Manifold Inspired feature extraction for hyperspectral image

Feature extraction is an indispensable preprocessing step for the large data, high redundancy hyperspectral remote sensing image (HSI). In this paper, a manifold inspired method, e.g., Laplacian Eigenmap (LE) is introduced for hyperspectral image dimensional reduction. In order to overcome the shortcoming of conventional manifold learning which could not deal with large data, linearization procedure for LE is proposed based on multiple linear regression analysis. Experiment on hyperspectral dataset demonstrates that the proposed manifold inspired feature extraction (MIFE) could preserve the local geometry of the samples in the original feature space. The low dimensional feature image could achieve a better classification accuracy rate.