Processing single-channel sparker data is not as complex and time-consuming as processing multi-channel seismic data. The data processing steps typically involve basic processes such as removing low- and high-frequency noise, muting, deconvolution and migration. Depending on the type and amount of noise components in the data, specific methods such as trace mixing and F-X prediction filter can be included in the processing. Figure 2 illustrates the proposed data processing flow used in the processing of single-channel sparker seismic data.
4.1. Data Loading
The raw data is typically in a binary format known as SegY or SegD. The process of loading raw sparker seismic data into the data processing system is referred to as “data loading” and also involves a conversion of the raw data format to the internal format of the data processing software. After data loading, the header information of the loaded raw data is checked and the accuracy of essential information for data processing, such as the shot number (Field File ID Number, FFID) and sampling rate, is confirmed. At this stage, necessary header information (e.g., CDP and channel numbers, offset value, etc.) is added to the trace header of the data if not present in the header section.
4.2. Spectral Analysis and Band-pass Filter
Different frequency amplitudes in the seismic data may not always be useful due to the incorporation of both signal and noise components in different frequency bands. Generally, seismic data consists of a linear summation of signal and noise components, including high-frequency random noise and low-frequency swell noise. Therefore, it is often necessary to filter both high- and low-frequency amplitudes from the raw seismics data using a band-pass filter. Spectral analysis is first applied to the data for this process, and the cut-off frequencies of the band-pass filter are then determined.
Sparker data typically includes dominant swell noise components below 60 Hz (Fig. 3a). Although the amplitude of this noise is extremely high, the frequency band is generally lower than the sparker signal frequency band and, therefore, it can be separated from reflection signals. Raw sparker data should also be limited in the high-frequency part of the spectrum due to the ghost reflection interference. Accordingly, the usable frequency band of the seismic data is the portion between the frequency where swell noise amplitudes die and the frequency where the ghost notch is observed (Fig. 3a). Both frequency values are determined from the average amplitude spectrum of the sparker data using spectral analysis.
The filter operator length is an important parameter for the performance of the band-pass filter. Too short filter operators leave residual swell noise amplitudes in the data. On the other hand, an excessively long operator length will unnecessarily extend the processing time. Therefore, tests can be conducted on this parameter to determine the most appropriate operator length. Figure 3 shows a raw single-channel sparker data before and after the band-pass filter and their corresponding amplitude spectra. Prior to the filtering, spectral analysis was used to examine the frequency content of the data, and the cut-off frequencies for the band-pass filter were determined. The high-amplitude swell noise extends up to 80 Hz in the amplitude spectrum of the raw seismic data. Additionally, the spectrum shows that the first ghost notch is around 400 Hz. Accordingly, suitable cut-off frequencies for the band-pass filter for this single-channel sparker data in Fig. 3a were determined as 80–420 Hz. The filter operator length was set at 400 ms. After filtering, the seismic data only contains the amplitudes of frequency components within the passband of the filter (Fig. 3b).
4.3. Top Mute
Muting is a method used to remove amplitudes in the seismic data that arise from noisy areas, typically regions where coherent noise is observed. In these areas, seismic amplitudes are simply multiplied by zero. The top mute process involves removing all events recorded in the water column before the seafloor reflection. With this process, direct waves in sparker data, along with subsequent ghost reflections and bubble effects in the water column, are removed. Figure 4a illustrates the top mute process applied to the sparker data given in Fig. 3b.
The only parameter to be determined for the top mute is the ramp of the mute transition zone. In the mute process, the part to be muted immediately above the seafloor is multiplied by 0, and the part to be retained in the data is multiplied by 1. For the transition zone just above the seafloor, an inclined transition zone should be defined to prevent the occurrence of Gibbs phenomenon in subsequent data processing steps. Since sparker data is high resolution, and a smaller sampling interval compared to standard multi-channel seismic data (e.g., 1 or 2 ms) is chosen for appropriate Nyquist frequency (generally 0.25 or 0.5 ms), the transition zone for the top mute process should be relatively narrow. Otherwise, amplitude of the seafloor reflection may be reduced after muting.
