PAdaptCD: Progressive Adaptive Thresholding and Bitemporal Image Augmentation for Semisupervised Change Detection

Change detection (CD) aims to identify pixel-level changes of interest in multitemporal remote sensing images (RSIs). Due to the high cost of pixel-level annotations, semisupervised approaches have gained attention by leveraging limited labeled data alongside abundant unlabeled data. However, most current semisupervised change detection (SSCD) methods ignore the severe imbalance between unchanged and changed pixels in CD tasks, which may lead to predictions biased toward the dominant category (unchanged pixels). In addition, there is a lack of data augmentation techniques specifically designed for SSCD, which consider the complexity of RSIs and the unique characteristics of bitemporal images. In this article, we propose a novel SSCD method, PAdaptCD, which incorporates a progressive adaptive dual-threshold (PADT) strategy designed for selecting relatively balanced and reliable classwise pseudolabels. The PADT strategy progressively adjusts thresholds for changed and unchanged classes in an online manner, selecting pseudolabels based on the training status. In addition, bitemporal image augmentation (BTIA) techniques are developed to better capture the specific properties of bitemporal data. BTIA first introduces a highly random intensity-based augmentation, which sets an appropriate degree of data augmentation. Subsequently, a simple yet effective approach called bitemporal image mixing augmentation is proposed to facilitate style interaction between different temporal images. Experiments are carried out on four public datasets, including two building CD datasets (LEVIE-CD and WHU-CD) and two datasets with natural and human-induced changes (JL1-CD and CDD). The proposed PAdaptCD achieves higher performance F1-scores with respect to literature methods with a small amount of labeled data. The results confirm the effectiveness of the proposed method.