A New Semisupervised Method for Detecting Semantic Changes in Remote Sensing Images

The growing availability of high-quality remote sensing imagery has led to increased interest in semantic change detection (SCD). Supervised methods for this task have shown significant performance improvements, but acquiring labeled data is often challenging and expensive. To confront this challenge, we propose a semisupervised approach for SCD in remote sensing images using an innovative teacher-student model. We use a convolutional neural network (CNN) in the teacher model and a fusion design combining CNN and vision transformer in the student model, with the rationale that the CNN requires fewer training samples compared with vision transformer, and fusion network allows us to leverage the advantages of both. To further enhance the model's performance, we propose a novel data augmentation approach by interchanging bitemporal images as well as their labels. The principle for that is the change from one moment to another and vice versa are two different changes and can, therefore, be used to augment the training dataset. More importantly, this method does not reduce its reliability, because no noise is brought to the remote sensing images. By adopting this approach, we are able to better utilize the small labeled dataset to increase the precision of the model while maintaining the robustness. According to the experimental results, the proposed method outperforms several state-of-the-art methods and achieves an improvement compared with bi-temporal semantic reasoning network (Bi-SRNet) in mean intersection over union (mIoU)/separated kappa (SeK)/overall accuracy (OA) of 3.25/4.42/1.78, 6.37/11.72/2.91 on the SECOND and Landsat-SCD datasets, respectively.