Research And Application On Seismic Noise Attenuation Based On Complex Diffusion Equation And Deep Learning

High quality seismic data is desirable for a variety of procedures of seismic exploration for probing subsurface structures and further imaging.During acquisition process,seismic signals are received along with seismic random noise,which severely degrades seismic data and makes it difficult to identify the seismic signals.Therefore,seismic noise reduction is a fundamental step in seismic data processing,which can effectively enhance the SNR of seismic data and improve the data quality.In addition,based on complex collection conditions,such as forest belts and desert areas,seismic random noise from different areas shows various characteristics and complexity,such as non-stationarity and non-Gaussianity.These kinds of complex noise that can be confused with seismic signals make it more difficult to identify seismic signals from seismic data,especially at low signal-to-noise ratio(SNR).Denoising seismic data contaminated by random noise with complex characteristics and preserving seismic reflections remains a challenging problem.This dissertation focuses on the researches of reducing non-stationary,non-Gaussian and low-frequency seismic random noise at low SNR,and proposes the denoising schemes based on the complex shock diffusion equation and deep convolution neural network.For enhancing seismic signals and filtering nonstationary random noise,we propose a structure-adaptive nonlinear complex diffusion method based on the complex shock diffusion by exploiting the structure-adaptive diffusion coefficients,as an enhancing-denoising process.The complex shock diffusion consists of a shock term for enhancing and a diffusion term for denoising,while the imaginary part is generated by complex diffusion process as a directional structure indicator.Moreover,the complex shock diffusion can adapt to non-stationarity seismic random noise which is based on the diffusion process.This dissertation analyzes the problems of complex shock diffusion in denoising the seismic data,and then proposed improved schemes for the existing problems.First,the complex shock diffusion provides the same diffusion intensity for random noise and seismic structures,which would lead to a significant signal loss for seismic data with complicated structural features.This dissertation utilizes the structure tensor to extract the structural texture of seismic data.The parameters of structure tensor are analyzed to make the structure tensor adapt the shallow and deep seismic data,in order to extract the structural features and orientations of seismic events,which are utilized to guide the diffusion coefficients.Based on the structural information obtained by the structure tensor,this dissertation chooses the appropriate function for constructing the threshold of the diffusion coefficient parallel to the gradient direction,and makes the threshold value being small at signal regions and big at noise regions.Then,the imaginary value as a smoothed directional second derivative combined with a spatially variant threshold is utilized to control the adaptive nonlinear diffusion coefficient parallel to the gradient direction,which penalizes the diffusion across the seismic features and thus preserve the seismic signals.Moreover,the orientations of seismic features is also used to guide the diffusion coefficient perpendicular to the gradient direction instead of constant,for reducing the signal loss of the seismic events,especially for those with steep and rapidly varying slopes.Hence,the diffusion coefficient perpendicular to the gradient direction holds lower values at the steeper seismic events for preserving seismic events.In order to remove non-stationary and strong seismic random noise,this dissertation proposes a patch-based denoising CNN model which is based on studying the deep convolution neural network and the local stationarity of seismic random noise.The deep convolutional neural networks have been shown excellent performances for image denoising.However,the denoising CNN model trained with a specific noise level cannot deal with the non-stationary seismic random noise which have spatiotemporally variant noise levels.This dissertation introduces the patch-based denoising into the deep CNN model,and combines the patch clustering and multiple CNN models to remove the non-stationary seismic random noise.The joint denoising strategy ensures the superior denoising ability and effectiveness of removing the spatiotemporally variant random noise with sudden variation of noise level.Furthermore,a model selection criterion guided by structural statistics is proposed based on the relevance of the CNN model to the noise level range.As a consequent,the matched CNN model can be automatically and efficiently chosen for each class,leading to outstanding denoising performance while preserving complicated morphology of seismic signals.Benefitting from the data intensity-based diffusion process,the CSD can adapt to the spatiotemporally variable levels.But,at low SNR,the waveform similarity between low-frequency seismic noise with seismic signals may seriously disturb diffusion process of the CSD.It is difficult for the CSD method to describe the characteristic difference between low frequency random noise and seismic signal.On the other hand,the deep learning network has the ability to extract the deep features of seismic data,which is a powerful tool to distinguish the low frequency seismic noise and seismic signals,but the performance of deep learning network is lack of adaptability to the current data to be processed due to the limitation of training data.Therefore,this dissertation combines the deep CNN model with complex shock diffusion,and proposes the deep complex reaction-diffusion model.Based on the CSD method,a reaction term is added to construct a complex reaction-diffusion model.Then,the deep CNN model is embedded to learn the additive reaction term from external seismic training data for more effective signal protection.During the evolution process of the deep complex reaction-diffusion model,signal compensation provided by learnable reaction term and denoising of diffusion process make the difference of low-frequency seismic noise and seismic signals become larger.The noise interference to the diffusion term and shock term becomes weaken,which helps both terms bring effective noise reduction and signal enhancement.Furthermore,the diffusion and shock terms also can rectify the inadaptability of CNN model in turn when the data to be processed deviates from the training data.The deep complex reaction-diffusion model can provide good performance in removing low frequency seismic random noise and preserving seismic events.Based on the complex shock diffusion and deep learning model,this dissertation proposes a series of optimized denoising schemes for dealing with the seismic random noise having complicated characteristics.The processing results of synthetic and field seismic data have proved these schemes have improvement in terms of increasing the SNR,reducing complicated seismic random noise and protecting seismic signals.These optimized denoising algorithms provided new ideas for the complex nonlinear diffusion methods and deep learning model in denoising seismic data.