数据增广和主动学习在波阻抗反演中的应用

在实际应用中，深度卷积网络以大量数据驱动模型进行网络训练，以获得地震数据与阻抗之间的映射关系，但需大量合成数据对网络训练后，再应用少量实际数据对网络进行迁移学习。为此，提出了一种基于数据增广和主动学习的地震波阻抗反演方法。数据增广首先通过同频率重采样对单道原波阻抗数据进行增广，再求取增广后的反射系数和随机核，最后计算增广后的地震数据。将增广后的地震和波阻抗数据作为训练集，结合主动学习思想选择最大误差样本对网络进行迭代训练。该方法不仅可以避免地震子波估计，而且能用少量的标签数据训练出预测精度更高的网络。Marmousi 2模型测试结果表明，该方法仅需十分之一标签数据和迭代次数就能达到与随机迭代训练方法相近的预测精度，且预测误差在剖面上分布更均匀。

Application of data augmentation and active learning to seismic wave impedance inversion

For an effective deep learning based seismic impedance inversion strategy, a deep convolutional network is trained by massive data-driven models to obtain the mapping between seismic data and impedance. After the network is pre-trained by substantial synthetic data, a small amount of real data is required for transfer learning of the network. In this paper, we propose a new method based on data augmentation and active learning for seismic wave impedance inversion. First, the original single-trace wave impedance data is augmented by resampling at the same frequency, and then the reflectivity and random kernel are calculated to generate the seismic data after augmentation. The augmented seismic and wave impedance data is taken as training sets, and the maximum-error samples are selected to train the network iteratively considering active learning. The proposed method can avoid seismic wavelet estimation, while training the network with higher accuracy using a small amount of label data. The test results from the Marmousi 2 model demonstrate that this method only needs one tenth of label data and iteration times to achieve the prediction accuracy similar to that of iterative random training, with the prediction errors distributed more evenly on the profile.