Neuron curve as a tool for performance evaluation of MLP and RBF architecture in first break picking of seismic data

• Robustness of MLP and RBF neural networks is compared in first break picking.
• In this paper a method to find the hidden layer size is described.
• The introduced neuron-curve is used to find the minimum size of training set.
• The effect of noise on the architecture in first break picking is also investigated.

Recently wide applications of neural networks are reported in geophysical scientific papers, mostly lack the consideration of their mathematical evaluation and performance. In these general estimators/regression function/classifiers, parameters to be tuned are the number of layers, neurons, type of transfer function, minimum size of training set, etc. These will be carefully tuned per each physical problem. Among all, the number of hidden layers and the number of neurons in each hidden layer are the two important parameters to be decided and normally no rules are available for finding them precisely. In this paper a method to find the hidden layer size is described beside the main purpose of the paper which is to compare the performance of the first break picker networks. We used a known learning-curve and introduce a measure named “neuron-curve” to find the optimal layer size & minimum size of training set. This paper shows the application of these two curves in finding the first break picks of seismic refraction data. Furthermore, the effect of noise on the architecture of two known neural networks (multilayer perceptron and radial basis function) in the first break picking is also investigated.