Automatic Interpretative Seismic Image-Focusing Analysis
- Thesis pdf
Table of contents
- Chapter 1: Introduction
- Chapter 2: Measuring image focusing with a CNN
- Chapter 3: 2D synthetic subsalt and field data example
- Chapter 4: Application to 3D Netherlands F3 data
- Chapter 5: Conclusion and future directions
- Appendix A: Automatic fault segmentation with a CNN
- Appendix B: Seismic image processing for domain shift reduction
- Appendix C: 4D cross-correlations with 3D kernels
The use of both the physical (midpoint and depth) and the offset/angle axes of seismic images during seismic image-focusing analysis enables a robust assessment of the errors within the migration velocity model. In fact, the use of the focusing information from all axes of the seismic image is crucial for imaging scenarios in which the subsurface is poorly illuminated. However, while the curvature information provided within the offset/angle axis is routinely used to assess velocity errors, the focusing within the physical axes of a seismic image is often neglected. This is due to the highly interpretative nature of the focusing of geological features within seismic images, therefore making focusing analysis within the physical space difficult to automate in a robust way.
I present a data-driven approach to automate interpretative seismic image-focusing analysis that is based on the recent success of supervised deep learning with convolutional neural networks (CNNs). I designed and developed a workflow that uses the output of a fault-focusing CNN as a focusing measure for geological faults within seismic images. As the CNN operates on prestack image patches, it makes use of both the spatial and angle/offset focusing provided by prestack seismic images. I demonstrate the effectiveness of this workflow by applying it to a 2D limited-aperture prestack image from the Gulf of Mexico and a 3D prestack image from the F3 block of the Dutch North Sea. For each application, I train the CNN on a small subset of real focused and unfocused image patches extracted from the images themselves, and also on synthetically generated focused and unfocused image patches to overcome the lack of real image patches for training. I then apply the trained CNN to each image during focusing analysis and demonstrate that the CNN-based approach is able to robustly estimate focusing maps from each of the images even in the presence of limited illumination.
Using the estimated focusing maps, I am able to improve the focusing and automatic interpretation of faults present within the unfocused images through an image refocusing procedure. When comparing the resulting refocused images from the CNN-based approach to those obtained from a traditional semblance-based approach that only uses the curvature along the aperture-angle axis, I show on both 2D and 3D field data examples that the automatically interpreted fault surfaces within the refocused image from the CNN-based approach are more coherent and geologically consistent with the faults observed within the image.