Predicting Calving Events in Antarctica using Machine Learning

Jacob Alexander Hay, Hamzeh Issa, David Parkes, Daniele Fantin, Jan Wuite, Amber A Leeson, Malcolm McMillan

Keywords: u-net, deep learning, machine learning, antarctic, antarctica, ice shelf, glaciology, dynamic system, cnn, calving

TL;DR: Forecasting and predicting calving events from an ice shelf in Antarctica using a U-Net architecture.

Abstract:

Monitoring the calving dynamics of the Antarctic ice shelves is central to
understanding a major driver for the changes to ocean levels on our planet.
Several physical models have been proposed as calving laws, with varying
predictive power. We propose an approach using Machine Learning (ML) to identify key
variables and parameters that may be used in future models of the ice shelf
calving dynamics. As part of an ongoing project, we have trained
a U-Net on samples from a set of Gaussian Random Field-represented Essential Climate Variables.
Ablation studies establish a few of the selected variables as having high
correlation with calving events, with an F1 score above 0.9.
Our first study site is the Larsen C Ice Shelf, on the northwest part of
the Weddel Sea, where in 2017 there was a massive calving event. We have found
strong correlations between the changes in ice-velocity leading up to this
event, which are further improved when accounting for basal melt rates in the
area.

Serve As Reviewer: ~Jacob_Alexander_Hay1, ~Hamzeh_Issa1, ~Daniele_Fantin1

Submission Number: 17