Integrated, adaptive and machine learning approaches to estimate the ghost wavefield of seismic data

In exploration geophysics, seismic measurements are used to obtain information about the subsurface. A large proportion of these measurements take place in oceans, seas and lakes, where the sources and the receivers are generally located somewhere between the water bottom and the water surface during data acquisition. The sources emit an acoustic signal into the subsurface and the receivers measure, amongst other things, the reflections of this signal. Some of these signals only reflect within the subsurface, but others may reflect at the water surface one or more times. The signals that reflect at the water surface disturb the reflections from the subsurface and have a destructive effect on the bandwidth. In this thesis the focus is on the removal of signals with the first reflection and/or the last reflection at the water surface. Correctly removing these so-called ghost reflections will improve the bandwidth.

In this thesis, three methods are covered, that aim to integrate the removal of ghost reflections into another process, or to improve the removal of ghost reflections under specific conditions. The first method integrates the removal of the receiver ghost into closed-loop surface-related multiple estimation. The results on modeled data and field data show that this is an efficient approach and provides a significant improvement over a sequential workflow. This first method, like many other methods that remove ghost reflections, requires accurate information about the depth of the receivers relative to the surface of the water. Due to a dynamic sea surface or movement of the cables this information about the depth of receivers is often not accurate, limiting the removal of the receiver ghost. The second method optimizes the removal of the ghost reflections by estimating and incorporating the depth of receivers relative to the dynamic water surface in this ghost removal process. On modeled data and field data, we show good results for cases where accurate information about the depth of the receivers relative to a dynamic water surface is not available.
The first two methods address the removal of the receiver ghost, and it is well known that the receiver ghost should be removed in the shot domain. This is different when removing the source ghost, which has to be done in the receiver domain. However, in practice, the receiver domain is often coarsely sampled, complicating the removal of the source ghost in this domain. The third method handles the removal of the source ghost in the coarsely sampled receiver domain by training a convolutional neural network. The training data consist of coarsely sampled shot records with and without the receiver ghost that can be obtained relatively easy because the corresponding densely sampled shot records are available as well. Using reciprocity, these training data are a representative data set for removing the source ghost in the coarsely sampled receiver domain. The modeled data and field data results show that this machine learning approach is able to accurately remove the source ghost in the receiver domain. The modeled data results also show that this approach significantly improves the removal of the source ghost compared to its removal in the densely sampled shot domain.