For economic and efficiency reasons, blended acquisition of seismic data is becoming increasingly commonplace. Seismic deblending methods are computationally demanding and normally consist of multiple processing steps. Furthermore, the process of selecting parameters is not always trivial. Machine-learning-based processing has the potential to significantly reduce processing time and to change the way seismic deblending is carried out. We have developed a data-driven deep-learning-based method for fast and efficient seismic deblending. The blended data are sorted from the common-source to the common-channel domain to transform the character of the blending noise from coherent events to incoherent contributions. A convolutional neural network is designed according to the special characteristics of seismic data and performs deblending with results comparable to those obtained with conventional industry deblending algorithms. To ensure authenticity, the blending was performed numerically and only field seismic data were used, including more than 20,000 training examples. After training and validating the network, seismic deblending can be performed in near real time. Experiments also indicate that the initial signal-to-noise ratio is the major factor controlling the quality of the final deblended result. The network is also demonstrated to be robust and adaptive by using the trained model to first deblend a new data set from a different geologic area with a slightly different delay time setting and second to deblend shots with blending noise in the top part of the record.

A B S T R A C TMarine seismic interference noise occurs when energy from nearby marine seismic source vessels is recorded during a seismic survey. Such noise tends to be well preserved over large distances and causes coherent artefacts in the recorded data. Over the years, the industry has developed various denoising techniques for seismic interference removal, but although well performing, they are still time-consuming in use. Machine-learning-based processing represents an alternative approach, which may significantly improve the computational efficiency. In the case of conventional images, autoencoders are frequently employed for denoising purposes. However, due to the special characteristics of seismic data as well as the noise, autoencoders failed in the case of marine seismic interference noise. We, therefore, propose the use of a customized U-Net design with element-wise summation as part of the skip-connection blocks to handle the vanishing gradient problem and to ensure information fusion between high-and low-level features. To secure a realistic study, only seismic field data were employed, including 25,000 training examples. The customized U-Net was found to perform well, leaving only minor residuals, except for the case when seismic interference noise comes from the side. We further demonstrate that such noise can be treated by slightly increasing the depth of our network. Although our customized U-Net does not outperform a standard commercial algorithm in quality, it can (after proper training) read and process one single shot gather in approximately 0.02 s. This is significantly faster than any existing industry denoising algorithm. In addition, the proposed network processes shot gathers in a sequential order, which is an advantage compared with industry algorithms that typically require a multi-shot input to break the coherency of the noise.

Processing marine seismic data is computationally demanding and consists of multiple timeconsuming steps. Neural network based processing can, in theory, significantly reduce processing time and has the potential to change the way seismic processing is done. In this paper we are using deep convolutional neural networks (CNNs) to remove seismic interference noise and to deblend seismic data. To train such networks, a significant amount of computational memory is needed since a single shot gather consists of more than 10 6 data samples. Preliminary results are promising both for denoising and deblending. However, we also observed that the results are affected by the signal-tonoise ratio (SnR). Moving to common channel domain is a way of breaking the coherency of the noise while also reducing the input volume size. This makes it easier for the network to distinguish between signal and noise. It also increases the efficiency of the GPU memory usage by enabling better utilization of multi core processing. Deblending in common channel domain with the use of a CNN yields relatively good results and is an improvement compared to shot domain.

No abstract