Multicomponent seismic datasets, such as PS (downgoing P-wave and upgoing S-wave), offer significant advantages over conventional PP (downgoing and upgoing P-wave) data for subsurface characterization. By integrating PP with PS data, a more complete understanding of the subsurface can be achieved. However, effective use of these datasets depends on registration, a process used to align different wavefields. Successful registration brings multicomponent seismic datasets into a common time domain, typically that of PP. This task becomes particularly challenging when misalignments are large, the data contain noise, or the frequency and phase content vary. In this study, we present a supervised deep learning method to estimate time shifts between stacked PP and PS data. A detailed workflow is introduced for registering PP and PS data, which begins by stretching and squeezing PP and PS datasets using a variable Vp/Vs ratio to create training data with time-shift labels. We then develop a modified version of the Recurrent All-Pairs Field Transforms (RAFT) deep learning architecture, which formulates optical flow estimation to predict time shifts. The model is trained using an L1 loss between predicted and true shift fields. Finally, we apply the workflow to the Big Sky multicomponent dataset, demonstrating its effectiveness and feasibility for registering PP and PS seismic data. Highlights : • Developed an integrated workflow combining training data preparation and deep learning for multicomponent seismic data registration. • Applied a modified RAFT (Recurrent All-Pairs Field Transforms) model to align PP and PS datasets. • Demonstrated the capability to handle large time-shift variations between different seismic wavefields. • Validated the workflow on the Big Sky multicomponent seismic datasets, confirming its effectiveness for data integration.
Hoque et al. (Wed,) studied this question.