A label optimization method for VSP wavefield separation with deep learning

Abstract

Vertical seismic profiling (VSP) has a wide range of applications in the field of earth sciences. It is utilized not only for seismic imaging in oil and gas exploration but also for the geophysical monitoring of CO2 reservoirs. Acquiring high-precision upgoing and downgoing waves from Vertical Seismic Profile (VSP) data is crucial since the majority of VSP applications use separated upgoing or downgoing waves, such as seismic imaging with upgoing waves or Q-attenuation estimation with downgoing waves. Traditional methods for wavefield separation typically depend on transform-domain techniques like f-k filtering or Radon transform, as well as time-domain methods such as median filtering and Singular Value Decomposition (SVD). However, transform-domain approaches face challenges like spatial aliasing; median filtering and SVD rely on manual selection of wave events, and are less effective for far-offset data. A critical aspect of using deep learning for wavefield separation is the preparation of the training dataset. Numerous studies utilize field labels derived from traditional methods or synthetic labels created using convolution operators. The significant difference between synthetic labels and actual field data, along with the inherent defects in labels generated by traditional methods, limits the effectiveness of the wavefield separation. To overcome these challenges in label creation, we introduce the label-optimized autoencoder (LOAE). The labels containing artifacts, which are produced by traditional methods, are trained through the LOAE network using unsupervised learning. After appropriate training, the LOAE can remove noise from the labels and output relatively pure upgoing or downgoing waves with consistent waveforms. The refined upgoing and downgoing waves are then merged to create a dataset for supervised learning in wavefield separation tasks. Both the synthetic and field data tests demonstrate that this label optimization method substantially improves the accuracy of wavefield separation.