Research Of Anomaly Detection Algorithms Of Hyperspectral Imagery

Hyperspectral imagery is a new type of remote sensing data. Its high spectral resolution makes it suitable for the detection of human-made targets surrounded by natural environment background. An anomaly detector can enable one to detect targets whose signatures are spectrally distinct from their surroundings with no prior knowledge, so it is practicable in real sences, and then it becomes a focus in the field of target detection. Based on the analysis of characteristics of hyperspectral imagery, the dissertation does some researches as following in order to solve the difficulties in anomaly detection, such as high dimensionality, nonlinear feature extraction, spectral variety, mixed pixels.Firstly, dimension reduction methods are studied and a new adaptive band fusion algorithm based on the second generation curvelet transform and pulse-coupled neural networks (PCNN) is proposed to solve the problems caused by the high dimensions of hyperspectral imagery. After the whole data space is divided into several subspaces by adaptive subapace decomposition method, every subspace is considered an independent processing unit and curvelet transform is performed in every unit. Then the coarse scale coefficients from curvelet transform are weighted fused based on the entropy of every band image and the fine scale coefficients are selected intelligently by PCNN. Finally the fused coefficients are reconstructed to obtain the fusion image of every subspace by inverse curvelet transform. Therefore, the way that uses the fused images to detect targets can not only reduce the data greatly but also extract the detail information of hyperspectral imagery effectively.Secondly, kernel methods are used to extract the nonlinear information of hyperspectral imagery. According to the study of the theory of kernel methods, a new kernel weighted RX algorithm is proposed for anomaly detection. The algorithm is implemented in the feature space of original hyperspectral data, and for the purpose of decreasing the portion of anomaly pixels in the background covariance matrix, each pixel in the covariance matrix is given weight by its distance to the data center. In addition, when the dot products in the high dimensional feature space are converted into the kernel computation in the low dimensional input space, the new spectral kernel function and the radial basis kernel function are mixed to conquer the difficulty brought by the spectral variety.Finally, a nonlinear anomaly detection algorithm based on the background error data is proposed on the basis of the analysis of linear mixed model. After the background endmembers are gained by the proposed fast endmember extraction method, spectral unmixing technique is applied to all mixed spectral pixels for the purpose to separate target information from complicated background information. Then the error data that include abundant target information were transformed into a high dimensional feature space to finish the target detection. This way the serious background interferences brought by the spectral mixing was overcame, and the exploitation of nonlinear information in hyperspectral imagery greatly improved the performance of the proposed algorithm.