Delve Into Hyperspectral Anomaly Detection Via Bayesian Gaussian Tensor Decomposition

Jiahao Qi, Xingyue Liu, Ping Zhong

Abstract: Tensor decomposition has been widely used in Hyperspectral anomaly detection (HAD). However, owing to factorization rules and optimization strategies, existing tensor-decomposition based HADs always possess limited performances in real-world applications. In this paper, we propose a novel HAD method based on Bayesian Gaussian tensor completion to address these problems. Firstly, the spatial correlation is utilized to transform the input HSI into a sparse form. Then, to further characterize the variants of spectral signatures, we describe the tensor decomposition process as a hierarchical probabilistic model on the basis of Bayesian Gaussian theory. With this special probabilistic model, variational inference strategies could help to tackle the optimization issue more precisely and effectively. Finally, the obtained factor matrices are used to reconstruct the hyperspectral background image, and the anomaly detection is derived from the reconstruction errors using channel-wise average means.