Applying machine learning models for decision support in maritime surveillance: Predicting illegal activities in national sovereignty areas
Nunes, A.F.T.S. (2025). Applying machine learning models for decision support in maritime surveillance: Predicting illegal activities in national sovereignty areas. MA Thesis. NOVA Information Management School, Instituto Superior de Estatística e Gestão de Informação, Universidade Nova de Lisboa: Lisboa. 62 pp.
|  |
| Available in | Author |
|
Document type: Dissertation
|
| Author keywords |
Maritime Surveillance; Machine Learning; Decision Support Systems; Illegal Activities Prediction; Spatiotemporal Data Analysis; SDG 14 - Life below water; SDG 16 - Peace, justice and strong institutions |
| Abstract |
This work explores how historical maritime inspection data can improve operational decisionmaking in maritime surveillance by applying machine learning models to predict the likelihood of illegal activities. In Portugal, the selection of patrol areas continues to be based on the experience of commanders and static views from previous inspections, which limits the efficiency of resource allocation in extensive maritime zones such as the Exclusive Economic Zone (EEZ). To address this challenge, a data-driven approach was adopted, using inspection data collected by the Portuguese Navy between 2014 and 2024. The dataset included over 15,000 records with spatial and temporal features such as inspection location, month, period of day, day of the week, and vessel type. Four classification models were tested—Naïve Bayes, Logistic Regression, k-Nearest Neighbour, and Classification Trees—following the CRISP-DM methodology. Model performance was evaluated using precision, recall, accuracy, F1-score, ROC-AUC, and Brier score, considering balanced and imbalanced training scenarios. Results showed that although the Classification Tree model achieved the highest precision (0.71), it had a very low recall, which indicates a conservative but limited detection capability. On the other hand, k-NN achieved a better balance with higher recall but lower precision. The results confirm that historical inspection data contains predictive value and could be applied to a decision-support system to assist in the selection of a patrol area. Probability maps generated from model outputs highlight high-risk zones and offer a practical tool for prioritising enforcement actions. Despite limitations related to data imbalance and geographic distribution of inspections, the proposed approach demonstrates the feasibility of improving maritime surveillance using structured data and machine learning. Future developments could include the integration of real-time data sources, such as AIS and weather conditions, and the use of more complex models to support wider maritime authority operations. |
|
