Ensemble technique to predict post-earthquake damage of buildings integrating tree-based models and tabular neural networks

Abstract

In this paper, we develop a novel ensemble model for seismic building damage prediction that leverages machine learning algorithms of two completely different mechanisms, tree-based models and tabular neural networks. The ensemble model can break the limitations of individual models and elevate the predictive power to a superior level. Ensemble techniques of stacking and boosting are utilized for model integration, and an optimization method is also proposed to eliminate the negative sub models based on the stacking coefficient map and the boosting process. On the database of the 2015 Nepal earthquake including 762,094 building samples, our developed ensemble model using boosting technique outperforms any other individual models and its accuracy also surpasses that of another research using the same database. Moreover, the technique of fuzzy prediction is proposed to address the inherent errors in the database caused by volunteers’ subjectivity, which preserves the uncertainty between two adjacent damage grades, and the accuracy in the testing set is increased to 76.37% by applying it. Finally, by using the Shapley additive explanations method, our ensemble model is shown to have explicit physical significance and provides valuable insights for seismic mitigation efforts.