Denoise 4D fast-track image using a conditional GAN: Brazilian pre-salt example

Time-lapse data processing can take many months, as reservoir changes have small magnitude and achieving the best possible repeatability is critical success factor. A fast-track processing is conducted sometimes to provide valuable intermediate inputs to interpretation and decision making, albeit those results may not be optimal compared to final results. In this study, we propose a novel method for denoising 4D fast-track images using a Conditional Generative Adversarial Network (cGAN) pix2pix. We frame 4D noise reduction as an image-to-image translation problem between baseline and monitor surveys, where the monitor image is transformed into the baseline image to minimize 4D differences caused by non-repeatability effects. The model is trained with the monitor fast-track image as input and the baseline fast-track as the target, leveraging the L1 distance as a conditioning factor to suppress 4D noise while preserving reservoir-related changes. We validate our approach using a Pre-Salt Brazilian 4D OBN dataset. The results demonstrate that our method provides significant uplift to fast-track 4D data quality enabling providing improved 4D seismic imaging in just a few hours of training.