Assessing the Key Drivers of Arctic Sea Ice Decline Through Machine Learning Techniques
The thesis project was developed by Angelica Iseni and Ellen Poli
In the following a short summary of the research work given by Angelica and Ellen:
The Arctic is warming at about four times the global average, a process known as Arctic Amplification. This rapid warming causes sea ice to melt more quickly, contributing to feedback mechanisms like the ice–albedo effect, where reduced ice cover exposes darker ocean surfaces that absorb more heat, further accelerating ice loss. The decline of Arctic sea ice is a major indicator of global climate change and has far-reaching consequences for the Earth’s climate system.
Read more
This thesis investigates the main factors driving changes in Arctic sea ice thickness. It uses machine learning for feature selection to identify the most important hydro-meteorological variables and climate teleconnections influencing sea ice. To handle the complexity and high dimensionality of the data, three automatic feature selection algorithms are applied: W-QEISS, IIS, and PyCRO-SL (combined with SHAP and alone).
The results show that all three methods yield broadly consistent outcomes, although PyCRO-SL tends to select a larger number of features. Among them, IIS produces the best predictive performance. Minimum and maximum air temperatures are found to be the most influential variables affecting sea ice thickness. Other variables like precipitation, snow conditions, and river discharge play a secondary role. Teleconnections—especially the Atlantic Multidecadal Oscillation (AMO)—are shown to significantly enhance predictive accuracy, demonstrating the importance of including large-scale climate patterns outside the Arctic when modeling sea ice processes.
Field of the research: Natural Resources Management
Supervisor: Prof. Andrea Castelletti