A Novel Evaluation Metric for Unsupervised Learning in AIS-Based Maritime Anomaly Detection: MADQI - Dataset and Python Code

This repository contains the dataset, Python notebooks, and supporting materials used in the study:

“A Novel Evaluation Metric for Unsupervised Learning in AIS-Based Maritime Anomaly Detection: MADQI”

The study proposes a novel evaluation framework named the Maritime Anomaly Detection Quality Index (MADQI) for assessing unsupervised anomaly detection performance in Automatic Identification System (AIS)-based maritime surveillance applications without requiring labelled data.

The repository includes:

• AIS-based maritime trajectory dataset used in the experiments
• Python notebooks for data preprocessing, feature engineering, anomaly detection, and evaluation
• Haversine distance-based spatial analysis implementation
• Isolation Forest-based unsupervised anomaly detection workflow
• MADQI metric calculation modules
• Visualisation and analysis scripts
• Example outputs and evaluation results

The proposed MADQI framework integrates four complementary evaluation components:

• ARC: Anomaly Rate Consistency
• PPS: Physical Plausibility Score
• SDS: Score Distribution Separation
• ECE: Extreme Case Evidence

The notebooks demonstrate the complete experimental pipeline, including:

1. AIS data preprocessing
2. Temporal and spatial feature extraction
3. Vessel movement analysis
4. Anomaly detection using unsupervised learning
5. Multi-component MADQI evaluation
6. Quantitative and visual interpretation of anomaly behaviour

The implementation is designed to support reproducibility, transparency, and further research in maritime anomaly detection, unsupervised learning evaluation, and AI-driven maritime surveillance systems.

Keywords

AIS, Maritime Anomaly Detection, Unsupervised Learning, Isolation Forest, MADQI, Maritime AI, Machine Learning, Haversine Distance, Vessel Behaviour Analysis, Maritime Surveillance, Explainable AI, Anomaly Detection Evaluation

Author

Ismet Gocer