Dimensionality Reduction Of Hyperspectral Data Based On Sparse Graph

With the development of imaging spectroscopy,the spectral resolution is improved unremit-tingly,by which more refined spectral features of landcovers can be obtained and the subdivi-sion can be realized.However,a series of problems are brought by high spectral dimension,such as higher computational cost,the effect of redundant information on model's stability and higher requirement of training data.Dimensionality reduction is the most direct tool for these problems.Sparse graph is learned by sparse representation,which has some advan-tages than6)-nearest-neighbor graph and-nearest-neighbor graph such as noise robustness,sparsity and adaptivity.Because of the nonlinearity of hyperspectral data,the preservation of local information in the low dimensional space is beneficial for classification.In this thesis,two points are focused,such as?I?sparse graph learning,by which a more discriminative sparse graph is learned,?II?consideration of the hyperspectral data's characters during de-signing the dimensionality reduction methods.From above two points,the following five dimensionality reduction methods are proposed:?1?Double sparse graphs based semi-supervised dimensionality reduction is proposed to deal with the limitation of training samples in hyperspectral image.Firstly,Aiming to ex-plore the discriminant information among unlabeled samples,joint k nearest neighbor se-lection strategy is proposed to select pseudo-labeled samples.In the following procedures,double sparse graphs are constructed by sparse representation,which contain the positive relationship and the negative relationship of data.Based on two different criterions,two double sparse graphs based semi-supervised discriminant analysis algorithms are designed,which also use different strategies to reduce the effect of pseudo-labels'inaccuracy.Fi-nally,the experimental results both on UCI datasets and hyperspectral images validate the effectiveness and advantage of the proposed methods compared with some classical dimen-sionality reduction methods.?2?An unsupervised sparse graph learning based dimensionality reduction method is pro-posed for hyperspectral image.In the proposed method,sparse graph construction and pro-jection learning are combined together in a unified framework and influence each other.During sparse graph learning,projected features are utilized to enhance the discriminant information in sparse graph.Likewise,in projection learning,the enhanced sparse graph could make projected features have high discriminant capacity.Besides,the spatial-spectral information in the original space combined with the structure information in the projected space is also exploited to learn the imprecise discriminant information.With the imprecise discriminant information,the projected space would contain abundant discriminant infor-mation,which is beneficial for hyperspectral image classification.Experimental results over two hyperspectral image datasets demonstrate that the proposed approach outperforms the other state-of-the-art approaches.?3?Based on the idea of SPP,an improved SPP is proposed in this paper for image classi-fication,in which an adaptive learning strategy is utilized to adjust the sparsity parameter of each sample in the sparse graph learning procedure.In addition,a new measurement is introduced into projection learning to control the effect of each sample on projection learn-ing.Adaptive sparse graph learning is an iteration processing,the projected information also is used in sparse graph learning as a guide.Experimental results on three datasets demon-strate that the proposed approach can achieve better classification performance over some available state-of-the-art approaches.?4?Multi-view graphs ensemble based graph embedding is proposed to promote the perfor-mance of graph embedding for hyperspectral image classification.By integrating multi-view graphs,more affluent and more accurate structure information can be utilized in graph em-bedding to achieve better results than traditional graph embedding methods.In addition,the multi-view graphs ensemble based graph embedding can be treated as a framework to be ex-tended to different graph based methods.Experimental results demonstrate that the proposed method can improve the performance of traditional graph embedding methods significantly.?5?Considering the characteristics of hyperspectral image,spatial-spectral regularized s-parse graph is proposed with enhanced discriminability.Based on the idea of sparsity p-reserving,spatial-spectral regularized sparse graph is applied to band selection for hyper-spectral classification,which is called spatial-spectral regularized sparse graph based band selection method.Besides,_2,1norm constraint is utilized to restrain the mapping from original space to the new feature space,by which just few bands are selected to represent new features.Therefore,instead of measuring the sparsity preserving ability of each bands directly,the importance score is defined by the total contribution of each band in the new features,according to which more important bands are selected.Two real hyperspectral images are used to validate the performance of the proposed method.