Machine Learning applied in Swell Noise classification

Abstract

Recorded seismic signals are inevitably contaminated by noise in field acquisition. Attenuating the high-amplitude noises, such as swell noise, is really a big challenge in the seismic data processing. High-amplitude swell noise is a common sort of noise in marine seismic survey. It usually affects a number of neighboring traces, and can be observed in seismic data as vertical stripes. One of the first tasks in seismic data processing is the swell noise attenuation. However this is not a trivial task, if not applied in adequate way can damage the signal and consequently affect posterior processing steps and therefore, compromise seismic interpretation. In this work, several approaches are proposed using machine learning to identify contaminated traces by this noise. Future research direction includes the attenuation of swell noise using these classified traces. The classification of the traces individually is given by the best AI model tested: it was tested seven different AI models with four different input variables. The architectures that had frequency-based features presented better results overall, especially the multi-layer perceptron.