Reconstruction of Missing Seismic Data Using Coarse Refine Network with Upsampling and F-K Loss

Abstract

Summary For successful seismic data processing, imaging and inversion, it is important to acquire evenly and densely sampled data. However, due to the malfunctioning of receivers and limited survey budget, acquired seismic data are often irregularly or insufficiently sampled along the spatial direction. To reconstruct missing or densely sampled traces, various seismic trace interpolation techniques have been proposed, and recently machine learning techniques have begun to be applied to enhance the performance of missing trace reconstruction. One of the most widely used machine learning techniques for seismic trace interpolation is Unet with mean-squared error (MSE). However, seismic trace interpolation with Unet has the limitations of aliasing due to the downsampling process of input data and over-smoothing due to the usage of MSE. In this study, we propose using coarse-refine Unet (CFunet) and frequency-wavenumber (F-K) loss for seismic trace interpolation to mitigate these problems. Numerical examples demonstrate that our proposed method reduces aliased features and precisely reconstructs missing traces, while accelerating the convergence of network.