Automatic Detection of Low-SNR Seismic Events by Pattern Matchingwith Automatically Generated Prototypes in an OSI Scenario

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

Science & Technology 2013

IntroductionThe reliable, automatic detection of low-SNR seismic events is not yet feasible withouta large amount of false-positive detections, i.e., false alarms. This applies especiallyto temporal local networks with exposed seismic stations and a-priori unknown noiseconditions and event signatures, as in an OSI seismic aftershock monitoring. To over-come this problem, we use a multi-path approach. As a first step, high-SNR eventsfrom noise bursts and seismic signals are detected by conventional STA/LTA triggeringand coincidence analysis. These events are grouped for similarity to define a set ofmaster events. Thus in step two the events are transformed into noise-adapted sono-grams, and further reduced in dimension by principal component analysis (PCA). Aself-organizing map (SOM) is used then to create event prototypes by event alignmenton a two-dimensional grid based on similarity. Prototypes which are based on noisesignals will positively identify repetitive noise sources, while the remaining signal pro-totypes are used to detect any low-SNR events in the full data set through adaptedpattern matching. This method allows to lower the automatic detection threshold sig-nificantly while identifying and discarding a large amount of false-positives. The dataof only one mini-array is used because the weak signals of interest might not be de-tected at more stations.

Seismic aftershock monitoring system (SAMS)

The seismic aftershock mon-itoring system (SAMS) of anOSI comprises up to fiftymini-arrays (Seismic Navigat-ing System, SNS) which canbe deployed in the 1000 km2

inspection area. The sought-after events can have a mag-nitude as low as ML -2.0, anda duration of just a few sec-onds which makes it particu-larly hard to discover them inthe large data set.

CenterEastWest

North

120°20 - 100m

Figure 1 : Tripartite mini-array with 1 central 3-componentseismometer and 3 satellite vertical seismometers.

Feature ExtractionPattern recognition on raw waveforms with cross-correlation can only be feasible forhigh SNR events and a compact source region. In case of the search for aftershockshowever the expected SNR of the events is so low that the changing noise conditionswould influence the cross-correlation significantly. A more noise robust representationfor pattern recognition which is called sonograms is used here. Sonograms are basedon spectrograms but apply several noise cancellation steps per frequency band whichmakes patterns robust even in changing noise conditions, Joswig (1993).I Power spectral density

through short-term-fourier-transformation(STFT)

I Half-octave frequency-and logarithmicamplitude-scaling

I Frequency dependantnoise-adaptation,muting andprewhitening

I Amplitude normalizationto provide anamplitude-invariantclustering

I Transformation withPrincipal ComponentAnalysis (PCA) of thefirst 5 principalcomponents

2µm/s

150.0 Hz

15.0 Hz

1.6 Hz

150.0 Hz

15.0 Hz

1.6 Hz

8 sec

STFT

PCA

Increasing

energy

Figure 2 : Transformation from seismograms to sonograms andfeature extraction by PCA.

Detection and Classification with only 1 Mini-ArraySuper-sonogram Compilation

Single stations of one SNS are within 200 m of distance whichmakes it possible to combine the four vertical traces of one SNSinto a super-sonogram. Each Pixel of the super-sonogram con-sists of four sub-pixels, each from one vertical trace of the SNS.Array-wide signal coherency can be checked fast and the dataof all stations can be displayed on one screen, Sick et al. (2012).

W N

C E

SUPER-PIXEL

CENTER EAST

WEST NORTH

SUPERSONOGRAM

20 sec20 sec

20 sec

Figure 3 : Combination of four single sonograms to one super-sonogram.

OSI Scenario and Event Examples

Data from OSI training campaigns is restricted to internal use which iswhy we take data from research projects with similar requirements asthe OSI SAMS would face. Here we use a dataset from a permanentseismic network of 2 years for the analysis of slidequakes on a creepinglandslide in Vorarlberg, Austria (Walter et al. 2011) with the followingproperties similar to SAMS challenges:I Detected events on the landslide with magnitudes down to ML -3.2.I Low event rate of approximately 1 slidequake in 5 days in average.I Multiple noise sources on the slope: ski-slope (lift, snowcat), holiday

village (people, cars), agriculture (cattle, tractor) etc.I Ground truth manual bulletin for verification of automatic analysis with

mainly three types of events:1. Slidequakes, small fracture processes from the movement of the

slope2. Frost-heave events, near-surface local events from freezing

processes which are only detected at one mini-array at a time3. Local earthquakes

noiselocalfrost-heaveslidequake

150.0 Hz

15.0 Hz

1.6 Hz

8 sec

Figure 4 : Super-sonogram pattern examples of the three event classes andtypical noise.

