Deep Learning For Seismic Data Denoising,Interpolation And Residual Statics Correction

In seismic exploration,especially in land and shallow marine areas,the complex near-surface structure leads to strong surface interference and low signal-to-noise ratio and resolution of seismic data.Therefore,obtaining high-quality shallow seismic profiles is a challenge for seismic data processing.The acquired seismic records are often interfered by various noises and usually have bad traces or missing traces.In near surface(tens of meters or several hundred meters),the weathering layer above the reflection layer could suffer from severe lateral velocity variations due to complex structures.All these issues bring difficulties for subsequent seismic data processing.With the increasing amount of seismic exploration data,the efficiency of seismic data processing is urgent to be solved.Facing the dual requirements and challenges of accuracy and efficiency in seismic data processing,we propose a new strategy to solve the problems of seismic data denoising,interpolation,and residual static correction by combining deep learning algorithms.In recent years,seismic denoising methods based on deep learning have been widely applied in seismic exploration.However,it is often difficult to obtain noise-free labels which are required for the training of supervised methods.We propose a new deep learning framework to denoise pre-stack seismic data without clean labels.This framework is based on a high-resolution residual neural network.The input is noisy data and the label is noisy data with different noise but containing the same valid signal.Since valid signals in noisy sample pairs are spatially correlated,while random noise is spatially independent and unpredictable,the model can learn the features of effective information while suppressing random noise.Synthetic and real data test results show that our method performs better than conventional denoising methods and is also more efficient and stable than other unsupervised methods.We have also extended the deep learning method to earthquake data denoising.We proposed a method that processes the livestream earthquake data in real-time using deep neural networks from a large seismic network.The neural network can attenuate various types of noise and non-earthquake signals and suppress noise in the frequency band that overlap with signals.We first created "clean" samples by scaling down the waveforms from ML 3.5-5.0 to ML 1.5-3.0 based on the Richter scaling relationship.We also select noise samples from the same seismic station and add to "noise-free" data to generate samples at different signal-to-noise ratio(SNR)levels.These data samples are randomly split into training,validation,and test sets.We verify the trained network to process data recorded in Sichuan and Yunnan,China from 2013 to 2018.Results show that the method can help improve SNRs from 5 dB to 15 dB in average.The number of detected small events at magnitude between ML 1.0 and 3.0 has been increased by 58.8 percent.It takes about 10 ms on average to process three-component 60-s data from 300 seismic stations on a single GPU.Due to the influence of topography or consideration of exploration costs,seismic traces are often missing in practice seismic data acquisition.Incomplete,irregular,and data with spatial aliasing can not meet the high-resolution requirements of seismic exploration.Considering the applicability of unsupervised methods in practice,and in order to improve the reconstruction results when the gap between missing data is too large,we proposed a deep learning method combined with deep image priors and a regularization constraint based on the convex set projection method,which combined with the feature extraction characteristics of neural network and data sparsity.In order to improve the efficiency of data reconsttruction,the threshold of convex set projection is selected with the decrement of(?).With the introduction of regularization constraints,the reconstructed results are improved in both synthetic and real data tests.Reflection residual statics is generally related to the analysis of stacking velocity.The residual statics of shots and receivers are often estimated by the stack power maximization method.In 3D seismic exploration,the 3D velocity analysis is very timeconsuming,and calculating the correlation between a large number of traces will cause a serious computational burden.we have developed a method for predicting the 3D surface-consistent residual statics using deep neural networks.Based on the continuity of reflections,residual statics are directly estimated from the 3D common midpoint(CMP)gathers.The CMP gathers of 3D reflections exhibit continuity characteristics similar to the common shot gathers,common receiver gathers,and common midpoint gathers of 2D seismic data.When predicting,we first interpolate the missing traces of the 3D CMP gathers to ensure the continuity of reflections and obtain more stable prediction results.We use high-resolution neural networks(HRnet)to learn the subtle residual statics features.In the real data test,the predicted 3D residual statics obtained by the trained model help to improve the quality of stacking profiles.