An Interoperable Framework for Real-Time Infrastructure Monitoring and Deep Learning–Based Post-Disaster Damage Assessment for Urban Resilience

Rapid infrastructure damage assessment is essential for effective emergency response, recovery planning, and urban resilience following extreme hazard events. Conventional post-disaster damage assessment relies heavily on manual field surveys, which are slow, labor-intensive, and often infeasible during early response phases due to safety risks and limited accessibility. These constraints delay decision-making and hinder timely resource allocation. Although recent advances in computer vision and deep learning enable scalable automation, existing approaches remain fragmented across systems and disciplines, inconsistently evaluated, and insufficiently aligned with engineering damage taxonomies, operational constraints, and interoperability standards.This dissertation presents an interoperable, multi-scale (system-level, image-level, and object-level) framework that integrates real-time infrastructure monitoring with deep learning–based damage assessment to support rapid disaster response and urban resilience. The research systematically examines how standards-based interoperability enables real-time infrastructure monitoring, and how architectural design choices, optimization strategies, and supervision formulations affect the accuracy, efficiency, and ordinal consistency of damage assessment models.The research makes four primary contributions. First, a comprehensive state-of-the-art review identifies a technology–resilience paradox in smart cities and demonstrates that hybrid approaches combining advanced analytics with traditional systems provide the most effective pathway to urban resilience. Second, a standards-based IIoT flood monitoring system is developed using low-cost sensors and OpenO&M OIIE™ specifications, enabling location-specific, real-time data collection and cross-platform interoperability for critical infrastructure resilience. Third, a large-scale experimental study evaluates 79 publicly available deep learning architectures across more than 2,300 controlled experiments, establishing the QSTD benchmark for post-disaster building damage classification. Results reveal that optimization strategies, particularly optimizer selection, can be more influential than architectural choice, with ConvNeXt-Base achieving 68.0% macro F1-score and maintaining 85.5% ordinal accuracy under cross-event generalization. Fourth, TornadoNet, a benchmark for real-time building damage detection incorporating ordinal-aware supervision, is introduced to explicitly model the graded nature of structural damage severity. Comparative evaluation shows architecture-dependent responses, with transformer-based RT-DETR achieving 4.8 percentage point improvement through soft ordinal targets, while anchor-free CNN detectors exhibit limited sensitivity to ordinal supervision.Overall, the framework outlined in this dissertation provides actionable guidelines for emergency management agencies and smart-city practitioners, while advancing transition toward an AI-enabled urban resilience.

Grantor dc:publisher

University of Alabama Libraries

Year dc:date.issued

2026

Author dc:creator

• Umeike, Chibuike Robinson

Advisor dc:contributor.advisor

• Dao, Thang N.

Contributors dc:contributor

• Crawford, Patrick S
• Wang, Jialai
• Hong, Xiaoyan
• Gong, Jiaqi
• Nguyen, Dang

• Automated Damage Assessment
• Computer Vision
• Deep Learning
• Real-Time Infrastructure Monitoring
• System Interoperability
• Urban Resilience

Statement dc:rights

• All rights reserved by the author unless otherwise indicated.

Language dc:language.iso

en_US, English

Dc Identifier Other

1234368

OAI identifier oai:identifier

oai:ir.ua.edu:123456789/18092