Building training data set for deep learning-based P- and S-wave separation: Field data case

Supervised learning-based seismic data processing has difficulty transferring synthetic data learning to real data applications. We report a distribution analysis method to build training datasets for the target field data, without modifying the network, and only a small volume of data is needed to train a well-performing network. We apply the distribution analysis on P/S wave separation of Vertical Seismic Profiling (VSP) data, where we train a fully convolutional network to separate P- and S-waves simultaneously from a two-component synthetic shot gather. Then we utilize the trained network directly to separate the field data, which are acquired from Liaohe River Basin, China. Clear separation in the data space and a high quality acoustic reverse time migration image demonstrated that the proposed training dataset building strategy is a key factor to the success of field data applications.