Multi-scale Contrastive Learning for Building Change Detection in Remote Sensing Images

Self-supervised contrastive learning (CL) methods can utilize large-scale label-free data to mine discriminative feature representations for vision tasks. However, most existing CL-based approaches focus on image-level tasks, which are insufficient for pixel-level prediction tasks such as change detection (CD). This paper proposes a multi-scale CL pre-training method for CD tasks in remote sensing (RS) images. Firstly, unlike most existing methods that rely on random augmentation to enhance model robustness, we collect a publicly available multi-temporal RS dataset and leverage its temporal variations to enhance the robustness of the CD model. Secondly, an unsupervised RS building extraction method is proposed to separate the representation of buildings from background objects, which aims to balance the samples of building areas and background areas in instance-level CL. In addition, we select an equal number of local regions of the building and background for the pixel-level CL task, which prevents the domination caused by local background class. Thirdly, a position-based matching measurement is proposed to construct local positive sample pairs, which aims to prevent the mismatch issues in RS images due to the object similarity in local areas. Finally, the proposed multi-scale CL method is evaluated on benchmark OSCD and SZTAKI databases, and the results demonstrate the effectiveness of our method.