Research On Feature Extraction And Feature Selection For Hyperspectral Remote Sensing Data

With the capability of collecting high-resolution spectral image data from visible to infrared spectrum to provide more abundant information for observed target, Hyperspectral Remote Sensing(HRS) technology makes it possible to detect, identify, and understand complex targets, which cannot be solved by the conventional sensor technologies. HRS can be viewed as the evolution of traditional remote sensing, and draws high attention in recent years. With the advance of hyperspectral sensor technology, the number of bands is increasing. In the high-dimensional feature space formed by the hyperspectral data, the strong correlationship existing among bands, the non-linearity of data distribution, and the insufficient training samples problems present new research topics for data classification and recognition. Feature Extraction and Feature Selection for hyperspectral data have been the crucial issues involved in hyperspectral data processing to preserve the key spectral information and perform dimensional reduction. The related research has been included in the special issues of IEEE Transactions on Signal Processing, IEEE Transactions on Geoscience and Remote Sensing, etc. This dissertation, supported by the key project of Natural Science Foundation of China, mainly focus on the Feature Extraction based on Manifold Learning and Matrix Factorization, and the Feature Selection based on data structure. The main contributions are as follows:(1) Traditional Manifold Learning methods ignore the correlationship between adjacent pixels in the image. In order to solve this problem, a Local Embedding based on Spatial Coherence(LESC) feature extraction algorithm is proposed. Local linear structure of data is constructed through an optimal local linear embedding in high dimensional space. Meanwhile, the spatial context of pixels is considered by adopting the difference between the surrounding patch of pixels. Then local linear structure of data is projected to the low-dimensional space to perform dimensional reduction by finding an optimal projection matrix. LESC takes into account not only the manifold structure but also the spatial information of hyperspectral image. Experiments on a benchmark hyperspectral dataset demonstrate the superiority of LESC. Furthermore, the projection matrix can be applied to the new samples rather than just on the training samples in Locally Linear Embedding methods.(2) Nonnegative Matrix Factorization(NMF) does not exploit the geometric structure of the data. In order to uncover the hidden topics, Regularized Nonnegative Matrix Factorization(RNMF) feature extraction algorithm is proposed. The algorithm imposes an additional constraint on NMF that both the local and non-local quantities are considered. RNMF seeks to find a projection that simultaneously maximizes the non-local scatter and minimizes the local scatter. An iterative multiplicative updating algorithm is proposed to optimize the objective. In addition, Semi-supervised Feature Extraction based on Concept Factorization(SFECF) is proposed for the real hyperspectral remote processing system in which plentiful unlabeled and few labeled data co-exist. The limited labeled samples are used adequately as a hard constrained into matrix factorization. Meanwhile, the local graph of data is constructed to exploiting the manifold structure and preserving the local invariance. Such matrix factorization process can provide the low-dimensional representation with more powerful discrimination. These properties have better effect on hyperspectral data classification than some other matrix factorization methods.(3) Due to the multi-modal characteristic of hyperspectral data, a segmented Feature Selection based on Local Fisher Discriminant(FSLFD) system is proposed, which calculates the scatters by adding the local information of data and selects features by maximizing the ratio between local between-class and local within-class scatters. This method enables to separate the different classes and preserve the local neighborhood structure of the same classes in low-dimensional space. The redundant information in bands can be removed effectually and the selected bands are uniformly distributed. In addition, in order to utilize unlabeled data to improve the generalization performance of just using the limited labeled data, Semi-supervised Feature Selection based on Manifold(SFSM) algorithm is proposed based on considering the prior information of labeled data with the local and non-local invariance of the whole data. In this method, the discriminate structure is optimized through simultaneously maximizing the between-class and minimizing the within-class variances. Meanwhile, the manifold structure is exploited from the constructed local and non-local graphs for the whole data. The experimental results demonstrate that these characteristics generate a better description of the distribution structure of the data in the reduced dimensional space so as to improve the classification accuracy.