An automatic and robust algorithm for the first-break picking is necessary to build the near-surface velocity model. We propose the algorithm based on a convolution neural network. The introduced first-break picking strategy and neural network architecture are suited for processing large volumes of seismic exploration data with reasonable computational resources. To develop an optimal neural network topology and architecture, extensive testing was performed. We compared several architectures of neural networks, including one- and two-dimensional approaches. Our tests justify that the one-dimensional approach (trace-by-trace processing) provides the most reliable results in the case of first-break travel-time variations typical of complicated near-surface structures. This study demonstrates that the four-layered neural network trained on 5,000 traces is enough for robust first-break picking. The algorithm is evaluated on two land-acquisition field datasets from West Siberia with a total used size of about 7 million traces. The first dataset is used for training, and the second one is used only for testing. For both datasets, the error between the original and the predicted first breaks is not more than three samples for 95% of traces. The final evaluation is done by a comparison of seismic stacks to prove the benefits of the approach and its robustness for offsets over 600 m. Finally, the influence of choosing the locations for the training dataset is discussed, and a strategy for using the proposed approach in production work is introduced.
