ABSTRACT
We consider the similarities and differences between earthquake and microtremor horizontal‐to‐vertical spectral ratios (eHVSR and mHVSR, respectively) using a dataset of 161 sites in southern California. Quantitative comparisons are made in terms of the eHVSR and mHVSR lognormal median curves, as well as the frequencies and amplitudes associated with the fundamental‐ and higher‐mode resonances where present. The results show only 58% of the eHVSR–mHVSR pairs agree in terms of their median curve and only 25% of the eHVSR–mHVSR pairs agree in terms of shared resonances, which increases to 68% if flat HVSRs are considered equivalent. Furthermore, while the shared resonances match very well in terms of frequency (root mean square error, RMSE, <0.11 Hz), the amplitudes of those resonances do not agree (RMSE >1.6). These findings demonstrate that while eHVSR and mHVSR agree at some sites, they are not equivalent at all sites. To investigate if the agreement between eHVSR and mHVSR could be related to features of the microtremor data, earthquake recordings, and/or the site conditions, three machine learning (ML) models at varying levels of interpretability are presented. The ML models—which include multivariate logistic regression, gradient‐boosted trees, and support vector machines—show only partial success at using site‐specific data to predict whether eHVSR and mHVSR will likely agree in terms of their median curve (accuracy of 78%) and number of resonances (accuracy of 84%). Therefore, we conclude that while eHVSR and mHVSR can be quite similar in terms of resonant frequencies at some sites, they are not identical at all sites. Furthermore, preliminary evidence shows that the agreement of eHVSR and mHVSR can be predicted a priori given features of the microtremor measurements, earthquake recordings, and site conditions, although a larger dataset will be necessary for developing a robust predictive model.