Earthquake risk assessment using artifical intelligence

Abstract

Earthquakes are considered one of the most unpredictable and destructive natural disasters, posing real threats to human life and socio-economic systems. While earthquakes themselves cannot be prevented, this study confirms that effective disaster management—supported by rapid and reliable post-earthquake damage assessment— can significantly reduce casualties and limit social and economic losses. Timely damage estimation is essential for guiding emergency response, prioritizing inspections, and allocating resources efficiently. In this context, machine learning has gained increasing attention over the past decade because of its strong ability to process large and complex datasets and to capture nonlinear relationships among seismic, structural, and geospatial factors. This thesis develops and validates a machine learning–based framework for postearthquake building damage estimation using real field inspection data that reflects actual observed damage and structural behavior. The field data are integrated with scenario-based ground-motion parameters generated using REDAS software, providing consistent seismic demand estimates at building locations. In addition, geospatial features extracted from open-source databases are incorporated to enrich the dataset. The proposed framework is evaluated through two major earthquake case studies in Turkiye: the 2020 Izmir earthquake and the 2023 Kahramanmaras earthquake sequence. The Izmir earthquake serves as a well-documented benchmark for model evaluation. Post-earthquake inspection data were combined with seismic and geospatial information to train and test multiple machine learning algorithms. Two dataset configurations were examined to assess the influence of feature selection. The results clearly demonstrate that including building age as an input parameter significantly improves model performance, increasing prediction accuracy to nearly 90% compared to approximately 80% when this feature is excluded. Confusion matrix analysis shows that models incorporating building age are far more effective at identifying severely damaged buildings, emphasizing the critical role of this parameter in representing structural vulnerability. The Kahramanmaras earthquake sequence represents an extreme seismic event affecting a wide region and diverse building stock. The corresponding dataset includes both masonry and reinforced concrete buildings and is evaluated using binary, threeclass (traffic-light), and four-class damage classification schemes at building, district, and city scales. While binary classification achieves relatively high accuracy, it results in an unacceptably high rate of unsafe buildings being misclassified as safe, limiting its practical use. The four-class scheme provides more detailed damage information but suffers from substantial confusion between adjacent damage levels. In contrast, the three-class traffic-light classification demonstrates the most reliable and balanced performance, with low confusion between safe and unsafe buildings. Among the tested algorithms, the Bagging classifier shows the most robust and consistent results for this classification scheme. At district and city scales, the proposed models successfully reproduce observed damage distributions, confirming their suitability for large-scale damage assessment and regional decision-making. Overall, this thesis presents a practical and transferable machine learning framework for rapid post-earthquake damage estimation based on real observed data. The findings indicate that the three-class traffic-light damage classification offers the best balance between accuracy, clarity, and operational relevance for emergency response. The proposed approach provides a realistic and scalable solution that can support disaster management efforts and contribute to reducing human and economic losses in earthquake-prone regions.