Machine Learning Using U-Net Convolutional Neural Networks for the Imaging of Sparse Seismic Data

Machine learning using convolutional neural networks (CNNs) is investigated for the imaging of sparsely sampled seismic reflection data. A limitation of traditional imaging methods is that they often require seismic data with sufficient spatial sampling. Using CNNs for imaging, even if the spatial sampling of the data is sparse, good imaging results can still be obtained. Therefore, CNNs applied to seismic imaging have the potential of producing improved imaging results when spatial sampling of the data is sparse. The imaged model can then be used to generate more densely sampled data and in this way be used to interpolate either regularly or irregularly sampled data. Although there are many approaches for the interpolation of seismic data, here seismic imaging is performed directly with sparse seismic data once the CNN model has been trained. The CNN model is found to be relatively robust to small variations from the training dataset. For greater deviations, a larger training dataset would likely be required. If the CNN is trained with a sufficient amount of data, it has the potential of imaging more complex seismic profiles.