Comparing Modern Deep Learning with Traditional Methods for Seismic Fault Interpretation

The primary method for interpreting 3D seismic data involves manually or automatically identifying faults and horizons within the seismic volume. Manual picking is subjective and time-consuming, relying heavily on interpreter expertise. While conventional automatic methods track horizons well, they struggle with fault detection, often producing inaccurate geometries. Seismic attributes help enhance subtle faults missed in the original data. Recent advances in deep learning, specifically convolutional neural networks (CNNs), enable automatic feature detection, as demonstrated in this study comparing CNN-based fault detection to traditional methods using seismic attributes and manual analysis. The study focuses on the Orange Basin offshore southwestern Africa, known for its complex fault system. Prior to CNN application, data underwent structural smoothing. The CNN approach proved more efficient, providing cleaner results in less time and with fewer hardware requirements compared to manual methods. Despite conditioning, seismic attributes struggled with noise removal. The CNN method, taking only hours compared to months for conventional techniques, demonstrated superiority in fault recognition for complex structures like the deep-water fold-and-thrust belts system. Our study demonstrates the effectiveness of an improved Fault-Net model in accurately identifying faults in complex 3D seismic data.