DistillSemiCD: DINOv2-Driven Distillation for Semisupervised Remote Sensing Change Detection
Although semisupervised learning (SSL) can effectively reduce the model’s dependence on large amounts of precisely annotated data for remote sensing change detection (RSCD), existing methods still face several challenges. First, existing lightweight architectures struggle to extract discriminative semantic features under limited supervision, while vision foundation models (VFMs) are computationally prohibitive for practical deployment. Second, the inherent noise in pseudolabels tends to accumulate and propagate during training, thereby degrading model performance. Third, changing boundaries frequently exhibit low confidence and blurred localization. To address these challenges, we propose DistillSemiCD. First, we construct a train–inference decoupled knowledge distillation (TID-KD) framework, where a frozen DINOv2 serves as the teacher during training to transfer high-level semantic knowledge to a lightweight student model. During inference, only the student model is deployed, achieving efficient performance comparable to DINOv2. Second, we propose a DINO-based adaptive pseudolabel calibration (DINO-APC) module and a dual-level contrastive learning (DLCL) mechanism. DINO-APC leverages VFM priors to dynamically adjust pseudolabel confidence for noise filtration, while DLCL suppresses noise propagation and feature distribution shift through pixel-level local contrast and batch-level class center alignment. Finally, we design a Taylor-inspired boundary refinement (TIBR) module that employs dual attention gates to localize uncertain regions and uses a discrepancy-guided feature refinement mechanism to sharpen prediction boundaries. Extensive experiments on the WHU-CD, SYSU-CD, CDD-CD, and LEVIR-CD datasets demonstrate that DistillSemiCD achieves competitive performance across diverse benchmark scenarios. The code will be publicly available at https://github.com/Sean1005-x/DistillSemiCD