Seismic Data Interpolation Using Dual-Domain Conditional Generative Adversarial Networks

Abstract

Seismic data interpolation is an effective way to reconstruct missing seismic traces and to improve the quality of the seismic data set. In the field of deep learning, generative adversarial networks are capable of data generation and interpolation and have been widely used for high-quality image generations and image interpolations. In this letter, we propose a dual-domain conditional generative adversarial network (DD-CGAN) for seismic data interpolation. The DD-CGAN consists of a generator network and a discriminator network and uses the seismic data set and discrete Fourier transformed data set in the frequency domain as input vectors. The loss function of the DD-CGAN is defined by the generative-adversarial loss, the data loss function, and the total variation loss. Thus, the DD-CGAN can be trained more accurately. The discriminator is used to calculate the feature differences between the interpolated seismic data set and the complete seismic data set to drive the generator network for learning optimal parameters. The numerical results on the test data set and field seismic data set demonstrate the effectiveness of the proposed DD-CGAN.