A Set of Data-Driven Models to Predict VS30 from the Horizontal-to-Vertical Spectral Ratio of Microtremors

ABSTRACT We investigate the use of microtremor horizontal-to-vertical spectral ratio (mHVSR) measurements to predict the time-averaged shear-wave velocity in the upper 30 m (VS30) using data-driven machine learning models. We develop a global dataset comprising 2861 three-component ambient noise recordings made at 536 sites. We consider two different types of predictive models: low-dimensional models, which use mHVSR features such as the fundamental site frequency (f0,HVSR) and amplitude (A0,HVSR), and high-dimensional models, which use the entire lognormal median mHVSR curve to predict VS30. In addition, we integrate two topographic features: binned elevation, as a proxy for geologic composition, and topographic position index, as a proxy for topography features. The low-dimensional models are shown to reasonably approximate VS30 when using mHVSR and the topographic features (coefficient of determination, R2, up to 0.69 on the testing set). The high-dimensional models outperform the low-dimensional models (R2 of 0.82 on the testing set) regardless of whether the lognormal median mHVSR curve is supplemented by the topographic features. In addition, we show that our top-performing model outperforms a recently developed model to predict VS30 from remote sensing data at sites from our testing set. Finally, we apply our models to predict VS30 from mHVSR computed at 1855 permanent broadband stations. We compare our data-driven VS30 predictions at a subset of 156 stations with measured VS30 values and find a reduction in the recommended model’s predictive power (R2 of 0.18). The reduction in performance at permanent broadband stations is shown to be related in part to differences in instrument type and data distribution between the temporary and permanent broadband datasets. Therefore, the proposed models are expected to perform best when mHVSR is measured using temporary deployments of broadband sensors. All the data-driven models developed in this study have been made publicly available.