Transfer learning model for estimating site amplification factors from limited microtremor H/V spectral ratios

Abstract

Summary Site amplification factors (SAFs) of seismic ground motions are essential in evaluating and estimating seismic hazards. In our previous study, the authors proposed a simple and cost-effective method to estimate a SAF based on a deep neural network (DNN) model and microtremor horizontal-to-vertical spectral ratio (MHVR). Since the previous DNN model was based on the observed SAFs and MHVRs within a limited district in Japan, the applicability of the previous model to non-source regions with different site conditions was limited. This study explored the application of a transfer learning (TL) technique to adapt an existing (pre-trained) DNN model to new regions and a different database. The SAFs obtained through generalized spectral inversion technique (GIT) at the seismic observation stations (K-NET and KiK-net) in Japan were collated as the ground truth for site effects. MHVRs recorded at the stations in several districts of Japan were collected to construct a dataset for the development of the TL model. Subsequently, a TL model was constructed, leveraging the neural network layers and their weights from the pre-trained model while incorporating additional neural network layers to enhance the performance. During the training process, a total data set of 112 sites was divided into training set, validation set, and external test set by 1:1:5. Utilizing a cross-validation strategy, the residuals between pSAFs (pseudo-SAFs) estimated by the TL model and the observed SAFs were analyzed for the external test set containing 80 sites. The results showed that the TL model outperformed the pre-trained DNN model. The cross-validation results demonstrated that almost consistent prediction results were obtained regardless of any combination of 16 sites selected as the training set. Furthermore, by contrasting the influence of varying training set sizes on the performance of the TL model and comparing the TL model to a DNN model with an extended training set, the effectiveness of constructing the model with the limited number of data (16 sites) was ascertained. Finally, the effectiveness and limitations of the TL model were evaluated using MHVRs with peak frequencies falling outside the training set's range.