Integrating Local and Global Features via CNN and Mamba for Semi-Supervised Change Detection

Semi-supervised change detection (SSCD) has become increasingly important in remote sensing image (RSI) analysis due to the scarcity of labeled data. While state-of-the-art SSCD methods have achieved notable results through pseudo-labeling and weak-to-strong consistency regularization with various perturbations, they face several inherent limitations: 1) learners (models or architectures with the same design) often focus exclusively on either local or global features, which fails to capture the intricate bi-temporal difference feature; 2) low-quality pseudo-labels, often resulting from inadequate learner diversity or suboptimal perturbations, struggle to reflect accurate changes, exacerbating confirmation bias; and 3) a predominant emphasis on pixel-level consistency overlooks broader image context, limiting the ability to capture complex, large-scale spatiotemporal changes. To address these challenges, we propose a novel SSCD framework, cross-supervision with convolutional neural network and Mamba (CSCM), which adopts the convolutional neural network (CNN) and Mamba as two independent learners within a cross-supervision mechanism, enabling the integration of local and global feature representations. The CNN excels at capturing fine-grained local details, while Mamba efficiently models long-range dependencies with linear complexity, making it particularly well-suited for processing large-scale RSIs. To enhance the collaboration between these architectures, we introduce the cross-architecture fusion module (CAFM), which fuses difference features extracted by the CNN and Mamba, combining local sensitivity with global awareness to produce refined pseudo-labels. Additionally, we incorporate instance-level global consistency to capture broader image context, ensuring a more comprehensive understanding of spatiotemporal changes beyond pixel-level consistency. Extensive experiments on three public datasets demonstrate that our approach significantly enhances CD accuracy and training efficiency, outperforming several state-of-the-art methods.