Improving the Low Accuracy of Traditional Earthquake Loss Assessment Systems

Abstract

Rapid earthquake loss assessment implies near real-time prediction of damage to the affected structures and monetizing loss at the community level. The accuracy and speed of prediction are the main quality features of rapid loss assessment systems. The problem with traditional systems is their low accuracy, making them unreliable and unusable in the recovery process, which is their primary purpose. Low accuracy is caused by significant uncertainty in analytical and insufficient data sets to create vulnerability curves in empirical methods. The root cause of low accuracy in analytical methods is assuming theoretical vulnerability relations before an earthquake. We propose a new kind of rapid earthquake loss assessment system which uses trained assessors to perform on-the-ground observation of actual damage on the representative sample after an earthquake. Machine learning methods are then used to predict damage for the remaining building portfolio, which is more accurate and still rapid enough. The contributions of this research are: the procedure of representative sampling for creating an informative and sufficient representative set, discovering the minimum building representation that uses only location and building geometry attributes, introducing the soft rule formula for monetizing loss, and combining those elements into a novel, usable framework. Using a building representation without earthquake data eliminates the need for analytical methods, shake maps and robust ground motion sensor networks, making the proposed framework unique and applicable in any region. All findings were verified using the M5.4 2010 Kraljevo earthquake data. Most importantly, a 14% relative error for predicting repair cost was obtained using a 10% representative set. This level of precision is acceptable and significantly better than in traditional systems for loss assessment.