Paper Summary
Before data from the multisensor streamers records are combined, it is essential to attenuate the noise from both records to ensure the creation of high-quality up- and down-going wavefields. In recent years, our industry has moved from using statistical and mathematical tools towards machine learning tools to attenuate noise in seismic data. The key motivation has been automation, consistency of output, and quality improvements. We present separate workflows for both pressure and particle motion records that use deep learning to directly attenuate the noise from the records. The heart of the workflows is a convolutional neural network called real image denoising network (RIDNet). The current workflows use a single RIDNet model with exact structure for both pressure and particle motion records to attenuate incoherent noise in the bandwidth where most of the noise exists. Both models were trained using data recorded in the field with supervised learning where the desired outputs were produced by previously developed machine learning based workflows. The new workflows have been extensively validated using records from surveys acquired with different survey geometry, water depth and sea conditions. The validation process confirms that there is no need for workflow modification or re-training of the models. Therefore, the workflows are automated and do not require user interaction.