The K-LLE Algorithm for Nonlinear Dimensionality Reduction of Large-Scale Hyperspectral Data

Abstract

This work addresses nonlinear dimensionality reduction by means of locally linear embedding (LLE) for large-scale hyperspectral data. The LLE algorithm depends on spectral decomposition to a great extent, resulting in computational complexity and storage-costing while calculating the embedding of the low-dimensional data, particularly for large-scale hyperspectral data. LLE is not applicable to dimensionality reduction of large-scale hyperspectral data using general personal computers. In this paper, we present a novel method named K-LLE which introduces K-means clustering into LLE to deal with this issue. We firstly utilize K-cluster centers to represent the manifold structure of data instead of all data points, and next regard the K-Cluster centers as a bridge between the manifold structure and all data in order to obtain the low-dimensional representation for each data point without handling the complex spectral decomposition. Finally, classification is explored as a potential application to validate the proposed algorithm. Experimental results on two hyperspectral datasets demonstrate the effectiveness and superiority of the proposed algorithm.