Site Classification Based on H/V Response Spectra, Using Image Processing and Neural Networks

In order to estimate the seismic hazard of a specific site, the classification of that site is of particular importance.
On the other hand, in order to interpret and analyze the ground motion data in different parts of the world, it is
necessary to know the site conditions in seismic stations. In some countries, including Iran, there is insufficient
information on the geotechnical and geological status of many seismic stations. The conventional methods to
characterize the site are based on shear wave velocity measurement such as SCPT measurement, downhole testing,
and seismic refraction. These methods have some limitations such as costs, maximum depth, execution problems,
etc. This research is a new and efficient approach in site classification using the data recorded from the seismic
networks, image processing techniques, neural networks and set of 5% damping spectral ratio reference curves of
horizontal to vertical component (H/V) for the four different site classifications. These reference sets, that include
four separate H/V curves for four different site conditions labelled as rock, dense soil, medium soil and soft soil and
classified as site I, II, III and IV, have been selected from the study of Zhao et al. [1]. The reference curves are based
on K-net seismic network data. The adopted soil classifications are based on Japan Road Association
recommendations. For the periods of interest, which were not presented in the Zhao et al. [1], the curves were
interpolated to come up with the values at the missing periods.
In this research, two types of basic radial functions (RBF) are called "probabilistic neural network (PNN)" and
"general regression neural network (GRNN)", as well as "convolutional neural network (CNN)" have been used. For
neural network input, the data from 182 seismic stations have been incorporated. The site condition at the location of
each station has been fully characterized. The horizontal to vertical spectral ratio for each recorded seismic event
was calculated. The ratio for each data was smoothed using the moving average function. Then, the smoothed H/V
ratio was normalized to match the sigmoid transfer function upper and lower range, which could minimize the
network training time. For the CNN network, the input H/V spectral ratio images were first unified using the exact
dimension of 150×300 pixels and then compared to the reference H/V spectral ratio using image processing
techniques implemented in MATLAB software.
To verify the proposed technique, H/V spectral ratio was calculated for all events recorded at all 182 stations and
then used as input for training the PNN, GRNN and CNN networks and then compared to the reference curves
proposed by Zhao et al. [1]. Two normalization methods were incorporated; in the first method, all the H/V spectral
ratios normalized to the maximum amplitude, and the second was to normalize the maximum to one and minimum
to zero. The results confirmed that the second normalization method could produce more accurate results due to a
better matching the sigmoid function.
According to the obtained results incorporating the second method of normalization and all 790 ground motion
data, which were recorded at 182 different stations., the PNN, GRNN and CNN networks have succeeded in
accurately predicting the site conditions in 73%, 71% and 81% of the stations, respectively. The results could prove
the applicability of the proposed approach, using neural networks, in site characterization.
References
1. Zhao, J.X., Irikura, K., Zhang, J., Fukushima, Y., Somerville, P.G., Asano, A., Ohno, Y., Oouchi, T., Takahashi, T. and Ogawa, H. (2006) An Empirical site-classification method for strong-motion stations in japan using H/V
response spectral ratio. Bulletin of the Seismological Society of America, 96, 914-25.