Using Gaussian Process Models for Dynamic Post-Earthquake Impact Estimation with Regional Risk Predictors

Abstract

The widespread damage to the built environment that may be caused by earthquakes, not only leads to direct consequences and losses, but also disrupts safe functionality of buildings and thus, potentially induces severe long-term societal consequences. Better community resilience may be achieved through well-organized recovery, which is complicated by the conditions, in which decisions need to be taken, characterized by intense time pressure and limited information on the severity and the spatial distribution of building damage. Approximate initial damage estimates are produced using regional risk models, which convolute early ground-motion data with indirect mapping schemes of buildings to typological classes and with generic building fragility functions. This domain-knowledge is integrated into Gaussian process models, a probabilistic machine-learning tool, to leverage the continuous data inflow from a building inspection campaign. Initial post-earthquake damage estimates are dynamically improved by simultaneously updating the distributions of ground-motion intensity, typological attribution, and building damage. Hence, the inspection information becoming available in the first days following an earthquake helps constraining uncertainties and provides reliable estimates of the geographical distribution of damaged buildings after a fraction of the time required to inspect all the buildings. The performance of the proposed methodology is demonstrated on a fictitious earthquake scenario and two real damage datasets from historic earthquakes and a comparison with purely data-driven methods shows that the underlying risk model reduces the number of building inspections that is required to provide reliable and precise predictions.