I demonstrated with the Sigsbee2B dataset that multiples can be reliably attenuated without introducing artifacts and/or damaging primaries as long as very accurate models for both noise and signal are provided. Because it is often not possible to obtain such a model for the signal, the Spitz approximation is used. As such, the Spitz approximation yields a very good separation of primaries and multiples with 3D filters. However, analyzing this separation in the image space, we notice that some weak primaries are attenuated.
This illustrates the necessity to evaluate multiple removal techniques on migrated images as much as possible. Different improvements are possible to make the noise removal better; for instance, by changing the trade-off parameter in equation (7). In addition, estimated primaries with 3D filters could be used as a signal model for a new iteration of multiple removal. Another possibility is to migrate the multiple model and apply the PEFs in the image space directly. There the effects of multiple removal could be quantified in a more interactive way.
Therefore, in addition to the fact that the image space should be used as much as possible for multiple removal, for quality control and/or attenuating multiples Sava and Guitton (2003), one important lesson learned with this dataset is that finding an accurate model for primaries and multiples before noise removal is crucial. We can relate this to the need for imaging to find the right velocity model; to paraphrase Claerbout (1999), everything depends on it.