A Study On Deep-learning-based Denoising Methods For Complex Land Seismic Data

Seismic exploration is an essential way to develop oil and gas reservoir resources.Compared with some other geophysical methods,seismic exploration has many significant advantages,such as high precision,high resolution,high efficiency and large depth of detection,so it is widely used in the development of many large oil and gas fields.In recent years,the focus of oil and gas exploration has gradually turned to the development of complex reservoirs;the depth of detection is further deepened;exploration areas are shifting to unconventional areas.Therefore,lots of seismic data is of poor quality;specifically,effective signals are often disturbed by numerous regular and irregular noise and difficult to be recognized,which greatly affects the signal-to-noise noise(SNR)and resolution of seismic data and also brings huge difficulties to the following inversion,imaging,interpretation and the final exploration of oil and gas.How to suppress the seismic background noise,recover the effective signal and thus improve the quality of seismic data is a technical problem to be solved.The deep-learning-based methods emphasize on starting from massive training data to solve the problems that are difficult to be dealt with by traditional machine learning algorithms,such as high dimension,miscellaneous and high-level noise in massive data..Convolutional neural network(CNN)is one of the most representative technologies in deep learning.It can extract the inherent high-dimensional features of data through multiple convolution operations and has two significant advantages of local perception and weight sharing.In recent years,CNN has achieved good application effects in many fields such as image denoising,image super-resolution,pattern recognition,automatic driving,data fusion,data enhancement,etc.However,at present,the application of CNN in seismic data processing is still relatively few,and its huge application potential is yet to be explored.Aiming at two kinds of complex seismic data: desert surface seismic data acquired from Tarim area and distributed optical fiber sensing(DAS)seismic data,we construct targeted seismic denoising systems based on CNN.We make use of network training,minimization of error,multiple losses,forward modeling,low-rank method,determination of high-order statistic to obtain corresponding CNN-based denoising models for the above two kinds of complex seismic data,so as to construct the optimal nonlinear mapping relationship from noisy seismic data to denoised seismic data.Moreover,The effectiveness of deep learning methods in seismic data denoising is verified by both synthetic and real experiments.In the desert surface seismic exploration of desert area,the effective signals are often contaminated by both irregular random noise and regular surface waves,which leads to the low signal-to-noise ratio(SNR)of seismic data.In addition,effective signals,random noise and surface waves seriously overlap in low-frequency bands.Aiming at the denoising of the desert seismic data,the feed-forward denoising convolutional neural network(DnCNNs)which has achieved excellent performance in image denoising is adopted in this paper.We utilize the determination of high-order statistic to construct an adaptive training dataset.Meanwhile,the suitable network depth and patch size are selected by using the optimization of SNR and root mean square error(RMSE).Finally,the CNN denoising model can suppress the random noise and surface waves completely,and also recover the effective signals completely.In addition,in order to solve the over-fitting of DnCNNs in dealing with desert seismic data,the DnCNNs is combined with low-rank matrix decomposition.The optimal low-rank and sparse matrices of noisy desert seismic data are obtained by nuclear norm minimization.Then,the effective signals in the two matrices are obtained by using two effective signal prediction model based on CNN.In this way,the network depth is deepened by using the low-rank matrix decomposition,so as to further improve the denoising performance of DnCNNs.DAS is a new type of fiber optical geophone.Compared with the traditional electronic geophone,it has many advantages,such as high spatial resolution,low cost and anti-electronic interference,etc.However,in the real seismic data received by DAS,the energy of background noise is obviously stronger than that of effective signals,and thus the DAS seismic data is usually characterized by low SNR.In addition,the coherent noise caused by the inferior coupling between optical cable and receiving surface and the optical noise from DAS itself are new seismic background noise;the two are not presented on the conventional seismic data.Aiming at the denoising of DAS seismic data,we propose two targeted denoising neural network.First,we adopt the basic strategy of generative adversarial network(GAN)and use a denoiser and a discriminator to form a convolutional adversarial denoising network(CADN).In CADN,the mean square error loss of denoiser and the adversarial loss between the two are combined to propose a new loss function,where the optimal weight relation between the two losses are determined by multiple tests.Finally,the denoiser is optimized by the minimization of the new-proposed loss function;we can use the trained denoiser to predict the effective signals from the noisy DAS seismic data,so as to achieve the separation of noise and effective signals.Second,the convolution unit,batch normalization unit,leaky rectifier linear unit,and residual learning are used to construct a CNN;also,we utilize the forward modeling and passive source data to construct the effective signal set and noise set with high authenticity,respectively.Meanwhile,an energy ratio matrix is introduced into the loss function,so as to enhance the adaptability of CNN to the DAS seismic data with different SNRs.Finally,we can realize the complete recovery of effective signals under low SNR and complex background noise.Based on CNN,we construct a series of denoising networks for two kinds of complex seismic data.Both synthetic and real experiments demonstrate that these proposed denoising networks based on deep learning can effectively suppress a variety of complex seismic background noise and completely recover the effective signals under the condition of low SNR,so as to lay a solid foundation for later inversion,fine imaging,formation interpretation,etc.