基于深度学习卷积神经网络的地震数据随机噪声去除

为了有效去除地震数据随机噪声,提出了一种基于卷积神经网络(CNN)的地震数据随机噪声去除算法。算法的关键在于构建一个适用于地震数据去噪的CNN,包含输入层、卷积层、激活层、输出层等。该CNN以含噪地震数据作为输入层,由多个卷积层提取并处理地震数据,激活层采用修正线性单元(ReLU)获取地震数据波动特征,再借助归一化层加速网络训练收敛速度。CNN通过残差学习获得随机噪声并由网络输出层输出。分别采用小波变换、双树复小波变换、曲波变换以及CNN对实际叠前海上地震数据、叠后陆地数据及复杂陆地叠后数据进行去噪,实验结果表明,CNN能有效去除随机噪声,且与常规去噪算法相比具有更强的去噪能力,验证了算法的可行性和有效性。

In order to effectively remove the random noise in seismic data,a denoising algorithm based on a convolutional neural network (CNN) within the deep learning framework is proposed.The key requirement of the algorithm is to construct a CNN that is suitable for seismic data denoising,which includes the input layer,convolution layers,activation layer,and output layer.The CNN uses noisy seismic data as inputs,extracts and processes the seismic data via the multiple convolution layers,extracts the fluctuation characteristics of the data using the rectified linear units in the activation layer,accelerates the training convergence using the normalization layer,and finally uses residual learning to obtain the random noise as the output via the network output layer.Tests using pre-stack marine seismic data,post-stack seismic land data,and complex post-stack seismic land data illustrated the feasibility and effectiveness of the CNN for seismic denoising.Furthermore,the CNN outperformed some traditional denoising algorithms,such as the wavelet,dual-tree complex wavelet,and curvelet transforms in random noise attenuation.

国家自然科学基金(51475136)、河北省引进留学人员资助项目(CL201707)和河北省研究生创新资助项目(CXZZSS2018012)共同资助。