Rapid Assessment of Blast-Induced Structural Damage Using a Physics-based Multimodal Network

Abstract

Accurate and rapid structural damage assessment (SDA) plays a vital role in post-disaster management, which supports emergency responders and decision-makers to prioritise resources, plan rescue operations, and support recovery efforts. Traditional field investigation and ground-based inspections, while offering high precision, are often constrained by limited accessibility, safety risks, and significant time requirements, particularly in the aftermath of large-scale explosions. Current large-scale SDA machine learning approaches typically rely on extensive human-annotated post-event remote sensing data, which can fail to identify structures that are internally compromised but remain outwardly intact. To tackle these challenges, this paper presents a physics-guided multimodal SDA pipeline that integrates numerically generated blast-loading information with pre- and post-event optical remote sensing imagery. Rather than proposing a wholly new architectural family, the key contribution is the event-specific physical guidance introduced into rapid blast-induced SDA through a transfer-learning strategy. By incorporating physical loading conditions, the assessment can help improve the identification of intermediate damage states, thereby enhancing the accuracy and realism of real-world structural damage evaluation. We evaluate the methods on both an image-interpreted dataset and an in-situ structural damage dataset derived from the aftermath of the 2020 Beirut explosion. The results indicate that incorporating blast-loading information significantly enhances SDA performance, outperforming state-of-the-art approaches. The code is available at: https://github.com/IMPACTSquad/Blast-Mamba.