Research On Kernel-based Spatio-spectral Information Miningfor Hyperspectral Image Classification

Whole trend of technical development in field of remote sensing is to achieve better the Earth observation with highly spatial, spectral and temporal resolution so as to provide more exact and finer information. Since1980’ years, after spectral imaging technique was proposed, hyperspectral imaging has become a very important method for remote sensing detection. Its inherent characteristic consists in providing spatial information and higher resolution of spectral information at the same time. Therefore, research on hyperspectral image processing and information mining technique has become one of hot topics in remote sensing imaging and detection, possessing important theoretical meaning and huge application potential.Taking landcover classification with hyperspectral remote sensing images as a research background, this paper has a theoretical basis of kernel learning theory and methodology and focuses on joint mining the spatial and spectral information in hyperspectral images with single/multiple-kernel learning methods. Aims of this thesis are to make full use of both spatial and spectral information and improve classification performance with hyperspectral images. Main works of this paper are shown as the following:First, this thesis deeply introduce kernel learning theory and its newest development——multiple-kernel learning (MKL) which is the important theoretical basis for investigation in this thesis. After summarizing kernel learning theory and corresponding methods, this work focuses on multiple-kernel construction and optimization learning methods. Furthermore, this work investigates composite kernel and multiscale kernel methods.Second, this thesis fully integrates characteristics of hyperspectral data and kernel design and proposes a method called subspace modulated kernel method which is used for feature extraction on hyperspectral data. According to hyperspectral imaging mechanism, this work investigates three approaches to subspace partition. In addition, this work designs subspace modulated kernel function so as to fully integrate the subspace characteristic into the kernel design and feature extraction. The experimental results directly prove that the features extracted by the proposed method can improve the classification performance, compared with standard kernel method.Third, taking landcover classification with spectral information as an orientation, this thesis focuses on multiscale multiple-kernel learning model and proposes optimal ensembling technique of multiscale multiple-kernels. To deal with the limitation of only single kernel in standard support vector machine (SVM), this work proposes a model of the multiscale MKL. The multiscale MKL problem can be decomposed into two sub-problems, i.e., unsupervised learning with basis kernels and optimization of SVM. Moreover, two new methods are proposed to deal with the problem, which use nonnegative matrix factorization (NMF) and kernel-based nonnegative matrix factorization (KNMF) with rank1. Compared with SVM and state-of-the-art MKL algorithm, the proposed method achieves better classification performance.Finally, this thesis builds a multi-feature and multiple-kernel learning model which can effectively integrate spatial features and spectral features under the MKL framework. This thesis proposes a multiple-kernel optimal ensembling learning method for optimally concentrating the multi-features into MKL model. Two cases of joint spatio-spectral information mining are considered. In the first place, with only the hyperpsectral images, three kinds of spatial feature which are statistical moments in local region,2D Gabor-based spatial contexture and multiscale morphological profiles, respectively. The classification performance and the applicability to different data sources of those three kinds of features are deeply analyzed. Then, the joint spatio-spectral classification is investigated with a visible image with highly spatial resolution and hyperspectral images with highly spectral resolution. At last, the experimental results from both two cases prove that the classification of hyperspectral images can be greatly improved by integrating the spatial and spectral features in the framework of the proposed MKL.