CoreNet: A Semi-Supervised Change Detection Framework With Correlation-Guided Feature Refinement and Regional-Wise Pseudo-Label Reassignment

With the advancement of deep learning, Semi-Supervised Change Detection (SSCD) methods in remote sensing have made significant progress in reducing the reliance on labeled data. However, current SSCD methods still face two key challenges. Firstly, in complex remote sensing scenarios, most approaches fail to effectively model the spatial structures and semantic similarities embedded in the encoded features, resulting in insufficient sensitivity to changed regions, especially at boundaries. Secondly, existing methods typically rely solely on high-confidence thresholds during the process of pseudo-label generation, overlooking the potential of encoder features to directly constrain and enhance pseudo-label quality. Moreover, the inherent class imbalance often biases model predictions toward background regions, thereby hindering the generation of high-quality pseudo-labels. To address these issues, we propose a novel Semi-Supervised Change Detection Framework with Correlation-guided Feature Refinement and Regional-wise Pseudo-label Reassignment (CoreNet). Specifically, we design a Correlation-guided Feature Refinement mechanism (CFR), which models regional-level correlations between pixels based on the structural and semantic similarities of encoder features. This helps identify truly changed pixels, guides the model in perceiving spatial structural relationships and enhances its ability to refine boundaries of changed regions. In addition, the Regional-wise Pseudo-label Reassignment strategy (RPR) utilizes region-level feature correlations extracted by the CFR mechanism and combines them with confidence scores to construct a dual-filtering scheme. This RPR strategy dynamically updates pseudo-labels and maintains a reliable region mask, effectively mitigating class imbalance and significantly improving pseudo-label quality. Extensive experiments on four public datasets show that the proposed CoreNet achieved excellent performance at the 5% labeled data and on the metric IoU, it is 2.98%, 1.43%, 2.64%, and 4.23% higher than the second-best method across the respective datasets. Code will be available at https://github.com/Dufangyu/CoreNet.