AEM-KD: Adversarial Entropy Minimization for Cross-Modality Knowledge Distillation in SAR Building Segmentation

作者：Hao Pei, Peishuang Ni, Yiguo Qiao, Jianlai Chen, Wenkang Liu, Hanwen Yu, Fan Wu, Gang Xu, Meng-Dao Xing · 发表于：IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing · 年份：2026 · DOI：10.1109/jstars.2026.3686377

Building segmentation in synthetic aperture radar (SAR) imagery is critical for numerous remote sensing applications. Nevertheless, the inherent limitations of SAR imagery, such as speckle noise, complex scattering effects, and geometric distortions, pose significant challenges to accurate segmentation. Existing research introduces auxiliary information or handcrafted priors to improve segmentation performance, but often suffers from increased model complexity and poor generalization. Recently, knowledge distillation has emerged as a promising alternative, as the distillation of optical image semantics provides transferable spatial structural knowledge to enhance SAR segmentation, particularly in complex imaging environments. Nevertheless, conventional distillation methods rely on pixel-wise constraints, failing to accommodate the significant heterogeneity between SAR and optical imagery arising from their distinct imaging mechanisms. To overcome these challenges, we propose adversarial entropy minimization knowledge distillation (AEM-KD), which distills logit-level knowledge via adversarial learning at the prediction layer. By leveraging structural–semantic correlations between heterogeneous modalities, AEM-KD encourages the SAR network to produce confident, low-entropy predictions consistent with those of the optical teacher, while avoiding rigid pixel-wise constraints. To further reduce feature-level discrepancy, a channel-wise knowledge distillation (CW-KD) module is intr...