ABSTRACT We propose an improved TransUNet‐based seismic data reconstruction method, designed to address the issue of seismic data loss resulting from factors such as instrument failures and environmental interference during acquisition, while overcoming the inadequacy of existing deep learning methods in capturing local details during reconstruction. First, the method applies amplitude consistency correction on the training data. By standardizing the distribution of data amplitude features, it effectively enhances the model's adaptability to seismic data under different scenarios and acquisition conditions, thereby improving generalization performance. On this basis, a dynamic mask prediction mechanism is introduced. Through accurately locating data‐missing regions, it guides the network to focus on the parts to be reconstructed, reduces excessive interference with the original valid signals and further prevents the destruction of effective information. Meanwhile, the method integrates the global feature extraction capability of the transformer and the local feature preservation advantage of the U‐Net. It can not only capture long‐distance structural correlations in seismic data but also finely retain local stratigraphic details. To further enhance the targeting of feature expression, the Convolutional Block Attention Module (CBAM) is incorporated into the method. Although it exhibits relatively low innovativeness in the current version, it can assist the network in more efficiently integrating local and global information and enhancing the collaborative feature expression capability—achieved through the channel attention module's weighting of key feature channels and the spatial attention module's focusing on important regions—thus holding certain necessity. The performance of the method was validated using SEG‐C3 synthetic data and real seismic data under both irregular and regular missing scenarios. Compared with commonly used methods, the proposed method can effectively recover detailed information and outperforms other methods in metrics such as signal‐to‐noise ratio (SNR), peak signal‐to‐noise ratio (PSNR), structural similarity index (SSIM) and root‐mean‐square error (RMSE). Further verification via FK (frequency wavenumber) spectrum analysis and single‐trace amplitude error analysis confirms the consistency between the reconstruction results of the proposed method and the original data. This method provides a high‐precision and strong generalization solution for missing seismic data reconstruction, with promising application prospects.
Yang et al. (Sun,) studied this question.