A back propagation neural network for identifying first-break times on cross borehole ground penetrating radar traces

Manually picking the first arrival of energy in a series of cross borehole GPR ray traces can be time consuming and subjective, especially when large data sets need to be processed. One possible remedy is the application of a back propagating neural network. Neural network applications have been used previously in seismic studies to pick the arrival of the P and S waves (Dai and MacBeth, 1997; McCormack et al. 1993; Murat et al. 1992). One particular method, which applied a moving window over the trace, is used here with slight modification. Noisy time-amplitude records were first normalized to range from -1 and 1. These data were then filtered such that values between -1 and a negative threshold were set to -1, values between 1 and a positive threshold were set to 1 and all other values were set to zero. The filtered wave was fed through a neural network that searched for a pattern related to a first arrival. Several filtering parameters were tested, including the size of the moving window, the values of the positive and negative thresholds, and neural network parameters pertaining to training and testing. With minimal training, the neural network performed very well compared to hand picking of arrival times on large data sets.