Seismic Data Interpolation Using Task‐Driven Pixel‐Level Masked Autoencoders

Supervised learning models, such as those based on convolutional neural networks, are widely used in seismic data interpolation and have achieved excellent results. However, the need for a large volume of complete data and corresponding sampled data pairs for model training presents a challenge in practical applications due to the limitations imposed by the sampling environment. We propose a self‐supervised seismic data interpolation method to address this issue, termed task‐driven pixel‐level masked autoencoders (TPMAE). Based on the traditional masked autoencoders (MAE) built for image classification and detection, we propose using the pixel‐level patching strategy to sample missing data to create the training data pairs. Moreover, the semantic reasoning module is suggested to complement the semantic information of the original missing seismic data. In addition, the local and global hybrid loss is utilized to enhance the accuracy of seismic data interpolation. Finally, qualitative and quantitative results on synthetic and field data show that our suggested TPMAE can reconstruct the irregularly and consecutively missing seismic data more accurately than the unsupervised models (i.e., MDAGAN and MAE), and its performance is comparable to that of the supervised model (i.e., U‐net).