Unsupervised seismic data random noise suppression method based on weighted total variation regularization and ADMM solution

Noise suppression is a crucial step for seismic data processing. In recent years，with the rapid development of deep learning，its application in seismic data denoising has achieved significant effects. For practical application，since it is difficult to collect a large number of labeled seismic data（noise‐free data），this paper proposes to suppress the random noise in two‐dimensional（2D）seismic data based on the unsupervised deep image prior（DIP）framework. Firstly，the influence of skip connection on network denoising performance is explored，and the network architecture is determined. Secondly，the weighted total variation（WTV）regularization term is added to the loss function. Different from that of the traditional total variation（TV）regularization term，the weight coefficient of the WTV regularization term is no longer a fixed hyper parameter but a learnable parameter related to the spatial structure of data. Finally，the alternating direction method of multipliers（ADMM）is used to solve the optimization problem. Synthetic and real data experiments show that the DIP method combining WTV regularization term and ADMM can reduce the effective signal loss while suppressing random noise in seismic data and has better denoising stability than DIP；the peak signal‐to‐noise ratio fluctuation of adjacent iterations is small，and it is easier to develop early stopping criteria and applied. © 2023 Science Press. All rights reserved.