CNN Based Automatic Fault Detection in 3D Seismic Images

Fault detection in seismic images is important in interpreting seismic changes because the faults confirm unexpected subsurface topographical changes. In the traditional approach, faults are manually detected from post-stack seismic data and identified as reflection discontinuities, which is a very tedious process. To increase efficiency and reduce time consumption, a number of automatic fault detection techniques have been developed, of which techniques based on deep learning have proven to be efficient. This work proposed a deep learning based method to generate a seismic attribute called fault probability to highlight fault zones in the 3D seismic images. Since the technique detects the faults directly from seismic volumes, pre-computed attributes such as those used in coherence or edge detection methods are not strictly necessary. The proposed technique is instigated in two stages- training and prediction. In the training stage, a CNN model is trained with real data taken from 7 annotated seismic volumes, in which every point is labeled as fault or no-fault. Then in the prediction stage, the trained network is used to calculate the fault probability at every location in the new seismic image volumes. Both synthetic and real data sets are used to validate the proposed method. The obtained results proved that the proposed deep learning-based fault detection method outperforms some existing methods and also achieves effective performance compared to humans on an expert labelled seismic image dataset, much exactly predicting subtle faults which cannot annotate by an expert interpreter.