Unmixing Of Hyperspectral Images Based On Spectral Prior And Collaborative Learning

Hyperspectral image has the characteristics of high spectral resolution and image-spectrum integration,containing abundance of spatial and spectral information.However,due to the limited spatial resolution and naturally mixture of different materials,mixed pixels combined by several materials are commonly exist in hyperspectral image,which severely limits the accurate analysis and application of hyperspectral image.Therefore,hyperspectral unmixing,which aims to extract the endmember spectra and estimate their corresponding abundances from mixed pixels,has become an important basic technique for accurate analysis and processing of hyperspectral data.Semi-supervised and supervised hyperspectral unmixing methods based on spectral prior information,such as sparse unmixing methods and convolutional unmixing networks,make full use of the prior information of hyperspectral image.This kind of unmixing technology has drawn more and more attention because they can avoid the problems of traditional unmixing methods and provide more accurate unmixing results.However,traditional spectral library-aided sparse unmixing algorithms have problems of high mutual coherence and high dimensionality in the use of spectral prior knowledge,and their unmixing results are susceptible to noise.In addition,traditional hyperspectral unmixing algorithms also have the problem of insufficient utilization of spatial and spectral information in hyperspectral image.To solve these problems in unmixing of hyperspectral image,this dissertation studies and proposes a hyperspectral image unmixing algorithm based on spectral prior and collaborative learning to improve the efficient application of spectral prior knowledge,the joint mining and collaborative learning capabilities of multi-view spectral and spatial information in unmixing,thereby improving the unmixing accuracy of hyperspectral image.The main contributions of this thesis are as follows:1.The high mutual coherence of spectral libraries,along with their ever-growing dimensionality,strongly limits the performance of sparse unmixing.For this purpose,we propose a joint dictionary framework for sparse unmixing to tackle such a limitation.The proposed approach combines the learning capacity and priori information to improve the performance of sparse unmixing by incorporating the spectral library into the dictionary learning method.A multi-dictionary learning model has been developed based on cluster analysis,and it takes advantage of the collaborative effect of endmembers in the local hyperspectral image to learn several local spectral dictionaries which comprise the joint dictionary.Moreover,the proposed joint dictionary framework can act as a dictionary pruning algorithm which provides a possibility that sparse unmixing algorithms could have higher accuracy and efficiency.Our research shows that,under certain conditions,the proposed approach can recover the optimal endmembers from the spectral library.2.Hyperspectral image is characterized by its huge contiguous set of wavelengths.It is possible and needed to benefit from the “hyper”spectral information as well as the spatial information.For this purpose,we propose a new multiview data generation approach that takes full advantage of the rich spectral and spatial information in HSI,by dividing the original HSI into several spatially homogeneous regions with different band margins.Then,a new region-based multiview sparse unmixing algorithm is presented to tackle such a multiview data model.The proposed algorithm combines the multiview learning and priori information to improve the performance of sparse unmixing by incorporating the multiview information and spectral library into the dictionary learning framework.We also show that the proposed algorithm can serve as a dictionary pruning algorithm,which provides a possibility that unmixing algorithms could have higher accuracy and efficiency.The experimental results on both simulated and real hyperspectral data demonstrated the effectiveness of the proposed algorithm both visually and quantitatively.3.While most methods focus on analyzing hyperspectral image by exploring the spatial information,it is known that hyperspectral data are characterized by its rich spectral information.This information can be naturally used to improve the representation of mixed pixels.In order to take the advantage of the “hyper”spectral information as well as the spatial information for hyperspectral unmixing,we explore and introduce a multiview data processing approach through spectral partitioning to benefit from the abundant spectral information in hyperspectral image.Some important findings on the application of multiview data set in sparse unmixing are discussed.Meanwhile,we develop a new spectral–spatial-weighted multiview collaborative sparse unmixing model to tackle such a multiview data set.The proposed model uses a weighted sparse regularizer,which includes both multiview spectral and spatial weighting factors to further impose sparsity on the fractional abundances.The weights are adaptively updated associated with the abundances,and the proposed algorithm can be solved by the alternating direction method of multipliers efficiently.The experimental results demonstrated the effectiveness of the proposed algorithm,which can significantly improve the abundance estimation results.4.Recently,deep learning has become a powerful tool for hyperspectral image analysis,such as hyperspectral image unmixing,classification and super-resolution.In order to promote the effective application of spectral prior knowledge in convolutional neural networks and improve the accuracy and utilization of unmixing networks,we propose a new unmixing algorithm that uses the convolutional neural network for hyperspectral data incorporating spectral library,which can be applied for a series of hyperspectral images after training.The proposed deep spectral convolution network extracts spectral features in mixed pixels and then executes the estimating process from these extracted characteristics to acquire the fractional abundances on a fixed spectral library.Meanwhile,considering the incorporation of spectral library,the network structure is improved and a deeper convolutional network is adopted to efficiently extract features and achieve better results.Moreover,we construct a new loss function,which includes pixel reconstruction error,abundance sparsity,and abundance cross-entropy to train the aforementioned network in an end-to-end manner.The experimental results indicated the advantage of the proposed unmixing network,which can obviously enhance the abundance estimation accuracy and has a good universality.