Research On Spectral-spatial Characteristics-based Anomaly Target Detection Algorithms For Hyperspectral Images

Target detection based on hyperspectral image(HSI)is one of the major research topics in the field of image processing and interpretation.Target detection can generally be divided into two categories: known and unknown with target spectra.Anomaly detection(AD)is an unsupervised case and no prior knowledge about the target spectral is required.In most cases,it is difficult to obtain the spectrum of ground surface objects in HSI,so AD generally has more extensive applications.The spatial and spectral resolutions have been increased with the recent development of hyperspectral remote sensing technology.In the meanwhile,new challenges are brought to hypersepectral AD.The major issues are:(1)High spectral dimensionality and data redundancy;(2)Complex nonlinear characteristics;(3)Anomaly targets with low probability and small sizes;(4)“Same spectral information,but different class”,“different spectral information,but same class”.In the issues,issue(4)will result in low detection accuracy since the AD algorithms are based on spectral information.Additionally,the spectral redundancy will bring about “Hughes phenomenon”.Therefore,the major focus of this dissertation is to investigate how to fully exploit the spectral and spatial information and how to develop efficient AD algorithms which will improve the detection accuracy.The major aspects of this dissertation are as follows:1.High dimensionality and nonlinear separability are two important properties of HSI.All kinds of hyperspectral AD based on sparse theory do not need to assume the background model distribution,and have achieved good detection results.Kernel collaborative representation-based detector(KCRD)maps HSI to kernel space and combines sparsity theory to further improve the detection result.However,these algorithms only consider the spectral sparsity of hyperspectral anomaly targets,and ignore the spatial sparsity,which results in low detection accuracy.To overcome this drawback,this dissertation proposes an algorithm using sparsity divergence index based on locally linear embedding(SDI-LLE)for hyperspectral AD.SDI-LLE uses the spectral and spatial sparsity of the hyperspectral anomaly targets and the better detection result than KCRD is obtained.Then,in order to reduce the effect of the anomaly targets in the background dictionary on detection results,an algorithm using spectral-spatial background joint sparse representation based on linear localtangent space alignment(LLTSA-SSBJSR)is proposed.Spectral BJSR is employed on the original hyperspectral data and spatial BJSR is employed on the low-dimensional data obtained by LLTSA.The final result is obtained by spectral-spatial BJSR.LLTSA-SSBJSR uses the residual energy ratio to suppress the anomalies in the background and avoids the high time complexity in the kernel learning algorithms.Synthetic and real hyperspectral datasets are employed to verify the effectiveness and superiority of the proposed algorithms.2.The hyperspetral AD algorithms based on sparse theory mainly focus on the sparse characteristics of HSI and ignore the background information.The AD algorithms based on low-rank and sparse matrix decomposition(LRaSMD)mainly divide HSI into low-rank background and sparse anomaly,and study the spectral information of HSI,but neglect the spatial information,which leads to the lack of high accuracy.To tackle this problem,an algorithm using spectral-spatial low-rank and sparse matrix decomposition(LS-SS)is first proposed.In LS-SS,spatial property is obtained using the low-rank matrix with spatial kernel collaborative representation and spectral property is obtained using the sparse matrix with sparsity divergence index(SDI).The final results are obtained by combining the two parts.Then,with the theory of deep learning,an algorithm using spectral-spatial stacked autoencoders based on low-rank and sparse matrix decomposition(LRaSMD-SSSAE)is proposed.In LRaSMD-SSSAE,stacked autoencoders are employed on low-rank matrix and sparse matrix for deep spatial features and deep spectral features,respectively,and the local Mahalanobis-distance detector based on spectral-spatial features is employed for the final result.Synthetic and real hyperspectral datasets are employed to verify the effectiveness and superiority of the proposed algorithms.3.The AD based on LRaSMD mainly decomposes HSI in spectral dimension,and does not make full use of the spatial information of HSI,which leads to the lack of high accuracy.To tackle this problem,since the three order tensor of HSI can describe the spatial information and spectral information of HSI equivalently,an algorithm using sparsity divergence index based on tensor decomposition(SDI-TD)is first proposed.SDI-TD algorithm uses Tucker decomposition to decompose the three order tensor of HSI into a core tensor and three factor matrices,and uses LRaSMD to obtain the SDI in the three tensor directions,and the final result is generated using the joint SDI.Then,in order to solve the problem of low detectionprecision in the case of the anomaly target is not sparse enough,an algorithm using tensor decomposition-based local Mahalanobis distance(Tensor-LMD)is proposed.In Tensor-LMD,the Tucker decomposition technology is used to decompose HSI tensor into a core tensor and three factor matrices.The minor PCs are used to eliminate anomaly and noise information along each mode and the more pure background data set is obtained.The sliding dual-window strategy is used for both the background data set and the original hyperspectral data set,and the local Mahalanobis-distance detector is employed for the final result.Finally,in order to fully exploit the spectral-spatial characteristics of HSI,and do not destroy the spectral-spatial spectrum structure,an algorithm using tensor-based adaptive subspace detection(TBASD)is proposed.In TBASD,the tensor blocks with the same size as the test tensor block are selected between the inner window and the local neighbouring window,and then in the whole HSI,respectively.Finally,the detection result is obtained using the tensor-based adaptive subspace detector.The three algorithms based on tensor make full use of hyperspectral spatial and spectral characteristics,and achieve better detection results,which are verified by experiments.To sum up,this paper mainly studies the hyperspectral AD algorithms based on spectralspatial characteristics of HSI,and the effectivenesses of the algorithms are verified by simulation experiments.