Paper Summary
Despite technological advancements in marine seismic multisensor acquisition and processing, noise attenuation remains a fundamental step in the early processes for producing high-quality upgoing pressure wavefield data. If we assume the main shortcoming of traditional methods is in the noise detection step, deep learning can be used in only the detection step and the selected noise attenuation engine can be automatically guided by the deep learning noise classification. We have created different deep learning models to detect a variety of noise types present in both marine hydrophone and geophone records. These models are used to automatically classify the samples in the noise attenuation workflows and pass the samples to the appropriate noise attenuation steps. Targeted noise detection lets us perform a better targeted noise attenuation with appropriate levels of harshness without undue concern over possible signal loss. Models can also be used at any step of the processing to classify the samples in both hydrophone and geophone records. The improvement in noise attenuation and its impact on the PUP generation is presented for a real dataset. The advantages to turnaround and quality that arise from the use of these workflows are discussed.