Automated Picking of Seismic First-Arrivals with Neural Networks

Abstract

Neural network implementations of first-arrival picking use a representative set of data and associated first-arrival pick times to create a multilayer perceptron neural network. The success of this technique depends on the statistical properties of the features input to the neural network, and the ability of the neural network to approximate a wide class of functions. Using methods from statistical pattern recognition, it is possible to determine the properties of the features input to the neural network that impact upon the reliability of the picking process. Interpreting the output of the neural network as the probability that the associated feature is the first-arrival, requires an appropriate probability model. This model is based on the multinomial distribution, since it can reflect the fact that there is only one first-arrival event on each data trace. Optimization of the neural network weights with an error function based on this probability distribution, produces a neural network that properly estimates the probability that the associated feature is a first-arrival event.