Host-to-target region testing of machine learning models for seismic damage prediction in buildings

Assessing or predicting seismic damage in buildings is an essential and challenging component of seismic risk studies. Machine learning methods offer new perspectives for damage characterization, taking advantage of available data on the characteristics of built environments. In this study, we aim (1) to characterize seismic damage using a classification model trained and tested on damage survey data from earthquakes in Nepal, Haiti, Serbia and Italy and (2) to test how well a model trained on a given region (host) can predict damage in another region (target). The strategy adopted considers only simple data characterizing the building (number of stories and building age), seismic ground motion (macroseismic intensity) and a traffic-light-based damage classification model (green, yellow, red categories). The study confirms that the extreme gradient boosting classification model (XGBC) with oversampling predicts damage with 60% accuracy. However, the quality of the survey is a key issue for model performance. Furthermore, the host-to-target test suggests that the model’s applicability may be limited to regions with similar contextual environments (e.g., socio-economic conditions). Our results show that a model from one region can only be applied to another region under certain conditions. We expect our model to serve as a starting point for further analysis in host-to-target region adjustment and confirm the need for additional post-earthquake surveys in other regions with different tectonic, urban fabric and socio-economic contexts.