Meta-Learning-Based Semi-Supervised Change Detection in Remote Sensing Images

Semi-supervised change detection methods with consistency regularization, which overcome the lack of labeled samples by using unlabeled samples and enforcing consistent predictions under weak perturbations. However, current consistency regularization methods lack randomness in their perturbation settings and treat all samples uniformly, limiting the model’s ability to leverage sample diversity to improve generalization. In contrast, meta-learning methods shift focus from individual samples to learning patterns across similar tasks, thereby enhancing information efficiency and model generalization. Inspired by these principles, we propose a Meta-Learning-based Semi-supervised Change Detection (MLSCD) method for remote sensing images, which aims to explore and leverage meta-learning methods to enhance the generalization capabilities of consistency regularization-based semi-supervised change detection. First, we set the degree of weak perturbation and the combination of different types of perturbations as random parameters to generate diverse and randomized weak perturbations. Second, we redefine consistency regularization-based semi-supervised change detection from a meta-learning perspective, which learns patterns from diverse perturbation tasks to improve sample utilization efficiency, thereby enhancing the model’s generalization capability. Third, we balance accuracy and efficiency by using AdamW for cross-task updates in the outer loop and SGD for single-task optimization in the inner loop, which experimental results demonstrate is an ideal method for applying meta-learning to remote sensing change detection. Finally, the superiority of the proposed method is validated on two datasets. The extensive experimental results demonstrate the superior performance of the proposed method compared to state-of-the-art methods.