Joint Representation With Spatial Constraints For Hyperspectral Target Detection

As an electromagnetic wave imaging technology,hyperspectral remote sensing can realize fine observation of ground objects in a narrow and continuous wave band,which provides a condition for the detection of hyperspectral targets.The spectral features are the inherent characteristics of different chemical compositions.Using the spectral information can effectively identify and distinguish the materials,which provides the conditions for the detection of hyperspectral targets.Hyperspectral image target detection has high research value and has practical application significance in many fields such as military target reconnaissance,maritime search and rescue,border detection,and mineral resources positioning.Due to the limitations of target size and resolution of hyperspectral sensors,the target of interest is often in the sub-pixel level or weak information state in hyperspectral images,which makes the traditional spatial-based target detection algorithm unable to achieve such goals.The detection of hyperspectral remote sensing targets is mainly performed through the spectral differences between the target and other background features.Therefore,it is particularly important to develop target detection techniques suitable for hyperspectral images.This article focuses on the characteristics of "spectrum integration" of hyperspectral data and studies how to effectively combine hyperspectral image spectral information and spatial dimension information in the framework of sparse representation to achieve hyperspectral image target detection.The thesis begins with the sparsity of hyperspectral image high-dimensional data,and reviews the basic model and solution method for sparse representation of hyperspectral image target detection.The hyperspectral image sparse representation target detection algorithm for joint spatial and spectral information is studied.This thesis describe a neighborhood joint multi-task sparse representation model,and joint tasks are screened according to spectral similarity to avoid false associations of dissimilar detection tasks.And this thesis also describes a pre-calculation method in binary hypothesis to avoid repeat calculation for spectral similarity.In addition,the collaborative representation algorithm in the anomaly detection is compared.The similarity between the collaborative representation theory and the bilateral filtering idea is discovered.The bilateral filtering is used to combine the spatial dimension and spectral dimension information of the hyperspectral image as a Tikhonov regular term for the collaborative representation model..Under the binary hypothesis,using the regularized constraint-based cooperative representation instead of the neighborhood joint multi-task sparse representation to model the background pixel,and finally,a spatial constrained joint sparse representation and cooperation is obtained.The paper uses simulation data and real hyperspectral remote sensing data to test the proposed algorithm.The experimental results are basically consistent with the theoretical analysis,and shows that the proposed method can effectively improve the accuracy of hyperspectral image target detection.AUC values reach more than 95% on all data sets used in the experiment,and all can achieve 100% detection rate at a lower false alarm rate.