The first-break detection for real seismic data with use of convolutional neural network

In this study, we appraise a convolutional neural network for the detection of the first breaks on the real 3D seismic data set. The use of convolution as a learning kernel is followed by an assumption that the seismic trace can be considered as a convolution of source signal with the reflectivity function. The investigation area includes mixed elevations, floodplains of the rivers and the regions of strong permafrost, where the shingling effect is observed. We consider the first-break detection for each trace independently to preserve the complicated structure of the arrival times. The proposed approach was apprised on real exploration 3D seismic data set with size over 4.5 million traces. This test showed that the error between the original and predicted first breaks is not more than 3 samples for 95 percents of data set. The final quality control of picking results was established by the calculation of static corrections and computing seismic stacks, which showed that the proposed approach provides better results.