Researches On Spatial-Spectral Based Dimensionality Reduction And Classification Of Hyperspectral Data

The hyperspectral remote sensing image contains abundant spatial, radiational andspectral information that increases the possibility of accuratelydiscriminating landcoversof interest. The spectra reflect the properties of differentmaterials, and the image revealsthe geometric characteristics of the material, so it is beneficial to the classification andidentification of the landcovers in the Earth’s surface. Recently making full use of thespectral information of hyperspectral for classification has become a hot research topic.However, hyperspectral sensors have been developed and widely used for observing theEarth’s surface by sampling a huge number of spectral bands, typically up to severalhundreds, and the high dimensionality of hyperspectral vector also cause challenges toimage classification. Therefore, it is advantageous to reduce the dimensionality ofhyperspectral data.The study mainly includes the following aspects:Firstly, this paper proposes a classification approach for hyperspectral data sparseempty based on spatial-spectral sensing graph.First, according to the spatial coherenceof hyperspectral data, we propose a hierarchical partial correlation matrix, combinedwith non-negative low-rank representation, constructed a hierarchical spatial-spectralsparse graphs combined with non-negative low rank representation. Meanwhile, thecombination of manifold learning, construct regularization term, then verify theeffectiveness of the proposed algorithm by semi-supervised learning. With rare labeledsamples in the Indian Pines, Salinas and Pavia University hyperspectral datasets,experiment results show that: the algorithm can effectively improve the classificationrecognition rate, especially in the case of small sample learning, itcan still achieved highrecognition rate.Secondly， this paper proposes a dimensionality reduction method based onspatial-spectral and representation learning for hyperspectral data. The basic idea of thealgorithm for the use of data selection method based on representation learning，choosing the most useful subdataset, combined with semi-supervised dimensionalityreduction method, constructsdiscrimination term, regularization term, hierarchicalspatial manifold regularization term, to find an optimal projection direction fordimensionality reduction. Experimental results on Indian Pines databset show that theproposed method can achieve higher classification accuracy than other availablealgorithms. Thirdly, an algorithmbased on spatial-spectral tensor and sparse coding forproposes hyperspectral data classification was proposed in this paper. On the basis ofsparse coding theory, the spatial coherence of hyperspectral images was taking intoaccount at the same time, the spatial and spectral information data informationcombined together and then promotedto the hyperspectral data tensor representation, wepropose an approach based onspatial-spectral tensorand sparse codingalgorithm forhyperspectral data classification.The experiments results show that the proposed methodin this chapter can obtain a high recognition accuracy.This work was supported by the National Basic Research Program of China (973Program) under Grant no.2013CB329402, NCET-10-0668, National ScienceFoundation of China under Grant no.61072108,60971112,61173090, and Higherschool subject innovation engineering plan (111plan), no. B0704, Specialized ResearchFund for the Doctoral Program of Higher Education of China under Grant no.20120203110005, the project of weapon equipment advanced research fund under Grantno.9140A24070412DZ0101.