Machine learning for enhancing tsunami damage assessment in coastal areas: insights from the 2011 Great East Japan event

Accurate assessment of tsunami-induced damage is crucial for effective disaster risk management. Traditional methods relying on univariate fragility functions often fail in capturing the complex interplay of factors influencing tsunami damage. To address such limitation, this study applies machine learning techniques to two extensive building and road damage datasets collected after the 2011 Great East Japan tsunami. The original datasets, compiled by the Japanese Ministry of Land, Infrastructure and Transportation, include detailed damage information and ancillary data for over 250,000 impacted buildings and more than 4,000 km of inundated roads. These datasets have been enriched in this study with additional explanatory variables, such as hydraulic features of the event, structural and geometrical properties of the exposed assets, and characteristics of the surrounding environment. A buffer-based approach is employed to account for reciprocal interactions between structures, such as shielding effects of buildings that might mitigate water impact on adjacent structures, and the presence of collapsed buildings that may generate debris. Developed models integrate these factors to improve tsunami damage predictions and the results show that while inundation depth remains a critical predictor, other variables also play significant roles, with coastal topography and interactions among neighboring structures crucial for buildings, and wave approach angle, road orientation and potential overflow from inland watercourses for roads.