Multi-level earthquake damage assessment using InSAR-derived displacement

Spaceborne Synthetic Aperture Radar (SAR) has gained prominence for post-earthquake damage assessment due to its ability to acquire high-resolution imagery under all weather and lighting conditions. However, most existing SAR-based studies rely on binary Machine Learning (ML) classification, distinguishing only between collapsed and intact buildings, while moderate levels of damage are typically overlooked. Although some recent studies have explored multi-level damage classification, their effectiveness is often limited by the lack of physically meaningful indicators that are directly linked to structural damage. To address this limitation, this study proposes the first implementation of an ML framework that integrates displacement information derived from Interferometric SAR (InSAR), providing a more direct and accurate measure of structural deformation and failure mechanisms shortly after an earthquake. A supervised learning model is developed using InSAR-derived displacement and coherence metrics, together with structural and site-related features. Two progressive components are implemented: (i) a region-specific learning scheme, in which models are trained and tested on individual earthquake events, and (ii) a combined learning scheme that integrates data from multiple events to assess the model’s ability to generalise across regions. The proposed framework is applied to post-event datasets from the 2015 Nepal earthquake and the 2016 Central Italy earthquake sequence. Results demonstrate the effectiveness of InSAR features in multi-class damage mapping. Region-specific models effectively capture event-driven damage patterns, while cross-event analyses suggest robust performance across the three studied earthquake events. These findings highlight the potential of InSAR-driven ML approaches as reliable and scalable tools for rapid post-earthquake damage assessment and improved disaster response.