小组新闻

Heterogeneous remote sensing change detection (CD) based on optical and SAR imagery is essential for rapid disaster monitoring, as it overcomes single-sensor limitations and enables all-weather observation. However, the inherent heterogeneity between optical and SAR modalities—stemming from their disparate imaging mechanisms—results in poor feature comparability, significantly impeding fine-grained, accurate change identification. To address the issue, this study explores a combination of data-level strategies and methodological enhancements. First, to alleviate the data scarcity, this study constructs two sub-meter (0.35 m) heterogeneous datasets tailored for earthquake building damage and flood assessment, thereby providing a high-resolution data foundation for fine-grained disaster monitoring. Second, a Siamese Spatial–Frequency Coupled Network (SSFCNet) is introduced to mitigate the cross-modal discrepancy. Specifically, a Spatial-Frequency Coupled Encoder (SFCE) is designed to simultaneously model spatial semantics and frequency directional features via wavelet transforms, thereby aligning heterogeneous representations while mitigating noise. To refine change representation, an Explicit Change Guidance Module (ECGM) is designed to incorporate distinct difference priors into the decoding stage. Subsequently, a Multi-scale Inverse Wavelet Decoder (MIWD) exploits the invertibility of wavelet transforms to restore high-frequency details, preventing information loss during upsampling. Finally, a Boundary-Aware Loss (BAL) is devised to enforce geometric consistency along change boundaries through gradient constraints. Extensive experiments on five public datasets and the two newly constructed datasets demonstrate that SSFCNet significantly outperforms state-of-the-art methods in terms of detection accuracy and boundary integrity. Notably, the proposed method achieves an optimal balance between performance and computational efficiency, demonstrating its potential for rapid, robust emergency response. For reproducibility, our code and data are available at https://github.com/yangyang12318/SSFCNet.