Fault detection in seismic data using a true 3D global attention convolutional network with self-supervised denoising...

Area of Science:

• Geophysics
• Artificial Intelligence
• Deep Learning

Background:

• Accurate fault detection in 3D seismic data is crucial for understanding subsurface geology.
• Current methods using hybrid CNN-transformer models fragment fault geometries, limiting the capture of continuous fault planes.
• Seismic noise often leads to misclassification of faults due to limited data and noise exposure.

Purpose of the Study:

• To present the first application of unfactorized 3D global attention for seismic fault segmentation.
• To improve the accuracy and generalization of fault detection models.
• To address the limitations of fragmented attention mechanisms in capturing continuous geological structures.

Main Methods:

• Developed a hybrid U-Net model integrating 3D convolution with unfactorized 3D global self-attention.
• Employed a two-stage training approach: self-supervised denoising pretext training followed by discriminative transfer learning.
• Utilized layer-wise learning rate decay and Unified Focal Loss for training on fault-labeled data.

Main Results:

• Achieved Dice coefficient of 0.853 and IoU of 0.744 on Thebe survey data, showing statistically significant improvements over state-of-the-art 3D CNN+Swin architecture.
• Demonstrated wider improvements of 9.9-16.7% on unseen test data, indicating superior generalization.
• Validated the effectiveness of unfactorized global attention and noise-aware pretext training in enhancing fault detection.

Conclusions:

• Unfactorized 3D global attention, combined with volumetric convolution, effectively models continuous fault planes in seismic data.
• Noise-aware pretext training significantly reduces misclassification of seismic noise as faults.
• The proposed approach establishes a new state-of-the-art in seismic fault detection accuracy and generalization.