2021 Volume 21 Issue 2 Pages 2_70-2_89
The purpose of this study is to generate nationwide maps for liquefaction susceptibility. Random forest, which is one of the machine learning methods, was selected to solve a binary classification of liquefaction as opposed to non-liquefaction. Our dataset consisted of 16 variables related to soil density, soil saturation and earthquake ground motion. Furthermore, the dataset has a highly imbalance problem of the classes, because the number of approximately 18,000 cells are liquefaction grid cells while the number of approximately 115 million cells are non-liquefaction grid cells. To solve the imbalance problem, we proposed an ensemble method combining under-sampling. As a result, the proposed method achieved an overall accuracy score of 95.1%, a recall score of 91.4% and a precision score of 7.6% with the imbalanced data. Finally, a parametric study based on seismic intensity was conducted to create liquefaction susceptibility maps.