Random noise suppression in seismic data: What can deep learning do?

In the past few years, deep learning has gained great success in image signal and information processing. What are the challenges of seismic denoising compared to image denoising when using deep learning? First, the clean training seismic data are usually unavailable. Secondly, most of the networks are set for two-dimensional (2D) image denoising. However, seismic data denoising is mainly aiming at three-dimensional (3D) or higher dimensional data. Whether it works when expanding the network from 2D to 3D or higher dimension for seismic denoising? Finally, the most importance thing is to learn what aspects has been enhanced from deep learning compared to conventional methods for seismic denoising. This abstract takes random noise suppression in 3D poststack seismic data processing as an example to discuss the application of deep learning in seismic denoising. Except for white noise, there are imaging noise and scattered noise due to near surface non-uniformity in seismic data. This means 3D seismic random noise is usually non-Gaussian. According to the features of seismic noise, we propose 3D denoising convolutional neural networks (3D-DnCnn) incorporating sample screening. The seismic data acquired in a survey covering 700 square kilometers are used in our test. A small part of data with high signal-to-noise ratio (SNR) is selected for training. Other data are used for testing. The test results show that the network has similar random noise attenuation performance to conventional methods. Moreover, it can extract features of noise from the overall situation and thus has better suppressing performance of imaging noise. In addition, we adopt residual learning and batch normalization for accelerating training speed. And after network training is satisfactorily completed, its processing efficiency can be significantly faster than conventional denoising methods.