectation Maximization/Markov Chain Monte Carlo(EM/MCMC)method is proposed to solve the segmentation model and obtain the optimal model parameters.The pixel labels can be obtained by maximizing the posterior probability as the final segmentation result.(3)To verify the proposed method,numerous experiments are carried out on high-resolution panchromatic,multispectral and synthetic aperture radar images.Results are qualitatively and quantitatively analyzed.The conclusion can be obtained that the proposed method can accurately model the complicated distribution of spectral intensities,and can accurately and effectively segment high-resolution remote sensing images.There are 59 figures,33 tables and 190 references.