Random noise suppression of seismic data based on deep convolution neural network
-
摘要: 本文以提高地震数据的成像质量为目标，提出一种智能的卷积神经网络降噪框架，从带有噪声的地震数据中自适应地学习地震信号。为了加速网络训练和避免训练时出现梯度消失现象，我们在网络中加入残差学习和批标准化的方法，并采用了ReLU激活函数和Adam优化算法优化网络。此外，Marmousi和F3数据集被用来对网络进行训练和测试，经过充分训练的网络不仅能在学习中保留地震数据特征，而且能去除随机噪声。首先充分地训练网络，从中提取出随机噪声，并保留学习到的地震数据特征，之后通过重建地震数据估算测试集中的波形特征。合成记录和实际数据的处理结果显示了深度卷积神经网络在随机噪声压制任务中的潜力，并通过实验验证表明了深度卷积神经网络框架有很好的去噪效果。Abstract: Random noise suppression of seismic data is essential in seismic data processing. Since the seismic data recorded by the geophone is usually noisy, this kind of noisy data can be regarded as a manifestation of low signal-to-noise ratio. Low SNR data will affect subsequent processing of seismic data, such as migration and imaging. In this paper, we aim to improve the imaging quality of seismic data and propose an intelligent noise reduction framework for convolutional neural network to adaptively learn seismic signals from noisy seismic data. In order to speed up network training and avoid gradient disappearance during training, we add residual learning and batch normalization methods to the network, and use ReLU activation function and Adam optimization algorithm to optimize the network. In addition, the two datasets, Marmousi and F3, are used to train and test the network. A fully trained network can not only retain weak features in learning, but also remove random noise. First, fully train the network, extract random noise from it, and retain the learned seismic data features, and then estimate the waveform features in the test set by reconstructing the seismic data. The processing results of synthetic records and field data show the potential of deep convolutional neural network in random noise suppression tasks, and experimental verification shows that the deep convolutional neural network framework has a good denoising effect.
-
-
![]()
图 1 深层卷积神经网络（DCNN）框架的结构
Figure 1. Structure of deep convolutional neural network framework
![]()
图 2 合成数据的去噪结果（图中方柱表示有效信号，下同）
（a） 干净数据；（b） 加噪数据；（c） NLM去噪结果；（d） BM3D去噪结果；（e） DCNN去噪结果；（f） NLM去噪残差剖面；（g） BM3D去噪残差剖面；（h） DCNN去噪残差剖面
Figure 2. Denoising results of synthetic data （Valid signals are in the boxes，the same below）
（a） Clean data；（b） Noisy data；（c） NLM denoising results；（d） BM3D denoising results；（e） DCNN denoising results；（f） NLM denoising residual profile；（g） BM3D denoising residual profile；（h） DCNN denoising residual profile
![]()
图 3 F3叠后数据的去噪结果
（a） 干净数据；（b） 加噪数据（PSNR＝20.4 dB）；（c） NLM去噪结果（PSNR＝28.97 dB，SSIM＝0.868 9）；（d） BM3D去噪结果（PSNR＝30.31 dB，SSIM＝0.942 8）；（e） DCNN去噪结果（PSNR＝32.43 dB，SSIM＝0.965 8）；（f） NLM去噪残差剖面；（g） BM3D去噪残差剖面；（h） DCNN去噪残差剖面
Figure 3. Denoising results of post-stack synthetic data for F3
（a） Clean data；（b） Noisy data （PSNR＝20.4 dB）；（c） NLM denoising results （PSNR＝28.89 dB，SSIM＝0.867 2）；（d） BM3D denoising results （PSNR＝30.34 dB, SSIM＝0.943 6）；（e） DCNN denoising results （PSNR＝32.41 dB，SSIM＝0.966 6）； （f） NLM denoising residual profile；（g） BM3D denoising residual profile；（h） DCNN denoising residual profile
![]()
图 3 F3叠后数据的去噪结果
（a） 干净数据；（b） 加噪数据（PSNR＝20.4 dB）；（c） NLM去噪结果（PSNR＝28.97 dB，SSIM＝0.868 9）；（d） BM3D去噪结果（PSNR＝30.31 dB，SSIM＝0.942 8）；（e） DCNN去噪结果（PSNR＝32.43 dB，SSIM＝0.965 8）；（f） NLM去噪残差剖面；（g） BM3D去噪残差剖面；（h） DCNN去噪残差剖面
Figure 3. Denoising results of post-stack synthetic data for F3
（a） Clean data；（b） Noisy data （PSNR＝20.4 dB）；（c） NLM denoising results （PSNR＝28.89 dB，SSIM＝0.867 2）；（d） BM3D denoising results （PSNR＝30.34 dB, SSIM＝0.943 6）；（e） DCNN denoising results （PSNR＝32.41 dB，SSIM＝0.966 6）； （f） NLM denoising residual profile；（g） BM3D denoising residual profile；（h） DCNN denoising residual profile
表 1 不同去噪方法的评价指标对比（合成数据）
Table 1 Comparison of evaluation indicators of different denoising methods （synthetic data）
噪声水平/dB 输入信号/dB 去噪算法 峰值信噪比/dB 结构相似性指数 10 28.13 NLM 34.64 0.893 7 BM3D 34.78 0.942 9 DCNN 35.12 0.959 3 15 24.59 NLM 31.74 0.850 2 BM3D 32.85 0.932 1 DCNN 34.16 0.951 7 20 22.09 NLM 30.94 0.795 1 BM3D 31.61 0.901 5 DCNN 32.78 0.938 2 25 20.15 NLM 29.56 0.742 4 BM3D 30.43 0.894 1 DCNN 31.54 0.927 8 30 18.61 NLM 28.53 0.696 7 BM3D 29.24 0.874 9 DCNN 30.19 0.906 3 
下载: 导出CSV
-
韩卫雪,周亚同,池越. 2018. 基于深度学习卷积神经网络的地震数据随机噪声去除[J]. 石油物探,57(6):862–869,877. doi: 10.3969/j.issn.1000-1441.2018.06.008 Han W X,Zhou Y T,Chi Y. 2018. Deep learning convolutional neural networks for random noise attenuation in seismic data[J]. Petroleum Geophysical Exploration,57(6):862–869,877 (in Chinese).
