应用部分卷积网络的频率域地震数据插值OA

Deep learning-based seismic data interpolation methods demonstrate significant advantages in recon-structing complex missing data by leveraging neural networks'ability to optimize non-linear and multi-scale fea-tures.However,these methods still face challenges in training efficiency,generalization ability and adaptability to large datasets.To address this,a frequency-domain seismic data interpolation method based on partial convolu-tion networks is proposed.This method transforms the 3D time-space domain interpolation problem into multiple feature-correlated 2D slice interpolation tasks in frequency-space domain,reducing the data volume and storage requirements for each training session to enhance training efficiency.A complex partial convolution module is constructed to capture the constraint relationships between the real and imaginary parts of frequency slice data while performing convolution operations and boundary weight optimization only on valid data,enhancing the model's generalization ability and applicability for interpolation problems of data with varying missing patterns and missing ratios.By introducing a minimum weighted norm constraint in the frequency-wavenumber domain into the objective function,the risk of overfitting of the network is reduced.The method was applied to the inter-polation of synthetic pre-stack data from theoretical models and field pre-stack seismic data.Compared with the projection onto convex sets,the minimum weighted norm method and the 3D U-Net deep learning method,it achieved higher interpolation accuracy and efficiency,verifying its effectiveness and superiority.