PRBCD-Net: Predict-Refining-Involved Bidirectional Contrastive Difference Network for Unsupervised Change Detection

• Published in: IEEE transactions on geoscience and remote sensing 2023, Vol.61, p.1-17
• Main Authors: Hu, Ling, Liu, Qichao, Liu, Jia, Xiao, Liang
• Format: Article
• Language: English
• Subjects: Change detection; Change detection (CD); convolutional neural network (CNN); cross-domain reconstruction; Detection; Distribution; Feature extraction; feature transformation (FT); heterogeneous images; image translation (IT); Iterative methods; Laser radar; Modules; Optical imaging; Optical sensors; Reconstruction; Robustness; Sensors; Transforms
• ISSN: 0196-2892; 1558-0644
• DOI: 10.1109/TGRS.2023.3314217

Heterogeneous bitemporal images have different visual appearances and inconsistent data distribution for the same scene, making it challenging to detect changes, which need to align the shared information and reduce various unwanted sensor-related noises for comparability. Mainstream methods usually adopt two types of techniques: feature transformation (FT) and image translation (IT). The former relies on handcrafted priors, while the latter lacks constraints on unwanted backgrounds, leading to limitations, such as a lack of robustness to nonintrinsic changes (e.g., seasonal and atmospheric changes, and sensor-related noise) and unsatisfactory detection performance. To overcome these drawbacks, we propose a novel unsupervised predict-refining-involved bidirectional contrastive difference network (PRBCD-Net) composed of a coarse prediction module and iterative refining modules. Each refining module utilizes feature extractors with a cross-reconstruction constraint and a bidirectional contrastive constraint to extract discriminative features and then generate a refined change map by change map optimizers. Two advantages of the proposed PRBCD-Net are as follows: 1) the cross-reconstruction constraint is used to promote the feature distribution consistency of the bitemporal images by using the forward and backward transformations and 2) the bidirectional contrastive constraint is used to improve the discriminability of features by narrowing the gap between nonintrinsic changes while widening intrinsic changes under the guidance of a coarse change map. Thus, the refining module can generate a finer change map than the coarse one, and the performance can be further improved through multiple iterations. Experimental results demonstrate the effectiveness and robustness of the proposed method compared with state-of-the-art methods.