In this study we use data from 14 days of the Heumoes slope which contains 37 local earthquakes, 60 slidequakes and 67 frost-heave events.The frost-heave events are only visible at one mini-array which will limit our analysis to the information of this array.

1st step: Template Creation

At first high-SNR events are detected by STA/LTA on the seismic data with a coincidence of at least 3 of the 4 vertical traces of the mini-array.With these parameters, 102 total triggers are made with 43 of these corresponding to events from the manual bulletin (3 slidequakes, 27frost-heave events and 13 local earthquakes). At each trigger position a pattern is created in the way described previously. These patternsare then used to construct the self-organizing map (SOM, Kohonen 2001). The SOM groups event classes without prior knowledge,i.e., unsupervised and creates a map of representatives for each event type arranged by proximity of features, giving us a synoptic andtopological overview of the triggered events. An analyst can classify regions in the SOM and the labeled events of these regions are usedin step two for a pattern recognition to classify the unlabeled events from triggers of a STA/LTA with a much lower threshold.

Figure 5 : Grid of the SOM, each black dot represents 1 SOM node. Thecolored rectangles represent patterns which are most similar to the node proto-type (blue=slidequake, pink=frost-heave, red=local earthquake). The thicknessof the rectangular border around the nodes indicates the similarity to a neigh-bouring node (thicker lines = less similar).

Figure 6 : The SOM with each node visualized as the according pattern pro-totype which is an average of the closest event patterns. In this view an ana-lyst can recognize event categories and label them appropriately and therebyclassify the underlying patterns. These labeled patterns are used in step 2 toclassify unknown events.

2nd step: Pattern Recognition

In the second step the STA/LTA parameters are changed to detect also low-SNR events. This results in 1728 detections of which 152 areactual events from the manual bulletin. Patterns are created again at all detections and compared by cross correlation and with additionalvarying amplitude shifts based on the area of maximum energy in the pattern, Joswig (1993). This allows to detect also patterns where thetrigger time is shifted and also takes into account similar patterns with different amplitudes.

ReferencesJoswig, M. (1993). Automated Seismogram Analysis for the Tripartite Bug Array: An Introduction. Computers & Geosciences, 19(2):203–206.Kohonen, T. (2001). Self-organizing maps. Springer Ser. Inf. Sci., 30:501 pp.Sick, B., Walter, M., and Joswig, M. (2012). Visual Event Screening of Continuous Seismic Data by Supersonograms. Pure and Applied Geophysics, pages 1–11.Walter, M., Walser, M., and Joswig, M. (2011). Mapping rainfall-triggered fracture processes, and seismic determination of landslide volume at the creeping Heumoes slope. Vadose Zone Journal, 10(2):487–495.

Results

In total we get 1576 false positive noise detections of which 1007 areclassified as such by the pattern comparison. All frost-heave events,almost all local earthquakes and the majority of the slidequakes areclassified correctly. 13 slidequakes are classified as local earthquakeswhich can be accounted to the high similarity of these events.

slidequake frost-heave local eq noisePredicted class

slidequake

frost-heave

local eq

noise

Act

ual cl

ass

42 0 13 9

0 60 0 0

0 0 26 2

246 125 198 1007

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0

Figure 7 : Confusion matrix for classification with 1 mini-array.

For most of the wrong classifications, it is also impossible to manuallyclassify the events if restricted to one mini-array. The 9 slidequakeswhich are discarded as noise events could be recognized by a coinci-dence analysis over multiple mini-arrays which could run on all “noise”detections. Furthermore only 3 slidequake were triggered in the firststep and only these were available as templates.In total we get a recognition rate of 80.6 % (86.2 % without the noiseclass) in bad noise conditions with very low-SNR events and only usingone mini-array with four single stations.

Acknowldgements: We would like to thank Marco Walter who constructed the permanent seismic network on theHeumoes slope and created the ground truth bulletin by manual analysis of the data with 3 mini-arrays and the softwareNanoseismicSuite (Sick et al. 2012). The software is also the official software of SAMS.

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