Anchor Dragging Risk Estimation Strategy from Supervised Cost-Sensitive Learning

Anchor dragging at anchorages poses a significant threat to marine traffic, potentially leading to collisions and damage to seabed infrastructure. This study analyzed a large dataset of ships in anchorage areas to develop a machine learning (ML) model that estimates the risk of anchor dragging using a binary classification system that differentiates between dragging and non-dragging incidents. Historical data from the automatic identification system (AIS), hydrographic, and meteorological sources were compiled for each case. Preliminary analysis revealed a significant class imbalance, with non-dragging cases far outnumbering dragging cases. This suggested that the optimal ML strategy would involve undersampling the majority class and cost-sensitive learning. A combination of data-undersampling methods and cost-sensitive algorithms was used to select the model with the best recall, area under the receiver operating characteristic curve (AUC), and geometric mean (GM) scores. The neighborhood cleaning rule undersampler paired with cost-sensitive logistic regression outperformed other models, achieving recall, GM, and AUC scores of 0.889, 0.767, and 0.810, respectively. This study also demonstrated potential applications of the model, discussed its limitations, and suggested possible improvements for the ML approach. Our method advances maritime safety by enabling the intelligent, risk-aware monitoring of anchored vessels through machine learning, enhancing the capabilities of vessel traffic service officers.