HCVT-Net: A hybrid CNN-Transformer network for self-supervised 3D seismic data interpolation

Abstract

Seismic data acquisition is an essential step for seismic exploration, constituting a substantial portion of the seismic exploration budget. To reduce the data acquisition overhead, it is an effective approach to acquire sparse seismic signals and interpolate the complete seismic data using designed interpolation methods. As a trending interpolation method, convolutional neural networks (CNN)-based methods have attracted much attention and shown some success in seismic interpolation. However, due to the local perception of CNN, these methods mainly focus on extracting local features, neglecting the global features of seismic data and limiting the performance. Additionally, most of these CNN-based methods work in a supervised manner, requiring high-quality paired training data and lacking generalization capability across different seismic data, which is challenging for 3D seismic data interpolation. Aiming at these problems, we propose a hybrid CNN-Transformer network (HCVT-Net) for 3D seismic data interpolation in this paper. Specifically, we design a CNN-based Encoder–Decoder structure to enable the network to learn local features at different resolutions. Meanwhile, we propose an improved Vision Transformer and deploy it to the CNN-based structure to enhance the extraction ability of global features. Finally, we adopt a self-supervised training strategy to alleviate the dependence on the high-quality paired data. Experimental results demonstrate that our method achieves better interpolation performance than competitive methods.