基于数据增广和CNN的地震随机噪声压制

卷积神经网络(Convolutional Neural Network,CNN)是一种基于数据驱动的学习算法,简化了传统从特征提取到分类的两阶段式处理任务,被广泛应用于计算机科学的各个领域.在标注数据不足的地震数据去噪领域,CNN的推广应用受到限制.针对这一问题,本文提出了一种基于数据生成和增广的地震数据CNN去噪框架.对于合成数据,本文对无噪地震数据添加不同方差的高斯噪声,增广后构成训练集,实现基于小样本的CNN训练.对于实际地震数据,由于无法获得真实的干净数据和噪声来生成训练样本集,本文提出一种直接从无标签实际有噪数据生成标签数据集的方法.在所提出的方法中,我们利用目前已有的去噪方法从实际地震数据中分别获得估计干净数据和估计噪声,前者与未知的干净数据具有相似纹理,后者与实际噪声具有相似的概率分布.人工合成数据和实际数据实验结果表明,相较于F-X反褶积,BM3D和自适应频域滤波算法,本文方法能更好地压制随机噪声和保护有效信号.最后,本文采用神经网络可视化方法对去噪CNN的机理进行了探索,一定程度上解释了网络每一层的学习内容.

Random seismic noise attenuation based on data augmentation and CNN

Convolutional neural network (CNN) has been widely adopted in various research fields of computer science. Combining the process of feature extracting and classification, CNN greatly simplifies traditional data processing task. However, as a data-driven algorithm, the generalization ability of CNN is limited in the problem of seismic noise attenuation which lacks labeled data. To solve this problem, we propose a CNN training framework based on data generation and augmentation for seismic noise attenuation. When processing synthetic data, we add Gaussian noise with different variance levels to clean seismic data and further augment training datasets to increase the diversity of features. For real seismic data, the clean data and corresponding noise are hard to acquire, thus we propose a method to generate labeled datasets directly from unlabeled noisy seismic data. In the proposed method, we apply existing denoising method to obtain the estimated clean data and estimated noise from real seismic data. The estimated data retains similar texture characteristics with clean data and the estimated noise has similar probability distribution with real seismic noise. We compare our method with F-X deconvolution, BM3D and adaptive frequency domain filtering method. The experiment results demonstrate that our method can efficiently attenuate random noise while preserving signals. Finally, we adopt neural network visualization methods to our CNN model and the visualization results explain the texture patterns learned by each layer of our network to some extent.