SUMMARY Ground roll attenuation of land seismic data is still an outstand-ing and challenging problem. Deep learning is a powerful tool for separating signal from noise. Recently, supervised deep-learning-based methods have been applied to ground roll attenuation. However, they require a large set of correspond-ing clean seismic datasets as labels. Constructing realistic training samples for network training is an unsolved problem. To circumvent it, we proposed an unsupervised deep learning method for attenuating ground roll where no training labels are utilized. The generator network first learns self-similar features before any learning. Therefore, if the reflections are self-similar in the time-space domain, but the ground roll is not, the network can extract the reflections before the ground roll. To make reflections look more self-similar than the ground roll, we apply the normal moveout (NMO) correction to flatten the reflections. Access to NMO correction makes the method also model-driven. The combinations of data-driven deep learning and a model-driven procedure are critical to the success of the proposed method. We use both synthetic and field shot data to illustrate the fidelity and validity of the proposed methods. The field data example shows that our proposed method can attenuate strong scattered ground roll.
Paper
Full text
Should we have labels for deep learning ground roll attenuation?
Semantic Scholar · Geology · 2020
Abstract
SUMMARY Ground roll attenuation of land seismic data is still an outstand-ing and challenging problem. Deep learning is a powerful tool for separating signal from noise. Recently, supervised deep-learning-based methods have been applied to ground roll attenuation. However, they require a large set of correspond-ing clean seismic datasets as labels. Constructing realistic training samples for network training is an unsolved problem. To circumvent it, we proposed an unsupervised deep learning method for attenuating ground roll where no training labels are utilized. The generator network first learns self-similar features before any learning. Therefore, if the reflections are self-similar in the time-space domain, but the ground roll is not, the network can extract the reflections before the ground roll. To make reflections look more self-similar than the ground roll, we apply the normal moveout (NMO) correction to flatten the reflections. Access to NMO correction makes the method also model-driven. The combinations of data-driven deep learning and a model-driven procedure are critical to the success of the proposed method. We use both synthetic and field shot data to illustrate the fidelity and validity of the proposed methods. The field data example shows that our proposed method can attenuate strong scattered ground roll.