Spatial Modeling of Maritime Risk Using Machine Learning.

Area of Science:

• Maritime Safety
• Spatial Risk Modeling
• Machine Learning Applications

Background:

• Navigational safety management is complex, with traditional risk models limited by scale and generalization.
• Existing maritime risk assessments often struggle with infrequent events, diverse causes, and large geographical areas.
• Current methods may lack scalability and inaccurately represent maritime risks.

Purpose of the Study:

• To propose a novel machine learning-based spatial modeling method for maritime risk.
• To enhance the characterization of navigational safety by leveraging extensive relevant data.
• To develop a scalable and accurate approach for predicting maritime accident occurrences.

Main Methods:

• Aggregation of historical accident data, vessel traffic, and other features into a spatial grid.
• Implementation of classification algorithms, including XGBoost and Random Forest, to predict annual accident occurrences for different vessel types.
• Application of the spatial modeling approach to assess collision and grounding risks in the United Kingdom.

Main Results:

• The machine learning model demonstrated high accuracy, with Area Under Curve scores exceeding 90% in most implementations.
• Ensemble tree-based algorithms like XGBoost and Random Forest showed superior performance compared to other tested machine learning methods.
• Risk maps effectively characterized maritime hazards, varying by hazard and vessel type.

Conclusions:

• The proposed machine learning method offers a scalable and accurate solution for spatial maritime risk modeling.
• The developed risk maps provide actionable intelligence for targeted risk mitigation strategies.
• This approach significantly improves the characterization of navigational safety compared to traditional methods.