← Back to Sessions

Global Geospatial Modeling of Earthquake-Induced Soil Liquefaction Using a System of Voting Machine Learning Classifiers

Description: 

Data-driven geospatial liquefaction models are useful tools for regional seismic hazard assessments. The models are based on liquefaction occurrence inventories, widely available geospatial variables, and earthquake-specific parameters. Our inventory is updated with geospatial data from non-liquefaction and liquefaction occurrence locations in 54 earthquakes around the world. Geospatial data includes 2 categorical and 28 continuous variables representing proxies for soil saturation, soil density, and earthquake loading. In our prior work, logistic regression was used to present an updated global geospatial liquefaction model. In this study, we evaluate the use of an ensemble of decision trees as an alternative advanced machine learning (ML) algorithm to find complex nonlinear patterns in a large dataset. The proposed methodology starts with an exploratory data analysis to remove highly correlated features and run feature transformations. Neighborhood component analysis is also implemented as an ML-based feature selection approach to remove the variables with low weight on the classification model. The liquefaction inventory is highly imbalanced in terms of liquefaction and non-liquefaction classes and in terms of the data provided by the event. The class imbalance issue is treated in an innovative manner by distributing the event datasets over several balanced subsets. On each data subset, the binary classification model is trained and validated via a K-fold cross-validation approach. Based on the designed voting system of classifiers, the final class assignment for each sample point is done by considering the majority votes of the classifiers as the final prediction. To check the model reliability and potential bias, the leave-one-out testing approach is used to independently test the model by removing one earthquake at a time. Comparing the results of the proposed method with the logistic regression model showed that the overall accuracy and spatial extent of the liquefaction predictions were improved for most of the earthquakes tested.

Session: Coseismic Ground Failure: Advances in Modeling, Impacts and Communication

Type: Oral

Date: 4/20/2023

Presentation Time: 08:45 AM (local time)

Presenting Author: Adel Asadi

Student Presenter: Yes

Invited Presentation: 

Authors