Frequency-Domain Heterogeneous Rank-Entropy Bipolarization Network for Remote Sensing Change Detection

Recently, deep neural networks have been extensively investigated for remote sensing image change detection and have achieved remarkable performance improvements. However, most change detection methods do not effectively disentangle content and style features, making them vulnerable to domain shifts caused by illumination and seasonal variations, thereby producing pseudo-changes. Moreover, existing methods tend to overfit background cues in uncertain regions, causing the decision boundary to rely excessively on background textures and introducing background-induced spurious correlations. To address these issues, we propose a frequency-domain heterogeneous rank-entropy bipolarization network (FHRB-Net) for change detection. Specifically, we propose a frequency-domain heterogeneous-modulation fusion module (FHFM) that enhances change information and suppresses style interference through heterogeneous frequency-aware modulation, and employs a cosine-gated fusion mechanism to enhance bi-temporal feature consistency and alleviate domain shifts. Furthermore, a rank-entropy bipolarization loss (REB-Loss) is designed to quantify uncertainty from the model's predicted probability map using a rank-based cumulative distribution function and Riemann-sum approximation and to reduce uncertainty by polarizing the probability distribution toward bipolar extremes to suppress background-induced spurious correlations. Experiments on several challenging remote sensing change detection (RSCD) datasets demonstrate that the proposed FHRB-Net yields the best overall performance. The code is publicly available at: https://github.com/lrz2025/FHRB-Net.