A neural network approach to first break picking

First-break picking is an extremely time-consuming task for manual operation since seismic data are so voluminous (e.g. a typical 3-D survey consists of more than half a million traces to be picked). Previous attempts to automate first-arrival picking have achieved only limited success. The authors describe a neural network (NN) solution to this problem using a back-propagation network. The NN-based application system achieved above 95% accuracy on picking several seismic lines based on a single training using only a few seismic records. The level of performance exceeded that achieved by an existing automatic picking program. Job turnaround time (compared to manual picking) improved by 88%. The approach appears robust and shows promise for automating other event-picking tasks in seismic velocity analysis and seismic tomography