Multiscale Residual Convolution Neural Network for Seismic Data Denoising

Obtaining high signal-to-noise ratio (SNR) data is significant for the subsequent processing and interpretation of seismic data. In recent years, the convolutional neural network (CNN) has been widely used in seismic data denoising. However, the existing CNN-based method usually has a single receptive field, making it difficult to effectively extract feature maps at different scales. Therefore, we propose a multiscale residual U-shaped CNN (MRUnet) by combining the multiscale structure, residual structure, and skip connection structure to cope with the random noise of the poststack seismic data. The network can use convolutional kernels of different sizes for feature extraction and transfer these features through more extensive skip connections. We construct a training set using existing seismic data and transfer the trained model to field data for denoising experiments. Experiments on synthetic and field data demonstrate that by training the network, a model that removes the random noise from the poststack seismic data can be obtained and outperforms the existing ones.