Abstract

First break picking is the essential part in the estimation of near surface velocity model building through the applications of delay time methods, refraction tomography and diving wave tomography methods etc. Being essential on one part it is quite laborious on other as it consumes a significant amount of processing time. In this study, we have demonstrated the use of neural networks in first break picking whereby we have picked first breaks picks manually for a small subset of data only and calculated certain seismic attributes in the vicinity of first breaks. Subsequently, the neural network then iteratively assigned internal weights based on the consistency of first break attributes values in correspondence to actual first break pick time. Finally, based upon these neural weights the first break pick time have been estimated for the whole seismic dataset. First break picking through neural networks have been found to achieve around 90 to 95 percent results against manual picking and thereby saved production time significantly for areas having almost linear first arrival trends. However, for areas having mixed first arrival trends (i.e. linear as well as non linear) using spatially varying neural networks have been found to serve the purpose fruitfully. Additionally, pre/post conditioning and optimum parameterization of seismic data/work flow have been found to serve the purpose to a larger extent.