Prediction of ground motion intensity measures using support vector machine in analysing spatial correlation

Abstract

Regional seismic risk assessments or engineering applications generally require the simultaneous estimation of ground motion intensity measures (IMs) at multiple stations. Spatial correlation plays a crucial role in characterizing the spatial distribution of IMs, typically as a function of the spatial separation distance. Previous studies using measured recordings have suggested that spatial correlation is additionally influenced by variations in site conditions or other earthquake parameters in the study regions. In this study, a machine learning method comprising multiple parameter inputs, called a support vector machine (SVM), is proposed to estimate the IMs of ground motions. This SVM model can inherit variations in the site conditions and earthquake information of the study region. Meanwhile, the generalization ability of the SVM model behaved distinctly, with a high similarity in the standard deviation between the training and test databases. Geostatistical analysis was performed to generate the spatial correlation of the IMs by introducing the SVM model. Considering the heterogeneous characteristics of the study region, the spatial correlation using the SVM model showed relatively low values, particularly for small separation distances and short periods, that is consistent with the results of previous studies. Finally, considering the high similarity of IMs at small separation distances (for example, less than 1 km in this study), the correlation was observed to be unsuitable for describing the spatial distribution of ground motion IMs at a small scale.