Seismic deconvolution using Hopfield neural network

Abstract

Seismic data processing, one of the most important stages in seismic exploration projects is very time consuming and expensive. Neural Networks (NNs), due to parallel processing and hardware implementation characteristics, have potential uses that greatly speed up seismic data processing. In this paper, a commonly used NN, Hopfield NN, is employed to implement seismic deconvolution. In this approach, deconvolution is decomposed into three subprocesses: reflectivity location detection, reflectivity magnitude estimation, and wavelet extraction. Hopfield NN is developed for each of the subprocesses. The obtained networks are combined by Block Component Method (BCM) for simultaneous estimation of reflectivity series and seismic wavelet. For sensitivity examination of the neural reflectivity estimator to noise, we implemented this algorithm on traces with variable signal to noise ratios. This approach is applied to synthetic and real seismic data and results are compared with those of spiking deconvolution. Obtained results indicate that: (1) unlike spiking deconvolution, deconvolution of seismic data using Hopfield NN is not sensitive to noise and provides much better results than that of spiking deconvolution for a noisy trace; (2) there is no assumption of randomness for reflectivity series; (3) the neural reflectivity estimator is not sensitive to frequency bandwidth of seismic source wavelet. We found that, neural estimator replaces wavelet with spike into the traces; this method increases temporal resolution of seismic section greatly.