Research On Hyperspectral Unmixing Algorithm Based On Sparse Optimization Methods

As a novel method for earth observation,hyperspectral imaging has gained in-creasing amount of attention with the development of remote sensing techniques and advanced imaging spectrometers.Hyperspectral imagery can record reflectance signa-tures of ground objects in a wide range of spectral domains with masses of narrow bands,which offers data support for object recognition and classification precisely.However,the imaging equipment is not able to balance the spatial resolution as well as the spectral resolution of the imagery,and masses of data impose pressure on data transferring heavily.Consequently,the imagery obtained has a low spatial resolution,and the spectrum of a pixel has high possibility to be a mixture of signatures from different substances,which brings difficulties for multiple tasks,including pixel fine-classification,statistical analysis and subtle target identification of image data.Hyperspectral unmixing can decompose the mixed spectrum of a pixel into endmembers and abundances by modeling the mixing process,thus it is possible to implement quantitative spectrum analysis and complicated relevant applications subsequently.Therefore,spectral unmixing technology is the key factor to achieve the quantitative study and promote a wide range of applications for hyperspectral imagery in remote sensing.The dissertation mainly focuses on hyperspectral imaging process,and it explores effective methods of semi-supervised and unsupervised endmember extraction and abun-dance estimation.The main problems for hyperspectral unmixing and image data re-construction are discussed,and the introduction of auxiliary information which includes spatial correlation and sparsity constraints is explored in different models to improve unmixing performance,and a series of sparse unmixing methods for hyperspectral data are proposed and investigated on different occasions.Experiments have been designed elaborately for the investigation of parameters setting,performance evaluation and ro-bustness to the noise respectively,and several state-of-art unmixing algorithms,including traditional sparse unmixing methods,the multiobjective optimization method and the network-based models,are selected using synthetic data generated from public spectral library and real-world hyperspectral data for comparison.The specific research content of this dissertation includes the following parts.As the unmixing performances of traditional sparse unmixing methods are limited,here a hypergraph regularized unmixing algorithm with reweighted?₁-norm minimization,named Re HGSU,is proposed.By constructing the hypergraph structure in hyperspectral image data,pixels with spectral similarity in the neighborhood can be enveloped in a hyper edge to represent the complex manifold relationship,and then the hypergraph is incorporated as spatial information in the model to improve unmixing performance.Additionally,the reweighted?₁norm minimization strategy is introduced in the unmixing framework to obtain a more sparse representation of the solution.Experimental results demonstrate the effectiveness and robustness of the method.Performances of regularization methods rely on the weighting parameters for vari-ous regularizers.Here a multiobjective sparse unmixing framework,named MOSU-SS,is proposed which incorporates semi-supervised sparse unmixing with multiobjective op-timization.In order to introduce spatial information,distance-based m SLIC algorithm is applied for hyperspectral image segmentation.Furthermore,a metric that measures local spatial uniformity is adopted in the framework,and the unmixing process on the whole image is divided into parallel processing on superpixels.By generating and evolving the population,the candidate sets are generated and combined through genetic algorithm.So-lutions with high adaptivity are survived,and a unique solution representing endmember selection with minimum objective conflicts is selected by the decision maker.Endmem-bers and abundances can be obtained subsequently.Experiments on hyperspectral image data reveal that the multiobjective unmixing framework is able to find endmembers and obtain the accurate abundance distributions under different noise disturbances,which shows its superiority among comparative methods.Deep networks can learn high order features from inputs with high dimensions in a semi-supervised or unsupervised manner,in the dissertation,effective sparse constraints are introduced in the model to enhance the representation of the network.Here multi-layer convolutional networks are constructed to investigate the performances for feature extrac-tion and spectral unmixing.First a 2-dimensional convolutional network is constructed on hyperspectral data.Spectral feature image consisting of signatures in the neighborhood of a pixel is fed into the network model as the input,and the network is trained with the inputs and corresponding labels.In addition,sparsity constraint is introduced in the optimization to improve the generalization of the network model.On the other hand,a denoising 3-dimensional convolutional autoencoder is proposed for the unmixing task.Hierarchical joint features in spatial and spectral domains are extracted for image recovery.?_2,1norm regularization is introduced in the optimization to gain better unmixing results,and endmembers and abundances can be obtained simultaneously.Experiments show that the convolutional networks exhibit good performances for both feature extraction and spectral unmixing on hyperspectral data.This dissertation aims at building unmixing models for the highly mixed hyperspec-tral imagery with sparsity constraints.The decomposition of mixed pixels in HSIs is investigated intensively in this dissertation,and several effective semi-supervised and un-supervised unmixing models are developed with affluent spatial-spectral information.The proposed Re HGSU method can acquire accurate abundance information by constructing a hypergraph structure in the data,which is suitable for unmixing problems where spatial information has not been fully exploited.MOSU-SS method is able to extract endmem-bers from the redundant library directly by image segmentation,and it can be applied for HSIs with spatially consistent features.Deep 3D CAE unmixing network,on the other hand,tries to learn endmembers and abundances by the cascaded autoencoder models,which can be applied for unmixing mixed pixels under high noise disturbance.The main work has important theory significance and research value for quantitative evaluation of surface feature components,improvement of hyperspectral data utilization,and facilita-tion of hyperspectral remote sensing techniques on precision agriculture,urban remote sensing,mineral exploration and many other fields.