Regularized Information Preserving Projections

In hyperspectral data classification, too many bands cause overfitting. Techniques for reducing bands are either informative or discriminant. Each technique has its strengths and weaknesses. Motivated by Gaussian processes latent variable models, we propose a linear projection technique that is both informative and discriminant. The technique optimizes a regularized information preserving objective, where regularization sets a preference for latent variables. Experimental results based on hyperspectral image data are provided to validate the proposed technique.