4.4. Surgical Mute
The surgical mute involves removing the noisy portions observed in the inner part of the data (below the seafloor reflection). For sparker data, this process is typically applied to remove high-amplitude transient spike-like noise. The time window in which the spike noise is observed, is picked, and this interval is multiplied by 0. Figure 4b illustrates sparker data containing spike-like noise (blue arrows), and Fig. 4c shows the result of applying the surgical mute process.
Since the regions where the surgical mute will be applied need to be manually determined and picked for each noise region, this process can be quite time-consuming. Additionally, after the process, the amplitudes of the muted portions in the section is zeroed out. To overcome this, it is recommended to apply trace mixing to the data after the surgical mute.
4.5. Trace DC Removal
The component corresponding to 0 Hz frequency in the amplitude spectrum is referred to as “DC component”. This component, exhibiting no oscillation, is added to the entire seismic trace as a constant amplitude value. Under normal circumstances, recorded seismic amplitude values are expected to oscillate around the 0 axis value. However, in some cases, due to a constant voltage interference in the streamer electronics, the DC component originates in seismic traces, and the amplitude oscillation moves to a constant positive or negative value away from the 0 axis. In such cases, removing the DC component from the traces ensures that the amplitudes of the traces oscillate around the 0 axis value. This process is typically done by subtracting the average value of the seismic trace from each seismic sample value. Figure 5 shows an example sparker data before and after DC component removal. After the process, each seismic trace oscillates around the 0 axis value.
4.6. Trace Edit/Kill
All seismic traces recorded as the seismic data may not be practically usable in some cases. The noise level in some traces can be excessively high, or the trace may belong to a missed shot. In such cases, the trace in question is simply zeroed out by multiplying its amplitudes by 0. This process is typically applied to sparker data in the following situations:
-
If the trace belongs to a missed shot
-
If the trace contains excessive amount of power line noise
-
If the trace contains high amount of spike-like noise
-
If there is static time shift in the trace
Trace edit is generally required in sparker data when there are traces from missed shots. In this case, the recording instrument records data, but the sparker source does not fire to produce seismic signal for that shot, and therefore, the recorded trace only contains noise. The issue often arises due to triggering problems with the sparker source. Figure 6 illustrates a sparker section containing missed shots and after killing those traces. The amplitudes of the killed traces are zeroed out. To prevent the zero-amplitude traces in the final processed section, it is recommended to apply trace mixing to the sparker data after the editing process.
4.7. f-k Dip Filter
Seismic sections are comprised of different events, including primary reflections and coherent noise with different slopes. Events with different slopes interfering within the seismics sections, can be separated from each other through 2D Fourier transformation, and events at undesired slopes can be removed from the data using an f-k filter in the 2D Fourier domain. The general purpose of the f-k filter is typically to eliminate linear events from the data, such as tail buoy noise or operational noise. It is commonly applied to shot gathers in multi-channel seismic data. However, the f-k filter can also be applied to stack sections and single-channel sparker seismic profiles.
For sparker data, the f-k filter is applied when it is necessary. Figure 7 shows the single-channel sparker section shown in Fig. 4a before and after the application of the f-k filter. In the zero-offset sparker data, the amplitudes of primary reflections typically exhibit a tendency to accumulate along the 0 wavenumber (central) axis. However, coherent linear noise amplitudes appear both at the negative and positive panels of the f-k plane. Therefore, such coherent noise can be eliminated from the data with an f-k filter that preserves amplitudes along the 0 wavenumber axis while suppressing amplitudes in other sections. In order to remove the noise amplitudes, an f-k polygon is picked, which encloses the primary reflection amplitudes (Fig. 7a). The amplitude values within the polygon is multiplied by 1.0 while the amplitudes outside of the polygon are zeroed out. After the inverse 2D Fourier transform, the seismics data contains only amplitudes in the picked polygon of the f-k plane (Fig. 7b).
4.8. Trace Mix
Trace mixing, commonly applied to zero-offset sections (stacked or single-channel seismic sections), constitutes a noise suppression and S/N ratio enhancement method. In this method, the amplitude samples of a specific number of adjacent traces are summed, and a resultant trace is obtained by simply averaging them. The process can be implemented as weighted or trimmed trace mix. In weighted trace mix, the weighted average of adjacent traces is computed. For sparker data, trimmed trace mix is often employed. In this procedure, after sorting the amplitude samples in ascending order, a certain percentage of the high and low amplitudes at both ends of this ordered series is excluded from the average.
