In spite of numerous advances in computational power in recent years, interpretation still requires a lot of manual picking. One of the main goals of interpretation is to extract from the seismic data geological and reservoir features. One commonly used interpretation technique that helps with this effort is to flatten data on horizons [e.g. Lee (2001)]. This procedure removes structure and allows the interpreter to see geological features as they were emplaced. For instance, after flattening seismic data, an interpreter can see in one image an entire flood plain complete with meandering channels. However, in order to flatten seismic data, a horizon needs to be identified and tracked throughout the data volume. If the structure changes often with depth, then many horizons need to be identified and tracked. This picking process can be time consuming and expensive.
Certain interpretation visualization products and auto-pickers seek to make picking and flattening processes as efficient as possible. However, they often suffer from weaknesses that prevent them from being truly practical. For example, 3-D volume interpretation packages allow interpreters to view their data with depth perception using stereo glasses. These products have an opacity ability James et al. (2002) that allows interpreters to make unwanted data transparent. Unfortunately, unless the zone of interest has a known unique range of attribute values, interpreters resort to picking on 2-D slices. Additionally, traditional amplitude-based auto-pickers can fail if the horizon being tracked has significant amplitude variation, or worse, polarity reversal. Other tracking techniques such as artificial neural networks are less sensitive to amplitude variations but are still prone to error if the seismic wavelet character varies significantly from the training data Leggett et al. (1996).
In this document, we propose a method for automatically flattening entire 3-D seismic cubes without picking that we first presented in Lomask (2003a). This is essentially an algorithm that is efficient enough to perform dense picking on entire 3D cubes at once. Our method involves first calculating dips everywhere in the data using a dip estimation technique Claerbout (1992); Fomel (2002). These dips are resolved into time shifts via a non-linear least-squares problem. The data are subsequently shifted according to the time shifts to output a flattened volume. Bienati et al. (1999a,b); Bienati and Spagnolini (1998, 2001) use a similar approach to resolve numerically the dips into time shifts for the purpose of auto-picking horizons and flattening gathers, yet solving a linear version of the problem and without flattening the full volume at once. Stark (2004) takes a full volume approach to achieve the same end yet is unwrapping instantaneous phase. Blinov and Petrou (2003) use dynamic programming to track horizons by summing local dips. Here, a version of the non-linear problem of summing local dips Guitton et al. (2005); Lomask (2003b) is solved iteratively using a Gauss-Newton approach. Each iteration utilizes the Fourier domain to invert efficiently a linearized operator much like the approach of Ghiglia and Romero (1994) for unwrapping two-dimensional phase. For faulted data, weights identifying the faults are applied within the iterative scheme, allowing reconstruction of horizons for certain fault geometries. As with amplitude based auto-pickers, amplitude variation also affects the quality of the dip estimation and will, in turn, impact the quality of this flattening method. However, the effect will be less significant because this method can flatten the entire data cube at once, globally, in a least-squares sense, minimizing the effect of questionable dip information. Additionally, flattening the entire cube at once should make the method more robust in noisy data or complicated structures. Once a seismic volume is flattened, automatic horizon tracking becomes a trivial matter of reversing the flattening process to unflatten flat surfaces. The prestack applications for this method are numerous and can be easily incorporated into an automatic velocity picking scheme Guitton et al. (2004).
In the following sections, we present an overview of the flattening methodology and a series of real world geological challenges for this method in order of increasing difficulty. The first is a 3D synthetic that is flattened to demonstrate how we can flatten faulted, folded data. Then we present a structurally simple salt peircement 3D field data from the Gulf of Mexico. We consider it structurally simple because the dip does not change much with depth. Increasing complexity, we flatten 3D field data from the North Sea that contains an unconformity and has significant folding. Lastly, we illustrate how this method is used to flatten datasets with faults on a field 3D Gulf of Mexico dataset.