N2N-WTV: Unsupervised Seismic Data Denoising Based on Weighted Total Variation Regularization

Seismic data has broad application prospects in many geophysical applications such as event detection, inversion, seismic diffraction separation. However, due to environmental factors such as waves and wind, or mechanical failures in receiver arrays, seismic data often experiences noise degradation, seriously affecting subsequent applications. Therefore, seismic data reconstruction research is an important prepossessing step that is beneficial for many geophysical applications. To effectively suppress noise, the visual quality of seismic data images can be improved, providing valuable insights to support reservoir characterization and optimal development decisions, thereby maximizing project value and reducing risks. We propose a simple and effective self supervised network based on weighted total variation regularization for seismic data restoration, which addresses the singularity of previous methods and combines the unique edge aggregation of seismic data. The model is a deep neural network composed of randomly adjacent sub samplers and weighted total variation. We have also developed an efficient ADMM optimizer to address the new optimization problem. The proposed network achieves good performance on seismic data images by randomly generating paired image training teams using adjacent sub samplers, training the network using the generated sub samples, and using weighted total variation as additional constraints to achieve better performance.