An Efficient Method of Hyperspectral Image Dimension Reduction Based on Low Rank Representation and Locally Linear Embedding

Abstract

Hyperspectralimages (HSIs) can provide powerful spectral discriminative information for the land-covers, thus is widely used in classification and target detection. However, HSIs always suffer from the curse of high dimensionality due the high spectral dimension, therefore dimension reduction and feature extraction are essential for the application of HSIs. In this paper, we propose an unsupervised feature extraction method for HSIs using combined low rank representation and locally linear embedding (LRR and LLE). LRR can structurally represent the intrinsic property of union of low-rank subspaces and LLE can employ the spatial correlation information. Two real HSI datasets are used in the experiments and the classification results using support vector machine (SVM) demonstrate that the features extracted by LRR LLE are more discriminative than the state-of-art methods. The classification accuracy of LRR LLE versus IR improved by an average of 4.47% and 2.97% on OA and AA, respectively; compared with the original data, it increased by approximately 12.07% and 7.35%.