Summary
Detecting faults in 3D seismic images is important in seismic interpretation since the faults reveal abrupt subsurface geology changes. We developed a method using a convolutional neural network (CNN) to generate a fault-probability attribute for highlighting fault zones in seismic amplitude images. The proposed method detects faults directly from seismic amplitude cubes, so that precomputed (e.g., coherence or edge detection) attributes are not required. This new method is implemented in two steps - training and prediction. In the training step, a CNN model is trained with annotated real seismic image cubes, where each point is labeled as fault or non-fault. In the prediction step, the trained network is applied to compute the fault probability at each location in the new image cubes. The method is verified with both synthetic and field datasets. Test results show that CNN-based fault probability outperforms coherence technology.
