Style transfer as Data Augmentation for Multiscale Data-driven Full Wavefrom Inversion
Abstract
Seismic full-waveform inversion (FWI), as a powerful imaging technique to reconstruct geological subsurface structures, is computationally expensive with the increasing size of seismic data nowadays. To overcome the computational issue, data-driven FWI is developed to learn a direct mapping from the seismic data to the velocity model from a large training set. However, the weak generalization ability of the data-driven FWI hinders its potential application on field data (Zhang et.al, 2020). Inspired by the artistic style transfer from the computer vision community (Johnson et.al, 2016), we develop a velocity model generation method that converts a large volume of existing natural images into subsurface structure models with pre-determined geologic styles. It increases the generalization of the neural network by generating a large number of synthetic subsurface velocity models with sufficient variability. With these synthetic models, we train a multiscale data-driven FWI network, called Multiscale InversionNet. It separately inverts transmission and reflection waves in the seismic data, which significantly improves the overall inversion accuracy. To validate the performance of Multiscale InversionNet and the effectiveness of the synthesized training set, these methods are tested on both synthetic and field data. The results show that once the network is fully trained using a properly designed training set, it can produce accurate subsurface velocity models more effectively and efficiently than conventional FWI.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFM.S35D0250F