| Summary: | American Geophysical Union Fall Meeting, 12-16 December 2022, Chicago Ice shelves play a pivotal role in controlling the evolution of Antarctic glaciology by restraining, buttressing, and modulating the flow of grounded ice into the Southern Ocean. The stability of the Antarctic Ice Sheet thus depends critically on the stability of the ice shelves that fringe the continent. It is therefore important to understand how these shelves respond to environmental stresses, especially those as common as gravity wave forcings. This study focuses on applying machine learning to automatically detect, classify, and catalog low-frequency (0-70 mHz) gravity wave events impacting the Ross Ice Shelf (RIS) by panoptically segmenting seismic spectrograms. The data used to supervise training was collected by a broadband seismic array deployed on the RIS from November 2014 to November 2016 and was used to generate spectrograms of up to 70 mHz that were examined for infragravity waves and swell events. Our modified U-Net architecture achieved a Dice similarity coefficient (DSC) of over 0.73 during event detection, and its corresponding post processing pipeline recorded an accuracy of 94.4% during classification, outperforming alternative rule based techniques. This work serves as a proof-of-concept for using deep-learning algorithms to detect and catalog gravity wave events, a development that would allow for an improved understanding of the long-term stability of Antarctic ice shelves Peer reviewed
|
|---|