M3DNet: A manifold-based discriminant feature learning network for hyperspectral imagery

Abstract

Feature extraction (FE) is an effective method for learning discriminant features from hyperspectral image (HSI). Recently, graph embedding (GE) framework has been widely applied in FE of HSI data. GE unifies many classical FE methods and explores the low-dimensional embedding of high-dimensional data by a projection matrix generated from undirected weighted graphs. However, GE is unable to adaptively optimize projection matrix due to the absence of an iterative strategy in a single mapping process. To address this issue, a unified optimization method termed manifold-based maximization margin discriminant network (M3DNet) was proposed to improve the performance of traditional FE methods. In M3DNet, an initial projection matrix is obtained from original FE method, and then a maximal manifold margin criterion (M3C) is proposed to maximize the margins among different classes, which enhances the discriminative ability of embedding features. After that, an iterative strategy is designed to optimize the projection matrix. Experiments on real-world HSI data sets indicate that the proposed M3DNet performs significantly better than some state-of-the-art methods.