Support Vector Regression for Developing Ground-Motion Models for Arias Intensity, Cumulative Absolute Velocity, and Significant Duration for the Kanto Region, Japan

Abstract

The Kanto region is an earthquake disaster-prone area where it is necessary to conduct regional seismic hazard analysis. Ground-motion models (GMMs) of Arias intensity, cumulative absolute velocity, and significant duration are developed by support vector regression (SVR) for the Kanto region, Japan. In contrast to traditional regression programs used in previous models, which are usually expressed as a mathematical function with a minimum observed training error as constraints, the SVR algorithm has one major feature: it minimizes the generalized error bound to improve robustness. In the database for this study, the regional ground-motion database contains 15,960 ground-motion records of 130 earthquake events from 2000 to 2020 with the Japan Meteorological Agency (JMA) with a magnitude MJMA 5.0–8.0 and a rupture distance less than 200 km. In developing SVR GMMs, the moment magnitude (Mw), rupture distance (Rrup), and shear-wave velocity averaged in the top 30 m of soil (VS30) were adopted to characterize the source, path, and site conditions. To verify the rationality and effectiveness of the SVR GMMs, the performance indices (e.g., correlation coefficients and slope coefficients) and residuals are analyzed. The residuals of the SVR GMMs have no significant deviation in magnitude, rupture distance, or VS30. The standard deviations of model residuals are calculated using the regional ground-motion database, and the standard deviations of SVR GMMs are less than those of previous models developed based on a Japanese or global database. Furthermore, the SVR GMMs are also compared with observed data and the previous GMMs. Data-driven SVR method constrains statistical theory and probability theory to develop GMMs, which can eliminate the problem that the specific form of the previous models may affect the prediction performance and capture the regional attenuation characteristics effectively.

Ground motion models for regions with limited data: Data‐driven approach

“…8 Furthermore, as the new era of machine learning started, many GMMs were developed using these techniques. [9][10][11][12][13][14][15][16] Nevertheless, there are several active regions where an exhaustive instrumentation network still needs to be developed, and as a result, the databases contain very few strong ground motion recordings. For instance, the Himalayan orogenic belt, formed due to the collision of the Indian plate and the Eurasian plate, is one of the most seismically active regions in the world.…”

“…Due to the abundance of ground motion records and corresponding metadata, several GMMs are developed for 5% damped response spectra (PSA), which include not only magnitude ( M w ), distance, and site amplification proxy as inputs but also hypocentral depth, fault classification, hanging wall effects, directivity, basin amplification, and linear and nonlinear site responses 4–7 using random‐effects regression algorithm 8 . Furthermore, as the new era of machine learning started, many GMMs were developed using these techniques 9–16 . Nevertheless, there are several active regions where an exhaustive instrumentation network still needs to be developed, and as a result, the databases contain very few strong ground motion recordings.…”

An All-in-One Rapid Prediction of Ground Motion Intensity Measures Hybrid Network for Multi-Task in The North-South Seismic Belt of China

“…Table S1 of the supplementary Information also shows the input information and data sources used for the data-driven model in Table 1. It is worth noting that many researchers are also employing data-driven methods to predict IMs, which can be categorized into two types: first type methods that use earthquake sources, paths, and site parameters as inputs (Fayaz et al, 2021;Hu et al, 2022;Fayaz & Galasso, 2022a;Fayaz et al, 2023), similar to GMPE methods, inability to meet timeliness requirements, and another type methods that use seismic wave or P-wave features as inputs (Dai et al, 2024;Hsu et al, 2013;Jozinović et al, 2020;Y. Liu et al, 2024;Y.…”

A hybrid non‐parametric ground motion model for shallow crustal earthquakes in Europe

“…Hu and Zhang 28 have developed GMMs for Western China to predict PGA and PSA using SVR. Hu et al 29 have developed SVR models for arias intensity (I a ), CAV, and significant duration (T 90 , 5%-95% energy of Husid plot) for the Kanto region, Japan. On the other hand, another machine learning technique that is as effective as SVR is the random forest regression (RFR), a supervised learning algorithm that uses an ensemble learning method for regression.…”

“…Hu and Zhang 28 have developed GMMs for Western China to predict PGA and PSA using SVR. Hu et al 29 . have developed SVR models for arias intensity (I a ), CAV, and significant duration (T 90 , 5%–95% energy of Husid plot) for the Kanto region, Japan.…”