3D9C seismic data reconstruction with multi-scale convolution neural network

Abstract

The nine components (9C) seismic data acquired with three-component (3C) sources and 3C receivers is beneficial to the inversion of lithologic reservoirs with high resolution. However, physical constraints and the trade-off between cost and data quality often result in under-sampled acquisition. For the 3D9C incomplete seismic data, conventional methods generally reconstruct each component data separately, which could not fully use the implicit relationship between the multi-component data and it is time-consuming. In this paper, we present a novel method based on a multi-scale U-Net (MSU-Net) to reconstruct 3D9C seismic data simultaneously. Compared with the U-Net, MSU-Net introduces multiple encoder and decoder sub-networks with different layers, which are connected through a series of nested and dense skip pathways. Through these re-designed skip pathways, the large scale feature maps with low-frequency information from shallow sub-networks and the small scale feature maps with high-frequency information from deep sub-networks are concatenated in each up-sample operator, which is more useful to reconstruct the seismic data with big gaps. Additionally, the three-channel input layer is designed in the MSU-Net according to the characteristics of the 3D9C data to implement 3C data reconstruction simultaneously. The MSU-Net is trained to learn the nonlinear relationship between the 3C data generated by the x-component source adaptively, and then simultaneously reconstructs the other 9C data in three times. Several examples of 3D9C seismic data reconstruction are used to evaluate the performance of the trained MSU-Net in comparison with U-Net and other traditional methods like the parabolic Radon transform. The reconstruction results demonstrate the effectiveness of the proposed method.