Elimination of multiples with neural networks

Abstract

Seismic signal processing today often still is based on very simple linear models. Nevertheless, traditional methods such as predictive and spiking deconvolution with optimum Wiener filters are implemented very successfully in many cases. On the other hand, the theory rests on assumptions that are often not met in practice. Hence, there exist plenty of cases where traditional predictive deconvolution filters fail. For example, they can encounter great problems with strongly reverberating seismograms from certain complex offshore areas. One major assumption is, that the series of reflection coefficients, i.e. the reflectivity is a sample of a random white noise process. If the seismogram originates from a periodically stratified medium, where the reflectivity series is all but random, predictive deconvolution removes the desired primary arrivals along with the multiples.