Non-negative Matrix Factorization Based Feature Extraction And Classification Of Hyperspectral Image

Hyperspectral data has abundant spectral information, which provides the possibility to conduct good classification through the spectrum characteristics. However, the bands of Hyperspectral image are so large that the amount of data is relatively big, huge volume of redundant information poses pressure on data storage and computing. In order to fully take advantage of the rich spectrum of hyperspectral remote sensing image information as possible while reducing the huge computational complexity due to large amount of data, feature extraction for hyperspectral data is necessary, in order to representatives of lower dimension to approximate the original hyperspectral data to achieve high accuracy with less data.In this paper, the characters of airborne hyperspectral imaging spectrometer image are described. The distributions are usually non-Gaussian, in extreme cases, the structure is extremely multimodal. Therefore the main contents of this paper are: 1)According to multimodal structure we choose LFDA as the dimensionality reduction method for classification, combining classifiers algorithm, which can be used for hyperspectral image classification operations, Simulation experimental conducted for the validity of the algorithm in hyperspectral images classification; 2) Explored NMFbased dimension reduction method through adding orthogonality constraints combined with the LPP linearization method, which can bring dimension reduction of hyperspectral data as orthogonal as possible, at the same time the samples within a class more close while the samples between classes more apart, Simulation experimental conducted for the effective of the algorithm; 3)Using the MATLAB/GUI to design hyperspectral image processing software, which is a part of data processing modules of the program, it is designed for hyperspectral image detection and recognition, MATLAB can be used in dealing with high-dimensional data matrix, we can use the algorithms after processing the hyperspectral data for classification.