Soil Velocity Models Informed by Remote Sensing and Artificial Intelligence

Abstract

Near-surface soil conditions can significantly alter the amplitude, duration, and frequency content of incoming ground motions – often with profound consequences for the built environment – and are thus important inputs to any ground-motion prediction. In current practice, the shear-wave velocity (Vs) time-averaged over the upper 30 m (Vs30) is widely used as a proxy for site effects, forming the basis of seismic site class and underpinning site-amplification factors in empirical ground-motion models (GMMs). Similarly, wave-propagation based GMMs rely on depth-continuous models of Vs, at least to where hard rock is found. Consequently, earthquake simulations rely on knowledge of either Vs30 or Vs-versus-depth (referred to herein as a soil velocity model, SVM), depending on which type of GMM is adopted. In either case, the need for these inputs at regional scale presents a challenge, given the infeasibility of subsurface testing over vast areas. At present, a patchwork of Vs30 models exists in the U.S., with the USGS National “baseline” model being a regression equation based on one parameter – topographic slope. Several regional models attempt to improve upon this baseline, generally by incorporating mapped geology into similar regression equations. Likewise, a sparse collection of SVMs, commonly called “Community Velocity Models,” exists in the U.S. These generally: (i) are available only in select urban regions; (ii) provide predictions with low spatial resolution; and (iii) focus more on deep geologic structure and less on near-surface conditions, even though the latter could alter motions most significantly. Given the growth of community geophysical and geotechnical datasets, satellite remote sensing, and artificial intelligence, unified and more accurate solutions, commensurate with the sophistication of emergent ground-motion models, are conceivable. Accordingly, this research develops U.S. National Vs30 and SVM solutions using machine- and deep-learning (ML/DL) techniques, wherein geospatial variables are used to predict subsurface wave velocities. These velocities lack theoretical links to above-ground parameters, but correlate in complex, interconnected ways – a prime problem for ML/DL. The resulting, trained models are tested against existing models using unbiased tests and shown to provide efficient predictions that match or best existing models. Ultimately, the proposed approach could be expanded using additional training data and new predictor variables, which were each relatively modest in the present study.