Anomaly Detection of Marine Seismic Airgun Signatures using Semi-Supervised Learning

Abstract

Marine Seismic sources (i.e., Airguns) play an important role in geophysical prospecting at sea. They produce seismic waves that propagate into the earth's surface, whereby sensitive detectors collect the reflected data for analysis. The reflected signal can be thought of as a convolution of the seismic emitter(Airgun) and the earth response plus noise. As such a thorough understanding of this Airgun signature is critical to the eventual signal processing used towards a representative image of the subsurface from these detectors. Airguns are typically deployed in the field as arrays of clusters varying in volume capacity. Together these Airguns produce a signature with distinct geophysical characteristics that are required to extract meaningful information of the earth's subsurface. Metadata of from these individual Airgun Arrays is used to assess their performance since anomalous signals can significantly impact interpretations of the subsurface. Features can be extracted from an Airgun signature via a hydrophone within close proximity to the individual array. These labelled features were then used to design a model that classifies Airgun clusters by volume capacity on data known to be free of anomalies. By analyzing features from metadata with an unknown anomalous status with this model, accuracy metrics were used to assess the health and performance of individual Airguns.