Updating Global Geospatial Liquefaction Models With a Focus on Feature Engineering
Description:
This work updates the global geospatial liquefaction models developed by Zhu et al. (2017) by significantly expanding the database to include fifty-four earthquakes and thirty candidate geospatial predictors. Zhu et al. (2017) developed global geospatial liquefaction models using liquefaction inventories for twenty-seven worldwide earthquakes and seventeen candidate geospatial predictors. In order to develop multivariate logistic regression models for liquefaction classification, multiple feature engineering techniques are employed to find the best combinations of geospatial predictors. First, we apply the feature filter to remove low-correlated and redundant predictors based on different correlation measures. Also, we select optimal predictors from a group of geospatial predictors with the same physical meanings but derived using different geospatial products. After the feature filter, ten predictors are retained in the predictor pool for developing the multivariate models. Then, two feature engineering strategies, exhaustive feature selection and principal component analysis (PCA), are applied to select the best predictor combinations and investigate the information loss. The results suggest that the exhaustive feature selection method achieves less information loss and better interpretability than the PCA method for this case. Based on exhaustive feature selection results, we recommend three models that use different geospatial predictors representing site vulnerability to liquefaction and have good classification accuracies. We conclude that such feature engineering methods could be applied to other geospatial modeling tasks that can benefit from finding the optimal predictor combinations.
Session: Coseismic Ground Failure: Advances in Modeling, Impacts and Communication
Type: Oral
Date: 4/20/2023
Presentation Time: 08:30 AM (local time)
Presenting Author: Weiwei Zhan
Student Presenter: No
Invited Presentation:
Authors
Weiwei Zhan Presenting Author Corresponding Author weiwei.zhan@austin.utexas.edu University of Texas at Austin |
Laurie Baise laurie.baise@tufts.edu Tufts University |
Babak Moaveni babak.moaveni@tufts.edu Tufts University |
|
|
|
|
|
|
Updating Global Geospatial Liquefaction Models With a Focus on Feature Engineering
Category
Coseismic Ground Failure: Advances in Modeling, Impacts and Communication