Contrastive Learning With Context-Augmented Transformer for Change Detection in SAR Images

Self-supervised contrastive learning can help alleviating the meet of large numbers of annotated samples and learning high-level representations from unlabeled data. However, the high diversities in ground objects make it difficult to learn the features at more robust and refined level in synthetic aperture radar (SAR) image analysis. To alleviate this issue, we propose a self-supervised weighted contrastive learning method with context-augmented transformer for change detection in multiresolution SAR images. First, a weighted contrastive learning framework is built by introducing a weighted contrastive loss, which can reduce the influence of changed pixels in the process of self-supervised feature learning and align feature representations of image pairs. Then, to model complex and rich context information, a context-augmented swin transformer is proposed to aggregate contextual information and compute hierarchical representations, which are beneficial for dense prediction. Specially, global channel-wise aggregation module and multiscale fusion structure are designed to enhance global features and capture fine-scale features, respectively. Thus, rich local, global and multiscale context information can be modeled jointly to achieve fine and robust feature expression. Compared with other network, our network gives full play to the advantages of CL and transformer to extract representations with rich context information in unsupervised scenes, with good generalization. Experiments on real SAR images with different resolutions demonstrate the effectiveness and superiority of the proposed method.