Deep learning velocity filtering for seismic data

Abstract

Seismic velocity filtering is a critical technique in seismic exploration, designed to enhance the quality of effective signals by suppressing or eliminating interference waves. Traditional transform-domain methods, such as frequency-wavenumber (f-k) filtering and Radon transform filtering, often introduce aliasing and artifacts, while time-domain filtering algorithms may oversmooth effective signals. Although deep learning-based velocity filtering has yet to be directly applied in practical seismic processing, studies have demonstrated its potential in specific tasks, such as linear noise attenuation and VSP wavefield separation. However, the limited functionality of existing models poses challenges in addressing diverse velocity filtering tasks. To tackle this issue, We develop a deep learning velocity filtering algorithm based on upgoing and downgoing wave separation. By employing linear transformation, problems associated with arbitrary velocity differences are transformed in terms of upgoing and downgoing wave separation tasks. A diverse training dataset was constructed, comprising 31 theoretical velocity models, over 200 forward modeling simulations, and 46 field seismic datasets. The model trained on this dataset, combined with the proposed linear transformation strategy, achieves results comparable to traditional f-k and Radon filtering while also supporting tasks such as median filtering through manual wavefield picking. Applications in VSP upgoing and downgoing wavefield separation, distributed acoustic sensor (DAS) VSP P- and S-wave separation, post-stack migration arc noise mitigation, DAS common mode noise attenuation, and pre-stack CDP gather optimization demonstrate the method's versatility and practicality. This approach offers a novel and effective solution for almost all velocity filtering tasks in seismic data processing.