Based On Rough Set Theory And Markov Random Field Of Remote Sensing Image Classification Algorithm

With the fast development of spatial technology,digital image processing and computer technology,remote sensing technology has made great progress.The remote sensing image grows from gray level image early,and color image to multispectral and hyperspectral image,and the appearance of large dimension,strong correlation among bands,high computational time of remote data followed.Besides that,the uncertainties existed in remote data request higher qualification to classification technology.The traditional remote sensing image cluster approaches are based on image statistical model of spectrum vectors,they ignored the spatial relationship between pixels so that the classification results are influenced.Directed towards the problems above,we apply acknowledge reduct in rough set theory to the multispectral remote sensing image to reduce its dimension first,and then obtain image statistical model parameters for unsupervised classification using equivalence class partition,finally introduce the spatial context relationship of Markov random field model to improve the cluster accuracy.The main contents of this paper are as follow:1.Try to apply the rough set attribute reduct to dimension reduction of multispectral image.Because of the wide value domain of remote sensing image,the discretization is needed and the choice of cuts set affects the reducts.Attribute reduction based on information entropy keeps the compatibility of remote data decision table well but need supervised sampling.We propose a method in this paper which can obtain remote data decision table unsupervised and automatically choose discrete cuts set to perform bands reduction,and finally compared the results with common bands selection approaches.Finally multispectral and hyperspectral images are used to validate the proposed method.2.To classify the remote sensing image unsupervised using rough set equivalence relation partitioning.The initial rough classes are obtained using rough set,and then mapping the rough classes to initial parameters of image statistical model which iterated by Expectation Maximization algorithm.The rough classes merged together according to the final cluster number which determined by the automatic detective the number of cluster method this paper designed,and finally the unsupervised image classification is completed with Maximum A Posterior probability.3.Directed towards the problem of omitting relationship between image pixels of traditional image classification,this paper designs a method to classify remote sensing image unsupervised based on rough set theory and Markov random field,which can improve the cluster accuracy.The likelihood energy function of image obtained from unsupervised cluster based on rough set,and spatial context relationship between pixels which decided by Markov prior model are combined.Then to get the final cluster through simulated annealing algorithm.Finally the remote sensing images and artificial synthetic images are used to demonstrate the validity and robustness of the method.All the processed of the algorithms presented in this paper are realized by MATLAB.Experiment results show the feasibility and validity of these algorithms through comparison and analysis.