Wave-mode separation is an essential step of multicomponent seismic data processing. Due to limited acquisition at sparse receiver locations, conventional physics-based wavemode decomposition methods are only feasible based on many strong assumptions such as homogeneous subsurface velocity and flat structures. We propose a new P-wave reconstruction method based on VSP data, aiming to learn the domain transformation directly from full waveform elastic VSP data to their P-wave components. This method is based on optimal transportation theory and implemented by Generative Adversarial Network. We train network on 22180 pairs of synthetic data produced on 2D subsurface model and test on 40596 pairs. The network outputs achieve an average accuracy rate of 97.87%. The results tested on the 2D synthetic data show the network is capable of learning the waveform of target separated data beyond phase recognition and wave classification.
Presentation Date: Wednesday, September 18, 2019
Session Start Time: 9:20 AM
Presentation Time: 10:10 AM
Location: Poster Station 1
Presentation Type: Poster
