Objective The presence of surface-related multiples results in a lower accuracy in seismic data interpretations, and the effective multiple attenuation is a key step in seismic data processing. Multiples are coherent noise signals with similar characteristics to effective signals. It is difficult impossible to distinguish multiple signals from full-wave field data using a traditional convolutional neural network (CNN). Additionally, since the surface-related multiples vary significantly with surveyed areas, the CNN method will face more severe challenges when being transferred to other networks.
Methods This study developed a CNN method based on the normal moveout correction (NMO) domain by introducing physical priors. The CNN was trained using the differences in curvature characteristics between the primary waves and multiples in the NMO domain, aiming to achieve effective multiple identification and attenuation. The performance of this method was tested using simulations and practical data.
Results and Conclusions Experimental results indicate that the CNN trained in the NMO domain can effectively identify and attenuate multiples while protecting the reflected signals of primary waves. Compared to the traditional Radon algorithm, the proposed method exhibits reduced manual parameter adjustments and calculation complexity, along with less leakage of effective signals. Compared to direct end-to-end CNN-based methods for surface-related multiple attenuation, the novel method is more adaptable to new data. The results of this study can provide a new philosophy for improving the accuracy of seismic data interpretations and reducing the calculation cost.
DownLoad: