Earthquake damage and rehabilitation intervention prediction using machine learning

Abstract

Predicting damage grade and rehabilitation interventions is important, especially in the aftermath of moderate to strong earthquakes as prioritization of post-earthquake housing recovery needs information regarding the damage extent. Damage prediction is generally performed using fragility functions, which are generally associated with large uncertainties. Moreover, availability and representativeness of fragility functions for a region affected by an earthquake is not always a given. A more realistic prediction of damage might be obtained from methods that rely on relevant attributes of affected buildings. Artificial intelligence-based formulations have huge prospect in this regard. Using the ground shaking intensity measure and detailed building specific features of 549,251 buildings affected by the 2015 Gorkha earthquake in Nepal, this paper assesses efficacy of four common machine learning algorithms for damage grade and rehabilitation intervention prediction. Decision tree, random forest, XGBoost, and logistic regression algorithms are used to prepare machine learning models and test their performance. The XGBoost algorithm is found to predict building collapse and strengthening more accurately than the other algorithms. Moreover, feature importance from the XGBoost model identifies 19 of the top 20 most important features as relevant for both damage grade and rehabilitation intervention prediction.