A Benchmark Study of Deep Learning Methods for Seismic Interpolation

Three distinctive deep learning algorithms have shown successful applications in the seismic interpolation task. The first, deep prior interpolation (DPI), trains a convolutional neural network (CNN) to map random noise to a complete seismic image using only the decimated image itself. The second, referred to as standard method, trains a CNN to map a decimated seismic image into a complete one using a dataset of both complete and artificially decimated images. The third is a generative adversarial network (GAN) that trains two CNNs; one generator and one discriminator, again by using a training dataset of complete and decimated images. Within this research, we compare the performance of these methods for different quantities of regular and irregular missing traces using 4 datasets. For the completeness of our benchmark study, we compare the methods with simple linear interpolation as a lower quality bound. We evaluate the results using 5 well-known metrics. Our research reports that overall the standard method performs better than the other approaches. The DPI method is competitive for a low level of regular decimation, and ranked second in the irregular cases. The GAN approach is the less effective of the three deep learning methods.