Seismic sparse time-frequency representation via GAN-based unsupervised learning

Abstract

Time-frequency representation (TFR) is an important tool to describe the local time-frequency (TF) features of seismic data, which is beneficial for seismic interpretation. Among a variety of TFR methods, the sparse TFR (STFR) is an effective one, which can obtain the sparse TF spectrum by solving the optimization problem. However, STFR is often based on a mathematical model designed with the domain knowledge. Moreover, it suffers from the expensive calculation and the difficult selection of the regularization term in real application. To address these issues mentioned above, we propose a GAN-based unsupervised learning (GANUL) model, containing a generator, a discriminator, and a reconstruction module. The generator builds the map between seismic signal to its STFR. Both the discriminator and the reconstruction module are used to promote the generator to compute correct STFR. To test the effectiveness of the proposed model, we apply it to synthetic and field data after the model training and make detailed comparisons with the traditional STFR methods.