Unsupervised multitemporal SAR image change detection via foreground-background collaborative optimization

Multitemporal Synthetic Aperture Radar (SAR) image change detection (CD) represents a significant focus in remote sensing interpretation research. Recently, matrix low-rank decomposition theory has gained popularity in this field to exploit inherent structural information without requiring annotated data. However, existing approaches predominantly rely on an idealized assumption that defines changed regions as spatially localized and sparse. This assumption introduces critical theoretical limitations: its sensitivity to change scales results in sparsity constraints failing to characterize large-scale continuous changes and accumulating decomposition errors, while the neglect of low-rank coupling between changed regions and backgrounds further undermines theoretical completeness. To address these issues, we propose a Foreground and Background dual-path collaborative optimization CD framework, namely FBCD. Specifically, a foreground change saliency model is constructed under generalized low-rank constraints, integrating low-rank consistency and local correlation mechanisms to capture complex change patterns. In addition, a background stability model based on low-rank self-representation learning achieves precise background separation through multi-view consistency constraint. Once generating reconstructed difference map, a self-supervised graph-optimized label propagation algorithm is designed to transform binary classification into a graph partitioning optimization problem, which further improves the CD accuracy. Extensive experiments on seven bitemporal benchmark datasets validate the superiority of the proposed method: Compared to state-of-the-art approaches, it achieves average Kappa coefficient improvements of 2.50% for large-scale continuous changes and 4.48% for small-scale localized complex changes. Furthermore, the method also shows strong applicability in short-term time series datasets. The source code will be made available at https://github.com/95xiaoli/FBCD.