The use of machine-learning methods for post-earthquake building usability assessment: A predictive model for seismic-risk impact analyses

Abstract

The assessment of building usability in the aftermath of an earthquake is mostly aimed at post-event emergency management, but it is also valuable for the planning of risk-reduction policies. In the seismic risk assessment field, the development of suitable consequence functions that correlate physical damage to usability and serviceability of structures is crucial to evaluate the expected social and economic losses in a region of interest. Predictive models for usability classification generally are calibrated on empirical data and provide the probability of loss of usability as function of the intensity measure, the building type and the severity of damage attained by the structure. Exploiting the large amount of data available in Italy, a decision tree-based approach is proposed in this study to assess post-earthquake usability of ordinary buildings. Thanks to its high interpretability coupled with reasonable predictive capability _, the selected machine learning algorithm allows investigation of the structural parameters that have a significant impact on building usability, while also accounting for the traditionally neglected uncertainty of subjective decisions. Finally, to show the potential of the proposed usability consequence models, a large-scale risk analysis is carried out to evaluate the spatial distribution of expected building-usability losses over time.