Twins Merge

Twins Merge is an application allowing to accurately merge seismograms from neighbour networks whose clocks are occasionally or permanently out-of-sync with the clock of a given network. The capability of accurately synchronizing and merging seismograms can be of significant importance to the seismic monitoring coverage of regional and international borders, and at global scale. It is also quite relevant to the integration of seismograms from temporary networks.

Twins Merge is an answer to the problem of merging seismograms recorded independently by the Geophysical Institute of Israel (GII) and by the Jordan Seismological Observatory (JSO) in the Dead Sea region. Since 1995 and as a result of a Peace treaty between the two countries, each of the two networks records also several stations from the other network. The innermost stations in Israel and in Jordan, however, cannot be directly recorded by the other country network due to telemetry limitations imposed by the topography. As a consequence, each network records all stations it owns and some stations from the other network. Until recently, common stations (stations recorded by both networks) displayed however a random pattern of clock differences ranging from a fraction of a second to several minutes.

When clocks from two networks are not synchronized, each network can only safely use signals from stations it is able to record with its own clock. As a consequence, the mutual exchange of unrecorded seismograms between two networks can be useless without a reliable way to known if their time codes are sufficiently similar and a good way to fix the problem if they are not. More generally, stations from network B can only be used to extend network A if they are also recorded by network A, and stations solely recorded by network B cannot be directly integrated into the database of network A.

In advanced studies like travel-time tomography for which the spatial coverage of stations can be of paramount importance, enlarging a network database by merging seismograms from one or several additional networks can sometimes make all the difference.

Naturally, the accurate merging of seismograms requires some conditions to be met. Sufficient conditions should be met if at least one of the two networks can directly record a few stations owned by the other network. Figure 1 diplays a schematic distribution of stations from two networks A and B where a few stations from each network are directly recorded by the other network.

Figure 1. Schematic distribution of stations from networks A and B


It is assumed that we want to merge seismograms of network B to those of network A. Common stations (bi-colored stations in Fig.1) do not need to be merged since they were directly recorded by network A. Candidate stations for merging are those displayed in black only, i.e. recorded by network B but not recorded by network A.

The principle to determine these time corrections is simple. Since common stations are recorded by both networks, it is possible to estimate clock differences from pairs of seismograms of common events (events recorded by both networks). One can expect these time differences to be nearly constant within the time window of most recorded events. Once these time differences are known for one or more common stations, an averaged correction can be determined and applied to the seismograms of network B that were not recorded by network A (stations displayed in black only in Fig.1). The synchronized seismograms can then be copied or moved into the seismogram database of network A and extend its receiver spatial coverage.

The minimal condition to allow some merging to be done is obviously to have one common station. This setup is however not recommended because for some important events the signal from the station might be corrupted, and the station might also not be in working order for extended periods of time. In addition, since the merging of seismograms from network B can have a significant impact on results from network A, some redundancy from other common stations is necessary to ensure the validity of such an important process.

A crude way to estimate clock differences is to derive them visually from some features clearly recognizable on seismograms. One or more phase arrivals, if present, can be used. This method is feasible but it is certainly tedious work, and it is also error-prone and rather inaccurate. A very fast and much more powerful way has been found.

Twins Merge is a program allowing to perform an in-depth analysis over time and to synchronize seismograms from pairs of networks with common stations. Twins Merge calls network A the "Permanent" network while network B is called the "Neighbour" network. The aim of the program is to extend the Permanent network by accurately merging seismograms from the Neighbour network.

In Search Mode, cross-correlation is used to detect "Twin" seismograms across the Permanent and the Neighbour networks. Twin seismograms are seismograms sufficiently similar over a significant time window to be considered as belonging to the same event recorded at the same station by two networks with common stations. The search is done for all pairs of stations of each event and one event at a time. If the absolute maximum value of a normalized cross-correlation function (correlation coefficient) between two seismograms exceeds some theshold, a twin pair is detected and further analyzed.

One common reason for which seismograms recorded simultaneously at the same station by different networks can be out-of-sync is because the pulse of their respective clock has not the same period. A simple translation of one seismogram with respect to the other one can then not be an adequate answer to the problem, especially for long recordings. Local earthquake data from GII and JSO show that twin seismograms solely corrected (from cross-correlations of entire seismograms) by a translation in time display noticeable differences already 20-30 sec away from the central part of seismograms. This means that by translation only, some parts of seismograms can be adequately synchronized while other parts can remain under- or over-corrected.

Once a twin pair of seismograms has been detected, a deeper analysis is performed by Twins Merge in order to provide an adequate correction valid for entire seismograms. This is done, in addition to simple translation, by finely modifying the sampling rate (between -0.3 % and 0.3 % in steps of 0.01 - 0.02 %) of the permanent seismogram until the best overall fit is obtained with the neighbour twin seismogram. This adjustment of the sampling rate is done by cubic spline interpolation of the initial time series but it is not applied on the recorded seismogram. The core assumption of Twins Merge is that the clock of each network has a constant but own pulse during any particular event. This pulse can vary from one event to another. Neighbour seismograms are then best corrected by a translation combined with a fine adjustment to their nominal sampling period. Seismograms from the Permanent network, however, are never corrected.

The potential accuracy of the method is clearly better than the sampling period but the practical accuracy for any particular event can vary according to the quality of the data, the duration of the event, and the stability of the clocks. For the Dead Sea region, correction uncertainties not exceeding 1 sampling interval of the smallest sampling period involved are usually observed when at least 3 to 5 common stations with acceptable data are available. The most important factor degrading the accuracy is usually the presence of strong electronic noise on one or both seismograms of a twin pair.

