A Change Detection Approach Based on Multi-Temporal SAR Images for Railway Perimeter Intelligent Perception

Accurately and promptly change detection (CD) of railway perimeter is significant to rail transportation cyber-physical systems (CPSs) safety and operating environment monitoring. Synthetic aperture radar (SAR) images have been demonstrated to offer advantageous information for change detection. However, it remains a challenge in SAR scene parsing due to the speckle noise limitations. Furthermore, fine-grained discriminant information needs to be introduced in high-similarity feature analysis. In this paper, an unsupervised multi-temporal contrastive enhancement dual-domain network (MCEDNet) is proposed. First, pre-classification is performed to generate pseudo-labels for training samples. Subsequently, the sample features are enhanced through the dual-domain network, in which the spatial feature attention module (SFAM) is used to improve the semantic representation of the central region, and frequency information is represented by multi-spectral attention. Finally, multi-temporal contrastive learning is developed to refine the final output. The experimental results on SAR datasets demonstrate the effectiveness and generalization of the proposed MCEDNet.