Research On Spatial-spectral Associative Classification For Hyperspectral Image

The appearance and technical improvement of imaging spectrometer forms a solid hardware foundation for the development of hyperspectral remote sensing technology.Hyperspectral remote sensing images have played a unique advantage in the field of remote object detection and interpretation in the matter of its multi-band, higher spectral resolution and rich information of ground feature, yet in the mean time also brings unprecedented challenges to the development of hyperspectral information processing technology. Being one of the key issues in the field of hyperspectral image processing, hyperspectral image classification technology gives the determination of the object category for each pixel, which provides quantitative symbol for the specific crop condition monitoring, military detection,water survey and other practical issues.At present, there are still many shortcomings in the traditional classifiers, that is, the traditional classifiers always purely depend on the spectral information but lack the explanation on the spatial structure of the ground objects or the depth of the spatial information mining, failing to give full play of spatial information. Besides, although support vector machine (SVM) play an uniquely important role in dealing with nonlinear problems and coping with dimension disaster, the selection of kernel functions is often fixed which is lack of good generalization ability for multi-classification problem. Therefore, this paper puts forward an effective method for spatial information mining on the basis of predecessors'research. For small sample problems, this paper introduces a multiple-kernel learning(MKL)algorithm of multiscale and multi-feature. The research works are described as follows:Firstly, the research background and previous works about hypersectral image analyzing are summarized. In addition, the theory of feature extraction of hyperspectral images is described. In view of the nonlinear problem, the theoretical framework of the kernel method is demonstrated, and the explanation of the basic theory as well as the algorithm of the support vector machine are also given in the paper.Secondly, the texture feature extraction based on the transform domain is proposed for supervised classification of hyperspectral image. For 2-D Gabor texture feature extraction,this paper introduces the technical principle and implementation of 2-D Gabor filter in the field of high spectral image texture extraction. For the two shortcomings of the original extraction strategy: (1) With texture feature analysis of original data, the noise and redundant information in the data may cause disturbance to the spatial structure; (2) Inadequate mining depth on the spatial structure after original texture extracting, and there still exists spatial information failed to be detected. A texture extraction method based on empirical mode decomposition(EMD) is proposed. The effectiveness of the proposed method is demonstrated by the experimental verification of two data sets. In addition, the anti-noise performance of the proposed algorithm is demonstrated by the experimental results of the noise data. The proposed algorithm can effectively extract the spatial structure of hyperspectral image.Finally, a multi-resolution and multi-feature multi-scale kernel learning technique has been studied. In view of the deficiency of the generalization performance of SVM's kernel, a non-negative matrix factorization algorithm is used to construct a multi-scale kernel learning machine with texture information. The new model makes use of the spatial information while optimizing SVM' s kernel parameters,and it makes the new kernel's properties better than each single kernel. The new kernel's identification ability of each class is analyzed by experiments and the potential of the algorithm is verified.