Anomaly Detection In Hyperspectral Images

The rapid development of hyperspectral remote sensing technology improves the information richness of the shooting scene.It opens up a new way for anomaly(small target)detection in remote sensing image,and gives its more important practical significance.In order to construct a complete anomaly detection system in hyperspectral remote sensing images,and realize the automatic location of anomaly,the methods such as band selection,feature extraction and anomaly detection need to be studied.The main work of this paper is as follows:Firstly,a band selection method based on optimal linear prediction of principal components in subspace is explored.Hyperspectral bands are divided into different subspaces by spectral clustering with the improved correlation measure.The principal component analysis(PCA)of bands is presented in each subspace,and main components are selected as the reconstructed targets.The subspace tracking method serves as the search strategy,and several bands are selected from each subspace to perform the joint optimal linear prediction of reconstructed targets.The selected bands in each subspace are combined to obtain the optimal band subset.Experimental results show that,the proposed method can reconstruct the original data more completely.Compared with original data,the band subsets obtained by adaptive band selection(ABS)method,linear prediction(LP)method,maximum-variance principal component analysis(MVPCA)method and auto correlation matrix-based band selection(ACMBS)method,the band subset of explored method has superior performance of anomaly detection.Then,a feature extraction method based on fuzzy supervising and manifold structure preserving is discussed.The likelihood of each pixel belonging to anomaly is roughly calculated by the RX method,and it serves as the fuzzy supervision information.On the basis of the information,the neighborhood graphs of manifold structure of local pixels with the same category,non-local pixels with different categories,and whole are established,respectively.And the objective function of low-dimensional mapping is constructed by combining the neighborhood graphs of the three manifold structures.In order to deal with new pixels and data with nonlinear structure,the linearization and kernel methods are given to obtain the linear and nonlinear projection vectors.Experimental results show that,compared with principal component analysis(PCA)method,kernel principal component analysis(KPCA)method,locally linear embedding(LLE)method and locality preserving projection(LPP)method,the feature extracted by the discussed method can get higher anomaly significance with better anomaly detection results.And then,an anomaly detection method based on background clustering and weighted iterative RX is studied.The hyperspectral pixels are divided into different background clusters by theimproved fast search and find of density peaks algorithm.On this basis,each pixel within the outer window in the RX method is weighted by combining three methods to obtain a more accurate background statistical model.In order to reduce the weights of abnormal pixels in the outer window,each pixel is weighted according to the Mahalanobis distance between the pixel and each background cluster.In order to reduce the influence of background pixels belonging to different clusters in the outer window,each pixel is weighted according to the contribution degree of each background cluster to the detected pixel.In order to achieve the purpose of further purification of background in the outer window,each pixel is weighted according to the initial detection results and iterative detection is performed.Experimental results show that,compared with the classical RX method,blocked adaptive computationally efficient outlier nominators(BACON)method,weighted anomaly RX(WARX)method and probabilistic anomaly detector(PAD)method,the studied method can improve the robustness of RX method to the size of detection window obviously with higher detection accuracy.Subsequently,a detection method based on adaptive parameter support vector machine(SVM)is proposed in this paper.The abnormal pixels are positioned rapidly and roughly by an unsupervised detection method,and the posterior information of SVM is got by the position result.The kernel parameter of SVM is determined adaptively based on the posterior information and the criteria of divergence in the kernel space.The best hyperplane in the kernel space for the segmentation of targets and background is found by the SVM.Abnormal pixels and background are separated by the best hyperplane.The accurate and stable anomaly detection result is obtained via iteration.A large number of experimental results show that,compared with the existing methods such as the RX method,kernel RX(KRX)method and support vector data description(SVDD)method,the proposed method is more effective to detect abnormal pixels accurately in the hyperspectral remote sensing image.Finally,a method using projection pursuit(PP)optimized by modified artificial bee colony(MABC)method and weighted K-nearest neighbor(WKNN)for detecting anomaly is proposed in this paper.The kurtosis and skewness jointly defined by neighborhood pixels are chosen to design the projection index.MABC method is applied to be the optimization method.These projection images of anomaly are obtained from the low dimensional hyperspectral images by using projection pursuit,and abnormal pixels are extracted according to the histogram of these projection images.The characteristic information and main structure of pixels are extracted based on preliminary examination results,and purification of the preliminary detection result is accomplished by WKNN method.Experimental results show that,compared with the RX method,independent component analysis(ICA)method and projection pursuit method based on chaotic particle swarm optimization(CPSO),the proposed method can not only obtain a detection result of lower false alarm,but also have a faster computing speed.