Scattered Ground Roll Intelligent Attenuation Based on Deep Learning

Siwei Yu,Wuyang Yang,Haishan Li,Xiaojing Wang,Jianwei Ma

DOI: https://doi.org/10.1360/tb-2020-0564

2020-01-01

Abstract:The Tarim Desert consists of loess and sand with strong heterogeneity. The heterogeneity will generate scattered ground rolls when meeting the ground roll. Conventional ground roll attenuation methods use the differences of ground rolls and effective signals in energy, frequency, and velocity to design filters. Then, ground rolls are suppressed in the transformed domain. However, the characteristics of scattered ground rolls and effective signals are mixed and cannot be removed by conventional filtering methods. Focusing on the scattered ground roll in the Tarim Desert, a deep learning method is proposed to improve the signal-to-noise ratio of the prestack data. A denoised convolutional neural network (DnCNN) architecture is used for removing scattered ground rolls. Compared with traditional seismic noise attenuation methods, the DnCNN is based on a large-scale data training set rather than the assumption of a signal and noise model. The typical strategies of the DnCNN are residual learning and batch normalization. The residual learning strategy uses noisy observation images as the input of the network, removes clean images implicitly through the hidden layers of the network, and takes the residual as output. The batch normalization layer normalizes the input data of each layer, which makes the data received by the model follow the same distribution in the training process. This strategy accelerates the convergence of the model and improves the generality of the network. The elastic wave equation is used to simulate the scattered ground roll on the synthetic velocity model and the field velocity model. In the simulation of scattered ground rolls, the following parameters are randomly generated: the velocity model, depth of the source and detector, number of scatterers, and depth and size of the scatterers, to generate a variety of scattered ground roll datasets. When the number of neural network parameters is fixed, increasing the number and diversity of the samples helps avoid the overfitting of network training and improve the generality of the network. The synthetic training set is constructed to show the feasibility of the deep learning method on the attenuation of the scattered ground roll. A field dataset denoised by Fourier filtering is used to construct the realistic training set. Fine-tuning of the network parameters is performed on the network trained from the synthetic dataset for further training on the realistic dataset. The numerical results show that the Fourier filtering method depends on the angle selection, while the DnCNN method can automatically remove the scattered ground roll. By comparing the signal-to-noise ratio, the denoising performance of the DnCNN method is higher than that of the Fourier filtering method. The field data are stacked to produce an underground image. The comparison of the stacked results shows that the DnCNN produces the best visual quality of the underground layers. The stacked result of the original data with scattered ground rolls shows no clear information about seismic events. The stacked results of the Fourier filtering and DnCNN methods clearly show the events and layered structures. The numerical results show that the deep learning method achieves a higher signal-to-noise ratio than the traditional filtering method. Transfer learning produces a similar result with fewer training samples compared to that directly trained on a large training set. The numerical results have indicated the feasibility of the deep learning method in geophysical data processing.