Feature Extraction for Hyperspectral Images Using Low-Rank Representation With Neighborhood Preserving Regularization

Abstract

Hyperspectral images (HSIs) usually contain hundreds of spectral bands. When they are used for classification tasks, HSIs may suffer from the curse of high dimensionality. To address this problem, the essential procedures of dimension reduction and feature extraction (FE) are employed. In this letter, we propose an FE method for HSIs using low-rank representation with neighborhood preserving regularization (LRR_NP). The proposed method can simultaneously employ locally spatial similarity and the spectral space structure, which comprises a union of multiple low-rank subspaces. The framework of LRR can structurally represent the union structure of a spectral space. Because spatial neighbor pixels always share high similarity in a feature space, an NP regularization item is introduced into the framework of LRR to consider the locally spatial correlation. Classification experiments are conducted on real HSI data sets; the results demonstrate that the features that are extracted by LRR_NP are more discriminative than the state-of-art methods, including both unsupervised methods and supervised methods.