Explainable Machine Learning-Based Prediction of CO2 Emissions From Passenger Vessels

Abstract

Accurate prediction of CO2 emissions from maritime transportation is critically important for environmental sustainability and effective regulatory action. In this study, hourly CO2 emissions were estimated using Random Forest and Extreme Gradient Boosting algorithms, trained on structural and operational data from approximately 2000 passenger ships-including Gross Tonnage, Speed, Length, Beam, Net Tonnage, Draft, and Deadweight Tonnage. Both models demonstrated high predictive accuracy (R2 = 0.92), with Extreme Gradient Boosting achieving lower Root Mean Square Error and Mean Absolute Error values. To enhance transparency, SHapley Additive Explanations analysis was applied, identifying Gross Tonnage, Speed, and Length as the most influential features. A simplified model using only these three variables achieved comparable performance to the full model, offering a more efficient alternative. Furthermore, the ability to predict emissions using only structural ship data provides a practical and cost-effective solution for shipowners and regulators to comply with the European Union's Monitoring, Reporting and Verification System and the International Maritime Organization's Data Collection System. It also supports strategic planning for carbon pricing schemes and Emission Control Areas, where stricter environmental regulations apply. These findings enhance the practical and policy relevance of explainable artificial intelligence-based emission models in the maritime sector.