SBGC: Bidirectional Graph Comparison-Based Self-Supervised Network for Change Detection in Heterogeneous Images

He Zong, Jinchang Ren, Xinyu Li, Hongming Zhang, Erlei Zhang

Keywords: Remote sensing, heterogeneous change detection, contrastive learning, bidirectional comparison

TL;DR: A bidirectional graph comparison-based self-supervised network is proposed for change detection in heterogenerous remote sensing images

Abstract: Change detection (CD) in heterogeneous images is a hot but highly challenging topic in the field of remote sensing. However, the significant imaging differences and varying visual appearances of heterogeneous images complicate the accurate detection of changes occurring on the land surface through direct comparison. To overcome this challenge, this paper proposes a self-supervised network based on bidirectional graph comparison (SBGC) for unsupervised heterogeneous CD, which exploits modality-independent structural relationships. First, pseudo-Siamese networks are established to extract discriminative and robust features from bi-temporal heterogeneous images based on self-supervised contrastive learning. Then, these learned features are utilized to construct graph structures that represent structural relationships. Second, we introduce bidirectional graph comparison to fully exploit the graph structures for exploring comprehensive change information. Specifically, we map the graph structures to their opposite image modality and perform a bidirectional comparison between the original and mapped graph structures to generate a difference image. Finally, the change map is obtained by applying the Otsu segmentation algorithm to the difference image. Experimental results on three public heterogeneous datasets with different modality combinations show that the proposed method achieves superior performance compared to seven state-of-the-art methods, achieving the best performance with an average overall accuracy of 96.69%.

Primary Area: applications to computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 2012