Hyperspectral Imagery Analysis Techniques Based On Low-rank Representation And Tensor Space

Analysis and processing of hyperspectral imagery is the core technique of the hyperspectral remote sensing.There are many problems in hyperspectral image processing such as small samples in high dimensional space,nonlinear data structure,the wide existence of mixed pixel and deficiency in utilizing spatial information in the hyperspectral imagery.These problems seriously restrict the development and application of hyperspectral imagery processing and analysis.To improve the accuracy and efficiency of hyperspectral images processing,this dissertation aims at solving the problems in hyperspectral image processing and analysis.Combining with low-rank representation,tensor space,manifold learning,the thesis focuses on hyperspectral image noise reduction,feature extraction and image fuzzy classification.The main contents and innovations in this dissertation are listed as follows:1.This thesis firstly illustrates the hyperspectral remote sensing technologies and several matters which need to be resolved in its application and summarizes the existing theories and methods in hyperspectral image analysis.The existing problems and potential of sparse kernel learning model,low-rank representation and tensor space in the application of hyperspectral image processing are analyzed.2.Low-rank representation is one of the state-of-art hyperspectral image denoising algorithms,but it suffers from ignoring the high-order relationships between data points in image.This thesis proposes a hypergraph laplacian regularized low-rank representation algorithm for noise reduction of hyperspectral images,which can represent the high-order relations between data points by using the hypergraph laplacian regularization.On the other hand,to further improve the ability to maintain the local information,the sparse and non-negative constraints have been added to the model coefficient matrix.The proposed method not only resumes the low-rank signal components,but also represents the high-order relations of the image data.The experimental results on hyperspectral images show that the proposed approach can maintain the spatial and spectral information of images better,which improves the hyperspectral image denoising results effectively.3.Aiming at the problem that current hyperspectral image tensor feature extraction methods can not make full use of the multiple spectral-spatial features of hyperspectral image,a new hyperspectral image tensor feature extraction method based on fusion of multiple spectral-spatial features is proposed in this thesis.Firstly,three-dimensional Gabor wavelets are used to get multiple texture features with different directions and different frequencies,and the multiple shape structural features are got by different morphological attribute filters.The tensor features are constructed by combining the spatial feature,multiple texture features and multiple shape structural features.Then,using the local tensor discriminant analysis,the proposed algorithm effectively increases the consistency of the same kind tensors and the difference of the different kinds tensors,which can get the lower dimensional tensors consisting of discriminating information and multiple spatial-spectral features.Experimental results of hyperspectral data sets indicate that the proposed method can maintain the spatial-spectral information and discriminating information,which has higher classification accuracy and better spatial continuity classification map than other algorithms when it is applied to the classification images.4.Though the current sparse kernel classification methods have been successfully applied in hyperspectral imagery fuzzy classification,they also have several limitations.In this thesis,a hyperspectral imagery fuzzy classification method based on the probabilistic classification vector machines is proposed.In the Bayesian framework,a signed and truncated Gaussian prior is adopted over every weight in the probabilistic classification vector machines,where the sign of prior is determined by the class label,and the EM algorithm has been adopted for the parametric inference to obtain a sparse model.This algorithm can solve the problem that the sparse kernel classification method is based on some untrustful vectors,which influences the accuracy and stability of the model.The experiments on the hyperspectral images were performed,and the results show the advantages of the hyperspectral imagery classification method based on probabilistic classification vector machines.