CausalCD: A Causal Graph Contrastive Learning Framework for Self-Supervised SAR Image Change Detection

In recent years, self-supervised synthetic aperture radar (SAR) image change detection methods have achieved remarkable results, particularly in reducing dependence on expensive supervised signals. However, some critical issues remain open: how to design the optimal unsupervised feature representation model, and how to utilize noisy pseudo-labels for self-training to improve the change detection model? In this article, efforts are made to find a principled and fundamental solution to the above issues from the new perspective of causal reasoning, proposing a self-supervised SAR change detection framework named CausalCD. Specifically, we first construct a structural causal model (SCM) to formalize the self-supervised SAR image change detection process, and carry out a principled analysis to assess the influence of the unsupervised feature representation module and the pseudo-label-driven self-training module on the performance of change detection model. On this basis, we present a feature representation module that employs graph contrastive learning (GCL) and leverages the causal invariant mechanism to extract the optimal augmented representation, thus improving the model’s generalizability and soundness. In addition, we employ a causal intervention strategy to mitigate the negative impact of confusing bias from noisy pseudo-labels by blocking the backdoor path. CausalCD’s strengths lie in reducing the model’s dependence on labeled samples through causal GCL and obtaining optimal feature representation. Concurrently, it helps disentangle the noisy pseudo-label’s impact, thereby improving the performance of the self-supervised SAR change detection model. Finally, comprehensive experiments on three bitemporal SAR scenes demonstrate that CausalCD significantly outperforms several mainstream change detection models, confirming the effectiveness of CausalCD.