Autoencoder neural network application for coherent noise attenuation in high frequency shallow marine seismic data

• Main Authors: Hamidi, R., Latif, A.H.A., Lee, W.Y.
• Format: Conference or Workshop Item
• Published: Offshore Technology Conference 2020
• Online Access: https://www.scopus.com/inward/record.uri?eid=2-s2.0-85097623343&partnerID=40&md5=5fc01a4e1049bdb22d8967d27ed6dfc7; http://eprints.utp.edu.my/24650/

Conventional noise attenuation methods involve transforming noisy data into a filter domain where noise and signal can be separated. Deleting the noise components and transforming back the data into original domain, the filtered data is achieved. Coefficients representing the noise in the filter domain are selected by thresholding or manually which can result in a time-consuming process and also introduce error to what should be considered as noise energy. In this study, a model is developed using Deep Neural Network with AutoEncoder architecture to select the noise energy automatically in the Frequency-Wavenumber domain. The objective is to train a model that can attenuate coherent noise with certain isolated frequencies and varying amplitudes while preserving all reflections (weak and strong). The network is only trained on synthetic data; but its performance is evaluated on real high frequency marine data. The synthetic data have very simple structure of high frequency reflections contaminated with sinusoidal noise; outstanding performance of the proposed method on real data, however, shows the exceptional capability of the Deep Neural Network based filters. Copyright 2020, Offshore Technology Conference