Integrating seismic first-break picking methods with a machine learning approach

Picking first breaks from seismic data can be very challenging, especially for land and shallow marine data. Over the past several decades, advanced picking algorithms that utilize seismic properties and the physics of wave propagation have been developed, but the challenge remains as data becomes larger, and data acquisition extends into more difficult territory. Developing a machine learning picking method to replace the previous seismic picking algorithms does not seem practical, since seismic data are not arbitrary signals, but following many laws of physics. Therefore, in this study, we still honor the conventional seismic picking methods by applying them first, then apply a machine learning method to identify and fix poor picks automatically. This involves constructing an architecture of convolutional neural networks (CNNs) in machine learning, helping identify poor picks across multiple traces. We train this network with 350,000 labeled seismic traces. The accuracy for classifying validation dataset is 97.9%. When testing with a new dataset, the accuracy for classification is above 95%.