Testing machine learning models for seismic damage prediction at a regional scale using building-damage dataset compiled after the 2015 Gorkha Nepal earthquake

Abstract

Assessing post-seismic damage on an urban/regional scale remains relatively difficult owing to the significant amount of time and resources required to acquire information and conduct a building-by-building seismic damage assessment. However, the application of new methods based on artificial intelligence, combined with the increasingly systematic availability of field surveys of post-seismic damage, has provided new perspectives for urban/regional seismic damage assessment. This study analyzes the effectiveness and relevance of a number of machine learning techniques for analyzing spatially distributed seismic damage after an earthquake at the regional scale. The basic structural parameters of a portfolio of buildings and the post-earthquake damage surveyed after the Nepal 2015 earthquake are analyzed and combined with macro-seismic intensity values provided by the United States Geological Survey ShakeMap tool. Among the methods considered, the random forest regression model provides the best damage predictions for specified ground motion intensity values and structural parameters. For traffic-light-based damage classification (three classes: green-, amber-, and red-tagged buildings based on post-earthquake damage grade), a mean accuracy of 0.68 is obtained. This study shows that restricting learning to basic features of buildings (i.e. number of stories, height, plinth area, and age), which could be readily available from authoritative databases (e.g. national census) or field-surveyed databases, yields a reliable prediction of building damage (4 features/3 damage grade accuracy: 0.64).

Machine Learning Assessment of Damage Grade for Post-Earthquake Buildings: A Three-Stage Approach Directly Handling Categorical Features

“…When Tr equals q, the values of Nfl are typically greater, and the SHAP value may be positive or negative, implying a weak interaction. The results and discussion in this stage provide a more informative and compact interpretation of the best-performing model compared to previous similar studies [15,16,59]. By directly coping with categorical features rather than employing one-hot encoding, information fragmentation attributed to transforming a categorical feature into several binary features is prevented.…”

Probing the Evolution of Fault Properties During the Seismic Cycle with Deep Learning

“…Existing work suggests that fault zone properties evolve during the seismic cycle in response to stress changes and microcracking prior to rupture with subsequent post-seismic healing 1,2 . Such changes are observed commonly in lab experiments [3][4][5][6][7][8][9][10][11][12][13][14][15] and őeld data conőrm these expectations in some cases, showing changes in elastic wave speed prior to earthquake fault slip, volcanic activity and landslides [16][17][18][19][20][21][22] . However, distinguishing subtle changes in seismic behavior or fault properties prior to and after earthquakes, even in locations with dense seismic networks, is challenging [23][24][25][26][27][28][29][30][31] .…”

Applying Machine Learning to Earthquake Engineering: A Scientometric Analysis of World Research

“…In the field of damage assessment, compared to pre-earthquake structures, the damage phenomena observed on post-earthquake damaged structures are more severe and intricate. The existing post-earthquake damaged structural databases suitable for deep learning are still insufficient [98], thereby constraining the progress of relevant research. In the field of residual performance assessment for post-earthquake damaged structures, researchers similarly encounter a deficiency of reliable training data to effectively train ML models.…”