Semi-supervised Dimension Reduction And Fusion Of Spectral And Textural Information For Hyperspectral Image

With the rapid development of remote sensing technology, hyperspectral remotesensing has become the important means for the earth observation. However, as agood input for the data analysis, the high-dimensional data also presents somechallenging problems. Because of the high costs at labeling samples, the ill-posedclassification problem is often encountered where the number of labeled samples issmall compared to the dimensionality of the data. Dimensionality reduction is aneffective solution to the problem. However, the supervised data reduction is alsomisled by the limited information from the labeled samples. The semi-superviseddimensionality reduction and classification based on the labeled and unlabeledsamples has become the inevitable trend of hyperspectral image interpretation. Onthe other hand, with the improvement of the spatial resolution of hyperspectralimages, the phenomenon of the same objects with different spectrum and thedifferent objects with the same spectrum has occurs more and more frequently.Combination of texture and spectral information for hyperspectral classification hasbecome a hot topic. In this background, the study of the hyperspectral imagesemi-supervised classification with fusion of texture and spectral information,labeled with unlabeled samples is carried out and the main contents are as follows:(1) Systematically introduce the related theories and methods involved insemi-supervised dimensionality reduction, texture extraction and image fusion.(2) Propose a semi-supervised dimension reduction method based on graphLaplacian. The algorithm uses the labeled and unlabeled samples to construct awithin-class graph and a between-class graph, and gets the graph-basedsemi-supervised feature score criterion. Then the linear transformation matrix in thecriterion is obtained by solving the generalized eigenvalue problem, and thesemi-supervised dimensionality reduction is conducted based on the transformation matrix. Experimental results show that the proposed is superior to a variety ofsupervised, unsupervised and semi-supervised dimensionality reduction methods.(3) Propose a fusion algorithm based on tri-training, through which a jointspectral and spatial analysis is done. In the method, three classifiers are used basedon the labeled samples with the spectral data and two spatial features, respectively.These classifiers are refined using the unlabeled samples in the tri-training process,and the final classification results are obtained by the decision fusion. Experimentalresults show that the proposed method out perform a variety of fusion methods, as itcan effectively integrate the information from the spectra and texture, labeled andunlabeled samples for classification.