Tensor subspace learning and folded-concave function regularization for hyperspectral anomaly detection

Hyperspectral anomaly detection focuses on identifying and localizing the anomalous targets in remote sensing. The complex scenarios in hyperspectral images make it more difficult to effectively distinguish anomalous objects from background data, especially in noisy environments. Currently, available low-rank representation models capable of effective denoising often unfold hyperspectral cubic data into two-dimensional form, but this causes the structural knowledge to be lost. To surmount the above disadvantages, we propose a tensor subspace-based learning strategy with folded-concave regularization for hyperspectral anomaly detection. First, hyperspectral data undergo initial preprocessing through dimensional reduction and robust tensor principal component analysis to generate a dictionary representing the background. Then, a tensor subspace learning approach aims to factorize hyperspectral data into the background and anomaly tensors, in which the folded-concave function is leveraged to minimize minor components for denoising. Next, \(l_{F,1}\) $l_{F,1}$ norm on tensor is used to extract abnormal information from hyperspectral data. Finally, comprehensive experiments on several real datasets show that the proposed algorithm performs better than the comparative benchmark methods in detection performance.