Active Learning-Enhanced Pyramidal Convolution Unet for Change Detection in Optical and SAR Remote Sensing Image Pairs

Change detection (CD) between optical and synthetic aperture radar (SAR) remote sensing (RS) image pairs is a crucial and challenging task. Because of their disparate imaging mechanisms, direct comparison between them is not feasible. To confront this challenge, numerous unsupervised and supervised deep learning (DL) approaches have been developed to map optical and SAR images into a unified feature space for identifying the changes between them. The performance of the unsupervised DL methods is often suboptimal. Although supervised DL methods tend to achieve better results, they require abundant annotated samples for training. However, obtaining abundant annotated data is typically expensive and time consuming. To reduce the cost of acquiring labeled data while improving the CD performance, this article proposes a method that combines pyramidal convolution Unet (PC-Unet) with active learning (AL) for CD between optical and SAR images. Specifically, the proposed PC-Unet integrates pyramidal convolution within a four-stage Unet framework, enabling the model to capture multiscale detailed information and contextual information from input maps, thereby enhancing CD performance. Furthermore, to minimize the cost of labeled data acquisition, AL method is combined with the proposed PC-Unet. In the AL method, we propose an average uncertainty of the top percentile of pixels sampling strategy to choose the most informative samples for annotation. To evaluate the performance of the proposed approach, three datasets are used to conduct experiments. The experimental results manifest that our proposed approach surpasses various state-of-the-art unsupervised and supervised heterogeneous RS CD methods.