Trace mixing can be conceptualized as a horizontal stacking method and can be effectively employed as a noise suppression and S/N ratio enhancement tool for single channel sparker sections where random noise cannot be suppressed through stacking. Additionally, it can be utilized in interpolating missing traces as well as samples resulting from the trace editing and surgical mute process of single-channel seismic data. Trimmed trace mix is highly effective and practical method for automatically removing spike-like noise observed in the sparker data. Figure 8a shows a part of an example single-channel sparker section containing several high-amplitude spike-like noise bursts. After 20% trimmed trace mix, all high-amplitude spike-like noises, indicated by blue arrows, have been removed from the data, and trace-by-trace consistency is significantly improved (Fig. 8b). Figure 8c shows the result of 20% trimmed trace mix applied to the sparker data shown in Fig. 7b.
4.9. F-X Prediction Filter (F-X Decon)
F-X prediction filter is an effective technique for random noise suppression, relying on time-distance prediction over small windows in seismic data. In the F-X prediction filter, a Wiener prediction filter is applied to a fixed-frequency series in the space direction in f-x domain to predict the next primary reflection amplitude. The difference between the predicted and actual waveforms is considered as noise and subtracted from the data. The effectiveness of the process is reduced with longer filter lengths and wider horizontal window widths, leading to potential artificial events in the output.
The F-X prediction filter is typically applied to post-stack or single-channel seismic data because the stacking process inherently suppresses a significant portion of the random noise amplitudes. However, in the case of single-channel sparker data where stacking cannot be applied, random noise amplitudes can be relatively high especially in late arrivals where the reflection amplitudes are relatively small. In such cases, the F-X prediction filter can be employed to increase the S/N ratio of sparker seismic data. Figure 9a shows the F-X prediction filter application to the sparker data shown in Fig. 8c. For this data, the time length of the filter is 1000 ms, and the horizontal window width is 200 traces. The effectiveness of the F-X prediction filter in suppressing random noise is significant particularly in the late arrivals of the sparker data (Figs. 9b and 9c). The amplitude spectra of the close-ups clearly illustrate the amplitudes of random noise removed by the F-X prediction filter (Figs. 9d and 9e).
4.10. Gain Recovery
The amplitude of the seismic signals travelling away from the source diminishes with distance due to the specific effects such as spherical divergence and absorption, in addition to the secondary factors such as scattering and energy partitioning at the layer interfaces. Consequently, reflection amplitudes in the late arrivals of the seismic sections are significantly small, which requires amplitude enhancement using appropriate gain functions. One of the most commonly employed gain control methods in the seismic data processing industry is known as Automatic Gain Control (AGC). AGC is applied to seismic data using a sliding time window along the trace. As AGC eliminates the relative amplitude changes from trace to trace in seismic data, it is not utilized in hydrocarbon analyses of the seismics profiles. If seismic data is analyzed for subsurface structural information, such as fault mapping, AGC facilitates interpretation by equalizing shallow and deeper amplitude levels.
The most critical parameter that needs to be appropriately chosen in AGC is the window length. As the window length decreases, AGC becomes more effective, and amplitudes in both shallow and deep regions of the data become more balanced. In practice, the suitable window length is selected based on the record length of the input data. Since the record length of sparker data is generally much shorter than that of multi-channel seismic data, relatively shorter AGC window lengths should be preferred for sparker profiles. Figure 10a shows the result of applying AGC to the sparker data given in Fig. 9a using a 75 ms window length. In the AGC output, the amplitudes at shallow and deep parts of the seismic section seem to be balanced.
4.11. Multiple Suppression
Multiple reflections in marine seismic data can pose significant challenges during the interpretation of the data collected with short streamers in relatively shallow waters. While various methods have been proposed to predict and suppress multiple reflections, these techniques (except predictive deconvolution) are effective on multi-channel seismic data only. Therefore, predictive deconvolution was employed to eliminate multiple reflections from single-channel sparker data in this study.
