Improving prestack migration of wide-azimuth data with a lightweight deep learning method based on transformers and convolutional neural networks

With migration methods, an appropriate aperture choice can reduce image noise by limiting the migrated data that contribute to each point of the output. However, subsurface geological features are often highly complex, with variations in layer properties and geological structures that require an adaptive aperture. Setting a time-varying, space-varying three-dimensional aperture for such areas presents a significant challenge, particularly for the migration of wide-azimuth seismic data. Recently, deep-learning methods have been extensively applied in scientific research due to their powerful feature extraction capabilities, nonlinear mapping, and ability to replace manual processes. Therefore, we propose a lightweight deep learning (LDL) method to extract effective seismic features from the full-aperture migration results of multiple azimuths and merge these features to output high-quality profiles, thereby avoiding the need for complex aperture settings. Specifically, convolutional neural networks (CNNs) are used as the first encoder to extract effective features and reduce migrated data size. Transformers are used as the next encoder to extract global and long-distance sequence features. These features are merged and decoded using another CNN block to generate high-quality imaging profiles. Tests on two field datasets show that the predictions of LDL have a similarity of 99.6% in normalized Rényi entropy proximity and reduce processing time at least five times comparing with stationary-phase migration results, and demonstrate the effectiveness and efficiency of the proposed method in extracting and merging migrated data features, eliminating the need for complex aperture settings.