Branching Convolutional Neural Network For Desert Seismic Random Noise Suppression

As an important energy source in China,oil and natural gas play an indispensable role in all walks of life.Seismic exploration is an effective means to explore oil and gas resources.Due to the influence of seismic exploration environment and acquisition instruments,the collected seismic data often contain a large number of random noises,which hinder the effective seismic signals and reduce the quality of seismic data.Desert noise characteristics are different from that of random noise from other areas.Desert seismic random noise often are non-Gaussian and non-stationary,and its waveforms are similar to the effective signals.The spectrum of random noise is overlapped with signal spectrum.These complex characteristics bring a significant challenge to the noise suppression.Therefore,it is of great significance to study the method for suppressing desert seismic random noise and restoring the structural information of effective signals in order to improve the quality of seismic data in desert areas and thus to ensure the accuracy of oil and gas exploration.Convolutional Neural Network(CNN)has been widely used in seismic noise suppression due to its powerful ability in feature extraction.The denoising convolutional neural network uses large data sets to train network parameters and obtains the mapping from noisy data to clean signals.Once the network training is completed,the CNN can self-adaptively complete the denoising task without the prior knowledge of signal and noise,and without the need to set the optimal parameters.However,the denoising convolutional neural network based on discriminant learning lacks flexibility for desert seismic random noise with the complex characteristics,making the network difficult to control the balance between noise suppression and signal fidelity.In order to suppress desert seismic random noise effectively,we optimize the network structure and learning mode of convolutional neural network.We propose a convolutional neural network(BCDNet)with a branch structure to enhance the ability to extract effective signal features from desert seismic data,so as to better restore the seismic events contaminated by desert seismic random noise.The BCDNet contains a main network and branch networks that is added on the top of the main network.The main network is used to denoise desert seismic data.Before denoising,the branch network learns the global context characteristics of the effective signals from the noisy data,and guides the denoising task of the subsequent main network after connecting with the noisy data,thus the main network can better extract the characteristics of effective signals from the noisy data and effectively suppress complex desert seismic noise.In addition,we construct a dataset for training a complex desert seismic random denoising network.The denoising results of the synthetic desert seismic data and the field desert seismic data not only demonstrate the effectiveness of the proposed BCDNET in desert seismic random noise suppression,but also remain competitive in terms of network training time and GPU memory utilization.The denoising convolutional neural network based on discriminative learning can only deal with noise of specific noise level,and it is difficult to achieve good denoising effect for the desert seismic noise with spatially various noise level.Using the estimated noise level diagram to guide the denoising convolutional neural network can effectively improve the suppressing effect of non-stationary random noise.However,the denoised results of convolutional neural network are robust to the overestimated error of noise level,but more sensitive to the underestimated error of noise level.In order to deal with non-stationary desert seismic noise and reduce the asymmetric sensitivity of denoising results to noise level estimation error,we proposed a branching denoising convolutional neural network(AL-BCDNet)based on asymmetric learning to improve the denoising ability of non-stationary desert random noise.In AL-BCDNet,noise level estimation subnetwork is constructed based on denoising convolutional neural network.The noise level estimation subnetwork can learn the noise level map from the noisy data.Therefore,denoising master network take the map as weight to control the balance between noise suppression and signal fidelity.In addition,the sub-network parameters of noise level estimation are trained by asymmetric learning method to impose greater penalty on the underestimated error of the noise level,and solves the noise residual problem caused by the underestimated of the noise level.The filtered results of the synthetic desert seismic data and the field desert seismic data validates that the asymmetric learning can improve the accuracy of noise level map estimation,thus promoting the AL-BCDNet to suppress noise while retaining effective signals.