LLE-Based Nonlinear Dimensionality Reduction of Hyperspectral Data for Forest Information Extraction

摘要

Dimensionality reduction is an indispensable preprocessing step for information extraction from hyperspectral remote sensing data. In this paper, we introduce a nonlinear dimensionality reduction algorithm, called Locally Linear Embedding (LLE), and customize it for hyperspectral data dimensionality reduction. Unlike the linear dimensionality reduction approaches based on eigenvectors of data covariance matrix, LLE preserves the original local geometry of the hyperspectral data in the reduced space. This geometric preservation is important to maintain the nonlinear properties of the input data. The objective of this study is to investigate LLE's effectiveness for information extraction from hyperspectral remote sensing data. In this paper, LLE was examined in terms of spatial information preservation, pure pixel identification, and forest species classification. Different approaches were adopted to downsize the input data volume to meet LLE's large memory requirements. One of the two open parameters in LLE, the number of nearest neighbors used for data point reconstruction, was also examined. The preliminary results of this study demonstrated that LLE did a better job than PCA on spatial information preservation and pure pixel identification. The classification results based on LLE features are comparable to those derived from the same number of PCA features.