A probabilistic framework with uncertainty decomposition for Arctic maritime risk prediction
Kandel, R., Baroud, H., DuVivier, A. K., Bennartz, R.. (2026). A probabilistic framework with uncertainty decomposition for Arctic maritime risk prediction. Reliability Engineering and System Safety, doi:https://doi.org/10.1016/j.ress.2026.113318
| Title | A probabilistic framework with uncertainty decomposition for Arctic maritime risk prediction |
|---|---|
| Genre | Article |
| Author(s) | R. Kandel, H. Baroud, Alice K. DuVivier, R. Bennartz |
| Abstract | Predicting Arctic shipping incidents is challenging not only because conditions are harsh, but because the environmental data driving those predictions are themselves uncertain. Existing Arctic risk models do not distinguish whether an unreliable forecast stems from limited training data, model imperfections, or inherent environmental variability that limits the usefulness of risk estimates for voyage planning and emergency response. This study establishes a probabilistic framework for Arctic maritime risk with uncertainty decomposition that explicitly separates and propagates model training (epistemic) and input data (aleatoric) uncertainties. The framework employs a two-stage cascaded XGBoost architecture to predict incident occurrence and, conditional on an incident, its incident type. Model uncertainty is represented through repeated training on stratified bootstrap resamples of the incident data. Data uncertainty is represented through an ensemble of climate model simulations, benchmarked against a historical reanalysis dataset, which together characterize the range of plausible environmental conditions for a given time and location. Uncertainty contributions are quantified and attributed using variance-based decomposition via the law of total variance and mutual-information-based decomposition, applied across the Northwest Passage and Northern Sea Route. Results show that inherent data and environmental variability dominate residual predictive uncertainty over model structure. This distinction carries direct operational meaning. Irreducible uncertainty should inform precautionary safety margins and preparedness planning. Reducible uncertainty, in turn, identifies targeted opportunities for model improvement and expanded environmental monitoring. Mutual information analysis identifies dominant environmental variables that drive uncertainty, including atmospheric variables, maximum wind speed, near-surface temperature, precipitation, and pressure, and snowfall rate on ice. A key finding is that the remaining sea-ice variables, ice fraction, snow fraction, and snow thickness on ice, contribute only marginal incremental information once the atmospheric state is known. This result has direct implications for model simplification and the prioritization of data collection efforts in Arctic risk assessment. The framework delivers interpretable, probabilistic risk estimates that can support voyage planning and route selection under uncertain Arctic conditions. |
| Publication Title | Reliability Engineering and System Safety |
| Publication Date | Aug 1, 2026 |
| Publisher's Version of Record | https://doi.org/10.1016/j.ress.2026.113318 |
| OpenSky Citable URL | https://n2t.net/ark:/85065/d7db86fn |
| OpenSky Listing | View on OpenSky |
| CGD Affiliations | PPC |