Improving quality control and data understanding of a large OBN survey through unsupervised machine learning

Summary Given the nature of operational complexity, raw OBN field data received from the acquisition crew often comes with quality problems such as noisy nodes, abnormal instrument response, and positioning errors. Even in relatively small amounts, these problems can become an issue in processing work if they are not detected and corrected at an early stage. In standard target-oriented OBN surveys, manual inspection of amplitude maps measured from the field records is often used to identify anomalously bad nodes for rejection, but this approach becomes impractical on large, regional-scale OBN surveys, such as the multi-year OBN acquisition campaign at offshore Abu Dhabi with hundreds of thousands of receiver gathers. Machine learning approaches can provide practical alternative quality control (QC) tools for identifying problems in field data. By combining the multivariate Gaussian distribution method, principal component analysis (PCA) and K-means clustering, we were able to identify bad nodes using receiver amplitude maps generated from field records. We also demonstrate the application of such tools for removing bad first-break (FB) picks to improve the first-break tomography result.