Machine-learning-aided regional post-seismic usability prediction of buildings: 2016–2017 Central Italy earthquakes

Abstract

This study addresses the development of machine learning models for predicting the post-seismic buildings usability within regions prone to frequent earthquakes. The analysis leverages on field data from 2016-2017 Central Italy earthquakes. Several machine learning techniques are employed for this task, namely K-nearest neighbors, linear support vector machine, radial basis function support vector machine, decision tree, random forest, neural network, adaptive boosting, naive Bayes, quadratic discriminant analysis, logistic regression, and linear discriminant analysis. The input variables include both building attributes and seismic intensity measures. Since the database turns out to be strongly imbalanced, the potential influence of two preprocessing techniques is examined, namely principal component analysis and synthetic minority oversampling technique. Several metrics are considered to evaluate the performance of the resulting predictive machine learning models. Moreover, this study investigates the optimal machine learning model’s robustness against uncertainties, quantifies the importance of its features, and investigates how usability classes clustering can impact its performance. Every step of the implemented procedure is deeply explained and discussed to provide useful guidelines for similar applications.