基于深度学习的相对地质时间体估计方法

常规自动解释方法难以正确追踪三维地震数据体中被断层错断的地震反射层位。为此，提出了基于深度学习的相对地质时间体估计方法。首先，针对相对地质时间体估计需求，设计一个由编码器—解码器框架组成的相对地质时间体估计网络；其次，采用结构相似性准则为损失函数，利用生成的准确标注合成训练数据集对相对地质时间体估计网络进行训练，使其具备准确地从地震数据体中估计相对地质时间体的能力；最后，通过提取多个恒定的相对地质时间体等值面实现多个地震反射层位的自动追踪。测试结果表明，该方法不仅在验证数据集上显示出优异性能，而且在实际地震数据体上也获得了较好效果；利用估计的相对地质时间体能够一次性获得多个能够表征地层空间形态的地震反射层位。

Relative geological time volume estimation method based on deep learning

Conventional automatic interpretation methods are difficult to correctly track seismic reflection horizons dislocated by faults in 3D seismic volume. Therefore, a relative geological time volume estimation method based on deep learning is proposed. Firstly, according to the estimation requirements of the relative geological time vo-lume, an estimation network of the relative geological time volume composed of an encoder-decoder framework is designed. Secondly, the estimation network of the relative geological time volume is trained by the generated accurately labeled synthetic training dataset with the structural similarity criterion as the loss function, and thus the network can accurately estimate the relative geological time volume from the seismic volume. Finally, the automatic tracking of multiple seismic reflection horizons is realized by extracting multiple constant iso-surfaces of the relative geological time volume. The test results show that this method not only shows excellent performance on the validation dataset but also achieves a positive application effect on the actual seismic volume. In addition, multiple seismic reflection horizons which can characterize the spatial shape of the stratum can be obtained at one time by using the estimated relative geological time volume.