Irregularly sampled seismic data interpolation with self-supervised learning

Supervised convolutional neural networks (CNNs) are commonly used for seismic data interpolation, in which a recovery network is trained over corrupted (input)/complete (label) pairs. However, the trained model always suffers from poor generalization when the target test data are significantly different from the training data sets. To address this issue, we have developed a self-supervised deep learning method for interpolating irregularly missing traces, which uses only the corrupted seismic data for training. This approach is based on the receptive field characteristic of CNNs, and the training pairs are extracted from the corrupted seismic data through a random trace mask. After network training, all target data are recovered using the trained model. This self-supervised learning interpolation (SSLI) method can be easily integrated into commonly used CNNs. Synthetic and field examples demonstrate that SSLI not only significantly outperforms traditional multichannel singular spectrum analysis and unsupervised deep seismic prior methods but also competes with supervised learning methods.