Spatial-constraint-based Sparse Coding For Hyperspectral Image Classification

In recent decades, with the progress of science and technology, the rise of hyperspectral remote sensing technology is one of the greatest achievements of remote sensing technology. Hyperspectral remote sensing technology has been widely used in military and civilian fields for the technology has the advantages of imaging and spectroscopy detection. Hyperspectral imagery(HSI) classification has become one of the hottest research fields of the remote sensing area. However, HSI classi?cation still faces some challenges, among which are the following : how to solve the problem of high-dimensional with small amount of samples, how to make use of spatial neighborhood information, how to choose the fittest classifier, and so on.Sparse representation is a kind of method based on a complete dictionary, which uses as far as possible concise and effective way to represent data. The emergence and rise of the sparse representation has made build a high-dimensional signal sparsity model becomes a hot research, which provides a new idea for the HSI classification. It is based on the observation that despite the high dimensionality of natural signals in the same class usually lie in a low-dimensional subspace. In recent years, sparse signal model by learning dictionary has been applied in the field of image processing and machine learning with state-of-the-art results. Sparse representation can capture the inherent structural features of signal and noise robust.It is clear that there is certa in limitation on the HSI classification performance for the traditional classification methods that only based on the spatial information or spectral information. To overcome this problem, we combine spectral and neighborhood characteristics to improve the accuracy of hyperspectral classification. In this paper, the sparse characteristic of the hyperspectral data is considered, and the sparse coding method is used to classify hyperspectral image. The main contribution can be summarized as follows:1. A new spectral-spatial dictionary learning method based on guided filter(GFDL) is proposed for hyperspectral image classification. The basic idea of this algorithm is touse the guided filter, which constraints spatial information. Then the corresponding coefficient of each pixel is calculated according to the learning dictionaries. Finally, the sparse coefficients are then used for classification by a linear SVM. The experimental results indicate that the proposed method has better performance than traditional methods which only consider individual spectral or spatial information.2. A new spectral-spatial sparse coding method based on nonlocal similarity(NSSC) is proposed for HSI classification. Inspired by the great success of nonlocal means, the nonlocal weighted method is proposed to build relationships between the neighboring pixels and the center test pixel through different weights. The nonlocal weights are determined based on the spectral angle mapping(SAM) algorithm, which is used for finding the difference between the neighboring pixel and the center test pixel. Experimental results on three real hyperspectral images(Indian Pines, Pavia University and Salinas) indicate the effectiveness of the proposed method for hyperspectral image classification.3. A new group sparse coding with a Multivariate Laplace scale mixture(MLSM) prior is proposed for HSI classification. We propose a class of sparse coding models that utilizes a Multivariate Laplace Scale Mixture prior to model dependences among coefficients. The group sparse coding model is used to combine the spectral and neighborhood information. The classification performance for hyperspectral image(Indian Pines) indicates the effectiveness of this algorithm.