Remote Sensing Image Segmentation With Regionalized Fuzzy Clustering Based On Voronol Tessellation

With the increasing of the resolution of remote sensors,the high resolution remote sensing image can provide more detail information on the earth's surface.On the other hand,the detail information also causes a series of problems such as weaker similarity among pixel spectral measurements in the same homogeneous region,stronger similarity between different homogeneous regions and the more notable geometrical noises.All of the characteristics increase the segmentation uncertainty and mis-segmented phenomenon,which make traditional segmentation algorithms invalid in high remote sensing image segmentation.Therefore,the regionalized fuzzy clustering algorithm based on image domain geometric tessellation for high resolution remote sensing image segmentation is researched in this thesis.It uses the Voronoi polygon instead of pixel as the basic processing unit to establish the regionalized image model and fuzzy clustering objective function.The proposed regionalized fuzzy clustering segmentation framework can effectively deal with the segmentation problem of high resolution remote sensing images.The main tasks can be summarized as follows.(1)The Voronoi tessellation technology is used to completely divide the image domain into a group of Voronoi polygons.Furthermore,the regionalized spectral fuzzy non-similarity model and regionalized membership fuzzy model are established to describe the spectral consistency in the homogeneous region and heterogeneity between heterogeneous regions,respectively.In addition,the applicability of different regionalized spectral fuzzy non-similarity models for multispectral,panchromatic and SAR(Synthetic Aperture Radar)remote sensing images is analyzed,which provides a basis to select regionalized spectral fuzzy non-similarity model for different types of remote sensing images.(2)Taking regionalized Shannon entropy,KL(Kullback-Leibler)entropy and spatial interaction KL(SKL)entropy as regularization terms,the regionalized fuzzy clustering objective function is defined to deal with the problem that the traditional fuzzy clustering objective function can not effectively overcome the noises.In addition,the traditional Euclidean distance is used for defining the regionalized spectral fuzzy non-similarity model based on the above three objective functions to analyze the effects of different entropy regularization terms in the regionalized fuzzy clustering segmentation algorithm.(3)Based on the above analysis results,according to the characteristics of multispectral,panchromatic and SAR remote sensing image,Mahalanobis distance,Gaussian distribution and Gamma Mixture Model(GaMM)are used for defining the regionalized spectral fuzzy non-similarity model.Then,the regionalized fuzzy clustering segmentation is achieved by combining regionalized fuzzy spectral non-similarity model and SKL entropy regularization term.In order to solve the parameters in objective functions,derivative,Lagrange function and sampling methods are designed to obtain the optimal solutions depending on the parameter properties.(4)In order to verify the effectiveness of proposed algorithms,a series of segmentation experiments are curried out on synthetic/simulated and real remote sensing images with the proposed and comparing algorithms.The qualitative and quantitative analyses of segmentation results demonstrate that the proposed regionalized fuzzy clustering segmentation method can effectively overcome the influence of segmentation uncertainty and geometric noises caused by high resolution,and achieve effective segmentation of different types of images.The overall accuracy of segmentation result is over 90%,and the Kappa coefficient is over 0.9.There are 80 figures,13 tables and 125 references.