Research On Kernel Machine Learning Based Anomaly Detection Algorithms In Hyperspectral Imagery

Hyperspectral data as a combination of high spectral resolution and two-dimensional image is a new type of remote sensing data. Utilizing the abundance spectral information to distinguish the tiny difference among ground cover, we can detect the targets which can not be detected by texture, edge in the image space. That is very useful for target detection. In spectral anomaly detection algorithms, materials that have a significantly different spectral signature from their surrounding background clutter pixels are identified as spectral anomalies. In spectral anomaly detectors, no prior knowledge of the target spectral signature is utilized or assumed. So it is a focus in the hyperspectral target detection area. The classical linear anomaly detectors are of poor performance due to the high dimensionality, nonlinear correlation between spectral bands, mixed pixels, spectral variety in hyperspectral imagery. Kernel method is regarded as an effective method to processing nonlinear information and is widely used in many research areas. However, there are some shortages and unsolved problems in kernel machine learning based anomaly detection methods in hyperspectral imagery. This dissertation considers anomaly detection technique as the research object, kernel machine learning as the method, improving the performance of anomaly detection algorithms as the goal, and dealing with the problems endured in kernel based nonlinear anomaly detection methods in hyperspectral imagery. The main innovation contributions of this dissertation are as follows.Firstly, in the traditional dimensionality reduction methods, only linear information between spectral bands is used and nonlinear information being wasted. For this problem, a nonlinear independent feature extraction method is proposed and applied to anomaly detection in hyperspectral imagery. With the input data is mapped into an implicit feature space, kernel principal component analysis (KPCA) is performed to whiten data and fully mine the nonlinear information between spectral bands. Then, independent component analysis (ICA) seeks the projection directions in the whitened feature space for making the distribution of the projected data mutually independent. As the nonlinear independent features are extracted randomly, in order to select the best feature for anomaly detection, a local negentropy measurement (LNM) method is proposed for nonlinear independent features selection. After nonlinear independent features ranking by LNM, the nonlinear independent feature with the max LNM score is used for RX anomaly detection. The method suppresses the false alarm rate and improves the performance of anomaly detector.Secondly, the kernel RX detector is of good nonlinear anomaly detection capability, but the degeneration of the background kernel matrix due to the background data blurred by anomaly samples lead to a high miss rate. In this dissertation, a spatial filter based kernel RX anomaly detection algorithm in hyperspectral imagery is proposed. Utilizing the spatial correlation of pixels in a hyperspectral band, the background data is optimized by depressing the anomaly data using spatial filter in every spectral band. The background kernel matrix after spatial filter is a better distribution representation of the real background data. Using this method, the detection probability of kernel RX detector is increased obviously in hyperspectral imagery.Thirdly, Gaussian radial basis function (RBF) kernel is used in most of the kernel base anomaly detection algorithms, but the width factor (or kernel parameter) selection for RBF is very difficult. In most of these algorithms, the kernel parameter is obtained through a large number of experiments. It is time consumed, heavy workload, and can not obtain the optimal kernel parameter objectively. For the problem an adaptive kernel parameter estimation method is proposed. Based on the relationship between kernel parameter and the local second-order statistic information in hyperspectral data, an adaptive kernel parameter estimation method is derived. The parameter estimation of kernel function can be obtained along with the shifting of the background clutter pixels automatically. The degeneration of detection performance brought by a global fixed kernel parameter method in a background of miscellaneous terrain is improved by the proposed algorithm. Numerical experiments are conducted and the results show that the detection probability of the proposed algorithm is better than the classical fixed kernel parameter method at the same false alarm rates. Finally, for the problem of the kernel function scarce in kernel base anomaly detection methods (RBF is selected in most algorithms) in hyperspectral imagery, a novel spectral similarity measurement kernel function is proposed and applied to anomaly detection in hyperspectral imagery. As the RBF is based on the Euclidean distance of two spectral vectors, it is sensitive for distance variations of two spectral vectors, but not for spectral curve variation. In order to dealing with the spectral curves variation of the same materials, a spectral similarity measurement kernel function is proposed according to the spectral curves similarity description. A theoretical analysis is expounded and the shift invariance property of spectral similarity measurement kernel is derived. Numerical experiments are conducted on real hyperspectral imagery. The detection result comparison of Gaussian radial basis function based and spectral similarity measurement (SSM) kernel based anomaly detector shows the SSM kernel can improve the performance of kernel base anomaly detection methods in hyperspectral imagery. Especially for small sub-pixel targets detection, the SSM kernel is of obvious superiority comparing with the RBF kernel. To solve the false alarm rate rising due to the sensitivity of SSM kernel for spectral curve diversification, RBF kernel and SSM kernel are composite to utilize the advantage of each other. Using the composite kernel, the results of numerical experiments show the false alarm rate declines at the same detection probability comparing to solo SSM kernel used.