Terrain Classification of Hyperspectral Remote Sensing Images Based on Kernel Maximum Margin Criterion

Hyperspectral remote sensing images have brought abundant spectral information for terrain classification. But the terrain classification of hyperspectral remote sensing images is confronted with the problems of high dimensionality and nonlinear separability, which lead to unsatisfied terrain classification rate. In order to raise the terrain classification recognition rate of hyperspectral remote sensing images, a new terrain classification method is presented based on kernel maximum margin criterion (KMMC), i.e., KMMC subspace method. Firstly, the original data are mapped to a high-dimensional kernel space by kernel method, and then the maximum margin criterion (MMC) is used to extract the nonlinear discriminant features of original data in the kernel space. Finally, the minimum Euclidean distance classifier is used to classify in the resulting KMMC feature subspace. Recognition results on account of an airborne visible / infrared imaging spectrometer (AVIRIS) hyperspectral remote sensing image show that, compared with the original space method, linear discriminant analysis (LDA) subspace method, MMC subspace method and kernel linear discriminant analysis (KLDA) subspace method, the proposed KMMC subspace method can significantly raise the recognition rate while reducing the dimensionality of the data.