Study On Key Techniques Of Hyperspectral Images Clustering Algorithm Based On Sparse Subspace Analysis

In recent years,the Hyperspectral remote sensing technology has developed rapidly,coupled with the Hyperspectral Image(referred to as HSI)can provide detailed coverage information of grpund features by itself,the researchers can be more efficient to extract the features of spectral information,the radiation energy and space information from the HSI.However,HSI contains rich spatial information and spectral information,it leads to the fact that HSI has hundreds or thousands of dimensions,which is easy to cause the dimensional disasters.Therefore,it is actually difficult to cluster it.At present,scholars can roughly divide HSI clustering into four methods:1)centroid-based method;2)density-based method;3)biological method;4)spectral-based method.The spectral-based method is more popular and has a good effect,and the Sparse Subspace Clustering(referred to as SSC)is one of the many spectral-based methods.Although SSC algorithm can process noisy data,it also has its limitations to use the algorithm for clustering directly.Based on this,the manifold regularization term is introduced into the SSC model because of the spatial feature information of HSI is unused.The spatial characteristics of HSI are better obtained with making use of the spectral information of HSI.The semi-supervised learning method"gaussian domain and harmonic function"(GFHF)is introduced for SSC algorithm without prior knowledge.Meanwhile,the class-probability initial value of unlabeled datas are obtained by using labeled datas instead of the label binary matrix in GFHF by using class probability.On the basis of SSC algorithm,this thesis makes the following improvements:l)The thesis puts forward a new algorithm,called the Laplacian regularized Sparse Subspace Clustering(LapSSC),it adds the new manifold regularization term characterized by Laplacian graph to reflect the local manifold structure.Because the manifold regularization term can capture the local geometry structure of data points.By adding a manifold regularization term to SSC model,which ensures that the global spatial structure of HSI is taken advantage of,as well as the local spatial geometry structure of HSI.The performance of the new algorithm is improved.2)The thesis uses the famous semi-supervised learning method,called gaussian fields and harmonic functions(GFHF),to transmit a small amount of supervised information to unlabeled data.However,there is a problem.Before utilizing the GFHF learning method,the method does not know the probability of unlabeled data points belonging to each class.The general method is to represent all initial labels of unlabeled data points as zero,which limits the clustering accuracy to a certain degree.So,the thesis proposed a new semi-supervised Subspace Clustering framework,called Semi-supervised Subspace Clustering for SSC via Class Probability(CPS4C).The framework uses the supervised information to make an initial class probability judgment for the unlabeled datas on which subspace they belong to by class probability propagation,which optimizes the initial datas of semi-supervised learning to a large degree and improves the clustering accuracy.Finally,for the different clustering algorithms proposed in this thesis a series of experiments were conducted on several well-known HSI data sets,namely Pavia University,Pavia Centre and Salinas.The experimental results further prove the effectiveness of the improved methods.