A DEEP LEARNING SYSTEM FOR BUILDING DAMAGE ASSESSMENT USING VHR COSMO-SKYMED IMAGERY FOR THE 2023 KAHRAMANMARAŞ EARTHQUAKE

Natural disasters like earthquakes require quick and accurate responses to minimize human and economic losses. This paper presents a method for rapid earthquake damage assessment using very high-resolution (VHR) COSMO-SkyMed (CSK) Synthetic Aperture Radar (SAR) imagery and Deep Learning (DL). In a two-phase pipeline, a lightweight, differencing-based Siamese U-Net with Dice regularization was first trained for binary change detection (CD) and then fine-tuned for multiclass building damage assessment. Training employed pre- and post-event COSMO-SkyMed (CSK) imagery over Malatya and Osmaniye (Türkiye), with ground truth (GT) obtained by combining Copernicus Emergency Management Service (CEMS) activation maps and OpenStreetMap (OSM) building polygons from the Humanitarian Data Exchange (HDX) Team. Notably, the Siamese U-Net achieved the best overall performance and outperformed larger models like MS4D-Net and transformer-based architectures such as SegTransformer. Ignoring the Dice term caused a clear drop in accuracy, highlighting its benefit for limited, imbalanced training data. These results confirm that a lightweight Siamese U-Net with Dice loss produces robust and accurate building damage maps, offering a scalable alternative to traditional labor-intensive visual assessments methods. Future work will expand the labeled dataset and investigate domain-adaptation techniques to ensure transferability across different regions and disaster scenarios.