Hyperspectral Target Detection Based on Macro–Micro Spectrum Contrastive Learning

Depending on single target example under the influence of spectral variation, deep learning-based hyperspectral target detection (HTD) methods are challenged by the problem of insufficient target knowledge. Inspired by self-supervised learning, this paper proposes a novel HTD method based on macro-micro spectrum contrastive learning, which enhance detection knowledge by unsupervised intra-contrast learning and supervised inter-correction learning. Firstly, to address spectral variation, a multi-scale view transformation strategy is proposed to learn intrinsic features by contractive learning between macro and micro spectrum groups. Secondly, to distinguishing target knowledge, a multihead task-driven network is designed to learn discriminative features by integrating target spectrum into the band- wise correlation calculation. Lastly, to signalize the target and suppress the background, a composite loss is designed to maximize the average distance between the target and background in feature space. Experimental results on four benchmark datasets demonstrate the superior performance of the proposed method over existing state-of-the-art algorithms.