On the Relation between Empirical Amplification and Proxies Measured at Swiss and Japanese Stations: Systematic Regression Analysis and Neural Network Prediction of Amplification

Abstract

ABSTRACT We address the relation between local amplification and site-condition indicators derived from in situ geophysical surveys for the estimation of the VS profile, and single-station recordings processed with horizontal-to-vertical spectral ratio technique. Site-condition indicators, or proxies (e.g., VS30), aim at “summarizing” the description of the local geophysical structure, with a focus on its relation to site amplification. The premise for our work was the compilation of two companion databases: one of soil condition proxies and the other of empirically derived Fourier amplification functions, for Swiss and Japanese stations. We investigated the connection between these two datasets, at first, with a systematic set of regressions correlating each proxy to amplification factors within the frequency band 0.5–20 Hz, second, with a neural network (NN) structure predicting site amplification from proxies. The regression analyses showed that, generally, site-condition parameters (SCPs) bear a better correlation with amplification within 1.7–6.7 Hz; the “best” indicators are the frequency-dependent quarter-wavelength (QWL) velocity and, among scalar parameters, VS30, the bedrock depth, and f0. Collating Swiss and Japanese datasets, the trend of variation of amplification with respect to most proxies is similar. Finally, we evaluated the prediction performance of various combinations of SCPs, for local amplification, using a NN. To attain a database large enough to constrain the estimation of the network parameters, we merged Swiss and Japanese stations into a single training and validation dataset, motivated by the similarities observed in the regression analyses. The outcome we obtained from the NN is encouraging and consistent with the results of the regressions; SCPs with higher correlation to amplification provide a better forecast of the latter (particularly within 1.7–6.7 Hz). More complete input information, such as QWL parameters (velocity, impedance contrast), or extended ensembles of scalar proxies (particularly, including f0), offer a better estimation of local amplification.