Seismic data interpolation using an anti-over-fitting mixed-scale dense convolutional neural network

Seismic data interpolation is a topic well suited for deep learning (DL) applications. Scaling operation-based DL neural networks, e.g., U-Net, have been popular since its booming development in the field of seismic data processing. Although many successful studies using U-Net on seismic data, scientists start to realize the downside of its implementation, i.e., large trainable parameters (normally larger than 1 million), the potential risks of over-fitting, and tedious hyper-parameter selection. Therefore, in this abstract, we introduce a mixed-scale dense convolutional neural network (MS-DCNN) for seismic data interpolation with relatively few trainable parameters to reduce the risk of over-fitting. This MS-DCNN was originally developed for biomedical image processing. In addition, this neural network can be trained with relatively small training set. Via a field data case study, the different behavior of U-Net and MS-DCNN is analyzed and compared for a specific interpolation problem, where 9 consecutive shot records were missing from a 2D line of marine seismic data.