The interpretation of seismic travel-time imaging depends on the accuracy of picking up first arrivals, and it is low efficiency and high cost to pick up arrivals manually. Previous studies showed that deep learning can be applied for automatically picking, but the methods often need a large number of arrivals as the training set. In this study, we trained the U-Net with multichannel images and found that the root mean square error (RMSE) of first P arrivals decreases with the increase in the number of training sets. The errors of 35 and 597 shots are 11.4 and 6.5 ms, respectively. Referring to data augmentation methods in semisupervised
learning, we selected methods (Random Crop, Random Erase, etc.) suitable for seismic data and applied them to the training set. After random erase of the training sets of 35 shots, the RMSE is less than 5.5 ms (about 3 sampling points), which is 51% less than the error with the original training set. Compared with previous deep learning methods, the augmentation methods can be implemented to pick up the first arrivals with higher precision in the case of small samples.