Fault enhancement comparison among coherence enhancement, probabilistic neural networks, and convolutional neural networks in the Taranaki Basin area, New Zealand

Abstract

Fault identification is a critical component of seismic interpretation. During the past 25 years, coherence, curvature, and other seismic attributes sensitive to faults improved seismic interpretation, but human interaction is still required to generate a complete fault interpretation. Today, image enhancement of fault-sensitive attributes, multiattribute fault analysis using shallow learning, and deep-learning algorithms based on extensive training and convolutional neural networks (CNNs) are promising fault interpretation workflows. We have compared three workflows to test fault-detection capabilities; these include image enhancement, probabilistic neural networks (PNNs), and CNNs. We compared results to human-interpreted faults as our ground truth for a merged 3D seismic survey acquired in the Taranaki Basin, New Zealand. We extracted fault surfaces from the results of the workflows using them as seed points for an active contour method. Extracted faults are then compared to the human-interpreted surface using the Hausdorff distance. Data conditioning, including spectral balancing and structure-oriented filtering, improved the performance of all three workflows. Although all three approaches produce enhanced fault volumes, we find differences in fault location and different artifacts (mispredicted faults). Because all three methods exhibit “false positive” predictions, the enhanced multispectral coherence method produces faults and stratigraphic edges in the final image, including residual stair-step artifacts. In our implementation, PNN produces many salt-and-pepper artifacts through the resulting image, suggesting that we might need to include better training data or reduce the volume size to reduce the number of relevant classes to obtain an improved classification. The CNN algorithm is trained with synthetic data that provide rapid results, correctly identifying larger faults, but missing smaller faults and, in some cases, misclassifying mass-transport deposits and angular unconformities as being faults.