Hyperspectral images contain space and spectral information. They provide both detailed spectral and spatial information for the analysis and recognition of physical materials. With the improvement of spatial resolution, hyperspectral instruments can obtain more detailed spatial information of land covers. Exploiting the spatial and spectral information of land covers, it is helpful to distinguish different land covers and to better achieve hyperspectral image analysis. However, the bottleneck of improving the hyperspectral image interpretation accuracy is how to effectively integrate the spatial-spectral features. Hyperspectral images usually have tens or even hundreds of spectral bands, interband spectral redundancy, strong noises. In addition, when considering the extracted spatial features, the dimensionality of features would become even more for analysis, incurring many problems for data transmission, storage and processing. Dimensionality reduction, which removes the redundant information and avoids the Hughes phenomenon, is an effective method to deal with these problems. Therefore, an important issue, which the hyperspectral image analysis is confronted with, is how to integrate the spatial-spectral features and reduce dimension. This thesis aims to develop spatial-spectral dimensionality reduction methods for hyperspectral images with the help of advanced technologies in pattern recognition and machine learning. More precisely, we will adopt manifold learning, sparse learning, and regularization theory to develop hyperspectral image spectral-spatial feature selection and extraction methods. The main contributions of this thesis are as follows: 1. Most existing feature extraction methods only consider spectral information,ignoring the spatial information of hyperspectral images in shapes, texture and other structural features. Thus, a novel hyperspectral image feature extraction technique is proposed; it is based on integrating spectral and spatial domain characteristics, termed ensemble manifold regularized sparse low-rank matrix approximation. The proposed method firstly stacks the spatial-spectral features of hyperspectral images into a long vector. Then, the stacked high dimensional features are decomposed into a low-dimensional representation by low-rank matrix factorization techniques. In order to exploit the complementary information provided by the spatial-spectral features, the proposed method employs ensemble manifold approach to reg...
修改评论