Testing machine learning models for rapid building damage assessment at regional scale.

Assessing or forecasting seismic damage to buildings is crucial for earthquake disaster management. Several classical damage assessment methods are available for seismic damage assessment by combining hazard, exposure, and vulnerability. However, during emergencies, collecting all the necessary data for seismic damage assessment may not be feasible due to time and resource constraints, as this information may not be readily available.

In this context, machine learning methods can offer a paradigm shift by reasonably assessing damage by relying on readily available data cost-effectively. In this study, we aim to study the damage prediction efficacy of machine learning models for regional scale damage assessment. Machine learning models were trained and tested on the post-earthquake building damage database.

Results show that the readily available building features such as the number of stories, age, floor area, and height can result in a reasonable assessment of damage at a large scale, mainly when using a traffic-light-based (green, yellow, and red) damage classification framework.

The machine learning models trained on past earthquake building damage portfolios can reasonably estimate damage during the future earthquake for a different region with similar building portfolios.