Proceedings of the
35th European Safety and Reliability Conference (ESREL2025) and
the 33rd Society for Risk Analysis Europe Conference (SRA-E 2025)
15 – 19 June 2025, Stavanger, Norway
Uncertainties in Iceberg Detection from Satellite Data: Error-Modelling for the Quantification of Total Uncertainty in Image Segmentation
Fraunhofer Institute for High-Speed Dynamics, Ernst-Mach-Institut, EMI, Germany.
ABSTRACT
In many real-life contexts, particularly those where the impact of potential errors is high, making maximum-likelihood predictions is not sufficient. Rather, one is interested in whether a prediction is certain or uncertain, that is how large one should expect the error to be. This paper focuses on a use case from the EU HORIZON Project AI-ARC (AI-ARC, 2024), which involves AIbased detection of icebergs from satellite images for naval navigation. The high-risk task carries significant consequences for incorrect predictions. Thus, an automated analysis of the uncertainty of iceberg predictions is essential.
Keywords: Predictive uncertainty, Monte Carlo dropout, Error-modelling.
