Research On Subspace Analysis-based Target Detection On Hyperspectral Imagery

Hyperspectral imagery is a three-dimensional imagery generated by imaging spectrometer simultaneously to the same surface scenery at dozens even hundreds bands, which has abundance of spectral information to detect the materials and greatly enhances the ability of target detection. Therefore, target detection by hyperspectral imagery has become a hot research area of remote application. However, owing to the low spatial resolution of the sensor, mixed pixels are widespread in hyperspectral imagery. The mixed pixel problem brings a great challenge for identify and recognition of materials, which is a problem to be solved urgently for target detection on hyperspectral imagery. Hence, how to effectively raise the ability of target detection on subpixels is a critical problem for further development of hyperspectral remote sensing imagery.Recently, most researches on subpixels target detection are based on subspace model. Therefore, this dissertation mainly study the subpixels target detection on subspace analysis. By systemly analyzing former studied, the dissertation focuses on how to improve the algorithms based on linear spectral mixture model, and mainly research on how to extend algorithms based on linear spectral mixture model to nonlinear feature space by kernel trick for resolving the difficulties in nonlinear unmixing.The major works and contribution of this dissertation are as follows:(1) Aiming at the problem of unsupervised hyperspectral image classification, a new unsupervised constrained linear discriminant analysis (UCLDA) approach to hyperspectral mixed pixel classification was introduced from the angle of feature extraction. We used vertex component analysis (VCA) to extract the endmembers, and then applied spectral angle mapping (SAM) to get the training samples and let constrained linear discriminant analysis (CLDA) extract features of the hyperspectral image, and finally implemented classification with least distance method. The method of endmember extraction was unsupervised, so the new algorithm was unsupervised, too. The proposed algorithm was studied using simulated and real hyperspectral data. Experimental results show that the UCLDA is slightly better than least square spectral mixture analysis method and significantly superior to spectral angle mapping classification.(2) Constrained linear discriminant analysis (CLDA) and orthogonal subspace projection (OSP) are both explored in hyperspectral image classification and have shown promise in signature detection, discrimination and classification. However, the two subspace projection approaches cannot directly estimate the signature abundance. The OSP has been extended by a least squares orthogonal subspace projection (LSOSP) to estimate the signature abundance while CLDA has not. The solution of CLDA turns out to be a constrained version of OSP implemented with a data whitening process and the means of samples as signatures. Due to this fact, following the same idea for extending OSP to LSOSP, in this paper, a modified constrained linear discriminant analysis (MCLDA) is proposed for unmixing hyperspectral data, which can directly estimate the signature abundance. Experiment results obtained from both artificial simulated and practical remote sensing data demonstrate that the MCLDA algorithm performs better than least squares method and the LSOSP.(3) Orthogonal subspace projection (OSP) approach is an effective way for detecting subpixel target, in linear spectral mixture model. In order for achieving subpixels target detection, kernel orthogonal subspace projection (KOSP) as the nonlinear OSP lets hyperspectral data in the original input space be mapped into high-dimensional feature space via kernel function without any knowledge of the actual nonlinear mapping function, which will result in the mixed pixels in feature space can be presented in linear mixing model, then apply OSP in feature space. KOSP exploits the higher order relationships between the mixed pixels in feature space, and improves robustness to noise. In this paper, we systematically analysis and deduce the theory of KOSP, and find out some mistakes existing in the conventional KOSP. Experimental results based on simulated hyperspectral data and real hyperspectral imagery show that KOSP method proposed in the paper outperforms the conventional OSP approach.(4) In this paper, a kernel-based signature space orthogonal projection (KSSP) technique is proposed for nonlinear subpixel target detection in hyperspectral imagery. As a nonlinear version of the signature space orthogonal projection (SSP), the SSP is used in a high-dimension feature space after the pixels of input space being mapped into the feature space via nonlinear mapping. Applying the kernel trick, KSSP need not know the actual nonlinear mapping. Experimental results of simulated and real data demonstrate that the proposed KSSP approach outperforms the SSP method in target detection, and improves robustness to noise.