StyleDe-CD: Dual-Level Style Disentanglement for Semisupervised Change Detection

Semisupervised change detection (SSCD) is a promising approach for reducing the annotation burden in bitemporal remote sensing imagery. However, learning robust representations remains challenging due to style variations across and within temporal images, which often dominate visual differences and interfere with semantic change modeling. These challenges can be attributed to two levels of style-semantic entanglement: 1) cross-temporal style discrepancies caused by acquisition variations (e.g., illumination shifts and seasonal changes), which may confuse the model and increase sensitivity to superficial appearance differences; and 2) intratemporal style heterogeneity introduced by diverse background characteristics (e.g., land-cover diversity and vegetation density differences), which can bias feature learning and obscure meaningful semantic changes. To address these challenges, we propose StyleDe-CD, a dual-level style disentanglement framework for SSCD. At the input level, we design a stochastic spectralphotometric style augmentation (SSPSA) strategy, which combines spectral style alignment and photometric transformations to mitigate cross-temporal discrepancies and enhance robustness to appearance variations. At the feature level, we introduce an intratemporal style mixing (ISM) mechanism that randomly mixes feature statistics across samples within a minibatch, encouraging the model to suppress background-specific styles and focus on semantic changes. Extensive experiments on the LEVIR-CD, WHU-CD, and CDD benchmarks demonstrate that StyleDe-CD consistently achieves strong and competitive performance compared to existing semisupervised baselines, particularly under lowlabel regimes, indicating its effectiveness in SSCD.