Hyperspectral Images Classification Using Sparse Representation Method

In 1980 s, earth remote sensing techniques evolved to hyperspectral remote sensing from multispectral remote sensing. Hyperspectral remote sensing immediately became one of the major earth-observing methods. This new technique, capable of providing nanometer-level spectral resolution, observes and reveals the ground object features hidden in the spectral curve that are undetectable by many traditional full-color sensing techniques. Hence, hyperspectral imaging has found a broad application in diversified fields, finally becoming an important component of the remote earth-sensing system in quite many countries. In the study of hyperspectral remote sensing images, their classification is the focus and basis. In the field of machine learning, numerous algorithms have been proposed for hyperspectral image classification. Sparse representation, a hot spot in machine learning field, has also been applied with success to the classification of hyperspectral images in recent years. Sparse representation based hyperspectral image classification method evolved, from the classic methodology that relies on spectral information, to the current philosophy that integrates spatial and spectral information. This thesis studies the classification of hyperspectral images based on sparse representation and includes the following:Firstly, considering that the classic sparse representation classification of hyperspectral images is low in precision and high in temporal complexity, another sparse representation classification method for hyperspectral images is suggested that is based on dictionary and band reorganization. This algorithm starts with pre-treatment using neighborhood equalization and applies multi-image k-Nearest Neighbor(KNN) fusion strategy when deciding the final classification results so that the spatial information in hyperspectral images is put to maximum use and the spatial-spectral information is integrated. In addition, band reconstruction and division are done by referring to the spectral signature of the ground target in the image, and the multi-image KNN strategy is ultimately integrated so as to maximize the utilization of spectral information. Moreover, it simplifies the classic sparse representation algorithm for hyperspectral images, greatly reducing the temporal complexity associated with the algorithm. The test results suggest this algorithm affords a low temporal complexity and a high classification precision.Secondly, this thesis proposes a joint sparse representation classification method with weighted pixel block for hyperspectral imagery. This method results from an in-depth study on the joint sparse matrix reconstruction algorithms: Simultaneous Orthogonal Matching Pursuit(SOMP) and Simultaneous Subspace Pursuit(SSP), major components of joint sparse representation classifier(JSRC) of hyperspectral remote sensing images, and improves on SOMP and SSP by giving consideration to their own properties. These two improved methods are named as Weighted Pixel Block SOMP(WPB-SOMP) and Weighted Pixel Block SSP(WPB-SSP). In WPB-SOMP and WPB-SSP, the classification results are assigned the image pixel blocks which have the test pixel to be classified as the center, and the pixel block weighting is estimated from the standard deviation of all the pixels in the block. In the following computation steps, all the image pixel blocks corresponding to the test pixel to be classified are decided for their weighting, obtaining the total weights of every pixel blocks belonging to each class, and obtaining ultimate classification results of the test pixel to be classified. This method takes advantage of the classification results of the neighborhood pixels, which are neglected by SOMP and SSP, and decides their weighting, and it provides a higher classification accuracy than SOMP or SSP does at the cost of a minimal growth in temporal complexity.