Kernel-Based Band Selection for Hyperspectral Image Classification

Hyperspectral image classification is a challenging task due to problems such as complex nonlinear structure of data, noisy and redundant spectral bands, and small-sample-size problem. The nonlinear support vector machine (SVM) is an excellent classifier to tackle these challenges. Selecting the most appropriate kernel and tuning its parameters for the data has a vital role in classification performance of the SVM. In this paper, an anisotropic Gaussian kernel is used. Anisotropic Gaussian kernel has a scaling coefficient for each band. The scaling coefficients are tuned by maximizing a classifier independent separability measure with l
1
regularization. The separability measure uses within-class and between-class dispersion information in the feature space. Some of the scaling coefficients will be zero after separability maximization due to l
1
regularization, that is equivalent to removing corresponding irrelevant bands from classification processes. Therefore, an embedded band selection is done to remove the destructive effect of irrelevant and redundant bands. The results show on average 5% classification accuracy improvement compared to an isotropic Gaussian kernel for classifying Indiana Pin Site hyperspectral image.