Study On Semi-supervised Dimensional Reduction Based On Sparse Representation For Hyperspectral Remote Sensing Image

Hyperspectral remote sensing imagery can provide richer information of earth surface than other remote sensing data. With the rapid development in the recent 20 years, it has been used in many applications. High dimensionality and large data of hyperspectral remote sensing imagery face to many difficulties in data process, therefore it should to do the dimensional reduction through feature extraction and feature selection in the preprocessing procedure. Based on the sparse representationbased dimensional reduction, we presents a semi-supervised dimensional reduction method for hyperspectral imagery. The results of this paper are shown as follows:(1) Analyzing the features of hyperspectral data and elaborating the theory of sparse representation on dimensional reduction of feature extraction, the wavelet-based denoising method is adopted in the sparse representation-based classifier(SRC). The experimental results show that combining the wavelet denoising method can improve the classification performance of the sparse representation-based classifier.(2) The unsupervised dimensional reduction method of sparsity preserving graph embedding(SPGE) and supervised dimensional reduction method block sparse graphbased discriminant analysis(BSGDA) based on sparse representation are analyzed in this thesis. With the low computing efficiency of BSGDA, we proposed a “little graph construction” method to construct the sparse similarity matrix for the BSGDA process. The results of experiments show that the “little graph construction” method can improve the computing efficiency of BSGDA and the “little graph construction” method can be performance well in BSGDA’s applications.(3) We proposed a semi-supervised dimensional reduction method based on sparsity representation by using the wavelet-based denoising in SRC and near regularized subspace(NRS). The results of experiments show that the proposed dimensional reduction method has better performance than the supervised dimensional reduction method BSGDA and the unsupervised dimensional reduction method laplacian eigenmap(LE) with limited training samples.