Hyperspectral Target Detection With Macro-Micro Feature Extracted by 3-D Residual Autoencoder

Abstract

Unsupervised autoencoders (AEs) have been demonstrated effectively to achieve robust performance in hyperspectral feature extraction. However, one-dimension inputs of AE cannot get precise detection due to spatial information loss. And the improved convolutional AEs (CAEs) perform outstandingly in microscopic spatial-spectral details learning, but lack macroscopic contextual information, which can provide meaningful sizes and shapes for target detection. To address this issue, a three-dimensional (3-D) macro-micro convolutional AE network is proposed to preserve both macro and micro information in this article. Furthermore, to make it easier to optimize, residual learning is induced for refining the network, and forming a novel 3-D macro-micro residual AE (3-D-MMRAE). For feature extraction, 3-D-MMRAE aggregates features from a macro branch and a micro branch, capturing discriminative and valuable structures. In addition, a promising loss combined with similarity measurement and distance constraint is proposed to improve the discriminative capacity and refine the reconstruction process. At last, the extracted features are fed to a novel hierarchical radial basis function (hRBF) detector for target preservation and background suppression. Extensive experiments and analysis on several classic hyperspectral benchmarks demonstrate that the macro-micro features have superior capacity of distinction, and our proposed 3-D-MMRAE detector outperforms several state-of-the-art detectors.