Automatic NMO correction and velocity estimation by a feedforward neural network

Abstract

We describe a new method of automatic normal moveout (NMO) correction and velocity analysis that combines a feedforward neural network (FNN) with a simulated annealing technique known as very fast simulated annealing (VFSA). The task of the FNN is to map common midpoint (CMP) gathers at control locations along a 2-D seismic line into seismic velocities within predefined velocity search limits. The network is trained while the velocity analysis is performed at the selected control locations. The method minimizes a cost function defined in terms of the NMO-corrected data. Network weights are updated at each iteration of the optimization process using VFSA. Once the control CMP gathers have been properly NMO corrected, the derived weights are used to interpolate results at the intermediate CMP locations. In practical situations in which lateral velocity variations are expected, the method is applied in spatial data windows, each window being defined by a separate FNN. The method is illustrated with synthetic data and a real marine data set from the Carolina Trough area.