Mainfold Learning Based Feature Extraction And Classification Of Hyperspectral Data

Hyperspectral data has abundant spectral information, which provides the chanceto make good material classification through the spectrum characteristics. However,huge volume of redundant information poses pressure on storage and computing. Takingthe inherent nonlinear characteristics of hyperspectral data into consideration, manifoldlearning as nonlinear dimension reduction algorithm is a good choice as featureextraction algorithm for hyperspectral data, which can effectively remove theredundancy and find internal essential characteristics of hyperspectral data. Then theoverall classification accuracy of hyperspectral data can be improved.In this paper, the existing prominent problems of manifold learning are studied,and the manifold learning is improved before appling to the visualization, featureextraction and classification of hyperspectral data in the concrete application:1) To theproblem that classic manifold learning such as LE is not with generalization ability, anovel generalization method (OSE-GLR) based on global linear regression is proposed.This algorithm improves the existing linearization algorithm, showing bettergeneralization effect;2) To the problem that linearization based generalization methodwill change the original manifold learning result, a novel generalization method (OSE-LLR) based on local linear regression is proposed. This algorithm keep the originalmanifold learning results, and can realize the generalization of any kind of manifoldlearning, with a very small generalization error;4) Considering that the hyperspectraldata distributes by clusters, a novel supervised manifold learning method terms asclasses encoding is proposed. The experimental result shows that this algorithm withbetter classification performance than the existing supervised manifold learningalgorithm;5) Considering that spectra of hyperspectral data have the characteristics ofaliasing, a novel supervised manifold learning method based on manifold segmentationis proposed. This method firstly use manifold segmentation algorithm to split manifoldin flat cluster. Then according to the double hierarchical category relationship betweensamples, hierarchical distance is defined. Experimental result shows that this methodoutputs relatively the best classification results.