Research On Seismic Noise Suppression Method In Tarim Based On Improved BM3D And DnCNN

The noise suppression of pre-stack seismic data is a critical step in seismic data processing and interpretation.Due to the complex surface and underground structures in Tarim,the noise in its seismic data has complicated components and dramatically different spatial-temporal characteristics,which does not satisfy the typical assumption of stationary white Gaussian noise in the current study.As a result,directly applying current methods in the target area has a relatively low performance,which brings difficulties for the follow-up work.To improve the signal-to-noise ratio of the seismic data in the target area,this paper defines the components of the noise in the target area and its characteristics in the space-time domain.Combined with traditional methods and deep learning technology,a noise suppression method for seismic data in Tarim is proposed based on an improved 3D block matching method and a feedforward denoising convolution neural network method.The main research contents are as follows:The existing research lacks the study on the noise characteristics of seismic data in Tarim,which makes the traditional methods incapable of suppressing all kinds of noise in Tarim seismic data accurately.To solve this problem,this paper first defines the components of the noise based on the characteristics of the surface and underground structure of the target area.Then it tests the Gaussianity,stability,power spectral density,similarity and linearity of the random noise using Shapiro Wilk method,improved generalized S transform,AR model estimation method,similarity coefficient,and delay vector variance method.Finally,it defines the space-time domain characteristics of the random noise in the target area.To solve the problem that the mainstream denoising methods do not consider the spatial variability of random noise in Kedong area,a random noise suppression method based on the block classification and 3D block matching method is proposed.Firstly,the seismic records in the target area are divided into several data blocks of the same size,and then the data blocks are classified so that the noise level difference of the data blocks in the same category is minimal,while the noise level difference between different categories is maximal,that is,the random noise in each category of data is considered as nearly stationary;then the 3D block matching method is applied to denoise each category of seismic data blocks,and the results are compared.Finally,the denoised data blocks are reorganized to suppress the non-stationary random noise in Kedong seismic data.Aiming at the problem that traditional algorithms cannot meet the needs of processing the tremendous amount of data from Tarim,a noise suppression method based on the traditional algorithm and an improved denoising convolutional neural network is proposed.The original seismic data of the target area and the noise data,which is removed by the domain transformation method and the improved three-dimensional block matching method,are used to construct a network training set.The network structure and hyper-parameters of the denoising convolutional neural network are optimized and fine-tuned according to the noise characteristics of the target area.In the initial stage of the algorithm,the transform domain method and 3D block matching method are used to suppress the noise and output the denoised result;the seismic data with a good denoising result is selected as the training samples to continuously train the neural network,and the trained network is used to suppress the noise;when the denoising accuracy of the neural network is higher than that of the traditional algorithm,it automatically switches to the neural network method for noise suppression.The results of theoretical analysis,numerical simulation,and real data processing show that the proposed noise suppression method is effective and feasible in Tarim.This paper enriches the theoretical methods of noise suppression and expands the technical means for high-precision processing of complex seismic data.