Imaging Ground Motions in the Tokyo Metropolitan Area Based on MeSO-net Using Lasso
Abstract:A rapid prediction of damage to structures due to a large earthquake is important to prevent secondary disasters. Ground motion at the base of each construction to be input to a numerical simulation is to be estimated from seismograms. An accurate damage prediction requires such ground motions with spatially-high resolution although the seismometers much sparsely distribute comparing with the required resolution. We have been developing a procedure based on sparse modeling to image the ground motions from seismic array data. Our target is the Tokyo metropolitan area of Japan, in which the seismic array “MeSO-net” (Metropolitan Seismic Observation network) is in operation.
Mizusako[2013, graduation thesis] applied an algorithm based on the Taylor expansion to MeSO-net data when the Earthquake off the Pacific coast of Tohoku occurred on March 11, 2011. This method was found to never account for ground motions in frequencies higher than 0.15Hz, which was insufficient taking into consideration that the typical eigenfrequency of a construction is usually between 1-10Hz. Moreover, this method requires a priori assumed truncation order in differential and groups of observatories called “cluster”, in order to determine the unknown partial differential coefficients. Mizusako suggested that the truncation order was one and each cluster included five nearest observatories.
We propose a new algorithm using lasso (least absolute shrinkage and selection operator, Tibshirani) in order to obtain an image of spatially-high-resolution ground motions objectively determining the truncation differential order and clusters. The truncation order for a given cluster is automatically determined by lasso owing to the L1-norm regularization, and an appropriate cluster is selected based on an information criterion. Our initial result indicates that the ground motion determined by using the information criterion EBIC (Chen and Chen) is successfully reproduced in frequencies lower than 2Hz, which is drastically improved comparing with the previous study.