The presence of random noise has a direct effect on the seismic signal-to-noise ratio (SNR). SNR is essential for several seismic exploration techniques, including AVO analysis, seismic attribute analysis, and micro-seismic monitoring.
The paper "Wenda Li et al., 2021, Residual Learning of Cycle-GAN for Seismic Data Denoising" proposes using Residual Cycle-GAN (RCGAN) for noise removal. In essence, they combine synthetic and real datasets and employ a variety of augmentation techniques during training. Since real data always has noise, they proposed an f-x adaptive prediction filter (APF) to produce clean seismic. They argue that using the APF technique for cleaning and then training the network with it would result in better denoising performance because of multiple augmentation methods and additional synthetic datasets.
FXDM - analytical denoising, DnCNN - deep learning-based denoising, RCGAN - Residual Cycle-GAN denoising
Link for the LinkedIn post