First Arrival Traveltime Picking Through 3-D U-Net

Abstract

In seismic exploration, picking first arrival traveltimes is an important step toward the estimation of subsurface velocity model, which has the direct impact on the well placement. In this letter, we present a 3-D deep learning method for automatic picking. In specific, we employ a supervised 3-D U-shaped full-convolutional network (3-D U-Net) to classify each sample in the 3-D seismic data into two categories: samples before and after the first break. Subsequently, we delineate a surface to separate these two categories, and such surface is corresponding to the desired first arrival traveltimes. In our training phase, besides the original normalized waveform data, we include the energy semblance feature as a channel of the input data, which helps to involve human understanding. The trained 3-D U-Net is applied to both synthetic and real datasets. The synthetic test shows that even if the labels contain some outliers, the predictions are still stable and reasonable in the sense that 3-D U-Net is capable of correcting some incorrect pickings. As for the field dataset test, the predicted results of 3-D U-Net are satisfactory based on human visual verification, and the picked traveltime surface is more consistent compared with the one predicted by the 2-D U-Net.