Deep-Neural-Network-Based Estimation of Site Amplification Factor from Microtremor H/V Spectral Ratio

Site amplification factors (SAFs) of S wave at ground surface are crucial for evaluating and predicting seismic ground motions. This study proposed a novel methodology for directly estimating S-wave SAF from microtremor horizontal-to-vertical spectral ratio (MHVR) based on deep neural network (DNN) model. We analyzed site amplifications obtained from generalized spectral inversion technique and microtremor data observed at Kyoshin net and Kiban–Kyoshin network sites in Chugoku district, western Japan. The DNN model was developed using peak frequency and the frequency-dependent relationship between MHVRs and SAFs. The sites were divided into training set, validation set, and test set. The training set and validation set were used in k-fold cross-validation technique to evaluate and select optimal model. Once the optimal model had been determined, the model was employed on the test set that was completely independent of the training and validation set for evaluating the generalization performance. Residuals and root mean square errors between the estimated and observed SAFs were evaluated to discuss the applicability of the proposed model. We also confirmed that the DNN model shows better performance in estimating SAFs com-pared with the existing double empirical correction method.