Seismic Horizon Picking Using Deep Learning With Multiple Attributes

Seismic horizon interpretation is a fundamental task in subsurface exploration, essential for understanding formation geometry, characterizing reservoir properties, and supporting drilling operations. Traditional methods and deep learning (DL)-based approaches often face challenges in complex geological settings, leading to potential instability or failure in horizon picking. To overcome these challenges, we propose a novel method that integrates deep learning with multiple seismic attributes for robust horizon interpretation. Initially, a sparse horizon grid is constructed, guided by well log data and incorporating geological knowledge, such as structural and sedimentary system information. Using the structural similarity criterion, we select key seismic attributes that effectively distinguish horizon and non-horizon features, enabling neural networks to learn horizon patterns more accurately. The developed horizon picking network establishes a nonlinear mapping between the selected attributes and horizon classification results, improving interpretation accuracy. The proposed method is demonstrated on 3D seismic data from the Bohai Bay Basin, where it successfully provides high-density lateral interpretations of multiple horizons with varying levels of complexity. In addition to this, the method is applied to an ongoing drilling well. The prediction of the deep Archaean buried hill interface, 165 m ahead of the drill bit, showed a depth error of 2.64‰. This approach enhances both the accuracy and reliability of seismic horizon interpretation, making it a valuable tool for supporting safe and efficient drilling in complex geological formations.