A convolutional neural network approach for ghost removal

Abstract

In towed-streamer acquisition, ghost reflections severely reduce the quality of seismic data. Ghost reflections limit the usable bandwidth, which hampers geophysical applications that require broadband data such as full waveform inversion. Different processing and acquisition solutions have been proposed to mitigate the ghost reflections problem. Satisfactory solutions often require complex acquisition configurations that come with complicated processing workflows. Receiver side deghosting is commonly carried on common shot gathers, whereas source-side deghosting is carried on common receiver gathers. Since shot gathers are sparsely acquired in towed-streamer acquisitions, the source-side ghost remains an unsolved problem. We propose a solution capable of removing source-side and receiver-side ghosts in the shot domain for pressure data conventionally acquired at a single reference surface. Our solution uses a supervised learning approach, namely, convolutional neural networks, to iteratively learn to map a ghost-contaminated input to a ghost-free output. This method does not require any particular acquisition configuration and does not come with an involved processing workflow. However, it requires the acquisition geometry of field data and the bathymetry of the ocean floor to be known to create training data with the approximate configuration. We demonstrate with a numerical example the ability of our solution to remove ghost reflections and show that it is robust to random noise, even when noise-free data are used for training.