Anomaly Detection in Maritime Ship Trajectory using Deep Learning Algorithms
Anomaly detection in maritime vessel trajectories is essential for maintaining safety, security, and operational efficiency in global shipping. Existing maritime anomaly detection methods often suffer from high false alarm rates and limited ability to model complex long-term vessel movement patterns, reducing their effectiveness in real world maritime surveillance. This study develops an ensemble deep learning framework that combines three unsupervised learning algorithms—long Short-Term Memory (LSTM), Autoencoder, and Recurrent Neural Network (RNN) to detect anomalies in vessel tracking data. The proposed framework was evaluated using an Automatic Identification System (AIS), AIS_Unicorn dataset obtained from kaggle consisting of 103,995 AIS records. From Automatic Identification System (AIS) information. Each model captures distinct patterns: LSTM learns long-range temporal dependencies in sequential vessel movements; Autoencoder reconstructs static navigational features to identify deviations; and RNN models temporal structures in trajectory sequences. Normalized reconstruction errors from each model are aggregated into a unified ensemble score, enabling robust anomaly classification through percentile-based thresholding. Model performance is evaluated using standard metrics including accuracy, precision, recall, F1-score, and ROC-AUC on synthetic anomaly-injected test datasets. Results demonstrate that the ensemble approach substantially outperforms individual models, achieving 95% accuracy, 0.8333 precision, perfect recall (1.0000), and an F1-score of 0.9091, compared to LSTM (90% accuracy, F1: 0.8333), Autoencoder (80% accuracy, F1: 0.7143), and RNN (85% accuracy, F1: 0.7692). By leveraging complementary strengths of the three base models, the ensemble successfully reduces false positives while maintaining complete anomaly detection capability. This framework provides a practical solution for detecting maritime anomalies including unusual course deviations, extended idle periods, and irregular speed patterns. The proposed ensemble methodology offers a scalable approach for real-time maritime surveillance and supports decision-making in port operations, coast guard activities, and autonomous vessel systems. Future work may incorporate temporal deep learning enhancements, adaptive weighting schemes, and integration of contextual maritime data such as vessel classification and geolocation information.
Download Paper