A convolutional neural network for seismic dip estimation

The seismic volumetric dip is widely used in the horizon and fault interpretation. One of the mostly used dip estimating method is the waveform similarity-scanning-based dip estimation which can deliver the reliable dip estimation. However, the waveform similarity-scanning-based (WSSB) dip estimation is computationally intensive. In this abstract, we tried to use deep learning to increase the seismic dip estimation's efficiency. We considered the seismic volumetric dip estimation problem as one convolutional neural networks (CNN) regression problem, and proposed a multi-layer convolutional neural network for seismic dip estimation CNN (SDE-CNN). The proposed SDE-CNN can estimate x-direction apparent dip and y-direction apparent dip spontaneously. Finally, we applied the proposed SDE-CNN to one 3D field seismic dataset. Part of the original seismic dataset and corresponding WSSB dip estimation results are adopted as the training pairs to train our SDE-CNN. The trained SDE-CNN was applied to the rest of 3D field seismic dataset to estimate apparent dips. The results show our SDE-CNN can obtain a similar result with the WSSB dip estimation but use less time than the WSSB dip estimation. The curvature estimations based on our SDE-CNN result and the WSSB dip estimation result also show the accuracy of our SDE-CNN-based dip estimation.