Nonnegative discriminative manifold learning for hyperspectral data dimension reduction

Manifold learning algorithms have been demonstrated to be effective for hyperspectral data dimension reduction (DR). However, the low dimensional feature representation resulted by traditional manifold learning algorithms can not preserve the nonnegativity property of the hyperspectral data, which leads inconsistency with the psychological intuition of “combining parts to form a whole”. In this paper, we propose a nonnegative discriminative manifold learning (NDML) algorithm for hyperspectral data DR, which yields a discriminative and low dimensional feature representation, with psychological and physical evidence in the human brain. Our method benefits from both the nonnegative matrix factorization (NMF) algorithm and the discriminative manifold learning (DML) algorithm. We apply the NDML algorithm to hyperspectral remote sensing image classification on HYDICE dataset. Experimental results confirm the efficiency of the proposed NDML algorithm, compared with some existing manifold learning based DR methods.