Project nº 2124-B

PWG - PAPERS
25 March 2009

A Chilean seismic regionalization through a Kohonen neural network.

Authors: Jorge Reyes (TGT) & Víctor Cárdenas (DFA - U de Valparaíso)

The following article is a review of a paper published in may 2010 in the scientific magazine "Neural Computing & Applications".
It's important to say that what it is presented in the present review are the latest results including data until april 2010.

Abstract.
Through this paper we are presenting a study of seismic regionalization for continental chile based on a neural network. A scenario with six seismic regions is obtained, irrespective of the size of the neighborhood or the range of the correlation between the cells of the grid.
Unlike conventional seismic methods, our work manages to generate seismic regions tectonically valid from sparse and non-redundant information, which shows that the selforganizing maps are a valuable tool in seismology.
The high correlation between the spatial distribution of the seismic zones and geological data confirms that the fields chosen for structuring the training vectors were the most appropriate.

1. The idea
the idea of a spatial classification of seismic sources can be traced back to 1941 when Gorshkov published one of the first studies on seismic regionalization for the USSR.
Other works related are:
* Richter seismic regionalization (1959)
* Gajardo and Lomnitz Chilean regionalization (1958)
* Barrientos Chilean regionalization (1980)
* Martin Chilean regionalization (1990)

Due to the fact that the self organized kohonen maps have the ability to associate n-dimensional vectors from a similar way the human brain classifies patterns, we have decided to investigate the performance of such a neuronal web applied into a seismic regionalization.

2. The vectors
Continental chile was divided in 156 cells of 1ºx1º and each cell was characterized with the following fields, including information from 1957 until april 2010:
x1: Mean deep of the seismic sources
x2: Number of earthquakes with magnitude > = 4.5 Ms
x3: Number of earthquakes with magnitude > = 5.5 Ms
x4: Number of earthquakes with magnitude > = 6.5 Ms
x5: Maximum observed magnitude
x6: Central horizontal coordinate of the cell
x7: Central vertical coordinate of the cell

3. The results
The Neuronal Network deduced the following seismic regions:

The seismic regions have the following characteristic related to a seismicity, where N = quantity of seisms with magnitude > = 4.5 Ms, observed in a year span in a cell of 111x111 km²:

It's important to say that the seismic regions are dynamic. It means that if the base of the data is extended, the seismic regions can change.

4. Comments
In this work, we have demonstrated that using a Kohonen NN, and a database with exclusive information from the seismic sources, can be possible to differentiate six seismic zones for continental Chilean territory, which are supported by the geological features of the sector and internal high correlation coefficients.
There are zones that have experimented big earthquakes, such as Valdivia (year 1960, 8.8 Ms or 9.5 Mw) and Cauquenes (year 2010, 7.8 Ms or 8.8 Mw), which are visualized respectively in the zone 6 and 4 and represent periods of recurrence longer than a hundred and fifty years. On the other side, the imminent danger corresponds to the coast zone of the North of Chile, which has an accumulated compression of 11 meters (similar to the one observed in Cauquenes-2010) which belongs to a zone 6.

5. Acknowledgments
JR wants to thank TGT for the support through grant number 2122 and 2123. VHC wants to thank Rafael Valdivia for useful discussions.
Download the paper.

HOME: www.geofisica.cl