Unsupervised 3D ground roll attenuation via continuous learning
Ji Li, Daniel O. Trad
Ground roll attenuation presents a significant challenge compared to incoherent noise inland seismic data. Traditional methods encounter limitations due to the substantial overlapbetween ground roll and reflections, resulting in a compromise between signal distortionand residual noise. We introduce an unsupervised deep learning framework for separatingreflections from ground roll. The network learns to flatten self-similar events in seismicdata before addressing those with steep dips and other incoherent noise. To enhance theself-similarity of reflections, we first apply normal moveout (NMO) correction to flatten thereflections and then use the network to extract these flattened reflections from the NMOcorrecteddata. Additionally, we incorporate a horizontal derivative regularization terminto the loss function to ensure convergence by penalizing horizontal variations. The networklearns a continuous function across sampling points in the seismic data, allowing themethod to attenuate ground roll in seismic data on both regular and irregular grids. Toevaluate the effectiveness of the proposed method, we apply it to a 3D field data set. Theresults demonstrate its superiority in noise attenuation and signal preservation.