Study Of Seismic Data Residual Statics And Al Denoising

In the field of seismic exploration,complex near surface structures bring great challenges to seismic data processing and underground structure inversion.Generally,the deep refraction layer is covered by a shallow low-velocity weathering layer in the near surface,which could suffer from severe lateral velocity variations due to complex structures.The acquired seismic signals are often contaminated by strong random noises and some kinds of coherent noises,and thus suffer from low signal to noise ratio(SNR).Besides,seismic records usually include bad traces and missing traces due to recording problems.All these issues bring difficulties for seismic data processing.If the effect of near surface structure and noise is not eliminated,we cannot obtain seismic data with high SNR,high resolution,and high fidelity.Poor processed seismic data will influence the results of velocity analysis,stack and migration,and finally mislead the geological interpretation.Residual statics could help eliminate the influence of small near-surface structures on reflection traveltimes and improve the quality of stacking.Traditional residual statics correction methods usually are based on reflections,which are not reliable for field data with low SNR.In the weathering layer,the raypaths of the reflections and refractions are almost the same,so residual statics values have similar influence on their traveltimes.Thus we could utilize refractions to calculate residual statics values and improve the quality of stacking.We apply refraction interferometry method to build virtual refractions which have high signal to noise ratio.By picking the max energy of the virtual records,we obtain the stationary time differences between the adjacent receiver pair/shot pairs.Then the residual static problem could be solved by a set of linear equations based on ray paths of forward and backward propagation.Our method does not need to pick the first arrivals of seismic records,which improves the efficiency.We validate the effectiveness and robustness of our method by applying it on both synthetic and field data.Traditional denoising methods usually are based on noise level estimation and require heavy manual work to select the data and parameters.However,field data often includes various noise distribution.Moreover,the frequency band of noise is usually similar to signals.In this cases,traditional methods could not eliminate noise and preserve signals effectively.We apply the deep learning method to solve the seismic data denoising problem.The proposed method could remove noise adaptively.Comparing to traditional wavelet and curvelet denoising methods,the deep learning method could obtain cleaner records and preserve more subtle information of signals.We perform synthetic and field data tests to validate the performance of the proposed method.In recent years,several deep convolution neural network architectures have been presented,such as Dilate convolution net,ResNet,Encoder-Decoder net,and so on.These network architectures show their advantages in different aspects.We perform a synthetic test to compare the performance of these models for seismic data denoising.The dilated convolution network could enlarge the receptive field.However,subtle information will be partly abandoned through the forward propagation.The Incoder-Decoder net could compensate the details through the skip connections.ResNet could mitigate the gradient dispersion problem,but it does not perform well when the architecture is not very deep.The residual learning strategy in DnCnn could highly improve the performance for medium deep architectures.Base on the above architectures,we present new network architectures which inherit the advantages of the old networks while compensating their disadvantages mutually.Synthetic tests shows that the proposed architectures could improve the denoising performance and fidelity while sparing training time.Based on the deep learning denoising method,we further propose an iterative alternating strategy,which could perform denoising and interpolation simultaneously.We separate the objective function of data restoring problem into two sub-problems.The sub-problems could be solved by the least square method and applying the pre-trained denoising models iteratively and alternately.The method only requires the training of denoising models,which could spare plenty of time.Furthermore,the proposed method could handle data with different SNR adaptively.We validate the effectiveness of the proposed method utilizing synthetic and field data tests.