Building damage detection from multi-feature fusion of Sentinel-1/2 imagery using variational autoencoder and MLP-Mixer network: insights from the Jishishan earthquake

Timely and accurate identification of earthquake-induced building damage is critical for effective emergency response and post-disaster reconstruction. However, existing methods often rely on single or limited Synthetic Aperture Radar (SAR) features, leading to reduced detection accuracy. To address this, we propose a novel framework that fuses multi-source features from SAR and optical imagery. The approach integrates a Variational Autoencoder (VAE) with an improved MLP-Mixer network to extract compact, discriminative representations. At the macro scale, an enhanced differential image method is first employed to delineate damaged areas. Subsequently, diverse building-level features—including SAR backscatter differences, polarimetric decomposition (H/A/Alpha), InSAR coherence and displacement, as well as the optical-based Normalized Difference Built-up Index (NDBI)—are incorporated into a comprehensive feature repository. Feature importance is further analyzed using XGBoost to assess each modality’s contribution. Validation on the 2023 Jishishan earthquake dataset shows that the proposed method achieves 82.4% accuracy, an F2-score of 76.6%, and a ROC-AUC of 0.862, outperforming baseline approaches. These results confirm that the integration of multi-modal features with generative representation learning significantly enhances damage detection performance. The proposed framework provides a robust and practical tool for post-earthquake building damage assessment, offering valuable support for disaster management and emergency response.