In the migrated sections, we see again the presence of the virtual multiple mentioned in the caption of Figure 2. This event is imaged at 120 meters depth which is the same as the separation between the two events in the model. This event exists because the two reflectors correlate with each other as well as the free surface. Real data may not be prone to this problem due to the saving grace of intra-bed multiples. Such an arrival would have opposite polarity from the virtual multiple and thus be destructed. Because of overly simplistic modeling code, such intra-bed multiples are not part of this test data.
Another feature of the section that deserves note is the ramping of amplitude of the reflectors from the edges to the middle of the model. This phenomenon is due to the experiment enjoying a linear, monotonic fold increase from one at the edges to half of the number of receivers at the center.
The size and cost differences for the two starting points described above for processing are significant. Making the correlation cube from the raw data squares the size of the data. However, after correlation, it is no longer necessary to maintain the extraordinarily long time series of the original data. We are free to discard all of the correlation lags computed after longest time the survey is actually interested in and only need migrate that many frequencies. By thus doing so, we shrink the data back down to about its original size. Therefore, the size of the data sets input to migration are roughly equivalent whether we consider the raw data or the correlated shot gathers. The large difference in processing time comes largely in sorting and write statements. By migrating all of the raw data as one shot gather, we enjoy operating on one entirely populated model space with only loops over depth and frequency. In contrast, the correlated data has the number of receivers equal to the number of shots to loop through, each of which populates only a small segment of the model space.
Comparisons of the time taken to migrate the two sections shown in Figure 3 show the correlated sections taking a bit more than twice as to compute with the same program. This does not include the time needed to produce the correlation volume from the raw data to use as input which makes the comparison even worse. Further, the raw data migrated section shown here was computed with a new parallel migration program that runs on our multi-node computer cluster. This architecture is well suited to the structure of real passive seismic datasets where we can expect a reasonably small model space and need to loop through a huge number of frequencies due to the multiple hours of recording.
Lastly, due to the source wavefield being completely full instead of incredibly sparse, as in conventional shot-profile migration, there is an opportunity to investigate better imaging conditions as discussed in Valenciano and Biondi (2002). This type of advanced imaging condition would also address the existence of the virtual multiples if the intra-bed multiples are of insufficient strength to cancel the multiple in real data considerations.