Predictive deconvolution allows obtaining the time-advanced version of the input trace. If the input trace is s(t), its time-advanced version is s(t + α), where α is referred to as the “prediction lag”. In deconvolution, the determination of prediction lag (α) and operator length (n) parameters is crucial. The maximum attenuation in the amplitudes of multiple reflections is achieved when the prediction lag is equal to the period of the multiple reflections, regardless of the operator length. The period of multiple reflections as well as the appropriate prediction lag value to suppress the multiples can be found by analyzing the autocorrelation of the input seismic data. To eliminate the multiple reflections from the data, the operator length should be chosen long enough to encompass the first isolated energy packet in the autocorrelation. Figure 10b shows the result of applying predictive deconvolution to the sparker data given in Fig. 10a. The autocorrelations of the sections are also presented beneath the sections. In the deconvolution process, the prediction lag (α = 120 ms, water depth dependent) and operator length (n = 30 ms) were selected. It is observed that predictive deconvolution successfully suppresses multiple reflections in the sparker data particularly originated from the seafloor.
4.12. Spiking Deconvolution
After applying predictive deconvolution to remove multiple reflections from the seismic data, spiking deconvolution is applied to enhance temporal resolution and eliminate the bubble effect as well as ghost reflections. This process eliminates the ringy character of the data, compresses the seismic wavelet, and provides a seismic section with higher temporal resolution.
Because sparker data is inherently mixed-phase, and spiking deconvolution is typically applicable only for minimum-phase data, it is necessary to first convert the data to its minimum-phase equivalent before applying spiking deconvolution. Figure 11 illustrates the application of spiking deconvolution to the single-channel sparker data given in Fig. 10b. In this study, the sparker data was transformed to its zero-phase equivalent, instead of minimum-phase, before applying zero-phase spiking deconvolution to the entire section. The operator length for the deconvolution process is chosen as n = 30 ms. The effectiveness of the process in eliminating ghost reflections and bubble effects in the data is particularly evident in the close-ups presented in Figs. 11b and 11c. The clinoforms of the delta structure, concealed by ghost reflections and bubble effect that extends parallel to the seafloor reflection (indicated by blue arrows in Fig. 11b), are now much clearer after deconvolution. The average amplitude spectra of the seismic sections before and after deconvolution are shown in Figs. 11d and 11e, respectively. From the spectra, it can be observed that spiking deconvolution enhances the amplitudes of high-frequency components, particularly between 300 and 400 Hz.
4.13. Migration
Seismic migration is a process based on the wave equation that migrates dipping reflections to their true locations along the time and distance axes, eliminating the diffraction energy. The goal of migration is to make the stacked section obtained along the seismic line resemble the geological section. After migration, the lengths of reflectors in the section shorten, and their dips increase. Synclines in zero-offset sections widen, while anticlines narrow. Migration eliminates diffractions resulting from sudden terminations of reflectors and makes fault planes in the section more distinct. Horizontal layers remain unaffected by the migration process.
The migration can be applied as pre-stack or post-stack to the multi-channel seismic data, either in time or depth. As they are zero offset, post-stack time migration algorithms can be applied to single-channel sparker seismic data. Although there are several migration algorithms used in the seismic data processing industry, the most commonly used for sparker data are Stolt and Kirchhoff time migrations. The Stolt method is valid for constant velocity medium and is often used for quality control purposes in multi-channel seismic data due to its fast migration algorithm. Kirchhoff migration is the most widely used algorithm today, which accounts for both horizontal and vertical velocity variations.
All migration algorithms require knowledge of the seismic velocity of the subsurface. However, obtaining the velocity distribution of the subsurface from single-channel sparker data is not possible. Therefore, constant velocity migration is typically applied to sparker data, assuming a seismic wave velocity of 1500 m/s for whole seismics section. Figure 12 illustrates the sparker data shown in Fig. 11a after Stolt and Kirchhoff migrations. Applications demonstrate that, in structurally complex environments, Kirchhoff migration algorithm makes the fault planes more distinct. However, for the shallow subsurface which is not structurally complex, both Stolt and Kirchhoff migrations produce approximately equivalent results for single-channel sparker data (Fig. 12). On the other hand, since it is the fastest and hence the most computationally efficient algorithm, the Stolt algorithm is often preferred for processing sparker data.