Development of Data-driven Models to Predict Vs30 from mHVSR

Abstract

This work investigates the potential of using microtremor horizontal-to-vertical spectral ratio (mHVSR) measurements to predict the time-averaged shear-wave velocity in the upper 30 meters (VS30) using data-driven models. We develop a dataset comprising 536 sites with 2,861 three-component ambient noise recordings from global regions, including New Zealand, Taiwan, Italy, Ecuador, Mexico and the United States. The identically processed three-component ambient noise recordings are used to make mHVSR measurements. To predict VS30 from mHVSR, we consider two types of models: low-dimensional models, which use features of the mHVSR curve such as the fundamental site frequency (f0,HVSR) and peak amplitude (A0,HVSR), and high-dimensional models, which use the entire mHVSR mean curve to predict VS30. In addition, we integrate topographic features from a 1 arc-second digital elevation model (DEM) for both model types using binned elevation as a proxy for geologic composition and relative elevation as a proxy for topography. The low-dimensional models are shown to reasonably predict VS30, coefficient of determination (R²) up to 0.69 on the testing set, when considering both mHVSR and topographic features. The high-dimensional models are shown to achieve improved accuracy to the low-dimensional models (R² of 0.82 on the testing set) when using the mean mHVSR curve regardless of whether the mean mHVSR curve is supplemented with the additional topographic features. These findings demonstrate that while low-dimensional features of the mHVSR curve are informative, leveraging the full shape of the mHVSR curve leads to improved prediction of VS30. Furthermore, the use of the mHVSR mean curve has the additional advantage that it does not require the extraction of features and can be used at all sites including those with and without resonant peaks. We compare the results with a model developed to predict VS30 from remote sensing data and demonstrate the utility of the models by predicting VS30 at 1,855 broadband recording stations across North America.