AG-Net3D: Embedding Attention Gate into U-Net for 3D Seismic Data Fault Detection

Automatic fault detection of seismic data plays a remarkable role in ensuring the efficiency even effectiveness of oil and gas field exploration. Deep learning methods such as U-Net are becoming competitive alternatives for traditional ones such as coherency and curvature. However, conventional U-Net may cause the loss of underlying features of faults, resulting in poor performance of complex field data in terms of details. Besides, the open data set we obtained is insufficient to train an excellent deep model. In this paper, we propose a new approach called AG-Net3D. First, AG-Net3D incorporates Attention Gate into the decoders of basic U-Net structure, allowing the network to focus more attention on the region where faults are located, effectively improving the accuracy of fault detection. Second, for the high imbalance between fault and non-fault in the image, we propose an improved BCE loss function to regulate the weights of the fault and non-fault images and increase reasonable attention to fault. Moreover, to increase the volume and diversity of training data, we synthesize images based on the given ones by changing the dip angle, amplitude, etc. Experiments are undertaken on synthetic and field data in comparison with several popular methods. Results show that our model has stronger noise resistance, higher generalization ability, higher accuracy and better fault continuity.