Noise Suppression For Desert Seismic Data Based On Feature Enhancement Denoising Network

Seismic exploration is an important mean of judging oil and gas reserves.At present,as there are fewer areas on the surface or other areas where oil and gas resources are easy to be exploited,people have shifted the focus of exploration targets to areas with complex geological structures and difficult mining,such as desert areas.Due to the geological conditions of the desert area and the particularity of the exploration environment,the seismic exploration data obtained is accompanied by strong desert random noise.In addition,due to the absorption and attenuation of seismic signal under the complex geographical conditions of the desert area,the seismic reflection signal energy is weak,and the signal-to-noise ratio(SNR)of desert seismic data is low.With the advancement of science and exploration methods,the requirements of seismic exploration missions for geological interpretation accuracy continue to increase,and seismic prospectors have put forward higher requirements for the quality of seismic data.Therefore,the current primary task is to develop a denoiser that can effectively recover seismic signal from the desert random noise background and improve the SNR of desert seismic data.Random noise in desert areas has complex characteristics such as low frequency,non-Gaussian,strong energy,high amplitude,and nonlinearity.Although traditional seismic data processing algorithms can improve the SNR of the data to a certain extent when applied to desert seismic data,the denoising effect still cannot meet the requirements of high SNR and high resolution.Since the convolutional neural network has the advantage of adaptively extracting features and codes on different training sets,this article is dedicated to using it to suppress the noise of seismic data in desert environment.However,as the depth increases,the network may face the problem of reducing the impact of the shallow layer on the deep layer,which is not conducive to the suppression of complex desert noise.In this paper,considering the inference speed of the network model and the feature extraction performance,a new method of desert seismic data noise suppression based on the Feature Enhancement Denoising Network(FEDnet)is proposed.First,this paper designs a new feature-enhanced connection method to increase the width of the network and fuse the feature information of different convolutional layers.This design makes full use of the influence of the original noisy input and the multi-layer feature information on the network,so it is helpful for the network to capture more desert noise features hidden in the complex background.Secondly,this paper incorporates the hybrid dilated convolution design into the network model to improve the receptive field,which plays an important role in obtaining more contextual information in the denoising task.Finally,this article also uses the residual learning technique to promote network training.In addition to designing the network structure,this article also builds a set of training sets suitable for the task of desert seismic data denoising.Because the type and characteristics of the noise will directly affect the training accuracy of the network,this article uses the desert random noise data collected in Tarim area to construct the noise set.For the pure signal set,this article applies Ricker wavelet and set appropriate parameters,including the dominant frequency and bending degree of wavelet,so as to approach and express the characteristics of the signal in the actual data as much as possible.The final experimental results prove the effectiveness of the training sets.This article chooses Shearlet,wavelet transform,band-pass filter three traditional classical denoising algorithms,as well as the classical denoising convolutional neural network(Dn CNN)as the contrast experiment.In the experiment of synthetic record denoising,the network of this article has good advantages in the SNR improvement and signal amplitude retention,especially in the case of low SNR of input record,the feature enhancement denoising network can improve the SNR by about 20 d B.In the actual recording experiment,the four contrast algorithms show different denoising disadvantages,and our algorithm can still show the best effect.