ACESemiCD: an adaptive change-aware enhanced semi-supervised change detection method based on consistency regularization

Abstract

ABSTRACT Semi-supervised change detection (SSCD) leverages limited labelled data in conjunction with abundant unlabelled data to reduce pixel-level annotation costs and enhance model generalization. However, consistency-regularized SSCD approaches still face two key issues: (1) limited initial annotations increase prediction uncertainty, causing noisy pseudo-labels to propagate and degrade performance during iterative self-training; (2) most augmentation strategies fail to provide scale-aware perturbations that adequately capture both subtle and large-scale changes, limiting detection robustness across diverse spatial scales. To address these issues, we propose ACESemiCD, an adaptive change-aware enhanced SSCD method. ACESemiCD employs a Mean Teacher architecture to generate high-confidence pseudo-labels, significantly suppressing noise propagation during training. A hierarchical Convolutional Block Attention Module (CBAM) is integrated into the feature extractor to enhance global semantic understanding and localize fine-grained changes accurately. Furthermore, we introduce Dual-scale Change-aware CutMix (DC-CutMix), which adaptively selects global contextual and core change regions based on prediction confidence, enabling scale-aware perturbations that enhance multi-scale sensitivity. Extensive experiments on the LEVIR-CD, GZ-CD and CLCD datasets demonstrate that ACESemiCD consistently outperforms existing SSCD baselines across various annotation ratios, confirming its robustness, effectiveness, and strong generalization capability for SSCD.