Hyperspectral remote sensing image feature extraction based on spectral clustering and subclass discriminant analysis

Hyperspectral remote sensing images (HRSIs) have the problems of high dimensionality and phenomenon of the same subject with different spectra. A class subdivision and feature extraction method based on spectral clustering (SC) and subclass discriminant analysis (SDA), namely SC-SDA, is presented. Firstly, when the overall separability is improved should a class be subdivided. Secondly, a generalized simple matching coefficient (GSMC) is proposed to evaluate the similarity of the clustering results in neighbouring dimensionality SC subspaces, and the SC subspace dimensionality corresponding to the maximum GSMC is selected. Then, SC is performed in the selected SC subspace according to the number of subclasses selected by intra-class separability. Finally, SDA is executed based on the class subdivision result. The experimental results of four real HRSIs datasets show that the classification results of the SC-SDA method are superior to those of linear discriminant analysis, separability-oriented subclass discriminant analysis and SDA methods.