Explainable decision support for icebreaker assistance estimation

Abstract

Safe and efficient navigation in ice-covered waters often depends on timely and accurate estimations of the need for icebreaker assistance. Currently, icebreaker assistance needs are assessed by experienced icebreaker captains based on their own judgment, which can be subjective. Data-driven models have been developed to support this non-trivial estimation, which involves several interconnected factors, including traffic restrictions, ice and weather conditions, and vessel characteristics. The existing study has investigated black box models that achieve great decision accuracy. However, black-box models are limited by poor explainability for end users. This gap reduces end-users’ trust and hinders the adoption of intelligent models in ice navigation. Our previous work (Liu et al., 2025) developed a deep learning-based ensemble model for estimating the need for icebreaker assistance and primarily focused on model predictive performance. This study aims to enable the model’s explainability without compromising the predictive accuracy. Employing SHapley Additive exPlanations (SHAP), we investigate how individual features affect the predicted probability of requiring icebreaker assistance relative to the model’s average prediction at both local and global levels. At the local level, SHAP illustrates how different input features contribute to a single prediction, while at the global level, it summarizes the contributions of these features across all predictions. The explainable results are verified using historical data in the Baltic Sea. The findings indicate that the model can achieve high predictive performance while ensuring explainability through SHAP-based explanations. The outcomes of this paper have the potential to support human-comprehensible explanations, which will help in the evaluation of trust in intelligent decision support systems in the near future.

Keywords: Human-centered Intelligent Decision Support, human-AI Interaction, Marine Technology

DOI: 10.54941/ahfe1007886