基于多分辨率U-Net网络的地震数据断层检测方法

断层检测是地震资料解释的一项重要工作。基于相干体、曲率等属性的常规断层检测方法不够直观，人工手动拾取断层无法高效处理实际生产中的海量地震数据。深度学习网络由于具有强大的特征提取能力和高效的特征表达能力，近年来被广泛应用于地震数据处理和解释中。为此，提出了一种基于多分辨率U-Net网络（MultiResU-Net）的断层检测方法，即引入多分辨率模块增强网络模型的多尺度断层检测能力，使用残差路径代替普通跳跃连接，缩小用于拼接的特征图之间的语义差别。相比于普通U-Net网络，训练完备的多分辨率U-Net网络模型测试结果具有更高的准确度，Jacard指数和Dice系数分别提高了0.027和0.136，并且断层检测错误率降低了0.094。通过网络中间层可视化分析直观地展示了网络模型对地震数据的特征提取、表达过程。将网络扩展到三维并与迁移学习结合后，同样在三维实际地震数据应用中取得了较好的效果。该方法对于实际生产中实现高效、自动化断层检测具有重要意义。

A fault detection method of seismic data based on MultiResU-Net

Fault detection is significant for seismic data interpretation. Conventional fault detection methods based on coherent volume and curvature are not intuitive enough. Using manual operation, it is impossible to deal with big seismic data in actual production. Deep learning is widely used in seismic interpretation in recent years because of its powerful ability of feature extraction and expression. This paper proposes a fault detection method based on Multiresolution U-net. It can enhance the multi-scale fault detection ability on network models by introducing multi-resolution blocks, and reduce the semantic difference between concatenate feature maps by using a residual path instead of an ordinary skip connection. The trained network model has higher accuracy than the conventional U-net. The Jacard index and the Dice coefficient were increased by 0.027 and 0.136 respectively, and the fault detection error rate was reduced by 0.094.Through visual analysis of the interlayer in the network, the feature extraction and expression process was displayed intuitively. When the network is extended to 3D, and combined with transfer learning, satisfactory fault detection in raw 3D seismic data can be obtained. It is of great significance to realizing efficient and automatic fault detection in actual production work.