Superior Ultra-High Resolution Seismic Deghosting through Deep Learning with a Custom Autocorrelation Loss

This paper introduces an innovative deep learning (DL) approach for deghosting Ultra-High Resolution Seismic (UHRS) data, addressing the limitations of traditional physics-based methods, which are computationally expensive due to the shallow receiver depths and fine temporal sampling in UHRS surveys. DL methods have also been developed for fast-track UHRS deghosting but risk learning residual ghosts if the ground truth data is poorly parameterized. To overcome these challenges, the paper proposes a custom loss function that combines the Mean Absolute Error (MAE) with an autocorrelation penalty, designed to suppress ghost residuals during training without relying heavily on perfect ground truth data. This approach enables efficient deghosting, even when ground truth data cannot be fully optimized. Results from a 2024 UHRS survey demonstrate that the proposed method (ACL-DL) outperforms conventional DL models, achieving superior ghost removal and cleaner seismic outputs, while minimizing the need for precise parameterization. This method provides a robust, automated, and fast alternative to deghost UHRS data, offering improved performance in real-world data conditions.