Hyperspectral Image Terrain Classification Based On Subspace Learning

In recent years, with the rapid development of remote sensing technology, people can get a lot of hyperspectral data, the spectral characteristics of hyperspectral data can providing spatial information of the ground objects and higher resolution of spectral information at the same time. The spectral information and spatial information of hyperspectral data is very rich, but the sample dimension is very high. Moreover the information between the band is redudant and the data is huge, but the number of labeled samples is small. To solve above problems, this paper carries out a more in-depth study.The concrete research content is as follows:1. A hyperspectral image classification method based on subspace learning and sparse coding is proposed. The algorithm is projected into the low dimensional subspace and group sparse encoding into a unified framework, which can ensure that the projected subspace is optimal. And the constraint of the spatial information is introduced, which makes the feature encoding not only has the overall sparsity, but also keeps the spatial information. The simulation experiment results show that the hyperspectral image classification based on projected group sparse coding algorithm can greatly improve the accuracy, and reduce the time.2. A hyperspectral data classification method based on sparse low-rank regression is proposed.we projecte hyperspectral data into the subspace of linear discrimination analysis and do Lasso regression in the subspace of linear discrimination analysis, which make full use of the category information of hyperspectral data and make the high-dimensional data have good separability after low dimensional mapping. The performance of the proposed algorithm is verified by simulation experiments on real hyperspectral data sets.3. A hyperspectral data classification method based on semisupervised classificion throngh virtual label regression is proposed.Semisupervised classification can make full use of unlabeled samples, effectively overcome the problem of insufficient sample label.But unlabel samples once are the wrong classified, this kind of mistake in the subsequent transfer process will be continue or even expand.The virtual labels by random walk can detect outliers and avoid that lable is wrong propagated.The experimental results show that the hyperspectral data classification algorithm based on virtual label regression, can greatly improve the accuracy, has the very good robustness.The experimental results prove that the classification of hyperspectral data can be greatly improved by the framework we propose. In the final, the paper summarizes the work, and points out the need for further research.