This work presents the progress of the development of a predictive model for carbon dioxide (CO₂) emissions from vessels transiting and arriving at the Panama Canal, based on Automatic Identification System (AIS) data. A reproducible workflow was developed that combines H3 hexagonal spatial indexing to segment each vessel’s route, pseudo-voyage generation through sliding windows, and a preprocessing pipeline (imputation, unsupervised correlation-based multicollinearity reduction, and scaling/encoding) fitted exclusively on the training set to prevent data leakage. During this stage of the project, a target-leakage issue affecting the previous iteration was identified and corrected: two of the three energy terms used to compute the emission target had been included among the model’s input features, which had produced artificially inflated preliminary results (R² ≥ 0.996). After removing these leakage sources and restricting the feature set to fixed vessel and operational attributes, three ensemble-based machine learning models were retrained and compared: Random Forest, LightGBM, and XGBoost. XGBoost achieved the best overall performance (R² = 0.6173, natural scale), followed by LightGBM (R² = 0.5585) and Random Forest (R² = 0.1691). Because the canal-proximity definition currently operates at the native H3 resolution of the dataset (resolution 1, hexagons of roughly 418 km per side), 91.78% of position records fall inside the defined neighborhood, so performance on this “critical segment” closely tracks global performance rather than isolating a distinct, harder navigational phase as in the earlier version of this work. The results confirm that, once leakage is removed, ensemble models still offer a viable though considerably more modest basis for emissions prediction and identify the refinement of the canal’s spatial definition as the priority for the next phase of the project.

Keywords

CO₂ Emissions, Vessels, Panama Canal, Spatial Indexing