EPLL With Mixed GMM For Nonstationary Seismic Random Noise Attenuation

Seismic denoising is the dispensable part of oil and gas exploration process.How to distill weak effective seismic signals from the strong noise and low signal-noise-ratio(SNR)background is the key question of the noise attenuation.Most traditional noise attenuation methods proceed from target seismic signals without using ground-truth statistic information of signals that is called prior knowledge in the statistic learning theory,which limits these methods’ performance and application in the real seismic denoising scenario.With the recent raising of machine learning,using machine learning to train prior statistical model while avoiding the shortcomings of the traditional denoising methods offers a new way to solve the denoising issues for the people,then improve the SNR of denoising results and higher definitions.Comparing to the traditional denoising methods,utilizing the machine learning methods to construct statistical prior model is a more precise way to extract statistical features operated as patch prior from target signals,which substitutes prior assumptions of traditional denoising methods.It also means these statistical models under machine learning is closer to the nature of target signals.A typical one among these statistical models is called the Gaussian mix model(GMM).Therefore,a method called expected patch log likelihood(EPLL)that operates GMM as the prior model has been given attention in the denoising field due to its remarkable performance.Due to the strength of EPLL in denoising,this article introduce EPLL algorithm to suppress the seismic random noise.It is necessary to consider the features of seismic signals while applying the EPLL.The random noise of the seismic signals is nonstationary,which causes the parameters inadaptation and signal distortion issues.Besides,the statistical prior model called GMM here is learning from the image patches and suppress the random noise efficiently.But it is incapable to restore the whole data someway.The lack of seismic samples,somehow,limits EPLL method to reach its highest performance.Therefore,this article get start with several aspects including seismic statistical model,adaptation to the nonstationary random noise,GMM model optimization etc,then propose EPLL seismic random noise attenuation method basing on weighting GMM model.In order to solve parameter inadaptation under nonstationary noise scenario,we firstly propose a patch classification model basing on patch noise level.The basic idea of this classification model is minimizing the within-class variance of all patches.We start with setting the initial threshold,then classify all patches into two patch groups.The classification process ceases if the patches of same group satisfy the stopping criterion.After this,we combine classification model with EPLL algorithm,propose EPLL method basing on patch classification,assign adaptive parameters to different patch group to solve the inadaptation issue,which aims to enhance the nonstationary reduction capacity of EPLL method.Furthermore,we found the single target GMM is incapable to attenuate the random noise in the background and leaves impulse-like residuals.Meanwhile,the generic GMM is incapable to describe the target signal but works in the noise attenuation in the background.Considering two models’ strength and shortcoming,we unite two models,operate them as prior model and further propose EPLL with mixed GMM on the basic of patch classification.The proportion of generic GMM is near to the low value while processing the patches of background.Then the proportion of target GMM is near to high while processing the signal patches.We combine EPLL method with two GMMs while avoiding shortcoming of each GMM.Finally,we verify the effectiveness of proposed method by applying it to the synthetic and real world data.The results of our experiments show that proposed method is effective to suppress the random noise while preserving the signals and details.Comparing to the traditional denoising methods,the proposed method achieves higher SNRs and resolutions,which sheds light on exploring underground structures and oil-gas distribution...