| dc.description.abstract | Accounting for the evolution of Antarctic ice shelves is crucial for predicting future sealevel
rise. While their role in buttressing tributary glaciers is recognized, modeling these
complex systems has not yet achieved the accuracy of many other components of the Earth
and ice sheet systems. Modeling accurately ice shelf evolution faces challenges, particularly
with ice calving, responsible for almost half of Antarctica’s ice shelf mass losses. The existing
approaches have seen a significant increase in their complexity, making their generalization
challenging on a large scale, and they remain confined to specific cases. Even more significant
limitations arise from our limited understanding of highly nonlinear processes such as
ice shelf damage, which is inherently connected to calving but is not represented in existing
methods. The massive influx of satellite imagery data, measuring physical parameters of ice
shelves, however, opens the door to new data-driven approaches. These new methods which
have largely been underused to model calving, are based on the use of machine learning, and
allows capturing highly nonlinear processes by jointly utilizing multi-source observational
data. In this study, we employed a data-driven approach to predict calving for Antarctic ice
shelves using machine learning models. We chose to predict ice shelf calving using crucial
parameters that can influence the rate of iceberg removal from ice shelf faces. These parameters
are summarized as: ice shelf basal melting, ice shelf thickness, ice shelf velocity, and
sea ice concentration. These were measured from space using multiple sensor remote sensing
data, which were analyzed and standardized. The standardization of the dataset includes
resampling all variables to the same spatial and temporal resolution, and the the design of an
interpolation scheme to assemble a comprehensive dataset. From this work, we assembled two
distinct datasets: a data cube incorporating spatial and temporal evolution, and a tabular
one integrating spatial information ice shelf-wise near the ice front. We then used the tabular
dataset to train a Random Forest model, implementing two cross-validation strategies.
The first ensured balanced data representation across folds, respecting spatiotemporal structures
in the dataset, while the second involved a random split of the data. Unfortunately,
the model failed to accurately predict ice calving in both cases. We attribute this failure
to substantial information loss resulting from the spatial information reduction during the
construction of the tabular dataset. Although allowing for simpler data representation, the
failure of this simple approach suggests the exploration of data driven methods that are able
to capture the strong spatial dimension of the dataset (e.g. how changes in the grounding
line or shear margins will affect ice shelf calving). For the next steps of this project, which is
to be continued beyond the scope of this MSc thesis, we propose using the data cube to feed
more complex machine learning methods, such as a Convolutional Neural Network (CNN),
which are specifically designed to account for spatial information. | |
