Fault Detection via 2.5D Transformer U-Net with Seismic Data Pre-Processing

Abstract

Seismic fault structures are important for the detection and exploitation of hydrocarbon resources. Due to their development and popularity in the geophysical community, deep-learning-based fault detection methods have been proposed and achieved SOTA results. Due to the efficiency and benefits of full spatial information extraction, 3D convolutional neural networks (CNNs) are used widely to directly detect faults on seismic data volumes. However, using 3D data for training requires expensive computational resources and can be limited by hardware facilities. Although 2D CNN methods are less computationally intensive, they lead to the loss of correlation between seismic slices. To mitigate the aforementioned problems, we propose to predict a 2D fault section using multiple neighboring seismic profiles, that is, 2.5D fault detection. In CNNs, convolution layers mainly extract local information and pooling layers may disrupt the edge features in seismic data, which tend to cause fault discontinuities. To this end, we incorporate the Transformer module in U-net for feature extraction to enhance prediction continuity. To reduce the data discrepancies between synthetic and different real seismic datasets, we apply a seismic data standardization workflow to improve the prediction stability on real datasets. Netherlands F3 real data tests show that, when training on synthetic data labels, the proposed 2.5D Transformer U-net-based method predicts more subtle faults and faults with higher spatial continuity than the baseline full 3D U-net model.