Consistent kernel size selection for seismic interpolation with an improved U-net

Abstract

Summary Seismic data in land acquisition cases always have large distances between adjacent shots, which causes the common receiver gathers (CRGs) with large trace intervals. The CRGs can result in spatial aliasing effects and decrease the accuracy of subsequent seismic imaging. Traditional interpolation methods have certain difficulties and limitations in anti-aliasing when processing seismic data with spatial aliasing, such as prior assumptions and human-computer interactions. Thus, we introduce a deep learning based method with adaptive training data generation and consistent kernel size selection. The spatial reciprocity of Green’s function is explored to construct the training dataset adaptively. Based on U-net, an improved U-net is introduced to accurately match the desired output, that is, completed data, with regularly sampled data as input. Considering the difference between seismic data and images, we propose a consistent convolution kernel size selection method to guarantee high accuracy of seismic data interpolation. Field data interpolation performance demonstrates the validity of the proposed consistent kernel size selection strategy and the improved U-net for intelligent seismic data interpolation.