Estimating S-wave velocity profiles from horizontal-to-vertical spectral ratios based on deep learning

S-wave velocity (Vs) profile or time averaged Vs to 30 m depth (V_S30) is indispensable information to estimate the local site amplification of ground motion from earthquakes. We use a horizontal-to-vertical spectral ratio (H/V) of seismic ambient noise to estimate the Vs profiles or V_S30. The measurement of H/V is easier, compared to active surface wave methods (MASW) or microtremor array measurements (MAM). The inversion of H/V is non-unique and it is impossible to obtain unique Vs profiles. We apply deep learning to estimate the Vs profile from H/V together with other information including site coordinates, deep bedrock depths, and geomorphological classification. The pairs of H/V spectra (input layer) and Vs profiles (output layer) are used as training data. An input layer consists of an observed H/V spectrum, site coordinates, deep bedrock depths, and geomorphological classification, and an output layer is a velocity profile. We applied the method to the South Kanto Plain, Japan. We measured MASW, MAM and H/V at approximately 2300 sites. The pairs of H/V spectrum together with their coordinates, geomorphological classification etc. and Vs profile obtained from the inversion of dispersion curve and H/V, compose the training data. A trained neural network predicts Vs profiles from the observed H/V spectra with other information. Predicted Vs profiles and their V_S30 are reasonably consistent with true Vs profiles and their V_S30. The results implied that the deep learning could estimate Vs profile from H/V together with other information.