利用反射地震资料和多尺度训练集的深度学习速度建模

随着地震勘探数据量的逐渐增大，常规地震速度建模方法在稳定性、精度和计算效率等方面均面临挑战。为此，提出一种利用反射地震资料和多尺度训练集的深度学习速度建模的方法，即将反射波形数据和速度谱联合作为全卷积神经网络的输入，并在网络中引入Dropout层提高泛化能力，结合多尺度训练集，实现从地震数据到速度模型的映射。为了测试该方法在不同地质构造条件下的效果和适用性，分别应用层状模型、孤立异常体模型和BP盐丘模型进行数值实验。实验结果表明，联合使用地震反射波形和速度谱作为深度学习特征数据集时，速度建模准确性优于仅采用地震反射波形或速度谱作为特征数据集的结果，并克服了单独使用反射波形导致建模不稳定和单独使用速度谱建模精度不足的缺陷；使用多尺度速度模型构建训练集的速度建模结果在异常体边界的准确性优于采用单尺度模型训练集；深度神经网络只需经过一次训练，就可以快速地对与训练集中速度结构相似的地下构造进行速度建模，比常规方法具有更高的计算效率。在构建大量速度模型时，该方法具有很好的推广价值。

Deep learning-driven velocity modeling based on seismic reflection data and multi-scale training sets

With the enlargement of seismic data volume, conventional methods of seismic velocity modeling are facing challenges to stability, accuracy, and computational efficiency. A deep learning-driven velocity modeling method based on seismic reflection data and multi-scale training sets was proposed in this paper. This method combined reflection waveforms and velocity spectra into the input of a full convolutional neural network and adopted the dropout layer in the neural network to improve the generalization ability. Moreover, it integrated multi-scale training sets to realize the mapping from seismic data to velocity models. To test the effectiveness and applicability of this method in different geologic structures, numerical experiments were carried out with layered models, isolated anomaly models, and the BP salt dome models. Experimental results show when seismic reflection waveforms and velocity spectra are combined as the feature dataset of deep learning, the accuracy of velocity modeling is higher than that in the case where they were adopted individually. It overcomes the defects of unstable modeling caused by using reflection waveform alone and low modeling accuracy induced by using velocity spectra alone. Furthermore, the accuracy of velocity modeling results in anomaly boundaries with the multi-scale velocity model to construct the training set is higher than that with a single-scale model as the training set. After only one training process, the deep neural network can quickly build the velocity model of the underground structure which is similar to the velocity structure in the training set. Therefore, compared with the conventional method, it has higher computational efficiency and deserves promotion when building large amounts of velocity models.