Prediction of Average Shear Wave Velocity of Ground in Japan Using Machine Learning Methods

The purpose of this study is to generate a nationwide map of average shear-wave velocity of ground upper 30 m (Vs30). Before estimating the nationwide Vs30, we predicted local Vs30 for expanding the data from a shallow velocity profile. Therefore, two types of predictions, which are nationwide Vs30 using terrain features and local Vs30 using a shallow velocity profile, were executed in this study. These predictions utilized machine learning methods: Support Vector Machine, Random Forest, and Gradient Boosting Decision Tree. As a result, relatively high accuracy was achieved. To explain the interpretability of the models for nationwide Vs30, three analyses were implemented. Analyses include important features for prediction; correlations between prediction and terrain features; and model sensitivity to slope and elevation. The important features are elevation, slope, distance from the tertiary mountains and geomorphologic classification similar to the explanatory variables utilized in Matsuoka et al. (2005). Correlation between results and features is reflected in the physical hypothesis with the exception being the result for tertiary mountain. In the sensitivity analysis results for elevation, there was a contradiction to the physical hypothesis in some geomorphologic classification such as tertiary mountain. Finally, the nationwide Vs30 map was generated after substituting the Vs30 estimated by Matsuoka et al. (2005) for the case of unintended results in the geomorphologic classification.