Improving faults continuity for extraction by transfer learning based on synthetic data

Xing, Liyuan (Norwegian University of Science and Technology) | Aarre, Victor (Schlumberger AS) | Theoharis, Theoharis (Norwegian University of Science and Technology)


Fault interpretation requires significant manual effort. Automatic fault interpretation is the process of automatically interpreting fault surfaces from 3D seismic volumes and usually involves the following steps: (1) compute edge attributes, (2) calculate fault likelihoods and (3) extract fault surfaces. However, a human editing or merging procedure is usually demanded in step (3) to extract large surfaces from small discontinuous regions. A possible step toward improving region merging is to enhance the continuity (refers to the extent of the fault regions, not to the seismic reflection itself) beforehand. The proposed transfer learning method is trained on synthetic circle data with different connectivity which is balanced over all dips and curvatures, and applied to fault likelihood data with discontinuity and no labelling. Specifically, two neural networks (single and combined), including convolutional and fully connected layers, are investigated. The first aims to train directly from discontinuous to continuous circle data, while the second approaches the same objective in two steps: blurring (which improves continuity) and skeletonization. Three synthetic circle datasets (discontinuity, continuity and blur) are generated and fed into the networks for training accordingly, and the trained models are applied on real ant tracking faults for testing. Continuity is evaluated quantitatively both on the synthetic circle datasets and on real fault datasets, showing promising results. The approach is generalizable to the continuity of other geobodies (e.g. salt bodies, horizon).

Presentation Date: Tuesday, October 16, 2018

Start Time: 8:30:00 AM

Location: 204B (Anaheim Convention Center)

Presentation Type: Oral