基于联合深度学习的地震数据随机噪声压制

噪声压制是地震资料处理中的一项关键任务.根据不同噪声的形成机制、特性,可以采用不同的压制方法,使地震资料的信噪比达到预期,提高后续地震资料处理和解释的效率和精度.现有基于深度学习的地震数据去噪方法,通常仅关注单一时域或频域的特征提取,导致局部过平滑或纹理模糊的现象;此外,传统卷积神经网络的卷积核往往采用固定较小的尺寸,...

Random noise suppression of seismic data based on joint deep learning

Random noise suppression is the key task of seismic data processing. To improve SNR (signal-to-noise ratio) and increase the efficiency and accuracy of following processing and interpretation, appropriate suppression methods should be used for noises induced by different mechanisms and with different characteristics. Applicable denoising methods based on deep learning usually focus on feature extraction in time or frequency domain, which result in over-smoothed or blurred textures in local zones. In addition, the kernel of a traditional convolution neural network is usually set to be a small and fixed block, which limits the size of the receptive field and reduces the diversity of the target characteristics extracted from seismic data. This paper proposes a method of random noise suppression based on joint deep learning. Firstly, features in both time domain and frequency domain are considered, and the joint error is used to define the loss function to improve effect of various extracted features. Secondly, by analyzing the influence of the kernel size and network depth on the size of the receptive field, the method of expanding convolution is used to extract more diverse features and reduce the loss of details of seismic data. Thirdly, according to the similarity between the input and output samples of the network, a residual learning strategy is introduced. Finally, the batch normalization (BN) algorithm is used to accelerate the convergence of the model and improve denoising efficiency. Compared with similar algorithms, the method proposed in this paper has a better effect on preserving the features of events and provides higher SNR.