Joint Self-Training and Rebalanced Consistency Learning for Semi-Supervised Change Detection

Change detection (CD) is an important Earth observation task that can monitor change areas at two times from the view of space. However, fully-supervised CD has a heavy dependence on numerous manually-labeled data, limiting their applications in practice. Beyond the fully-supervised setting, semi-supervised change detection (SSCD), which uses a few labeled data to guide the unsupervised learning of dominant unlabeled data, has attracted increasing attention for its significant advantage in alleviating the demand for annotations. To this end, in this paper we propose a joint self-training and rebalanced consistency learning (ST-RCL) framework for SSCD, which consists of a basic supervised branch for the labeled data and a novel unsupervised branch for the unlabeled data. To make full use of the unlabeled data, the unsupervised branch generates pseudo-labels from weakly-augmented unlabeled remote sensing image (RSI) pairs to supervise the CD of two strongly-augmented counterparts, including an unrotated version and a rotated version. On one hand, the unrotated unlabeled RSI pairs are pseudo-supervised with the pseudo-labels by confidence-based self-training. On the other hand, to further enhance model robustness to rotation non-equivariance and imbalanced distribution, the predictions of rotated unlabeled RSI pairs are aligned to the pseudo-labels by a well-designed rebalanced consistency learning strategy based on uncertainty-based class weighting. Extensive experiments are performed on four widely-used CD datasets, and the proposed ST-RCL yields new state-of-the-art results on all these datasets in comparison with some other SSCD methods, demonstrating its effectiveness and generalization. Our code will be available at https://github.com/zxt9/STRCL-SSCD.