基于密集扩张卷积残差网络的地震数据随机噪声压制方法

地震数据处理过程中压制随机噪声是提高地震数据质量的重要环节之一, 其关键是有效压制噪声并尽可能地保留有效信号。针对深度学习方法在地震数据去噪处理时局部特征提取的局限性, 提出了一种基于密集扩张卷积残差网络(DDCRN)的去噪方法。DDCRN主要由多个密集扩张卷积特征融合块(DDCFFB)构成, DDCFFB内部的密集块和多尺度扩张卷积可以用来并行提取特征, 融合结构可以用来融合特征, 残差结构则跳跃连接通道数。其中, 密集块连接不同的卷积层来学习特征, 关注局部特征的传播和重用, 高效提取复杂信息; 多尺度扩张卷积扩大感受野, 增加特征提取范围; 残差学习则加快网络训练的收敛速度。分别采用K-奇异值分解(KSVD)、频域-空间域反卷积(f-x decon)、去噪卷积神经网络(DnCNN)、U型网络(Unet)以及DDCRN去噪方法对合成地震数据和实际地震数据进行去噪处理。结果表明, DDCRN去噪方法不仅能更有效地压制随机噪声, 同时还能更完整地保留同相轴的连续性。

Seismic data denoising is an important processing step to improve data quality. The key to the denoising method is to retain effective signals as much as possible while suppressing the noise. Currently, denoising methods based on deep learning primarily use fixed-scale convolutional kernels to extract local features, which may result in incomplete events. Therefore, we proposed a denoising method based on Dense Dilated Convolutional Residual Network (DDCRN). In this method, multiple dense dilated convolutional feature fusion blocks (DDCFFB) cascade to form a deep network, thereby increasing the information reception range of DDCRN. DDCFFB is mainly composed of two parts that extracted features in parallel. The first part was a dense block that connected different convolutional layers to learn features. Complex information could be extracted efficiently by propagation and reusing of local features. The other part was multi-scale dilation convolution that could access a wide range of information windows. Dilated convolution provided a more comprehensive range of information. The fusion structure combined the features extracted from the two parts. The residual structure accelerated the network training convergence and avoided network degradation by skipping the connection channel. We evaluated the k-singular value decomposition (KSVD), f-x deconvolution (f-x decon), denoising convolutional neural network (DnCNN), u-shaped convolutional neural network (Unet), and DDCRN denoising methods, on the synthetic and real seismic data. The results show that the DDCRN effectively suppresses random noise while preserving the continuity of events compared to the other method.