Predicting Carbon Emissions from Maritime Transport using Machine Learning Techniques: Predictions in Small Sample Environments

This study aims to compare the predictive performance of various models using a small-sample maritime time series dataset and to identify the key drivers of carbon dioxide (CO₂) emissions in the shipping sector through principal component analysis (PCA)-based variable contribution analysis. The dataset comprises 25 years of annual maritime cargo volumes, ship specifications, and vessel counts, with CO₂ emissions as the dependent variable. After dimensionality reduction via PCA, five predictive models —Linear Regression, Lasso, Ridge, Random Forest, and XGBoost—were applied. Model performance was evaluated using holdout validation, Leave-One-Out Cross Validation (LOOCV), and 5-fold Time Series Split cross-validation. The results show that Ridge regression achieved the highest prediction accuracy with a Mean Absolute Percentage Error (MAPE) of 1.44%, while Lasso and Linear models also demonstrated stable performance. In contrast, tree-based models suffered from overfitting in the small-sample environment, leading to performance degradation. The variable contribution analysis revealed that Gross Tonnage (TT), Container Tonnage (CT), and the number of container vessels (CF_NO) were the most influential predictors, followed by indicators related to LNG/LPG carriers and bulk carriers. This study provides a methodological framework for model selection and variable interpretation in small-sample time series contexts, contributing to both academic research and practical policy-making for carbon emission reduction in the maritime industry.