Dimensionality reduction and feature extraction from hyperspectral remote sensing imagery based on manifold learning

Manifold learning, as the novel nonlinear dimensionality reduction algorithm, is applied to dimensionality reduction and feature extraction of hyperspectral remote sensing information. In order to address inherent nonlinear characteristics of hyperspectral image, Isometric mapping (Isomap), the most popular manifold learning algorithm, is employed to dimensionality reduction of hyperspectral image, and the experimental results show that it outperforms traditional MNF transform. In order to include spectral information into manifold learning, spectral angle (SA) and spectral information divergence (SID), instead of Euclidean distance, are applied to derive the neighborhood distances in Isomap algorithm, and the result is better than that using Euclidean distance in terms of residual variance and normalized spectral eigenvalue. It is concluded that manifold learning is effective to dimensionality reduction and feature extraction from hyperspectral remote sensing imagery.