AMS-SemiCD: A Synergistic Framework for Semisupervised Change Detection via an Adaptive Pseudolabeling Strategy and a Multibranch Network
Abstract
Semisupervised change detection (SSCD) aims to reduce the dependence on extensive labeled data. However, existing methods face three critical challenges. First, class imbalance and boundary ambiguity induce class-dependent confidence distributions, hindering reliable pseudolabel selection. Second, uncertain pixels may still contain useful structural cues, while their indiscriminate use can amplify pseudolabel noise. Third, achieving complementary local–global feature representations in dual-branch architectures typically incurs prohibitive computational costs. To address these issues, we propose the adaptive multibranch synergistic SSCD (AMS-SemiCD) framework, which uses teacher–student consistency regularization to supervise unlabeled bitemporal pairs. Within this framework, the adaptive statistical pseudolabel refinement (ASPR) strategy improves the reliability and utilization of consistency targets through class-adaptive dynamic thresholding and contribution-adaptive soft weighting. As a result, ASPR accommodates class-dependent confidence distributions while effectively exploiting uncertain but informative pixels. Meanwhile, the multibranch synergistic enhancement network (MSE-Net) enhances discriminative feature representation by integrating local CNN and global Mamba branches with manageable computational costs. Specifically, MSE-Net consists of the multibranch spatiotemporal synergistic fusion (MSSF) for local–global fusion, the multibranch spatiotemporal difference booster (MSDB) for temporal-difference enhancement, and the adaptive gated attention refiner (AGAR) for cross-layer refinement. Extensive experiments on the GoogleGZ-CD, WHU-CD, and LEVIR-CD datasets demonstrate that AMS-SemiCD consistently outperforms advanced SSCD methods. Notably, with only 20% labeled data on LEVIR-CD, our method achieves an $F1$ score of 91.17% and an intersection over union (IoU) of 83.78%. The code will be available at https://github.com/RSII-NWNUAMS-SemiCD