刘婷婷,陈阳康. 2016. f-x域经验模式分解与多道奇异谱分析相结合去除随机噪声[J]. 石油物探,55(1):67–75. doi: 10.3969/j.issn.1000-1441.2016.01.009 Liu T T,Chen Y K. 2016. Random noise attenuation based on EMD and MSSA in f-x domain[J]. Petroleum Geophysical Exploration,55(1):67–75 (in Chinese).
王钰清,陆文凯,刘金林,张猛,苗永康. 2019. 基于数据增广和CNN的地震随机噪声压制[J]. 地球物理学报,62(1):427–439. Wang Y Q,Lu W K,Liu J L,Zhang M,Miao Y K. 2019. Random seismic noise attenuation based on data augmentation and CNN[J]. Chinese Journal of Geophysics,62(1):427–439 (in Chinese).
Buades A,Coll B,Morel J M. 2005. A review of image denoising algorithms with a new one[J]. Multiscale Model Sm,4(2):490–530. doi: 10.1137/040616024
Chen W,Chen Y K,Cheng Z X. 2017. Seismic time-frequency analysis using an improved empirical mode decomposition algorithm[J]. J Seism Explorat,26(4):367–380.
Chen W,Bai M,Song H. 2019a. Seismic noise attenuation based on waveform classification[J]. J Appl Geophys,167:118–127. doi: 10.1016/j.jappgeo.2019.05.014
Chen Y K,Chen W,Wang Y,Bai M. 2019b. Least-squares decomposition with time-space constraint for denoising microseismic data[J]. Geophys J Int,218(3):1702–1718. doi: 10.1093/gji/ggz145
Chen Y K,Zu S H,Chen W,Zhang M,Guan Z. 2019c. Learning the blending spikes using sparse dictionaries[J]. Geophys J Int,218(2):1379–1397. doi: 10.1093/gji/ggz200
Dabov K,Foi A,Katkovnik V. 2007. Image denoising by sparse 3D transform-domain collaborative ltering[J]. IEEE T Image Process,16(8):2082–2094.
Duchi J,Hazan E,Singer Y. 2011. Adaptive subgradient methods for online learning and stochastic optimization[J]. J Mach Learn Res,12(7):2121–2159.
He K M, Zhang X Y, Ren S Q, Sun J. 2016. Deep residual learning for image recognition[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition （CVPR）. Las Vegas: IEEE: 770–778.
Li P,Chen Z K,Yang L T,Gao J,Zhang Q C,Deen M J. 2017. Deep convolutional computation model for feature learning on big data in internet of things[J]. IEEE Trans Ind Inform,15(3):1341–1349.
Liu J L,Lu W K,Zhang Y Q. 2017. Adaptive multiple subtraction based on sparse coding[J]. IEEE Trans Geosci Remote Sens,55(3):1318–1324. doi: 10.1109/TGRS.2016.2622399
Ioffe S, Szegedy C. 2015. Batch normalization: Acceletating deep network training by reducing internal covariate shift[C]//International Conference on Machine Learning. Lille, France: PMLR: 448–456.
Stumpf A,Lachiche N,Malet J P,Kerle N,Puissant A. 2014. Active learning in the spatial domain for remote sensing image classification[J]. IEEE Trans Geosci Remote Sens,52(5):2492–2507. doi: 10.1109/TGRS.2013.2262052
Ullah A,Ahmad J,Muhammad K,Sajjad M,Baik S W. 2017. Action recognition in video sequences using deep Bi-directional LSTM with CNN features[J]. IEEE Access,6:1155–1166.
Vincent P,Larochelle H,Lajoie I,Bengio Y,Manzagol P A. 2010. Stacked denoising autoencoders:Learning useful representations in a deep network with a local denoising criterion[J]. J Mach Learn Res,11(12):3371–3408.
Wang F,Chen S C. 2019. Residual learning of deep convolutional neural network for seismic random noise attenuation[J]. IEEE Geosci Remote Sens Lett,16(8):1314–1318. doi: 10.1109/LGRS.2019.2895702
Yuan S Y,Liu J W,Wang S X,Wang T Y,Shi P D. 2018. Seismic waveform classification and first-break picking using convolution neural networks[J]. IEEE Geosci Remote Sens Lett,15(2):272–276. doi: 10.1109/LGRS.2017.2785834
Zhao Y X,Li Y,Dong X T,Yang B J. 2019. Low-frequency noise suppression method based on improved DnCNN in desert seismic data[J]. IEEE Geosci Remote Sens Lett,16(5):811–815. doi: 10.1109/LGRS.2018.2882058
-
期刊类型引用(0)
其他类型引用(4)
计量
- 文章访问数: 592
- HTML全文浏览量: 235
- PDF下载量: 89
- 被引次数: 4
