Interpreter-driven automatic image segmentation and model evaluation

by Adam Halpert

Full thesis PDF

Table of contents

Abstract This dissertation addresses a major bottleneck in iterative seismic imaging projects: interpretation and modeling of large, subsurface salt bodies. These salt structures are ubiquitous in many of the world's most active areas for seismic exploration, and are extremely time-consuming and difficult to interpret – a process that is conventionally undertaken manually, with little automation. In this thesis, I propose two tools that can help interpreters increase the degree of automation in the salt interpretation workflow, while allowing them to maintain control of the process through expert guidance and scenario testing. These tools are an interpreter-guided image segmentation algorithm, and an efficient, wavefield-based scheme to evaluate several potential velocity models.

The image segmentation method presented here is modified from a graph-partitioning algorithm designed to be both highly efficient and globally accurate when segmenting an image. To account for the unique nature of seismic images, changes to the algorithm's input data, procedure for building graph edges, and subsequently for weighting those edges, are required. The modified algorithm is capable of performing accurate, fully automatic segmentations of salt bodies in 2D and 3D seismic images; however, a fully automatic approach can fail when the salt boundary is poorly imaged. In these cases, expert interpreter guidance can be supplied, either directly on the 2D image, or on one or more 2D slices of a 3D volume. In the 3D case, the interpreter's picks are projected into the third dimension prior to the automatic 3D segmentation. By incorporating interpreter guidance, highly accurate segmentations of a 3D field image are obtained with limited manual intervention.

If multiple salt scenarios lead to several possible velocity models being created, these models can be quickly tested by synthesizing new source and receiver wavefields from an initial image. Both wavefields are generated using prestack velocity information from the initial image, via a generalized form of Born modeling. This allows the velocity inaccuracies present in the initial model to be identified and, ideally, corrected in future model iterations. Because the synthesized receiver wavefield is imaged with an areal source function (both obtained with the initial velocity model), a single shot can be migrated using any other velocity model to provide an image of targeted locations within the model. Crosstalk issues arising from interfering events in the subsurface offset domain can be mitigated by imaging only sparsely-spaced model locations; alternatively, several separate experiments can be performed at different model locations, and the resulting images summed to provide a more detailed final image. If qualitative inspection is insufficient to determine the most accurate model among those being tested, a measure of image focusing can provide a quantitative comparison based on the proportion of an image's energy focused at or near zero-subsurface offset. This strategy is shown to be effected for synthetic and field datasets, in 2D and 3D.

The image segmentation and model evaluation tools are designed to work together to alleviate the salt interpretation bottleneck in model building. Using a wide-azimuth survey from the Gulf of Mexico, image segmentation is used to isolate a sedimentary inclusion within a salt body, and to define two alternate interpretations of the base of salt. The efficient model evaluation scheme is then used to test these two models, along with one provided with the dataset, in a fraction of the time required for full migrations. Qualitative and quantitative analysis of the results indicates that one of the alternate models is most desirable, and a subsequent remigration of the full dataset with this model provides an image with improved clarity and continuity of subsalt reflectors.