Personal tools

You are here: Home / Science and society / Science for all / Just published! / Using hydrophones for the automatic recognition of oceanic earthquakes

Using hydrophones for the automatic recognition of oceanic earthquakes

Moored hydrophones are recovered after one year or more and during this period store huge amounts of sound recordings. How to perform the automatic analysis of these data to reliably detect the different types of seismic waves?

Bandeau Sukhovich.jpg

 Oceanic ridges undergo a seismic activity of low magnitude which is studied with ocean bottom seismometers or moored hydrophones placed in the water column. The acoustic waves generated by oceanic earthquakes (T-waves) propagate long distances with almost no attenuation whithin a water layer which acts as an acoustic waveguide ("SOFAR channel"): hydrophones are thus placed at this depth, and form networks enabling the localization of seismic events.

Besides these T-waves, hydrophones also record waves produced by far distant events (teleseismic P-waves) as well as other noises: it is therefore necessary to discriminate among them within the recordings. Although easy for an trained human observer, this task is very time-consuming, which prevents using it for the large amounts of data acquired by hydrophones between their mooring and their recovery, i.e. during generally one year or more. Automatic signal identification methods are available but so far can only discriminate one signal type from the rest. This paper describes a new method to automatically recognize T-waves, P-waves and the signals of other origins recorded by hydrophones.

 

mouillage-hydrophone.jpeg

Hydrophone mooring in the Indian Ocean on R/V Marion-Dufresne

The method follows several steps:

- the whole data set is analyzed with a classical technique whereby the beginning and the end of each signal are detected by monitoring the ratio of short-term (10 s) to long-term (100 s) average power received.

- using a transformation called "in wavelets", the acoustic properties of the signal are described by the proportion of total power among seven frequency bands.

- these seven proportions characterize a signal. Comparing them to a database of signals of known origin, it is possible to deduce which kind of signal it is: teleseismic P or T waves, noises produced by ships or icebergs. This comparison is performed through a series of "decision trees", sets of binary rules leading to the classification of signals into groups of identical origin. Decision rules are first elaborated during a "training phase" from model signals, for instance signals identified manually by an observer.


Sukhovich_Fig 1.jpg

Examples of a teleseismic P-wave (left) and a T-wave (right): the signal (top), its wavelet transform (middle) and the spectrogram (bottom)

 

The method was applied to recordings collected from June 2002 to April 2003 by two hydrophones (S2 and S5) located on each side of the Mid-Atlantic Ridge; as these data had already been analyzed, the results of the automatic recognition can be compared to the catalogue of known seismic events. All signals detected by the automatic method were identified using the catalogue. In most cases they were T-waves, more rarely P-Waves ; signals that were not related to any seismic event were produced by ship traffic or icebergs (noises emitted by cetaceans are too short to be detected by the automatic method).

.

Sukhovich_Fig 3.jpg

Characterization of the four main types of signals by the proportion of energy received within each of the seven frequency bands (from left to right : T-waves, P-waves, ships, icebergs)

 

The maximum predictive ability of the model can be assessed by applying it to the training data set which was used to build it. The results of the tree constructed with the whole set of data are excellent: wrong classifications were obtained for only 1 and 5 of the 1570 and 1329 signals of the two hydrophones.

The decision rules actually derive from a subsample and then applied to the whole data set. The subsample contains signals whose types are manually identified by an observer; how it is built is therefore important. Its size is a compromise between the quality of the predictive model and the time required to identify signal types: for the data of a single hydrophone, tests showed that a proportion of 10 % of the total is enough. It must also be representative by containing signals of every type and integrating the variability of signals within a given type. As by definition signals are not identified, the subsample was randomly drawn from a automatic classification using only the acoustic properties of the recordings without involving any prior knowledge. For each hydrophone, the predictive power of the model was tested by its application to 100 random subsamples drawn independently.

 

True positives
(correct identifications)

False positives
(incorrect identifications)

False negatives
(false rejections)

T-waves

1471

19

5

P-waves

17

2

0

Ships

0

0

14

Icebergs

58

3

5

Average performance (on 100 subsamples) of the classification of hydrophone S2 signals

For both hydrophones, P- and T-waves are most often correctly identified, but all signals of ships and some of icebergs were wrongly identified as T-waves; these false positives are few and will be eliminated when localizing the events. Very close results were found when combining the date of both hydrophones instead of analyzing them separately.

In all cases, no ship signal was correctly identified; they were most often confused with T-waves. The scarcity or even absence of ship-generated signals in the training sets explains their poor classification in the operational prediction. Forcing the inclusion of randomly drawn ship-generated signals in each training set greatly improves this result, which means that the classification will be better if data are collected in a region with heavy shipping.

 

True positives
(correct identifications)

False positives
(incorrect identifications)

False negatives
(false rejections)

Ondes T

2687

8.5

13

Ondes P

38

5

1

Navires

23

2

3

Icebergs

125

9

3

Average performance (on 100 subsamples) of the classification of hydrophone S2 and S5 signals, with subsamples simulating heavy shipping conditions

 These results demonstrate the high potential of the presented automatic signal identification method, which could prove useful for other geophysical applications.

 

The paper

Sukhovich, A., Irisson J.-O., Perrot J., and Nolet G., 2014. Automatic recognition of T and teleseismic P waves by statistical analysis of their spectra: An application to continuous records of moored hydrophones, J. Geophys. Res. Solid Earth, 119, 6469–6485.

See the first page

 

The authors

The authors of this paper are members of the Domaines Océaniques laboratory (LDO- IUEM, Brest), of the Laboratoire d'océanographie de Villefranche (LOV) and of Géoazur (University of Nice)

 

The journal

Journal of Geophysical Research is the flagship journal of the American Geophysical Union. Over its 115 years of continual publication, it has adapted to meet the needs of multidisciplinary science. It now has seven disciplinary sections, one of which is dedicated to the Earth. JGR-Solid Earth focuses on the physics and chemistry of the solid Earth and the liquid core of the Earth, geomagnetism, paleomagnetism, marine geology/geophysics, chemistry and physics of minerals, rocks, volcanology, seismology, geodesy, gravity, and tectonophysics.

 

Contacts

Authors : browse IUEM staff directory
Communication and scientific mediation service: communication.iuem@univ-brest.fr


Back to publications list

 

 

 

Picture of the month

Drague.jpg