Deep learning for prestack strong scattered noise suppression

Abstract

As a robust feature extraction method, deep learning has made significant progress in attenuating noise from seismic datasets. One critical assumption of deep learning for prediction is that test and training data should arise from the same distribution. Poststack data from a given survey can approximately meet this assumption because of subsurface structure similarities. Also, poststack data has a high signal-to-noise ratio (SNR) and a moderate amplitude variation, which makes the network training relatively easy. Therefore, denoising poststack seismic data with deep learning techniques appears to be a solved problem. However, noise is often prevalent in prestack data. The noticeable amplitude decay and significant waveform changes of reflections make network training unstable. More than that, the strong near-surface scattered noise on land data, which often overlaps with valuable signals in the t – x domain and f −k domain, poses a severe challenge to the conventional suppression strategy in common shot and common receiver gathers. Supervised deep learning methods, which require access to realistic learning samples, have failed to make a breakthrough on prestack seismic data denoising. Some unsupervised deep learning methods have made progress in areas with weak scattered noise, but further work is required to meet the industry’s requirements. To deal with the problem above and make deep learning more generalized and tractable for processing prestack denoising, we propose to train a denoising network on the offset vector tile (OVT) domain. OVT is a particular prestack seismic gather type that can faithfully represent a continuous wavefield; hence, it is an excellent domain to extract seismic data features. We use a 3D survey containing 1260 OVT volumes to illustrate the validity of the proposed methods. Only two OVT volumes with the same azimuth value are used to train the network, and the rest of the 1258 OVT volumes are adopted as test data. The results show that our method can effectively attenuate random and scattered noise. Moreover, our approach ameliorates the poor denoising performance on the boundary of OVT gathers compared to the conventional method used to construct training samples. It is worth mentioning that our calculation time for an OVT volume (200 × 200 × 3001) is only about 6 minutes, which is about one-tenth of the comparison conventional method. This work focuses on the following aspects: We study the suppression of intense prestack scattered noise based on deep learning, we also examine which is the best domain to perform deep learning denoising.