Spatially aware supervised nonlinear dimensionality reduction for hyperspectral data

In this paper we study the effect of injecting spatial information of image patches directly in the process of supervised dimensionality reduction. In particular, we adopt an approach derived from the mean map kernel framework to map image patches of variable size into a reproducing kernel Hilbert space. In that space, the orthonormalized partial least squares performs supervised dimensionality reduction to a discriminant subspace. Advantages of the proposed approach are discussed by studying two well known hyperspectral image benchmarks and by comparing it to composite-kernel feature extraction framework.