Feature Extraction Based Multi-Structure Manifold Embedding for Hyperspectral Remote Sensing Image Classification

Abstract

Hyperspectral remote sensing image is a typical high-dimensional data with a large number of redundant informantion, which will impact the classification accuracy. Feature extraction is an effective method to reduce the redundancy of hyperspectral image (HSI) and improve the classification performance. However, most feature extraction methods just consider a single structure information of HSI that will lose some valuable information. To address the drawback, we proposed an unsupervised feature extraction method termed multi-structure manifold embedding (MSME) for HSI classification. First, MSME utilizes sparse representation to obtain the sparse coefficients of HSI data. Then, it constructs a sparse graph and a sparse hypergraph with the sparse coefficients. We use the sparse graph, the sparse hypergraph, and the local linear property to represent different intrinsic structures of HSI. Finally, we construct a feature learning method with these structures to achieve an optimal projection matrix for feature extraction. MSME makes full use of the complementarity of different structures to reveal the intrinsic properties of HSI and improve the discriminating power of features for classification. Experiments on the Salinas and PaviaU data sets show that the proposed MSME algorithm achieves the best classification results than other state-of-the-art methods.