The search for twin seismograms and the determination of the sampling rate correction are the most time-consuming tasks of Twins Merge. Several options allow however to speed-up the search by reducing the number of cross-correlations to be performed. For instance, the search for twins can be restricted to seismograms for which the first sample occured within a few seconds or minutes of each other. Other selection options are also possible.

Besides the central property of accurately synchronizing and merging seismograms, Twins Merge provides also for two related networks the automatic recognition of common events, and an automatic recognition of common stations for each common event. These beneficial side-effects also increase the overall quality of merged databases of seismograms compared to any manual approach of the synchronization problem.

Twins Merge provides a detailed log of the search results, and a summary file that will be used for merging. For each event, the summary file displays the time correction and the sampling interval adjustment for each pair of detected twins. In this version of the program, the user is required to somewhat analyze the results in order to decide whether or not to include any particular event for merging, and to determine average synchronization and sampling rate corrections to apply to the Neighbour seismograms. This is usually easy to do from the results displayed in the summary file.

In Merge Mode, Twins Merge reads the edited summary file and performs the merging. Files for merging are copied into the appropriate directories of the permanent network data where their SAC reference time and sampling period are then corrected. The header of all seismograms of a particular event are also adjusted in a similar way as with the Synchronize function of Sac2000. Clock corrections, however, are only applied to seismograms from the neighbour network.

Naturally, this process can be repeated as many times as there are neighbour network databases to merge.


Twins Merge screenshots

Figure 2 displays the Neighbour seismogram (in red) from GHLI superimposed over the Permanent seismogram (in blue) form GLH . The two corrections determined automatically by Twins Merge are a synchronization of 200.156 sec and a small Sampling Period adjustment of -0.02 %. A positive synchronization correction means that the Neighbour seismogram occured too early (by 3 minutes 20 sec and 156 msec). The positive and important value of the main correction fits with the fact that, although being much shorter than the Permanent seismogram and requiring a shift of 63.977 sec between first samples (see seismograms below), the Reference Time (first sample) of the Neighbour seismogram has been recorded as occuring earlier than the Reference Time of the Permanent seismogram. A fine correction of -0.02 % (dDT) applied to the Neighbour seismogram sampling rate provided the best match between the two seismograms. This means that with respect to the Permanent seismogram, the Sampling Period of the Neighbour seismogram has been evaluated as being 9.998 msec instead of its nominal value of 10.000 msec.


Figure 2. Twin seismograms (Vertical component) superimposed after synchronization and sampling rate correction. Neighbour seismogram (GHLI) displayed in red, Permanent seismogram (GLH) displayed in blue.


Figure 3 displays a detail of Figure 2 around the time of the first P and S arrivals. The seismograms in Figure 3 were also de-trended and equalized.


Figure 3. Detail of Figure 2 after de-trending and equalization


The content of the Summary file for the sample seismograms provided with Twins Merge is displayed below. Results are typical for the method and give a good idea about the reliability of the determined time corrections. The sampling period of the Permanent data is 20 msec while the sampling period of the Neighbour data is 10 msec. First, one can see that true common stations have been identified during the search. One letter is sometimes different at the end of the station name between the Permanent stations (P) and the Neighbour stations (N) but this results from station label differences between GII and JSO. DRGI and DSII refer correctly also to the same station.

    Permanent : D:\Samples\Permanent\9601012149\
    Neighbour : D:\Samples\Neighbour\19960101214717\

    PStationsN     Correl     Synchro Corr     Samp Corr
    GLH   GHLI      0.888        200.156 s       -0.02 %
    HMDT  HMDI      0.913        200.131 s        0.05 %
    MMLI  MMLI      0.930        200.145 s        0.01 %

    SYNCHRO CORR: 00000.000 SEC   SAMP CORR: 00.00 %


    Permanent : D:\Samples\Permanent\9603241955\
    Neighbour : D:\Samples\Neighbour\19960324195637\

    PStationsN    Correl    Synchro Corr    Samp Corr
    DHLJ  DHLJ     0.497        -0.984 s      -0.02 %
    DRGI  DSII    -0.895        -0.994 s      -0.02 %
    KFNJ  KFNJ     0.923        -0.989 s      -0.01 %
    LISJ  LISJ     0.694        -0.989 s      -0.01 %
    MASJ  MASJ    -0.886        -0.985 s      -0.04 %
    MKRJ  MKRJ    -0.871        -0.989 s      -0.01 %
    SALJ  SALJ     0.907        -0.984 s      -0.02 %

    SYNCHRO CORR: 00000.000 SEC   SAMP CORR: 00.00 %


For the first event, the synchro correction is high, about 3 minutes and 20 seconds. The high synchro correction does not prevent, however, a dispersion not greater than +/- 14 msec around the central value of 200.145 sec. The dispersion is smaller than +/- 1 sampling period of the Permanent data set but there are only 3 observations. The sampling period correction is low but not very stable. A possible central value would be +0.01 %.

For the second event, there are 7 observations and the synchro correction of -0.989 sec. appears frequently, with a very low dispersion of only +/- 6 msec at the most. The sampling period correction can be set at the very low value of -0.01 %. These are excellent results.


Download Twins Merge 1.1.0

Operating System
Executable and Help
Download (ZIP 9.90MB)