KAN-UG: Kolmogorov–Arnold Network-Based Difference Feature Enhancement Method With Uncertainty Guidance for Change Detection

Although existing semi-supervised methods for remote sensing change detection (RSCD) have shown promising performance with limited labeled data, they still face two key challenges. First, due to the complexity of remote sensing (RS) scenarios, these methods often struggle to capture subtle changes of ground objects and exhibit limited capacity to model the nonlinear characteristics of complex objects, thereby restricting detection accuracy. Second, existing methods fail to adequately exploit the collaboration between labeled data and unlabeled data, which tends to cause error accumulation and consequently leads to insufficient representation of challenging regions. Recently, Kolmogorov–Arnold network (KAN) has attracted increasing attention for its strong capability in modeling complex nonlinear data distributions. To overcome these challenges, we propose a Kolmogorov–Arnold network (KAN)-based difference feature enhancement method with uncertainty guidance (KAN-UG) for semi-supervised RSCD. First, a KAN-based difference feature enhancement module (K-DFEM) is constructed to better capture subtle changes in complex scenarios. This module enhances the discriminability of features in complex scenarios through learnable nonlinear activation and progressive multiscale fusion. Second, an uncertainty-guided adaptive learning (UAL) mechanism is designed to prevent the model from being misled during optimization. This mechanism promotes effective collaboration between supervised and unsupervised learning by improving the utilization of challenging regions in labeled data and mitigating error propagation from challenging regions in unlabeled data. Experimental results on four widely used datasets demonstrate that our method outperforms state-of-the-art semi-supervised change detection (SSCD) methods across diverse and challenging scenarios. In particular, with only 5% labeled ratio, our method achieves intersection over union (IoU) improvements of 7.42% and 3.70% over the best competing method on the CDD and SYSU-CD datasets, respectively. The code will be available at https://github.com/CAG-kong/KAN-UG.