Study On Application Of Nonlinear Manifold Structure In Anomaly Detection For Hyperspectral Imagery

With the appearance of nanometer resolution imaging spectrometer, the abundant spectral information of the hyperspectral imagery provides the detection of earth targets with a strong support. Anoaly detection technology which can detecte the surface substance dissimilar to ambient spectrum has been widely studied.The traditional linear model based anomaly detection has been recognized with many shortcomings. In this paper we studied the distribution characteristics of the image data sets in the high-dimensional from the perspectives of anomaly detection, analyzed the nonlinear manifold structure comprised in the hyperspectral image data, based on which the research of anomaly detection algorithm for hyperspectral image is carried out.(1) Starting from the spectral imaging sensor mechanism, we analyzed the causes of spectral changes and physical factors affecting the spectral characteristics and showed the performance of the spectral image of uncertainty. Simultaneously, we studied mature non-linear mixed model for the actual data using principal component analysis and geodesic distance to analyze the manifold structural characteristics of data sets in the high-dimensional spectral space which support the derivation of the data sets nonlinear manifold structural properties and to provide the guiding principle for the determination of the non-linear structure based anomaly detection.(2) Hyperspectral image data is considered to appear with nonlinear manifold geometry characteristics in high-dimensional spectral space, on the basis of which we proposed an algorithm of spectral space window anomaly detection(SSW-AD). In the spetral space, each spectral curve corresponds to one point in the space. Firstly a sliding window is established in the in the spectrum space. Based on local linear concept, pixels covered in the slide window should own a linear manifold structure. Then, on the basis of this anomaly detection algorithm for local linear structure is carried out in order to achieve effective handling of the global non-linear manifolds.(3)According to the high dimensional characteristics of the hyperspectral imagery, the necessity of reducing the dimension of the hyperspectral data with insuffient sample points has been discussed. The differences between traditional linear dimensionality reduction method and the nonlinear dimensional reduction methods have been compared and analysised. The effectiveness of the nonlinear manifold methods is verified. A new anomaly detection algorithm based on manifold learning methods is proposed to solve the neighborhood selection uncertainty in the traditional methods(4)A new spectral measure is proposed according the characteristics of the hyperspectral imagery which combines the spectral angle and the spectral gradient. The new measure can avoid the measure uncertainty which is caused by the light change and terrain diversification and the new measure and improve the spectral resolution capability of the traditional measure.(5) Anomaly detection based on kernel funcion is studied in the thesis. A new algorithm based on the backgroud endmember extraction algorithm and the kernel RX algorithm is proposed. The new algorithm improvs the generation pattern of the background kernel Matrix, inhibits the impact of the abnormal pixels while making use of the intrinsic nonlinear characteristics of the hyperspectral imagery, and increases the applicability of the algorithm while improving the anomaly detection results.