Application of Extreme Learning Machines and Echo State Networks to Seismic Multiple Removal

• Main Authors: Carvalho, Heitor S., Shams, Farzin, Ferrari, Rafael, Boccato, Levy
• Format: Conference Proceeding
• Language: English
• Subjects: Deconvolution; Distortion; echo state networks; extreme learning machines; Multiple removal; predictive deconvolution; Sea surface; Sensors; Surface impedance; Task analysis; Transforms
• ISSN: 2161-4407
• DOI: 10.1109/IJCNN.2018.8489620

In seismic exploration, the collected signals not only contain the primary reflections of the subsurface structures, but also multiple reflections, which may hamper a proper and reliable analysis of the seismic image. In order to attenuate the effect of the multiple events, a predictive deconvolution approach can be used, which is particularly adequate for shallow water, and involves the design of a filter to predict the multiple events. In this work, we investigate the possibility of employing nonlinear structures in lieu of the usual linear filter. The chosen structures in this study refer to extreme learning machines and echo state networks, which offer an interesting trade-off between compu- tational cost and processing capability. The obtained results in the context of synthetic data clearly reveal the potential of this approach and motivate the continuity of this investigation.