Research On Sparse Unmixing Algorithms Of Hyperspectral Image

Spectral unmixing plays an important role in hyperspectral image applications.Due to the low spatial resolution of the sensor and the complexity of the feature distribution,each pixel in a hyperspectral image typically contains more than one feature object.Spectral unmixing,a great challenging task underlying many hyperspectral image applications,is designed to decompose the measured spectrum of each mixed pixel into a constituent spectrum(endmembers)and a corresponding set of fractions(abundances).As a semi-supervised unmixing strategy,sparse unmixing has received extensive attention and research.Compared with the unmixing algorithm based on geometry and statistics,sparse unmixing avoids the problem of extracting virtual endmembers without physical meaning.In this paper,the research status of hyperspectral image sparse unmixing algorithm in recent years is summarized.In response to that sparse unmixing via variable splitting augmented Lagrangian and TV(SUnSAL-TV)leads to the phenomenon that the abundance iamge of the solution is over smooth and the boundary is blurred,a structure tensor total variation(STV)re-optimization(SUnSAL-TV-STV)sparse unmixing algorithm is proposed.STV has the ability to capture the first-order information around the local neighborhood.The STV regularizer is introduced in the SUnSAL-TV unmixing model to correct the abundance matrix of the solution and improve the unmixing precision.It has been proved in experiments on synthetic data and real hyperspectral data that the proposed algorithm obtains better unmixing performance and can effectively overcome the abundance matrix over-smoothing and edge blurring.Furthermore,considering that the local collaborative sparse unmixing algorithm uses a fixed window to contain local spatial information is not rigorous,a super-pixel based local collaborative sparse unmixing algorithm is proposed.The superpixel segmentation algorithm based on the quaternion color distance theory and simple linear iterative clustering can divide the image into multiple homogeneous regions,and the pixels contained in each homogenous region obtained by the segmentation have similar spectral characteristics.Performing collaborative sparse unmixing in each homogeneous region can more accurately contain local spatial information.At the same time,considering the existence of non-local similar blocks in natural images,as an extension of the super-pixel segmentation algorithm,a non-local super-pixel segmentation algorithm is proposed to improve the simple linear iterative clustering(SLIC)algorithm.The proposed algorithm will contain more abundant spatial prior information.