Hyperspectral Target Detection With Target Prior Augmentation and Background Suppression-Based Multidetector Fusion

Abstract

Hyperspectral target detection (HTD) methods aim to exploit the abundant hyperspectral information to distinguish the key target pixels from multifarious background pixels. However, the performances of existing HTD methods are limited by the dilemmas of scarcity of target prior spectra, imprecise estimation of background spectra, as well as noise pollution. For the issues, this article proposes a novel target prior augmentation and background suppression-based multidetector fusion (TBMF) method for HTD, based on the joint optimization of the target prior spectra augmentation, low-rank pure background spectra separation, and nontarget nonbackground noise component removal. Specifically, a constrained linear spectral mixture model is seamlessly incorporated to implicitly augment the target's prior spectra. Also, the nontarget nonbackground components of HSI, i.e., noise with complex distribution are removed by a noise-robust l1,1-norm-based regularization. Subsequently, multiple basic constrained energy minimization detectors are trained using the augmented diverse target spectra in the background-suppression subspace derived by the separated background spectra. The detection results of these basic detectors are fused with a winner-take-all strategy to acquire the final detection result. Plenty of experimental results on four HSI datasets show that the proposed TBMF method performs promisingly when compared with several classical and recently proposed HTD methods.

Keywords

• Hyperspectral target detection (HTD)
• linear spectral mixture model
• matrix decomposition
• noise modeling