Probabilistic Maritime Trajectory Prediction in Complex Scenarios Using Deep Learning

Area of Science:

• Maritime Technology
• Artificial Intelligence
• Oceanography

Background:

• Increasing maritime activity necessitates enhanced surveillance and safety measures.
• Automatic Identification System (AIS) data aids in tracking, but 'dark ships' pose a surveillance challenge.
• Predicting ship trajectories is complex due to numerous possible routes and inherent uncertainty.

Purpose of the Study:

• To develop a probabilistic deep learning model for predicting future ship locations.
• To improve maritime surveillance by characterizing ship trajectory distributions.
• To aid in identifying 'dark ships' by predicting probable future positions.

Main Methods:

• Implementation of a Bidirectional Long-Short-Term-Memory Mixture Density Network (BLSTM-MDN) deep learning model.
• Utilizing AIS data from 3631 cargo ships in a region west of Norway.
• Characterizing conditional probability using an 11-dimensional Gaussian distribution for probabilistic trajectory prediction.

Main Results:

• The BLSTM-MDN model achieved a test Negative Log Likelihood loss of -9.96, with a mean distance error of 2.53 km at 50 minutes into the future.
• The model successfully predicted multiple probable trajectories from a single input trajectory.
• Performance was comparable to deterministic deep learning models on straight paths but superior in complex navigational scenarios.

Conclusions:

• The BLSTM-MDN provides a robust method for probabilistic ship trajectory prediction, outperforming deterministic models in complex situations.
• This approach enhances maritime safety and surveillance capabilities, particularly for identifying unknown or 'dark' vessels.
• Probabilistic predictions offer a more realistic representation of future ship movements compared to single deterministic forecasts.