Enhancing the resolution of seismic post-stack data based on deep learning

Abstract Seismic interpretation is often restricted by the low resolution of field seismic data, which limits the characterization of subsurface geological structures and stratigraphic features. Improving the vertical resolution of seismic data has long been a challenge in seismic exploration. Recent studies have employed deep learning techniques to enhance seismic vertical resolution; however, limited attention has been paid to the role of loss function design in such enhancement. To address this gap, we propose a modified U-Net – based approach that employs a hybrid loss function to improve the vertical resolution of post-stack seismic data. Specifically, we first synthesized a substantial training dataset by convolving reflectivity models with wavelets of varying dominant frequencies. The prepared dataset was then used to train a modified U-Net architecture. The training process was guided by a loss function combining the Structural Similarity Index Measure and mean square error to optimize network parameters. We subsequently applied the trained network to a synthetic seismic image and two field datasets. The results confirm the efficacy of our proposed method in significantly improving the vertical resolution of seismic data while simultaneously attenuating random noise to a certain degree.