Tensor ring with alternative change mask for multitemporal hyperspectral image change detection

Abstract

• The proposed method, TRACM-CD leverages tensor ring to extract temporal-symmetric features in multitemporal HSIs. • It also uses alternative change masks to constrain change representation and improve detection accuracy. • Experiments on four data sets show that TRACM-CD outperforms several classic and new change detection methods. Multitemporal hyperspectral image (HSI) change detection (CD) is a prevalent topic in remote sensing image processing. HSI CD usually consists of change feature extraction and classification. Although the high dimensionality of HSIs provides rich spectral information, HSIs are prone to spectral-spatial variability that degrades change detection accuracy. Recently, tensor decomposition has been successfully applied to CD. However, there is still room for improvement. We propose a tensor ring-based CD model with alternative change masks (TRACM-CD) for multitemporal HSIs. TRACM-CD extracts temporal change features using TR decomposition applied to different temporal change vectors. The alternative change masks constrain the temporal change representation and guarantee the temporal symmetry for change features to facilitate recognition of background and changes. Experimental results on four real-world multitemporal HSI datasets confirm the effectiveness and superiority of TR-based CD. The proposed model outperforms its tensor counterparts and classic approaches for multitemporal HSIs.