Paper submitted to IMAGE 2026, by Mark Roberts, Leandro Gabioli, Olga Brusova, David Brookes, Xin Zhao, Alejandro Valenciano (TGS).
Abstract
This study presents a hybrid ML workflow for dual-sensor towed-streamer data that tackles three of the most computationally intensive steps in marine seismic preprocessing: deblending, denoising, and deghosting. Each ML step is designed to minimize manual parameter tuning: denoising models are trained on pseudo-synthetic examples generated from real noise and cleaned records; deblending uses an ML-created seed to accelerate sparse iterative inversion; and deghosting incorporates both hydrophone and geophone inputs through a modified DuckNet architecture. By utilizing the complementary ghost responses of co-located geophones, the deghosting model achieves a validation loss of 0.017, compared to 0.025 for a hydrophone-only model, demonstrating that the geophone supplies information that the network cannot extract from pressure data alone. The workflow is tested on a 3D survey along the Equatorial Margin of Brazil and processes individual sail lines within hours, greatly reducing the parameterization effort compared to traditional processing methods.
Read the full article here.