Supervised Feature Extraction of Hyperspectral Image by Preserving Spatial-Spectral and Local Topology

Abstract

Listen

Translate article

Manifold learning, as a promising tool for nonlinear dimensionality reduction of hyperspectral image (HSI) data, has drawn great research interests in the remote sensing community. It can extract meaningful and low-dimensional features underlying complex HSI data, which is useful in classification of ground targets. However, there are two limitations with current approaches, few considerations of spatial information and lack of explicit mapping relationship. In this paper, we propose a supervised spatial-spectral local topology preserving embedding (sssLTPE) method for efficient feature extraction of HSI, which owns two merits. First, spatial and spectral information at each pixel is integrated by an intuitive strategy. Second, an explicit and nonlinear mapping relationship is provided to effectively map unlabeled data to learned feature space. Experiments conducted on benchmark data set demonstrate that high classification accuracy can be obtained by using the features extracted by sssLTPE.