Removing spurious noise from today's large 3D seismic data volumes is often a less than straight forward proposition. The identification of one or more time or frequency domain attributes has, in the past, allowed for some measure of automatic editing [see trstat, skill, glitches, clean]. Using time - frequency methods [see stft] it is now possible to generate sub-banded time-frequency attributes for much greater control over automatic trace editing. The routine tfskill [time-frequency spike kill] has been written to provide a first cut at this capability.
The routine utilizes a two pass approach. During the first, or scoping pass, a time - frequency statistics table is constructed. The following algorithm is used to transform the data and tabulate the statistics:
2. A sliding sampling window of user defined length is centered over each input sample of the trace in turn. Data laying within the limits of the window is extracted.
3. A Gaussian weighting function is applied to the extracted data.
4. A Fourier Transform is applied to the windowed and weighted data resulting in nwin / 2 frequency samples [sub-bands] where nwin is the number of samples in the sampling window.
5. Spectral amplitude information for all sub-bands is stored for that sample position.
6. The sampling window is then moved down one sample and the above process is repeated until the entire trace has been transformed. At this point a time - frequency amplitude spectrum is contained in memory.
7. The frequency sub-band containing the greatest range of spectral amplitudes is identified.
8. The maximum and mean spectral amplitudes are calculated for that sub-band. The range between the maximum spectral amplitude and the mean is also computed as is the ratio of the above range to the mean.
9. The calculated statistics are output for each trace in tabular form for use in the second or processing pass.
The second, or processing pass, utilizes the above table in concert with a user supplied parameterization file derived through examination of the statistics table to actually edit the data. As an illustration consider the following field data example:
The data in this example (fig.1) was acquired in the early 90's by GECO during a marine acquisition project in the waters offshore from Trinidad. There were a myriad of difficulties associated with this acquisition not the least of which was instrumentation noise. Throughout the observer notes is listed occasional spiking on various traces. In fact the spiking existed on every trace at all times and at amplitude levels covering a range from much less than to much greater than the recorded seismic amplitudes.
The frequency signature of the noise is similar to the direct arrivals. It actually spans a frequency range equivalent to the recorded seismic bandwidth with most of it's power concentrated in the lower frequencies (fig's 2, 3) rendering classical whole trace FFT methods ineffectual in separating this particular instrumentation noise from the desired seismic signal.
Attempts at discrimination using time domain sample amplitude alone failed due to overlap between signal and noise amplitude ranges. (fig's 4, 5, 6) . When the amplitude threshold was lowered sufficiently to remove all pathological noise events most of the desired seismic signal was unfortunately also eradicated.
Using tfskill to allow discrimination based on frequency content at a given time was required. The scoping pass of the routine [the first pass] was run using the command line:
The resulting tfskill time - frequency statistics file was then examined graphically. Plots of all statistical parameters [mean, range, ratio] were made (fig's 7, 8, 9) using:
FreeFormat -Axy < tfskill_stats | xgraph
to determine if selective identification and hence editing of the offensive traces was possible. It was seen that both the mean and ratio discriminators provided unique identification of the noise events. Focusing on the mean and replotting the data over a range of amplitudes consistent with the seismic data (fig. 10) revealed an apparent correlation between geologic changes and variations in mean spectral amplitude along the line. A limiting function file (table 1) was then created to bracket only the seismic signal. This file was supplied to tfskill during the production pass [the second pass] at which time the data was edited (fig. 11).
Any trace whose mean amplitude parameters fell outside those supplied in the limiting function were killed and flagged as a dead trace. A flat file documenting the affected traces is output [tfskill_kills]. This file may be plotted using:
FreeFormat -Axy < tfskill_kills | xgraph
to give a graphical summary of the affected traces.
In the case of pre-stack data you will be able to see if you have preferentially affected the near or far offsets in some unpredictable fashion. The trace kill information is also tabulated in the tfskill printout file along with the total number of traces killed.
This is the minimum tfskill quality control recommended. As with any automatic editing algorithm, tfskill may yield unpredictable results. Since parameterization of the routine is interpretive, the veracity of the result is directly related to the quality control process used during application.
Trace Index(SrcPnt) | Minimum Mean Value | Maximum Mean Value |
|---|---|---|
The effect of geologic variations on mean spectral amplitude along the line are now more evident (fig. 12). If this type of information can be used to identify interesting or anomalous areas and you would like to make a scoping run on a large volume of data see the USP routine scope3d which offers such capability.
In cases where time - frequency separation of signal and noise are possible tfskill may be used to automatically remove the unwanted noise from large volumes of seismic data. The routine successfully eradicated severe instrumentation noise from the Tringas 3D survey which otherwise would have made the survey unusable.
Return to the USP Home Page