Mathematik | Informatik
Fabienne Irene Good, 2007 | Münchenstein, BL
Due to global warming, glaciers globally have been retreating at a fast rate since the mid-20th century. Because of their short response times and accessibility to observation, glaciers are some of the most effective indicators of climate change.
This study develops and tests a novel approach to simulate glacier retreat. The recent development of two mid-latitude mountain glaciers which are subjected to rapid retreat – the Glacier de la Plaine Morte in the Swiss Alps and the Smørstabbrean in the Norwegian Jotunheimen – is analysed using remote sensing data. The observations are used to train a random forest classifier – a type of machine learning model – that is then combined with the existing Open Global Glacier Model (OGGM) to make predictions about the glaciers’ dynamical two-dimensional development in the near future.
The random forest classifier adds to the OGGM projections, as it allows to analyse the spatial patterns which drive the overall changes in glacier area.
Introduction
This study aims to accurately depict the outlines of two European mountain glaciers in annual intervals over the past seven years. It investigates whether a random forest classifier can make predictions about the development of those glaciers’ geometries in the near future. The final objective is to evaluate what these predictions signify, how well they align with reference predictions and how plausible they are.
Methods
Sentinel-2 satellite imagery is processed on Google Earth Engine to analyse the past development (2018-2024) of the studied glaciers. A thresholding method is used to identify glacier ice and create glacier outlines. A random forest classifier (RFC) is programmed and trained using Python in Jupyter Notebooks. It is then combined with the existing Open Global Glacier Model’s (OGGM) standard projections using the SSP2-4.5 climate change scenario to simulate the development of the studied glaciers in the near future (2025-2034). The OGGM provides the predicted total glacier area while the RFC simulates the spatial distribution of that area. The predictions are validated against a different set of projections from the OGGM under the same scenario and their plausibility is discussed. In addition, the predictions for 2025 can be validated with now available ground truth data.
Results
With the used methods, past outlines of both glaciers could be derived with an accuracy of >98%.
The predictions show a steady retreat of 12.3% (0.89 km2) from 2025 to 2034 for the Plaine Morte and 6.7% (0.94 km2) for the Smørstabbrean. All individual glacier parts listed in the Randolph Glacier Inventory were predicted to retreat by between 3.2% and 85%. The most retreat was predicted for both glaciers’ Southwestern borders and the termini of glacier tongues.
Annually, an average of 0.31 km2 (Plaine Morte) and 0.42 km2 (Smørstabbrean) were predicted to change their label in comparison to 2024, contributing to a dynamic change in glacier shape and size. Further into the future, the predictions became gradually less dynamic.
The predictions corresponded well to reference predictions in case of the Smørstabbrean and poorly in case of the Plaine Morte. However, for the RFC’s ability to predict the location of annual changes to the glacier’s area, the ground truth validation of the predictions for 2025 showed a recall of 77.8% and a precision of 85.5% (Plaine Morte) and a recall of 59.9% and a precision of 62.0% (Smørstabbrean), respectively.
Discussion
The method applied to analyse the glaciers’ past is well suited and provides scope for refinement, as shown in other studies.
The random forest classifier and the OGGM complement each other well, as they have opposing strengths and weaknesses. The predictions meet fundamental quality requirements and the glaciers behave in a seemingly reasonable, explicable way. The partially high deviations from the reference predictions show the need for further improvement but can be relativised considering the different format of the reference predictions and the irregular sizes of the glacier parts. The ground truth validation advocates for the RFC’s ability to correctly locate the majority of annual changes, although more data has to be collected in the future to deduce clear statements.
Conclusions
The combination of a random forest classifier with the OGGM to create dynamical simulations of glacier outlines gives the model access to climatic data, glaciological theory, and global trends as well as the unique features and past behaviour of an individual glacier, making it a promising new approach.
Würdigung durch den Experten
Martin Gauch
Fabienne Good vereint Methoden der Erdwissenschaften und des Machine Learning, um Änderungen in der Ausdehnung von Gletschern zu modellieren. Auf Basis historischer Satellitenbilder trainiert sie Modelle, die pixelbasiert die zukünftige Gletschergeometrie vorhersagen. Um trotz begrenzter Messdaten möglichst genaue Vorhersagen treffen zu können, wird das Modell von existierenden physikalischen Simulationen der erwarteten Gesamtgletscherfläche geleitet. Die Qualität und Plausibilität der Ergebnisse wird im Vergleich zu realen Messungen und anderen Simulationen detailliert analysiert.
Prädikat:
Silber
Sonderpreis «Forschung auf dem Jungfraujoch» gestiftet von der Akademie der Naturwissenschaften Schweiz & dem Paul Scherrer Institut
Gymnasium Münchenstein, Münchenstein
Lehrer: Alexandre Warin