Stabilized super resolution deep generative networks for seismic data

High-resolution seismic data enable us to characterize the reservoirs with higher accuracy and/or detect smaller targets. Enhancing the seismic bandwidth can be achieved with broadband acquisition, various processing algorithms or a combination of both. In contrast to classic spectral matching type processes, we propose to take a different approach by using deep Generative Adversarial Networks (GANs). In theory, they can reconstruct the seismic data both temporally and spatially. This is inherent by design given the convolutional architecture of the GANs. That means GANs allow recovering the frequency content or the missing traces in seismic data. We propose amplitude encoding and histogram equalization to stabilize the performance of GANs on seismic data and show promising preliminary results for typical seismic processing and interpretation applications.