← Back to Sessions

General Regression Neural Network (GRNN)-Based Seismic Site Classification Scheme for Chinese Seismic Code Using HVSR Curves

Session: How Well Can We Assess Site Effects So Far? II

Type: Oral

Date: 4/20/2021

Presentation Time: 03:15 PM Pacific

Description: 

Seismic site classification, which is fundamental for site-specific seismic hazard assessment, also plays an important role in accurate interpretation of ground motion data. However, detailed borehole information is not always available in many countries, e.g., China. Therefore, this study investigated application of the generalized regression neural network (GRNN) method to seismic site classification according to Chinese seismic code. First, stations from KiK-net in Japan were classified based on their borehole information and individually assigned to Ⅰ, Ⅱ, Ⅲ, and Ⅳ site classes as defined in Chinese seismic code. Then, mean horizontal-to-vertical spectral ratio (HVSR) curves for each site class were calculated. Owing to the wide ranges of shear wave velocity and thickness of the soil layer for the Ⅱ site class, an unsupervised K-means clustering algorithm was proposed to separate this class into two groups and two corresponding reference curves were derived. Receiver operating characteristic curves indicated that non-normalization of the reference curves and the K-means clustering strategy in our proposed scheme could improve overall classification performance. After exclusion of HVSR curves without significant peaks, the overall recall rates for Ⅰ, Ⅱ, and Ⅲ sites could reach 66.60%, 67.57%, and 68.42%, respectively, regarding use of KiK-net stations. The GRNN-based classification scheme was validated using borehole information of K-NET stations, with recall rates for Ⅰ and Ⅱ site classes reaching 68% and 60%, respectively. Finally, based on HVSR curves calculated using strong ground motion data acquired during 2007–2015 in China, the site conditions of 165 National Strong Motion Observation Network System stations were estimated using the GRNN-based classification scheme. The results were partially validated using borehole information of 73 stations. The similarity between the mean curves and reference curves indicated that the GRNN-based seismic site classification scheme is robust and could produce plausible results succinctly.

Presenting Author: Kun Ji

Student Presenter: No

Authors