View article

Compared with medium and low resolution Synthetic Aperture Radar (SAR) images, high resolution SAR images are more seriously interfered by target complexity and imaging noise, which present a greater challenge for using high resolution SAR images to detect changes. In this paper, we proposed a self-supervised contrastive representation learning method based on convolution-enhanced Transformer to identify the changes from high resolution SAR images. First, contrastive self-supervised representation learning framework is constructed on unlabeled SAR data without introducing any human-annotated labels, where the hierarchical representation is computed with convolution-enhanced Transformer. The key design of this Transformer lies in dividing patches and limiting self-attention calculation within local windows, which lead to linear computational complexity. However, the lack of across-window connections limits its representation power. Thus, we propose convolution-based module to build the information interactions across windows. As a result, rich local-global context information can be modeled more powerfully. Then, the changes are identified in the learned representation space. To improve the adaptability of the algorithm to different data sets, the final detection map is generated by decision level fusion. In addition, we propose sparse sampling strategy to reduce the amount of data and enable our method to be friendly to large-scale HR images. Experiments on simulated and real data sets demonstrate the effectiveness and superiority of the proposed method.