First break picking with machine learning

Bernard K. Law, Daniel O. Trad

First break (FB) picking is a laborious task for land data processing. In this report, we experiment with two Machine learning approaches 1) unsupervised automated editing of outlying picks by clustering. 2) supervised deep learning by training the networks with manually edited FB and classifying the first arrival energy waveforms as pre-FB and post-FB. The first approach is easier to apply but more limited. The second approach requires a catalogue of images and their first break picking for training. With enough training samples, the deep neural works will be able to classify the first arrival energy waveforms of new datasets as pre-FB and post-FB as accurately as the trained technicians.