automatic clustering of seismic events in an on-site ...€¦ · ’94 project which measured...

Automatic clustering of seismic eventsin an On-Site Inspection scenario

Benjamin Sick (benjamin.sick@geophys.uni-stuttgart.de), Manfred JoswigInstitute for Geophysics, University of Stuttgart

Science & Technology 2011

IntroductionDuring an On-site Inspection (OSI) it is essential to get a fast overview of recorded seismicsignal classes to evaluate the local seismicity, and in particular to investigate on suspiciousevents eventually representing aftershocks from an underground nuclear explosion (UNE).The seismic aftershock monitoring system (SAMS) of an OSI comprises up to fifty mini-arrayseach having six traces of continuous waveform data. The sought-after events can have amagnitude as low as ML -2.0, and a duration of just a few seconds which makes it particularlyhard to discover them in the large data set.

Event detectionTo overcome the first challenge of event detection we use a special form of spectrogram, thefour-traces Super-Sonograms to rise these signals from stationary background noise and totest on array-wide signal coherency. The Super-Sonograms proved to be indispensable inmanual screening of IFE08 data and got implemented into SAMS.

Mini-arrays (SNS)

Seismic measurement during an OSI is donewith mini-arrays (Seismic Navigating Systems,SNS).One SNS consists of:I 1 central 3-component seismometerI 3 satellite vertical seismometersThe inspection area for an OSI is limited to1000 m2 and each SNS covers an area of ap-proximately 30 m2 to detect events of magni-tude ML < 2.⇒ It requires approximately 35 SNS to

cover an inspection area

CenterEastWest

North

120°20 - 100m

Sonogram

Aftershocks of UNEs are often too weak and too short to be recognized in the seismogramsof the huge waveform datasets which have to be analyzed in nearly real-time. Sonograms[Joswig(1990)] are based on spectrograms and allow a fast visual interpretation of energiesover time and frequency. By applying multiple signal enhancing steps the identification of verysmall scale events is greatly improved and event patterns are prominent even when analysinglarge datasets with varying noise conditions:I Power spectral density

through short-term-fourier-transformation (STFT)

I Logarithmic frequency- andamplitude-scaling

I Frequency dependant noise-adaptation, muting andprewhitening

I Color scale optimized for vi-sual inspection

Time

Amplitude

Time

Frequency

Increasingenergy

⇒ For STFT(ω, t) > µ(ω) + σ(ω): Sonogramm(ω, t) = (int)log2

(STFT(ω,t)−µ(ω)

σ(ω)

)+ 1

Super-Sonogram

Single stations of one SNSare within 200 m of distancewhich makes it possible to com-bine the four vertical tracesof one SNS into a Super-Sonogram. Each Pixel ofthe Super-Sonogram consists offour sub-pixels, each from onevertical trace of the SNS. Array-wide signal coherency can bechecked fast and the data of allstations of one campaign can bedisplayed on one screen.

W N

C E

NWEC

SNS

1C

3C

1C

1C

Center

EastWest

North

Further information

http://www.nanoseismic.net

SoftwareTo be able to meet the hard time schedules of an OSI which re-quire to analyse the seismic data of all SNS of one day within thefollowing day, a software suite was developed which focuses onthe fast scanning for small scale events in huge datasets, calledSAMS - GeophysSuite. It is written in Java to be multi platformcompatible and consists of multiple modules which synchronizecurrent inspection times:

I SonoView, fast scanning of Super-Sonograms of numerousSNS

I TraceView, map overview of stations and visualization of seis-mograms of neighbouring SNS

I HypoLine, interactive localization and magnitude estimation

TraceView

HypoLine

Timesynchronization

Timesynchronization

SonoView

13 minutes

Overlap

25S

NS

Unsupervised classificationWe test unsupervised classifica-tion on the dataset of the PISCO’94 project which measured seis-micity in the central Andes ofnorthern Chile. We try to discrimi-nate five types of events which oc-cur in different regions of the mea-surement area.

−72˚ −70˚ −68˚ −66˚

−24˚

−22˚

50 km

N

−250

−200

−150

−100

−50

0

Hyp

oce

nte

r d

ep

th (

km

)

QUARRY

NORTH

MAINDEEP

CRUST

Station A04Cluster borders

SOUTH AMERICA

Feature extraction

The main feature extraction is done by the Sonogram computation which is enhancedby the following methods:I Normalization to provide an

amplitude-invariant clusteringI Transformation with Principal Com-

ponent Analysis (PCA) of the first 5principal components

I 2-D similarity function optimizedfor fuzzy comparison of Super-Sonograms

Types Seismograms FeaturesCrust

Main

North

Deep

Quarry

Visualization of single principal components

The dominant features of the SOM clustering andsonogram compilation of the arranged representa-tives can be analyzed with PCA. Here we show theadjustment of Sonograms along the 5 principal com-ponents with the largest variance. It can be seene.g. that the first principal component is mainly re-sponsible for energy amplitude variations and ex-tension of S-P-onset times.

PC1 PC2 PC3 PC4 PC5

Self-organizing Map [Kohonen(2001)]

Grouping event classes without prior knowledge, i.e., the task of unsupervised classification is handledby Self-organizing Maps (SOM). The SOM creates a map of representatives for each event type ar-ranged by proximity of features, giving us a synoptic and topological overview of the acquired seismicdata. Moreover first tests showed that the obtained SOM is well applicable for classification of newevents that are passed to it (94.2 % success rate with the given dataset).

0 2 4 6 8

0

2

4

Map width

Map

heigh

t CRUST

MAIN

NORTH

DEEP

QUARRY

DISCARDED

Outlook

The data of only one station of the dataset was used for the unsupervised classification. As a nextstep surrounding stations will be integrated in the classification process and decisions based onsignal coherency and logic will be involved.

Acknowledgements

The GeophysSuite software development is funded by the CTBT in line with the OSI SeismicAftershock Monitoring System development. The PISCO ’94 dataset was used for this study whichis free available through the GEOFON (http://geofon.gfz-potsdam.de/geofon/) project,Hypocenters were calculated by Graeber [Graeber(1997)].

References

I Graeber, F., 1997.Seismische Geschwindigkeiten und Hypozentren in den sudlichenzentralen Anden aus der simultanen Inversion von Laufzeitdatendes seismologischen Experiments PISCO’94 in Nordchile, Diss.,Freie Universitat Berlin.

I Joswig, M., 1990.Pattern recognition for earthquake detection, Bull. seism. Soc. Am., 80(1),170–186.

I Kohonen, T., 2001.Self-organizing maps, vol. 30 of Springer Ser. in Inf. Sci., Springer.

Institute for Geophysics, University of Stuttgart http://www.geophys.uni-stuttgart.de