Sampling and Machine Learning Methods for a Rapid Earthquake Loss Assessment System

Abstract

Rapid earthquake damage and loss assessment is crucial both for insuring the safety of inhabitants in the immediate aftermath of an earthquake and for the recovery of the stricken communities in the long run. This paper investigates the potential of different machine learning methods for building a rapid earthquake loss assessment system intended for residential houses in a municipal area. The system is trained on a pre-earthquake selected representative set of residential houses, after observing their damage and loss states. Two representative sampling strategies and three machine learning algorithms are described and evaluated on the 2010 Kraljevo M5.4 earthquake data set. The proposed models showed satisfactory accuracy in predicting the total expected repair cost (less than 20% error with the representative sample size of 10% of the inventory). The approach is independent of geological and earthquake data and does not require local peak ground acceleration values.