Core Concepts
This study proposes a probabilistic feature engineering approach and a deep learning model architecture to accurately forecast long-term multi-path vessel trajectories using Automatic Identification System (AIS) data.
Abstract
The key highlights and insights of this content are:
The authors developed a sophisticated AIS data analysis tool called the Automatic Identification System Database (AISdb) to handle the challenges of dealing with erroneous and noisy AIS messages.
The authors split the Gulf of St. Lawrence into a hexagonal grid and extracted probabilistic features that indicate the potential route and destination of vessels based on historical AIS data. These features were engineered using conditional probabilities and Euclidean distance calculations.
The authors proposed a deep learning model architecture that uses parallel convolutional neural networks to extract spatial features, a Bidirectional LSTM with a position-aware attention mechanism to capture temporal dependencies, and a decoding phase to generate the final trajectory predictions.
The authors conducted extensive experiments to test the effectiveness of the probabilistic features and the deep learning model. They found that the probabilistic features had an F1 score of around 85% and 75% for predicting the route and destination, respectively.
When using only coordinates, speed, and course information, the model struggled with complex and curved paths, resulting in mean and median errors of 13 km and 8 km, respectively. However, incorporating the probabilistic features and applying trigonometric transformations reduced the errors to 11 km and 6 km, respectively.
The authors achieved an R2 score higher than 98% for cargo and tanker vessels, demonstrating the model's ability to make accurate path decisions, even in the face of curvy patterns and sudden route changes.
The proposed approach has the potential to improve maritime safety and reduce the risk of vessel-whale collisions by enabling proactive adjustments to shipping routes and speed restrictions in areas with high whale activity.
Stats
The average and median forecasting errors were reduced from 13 km and 8 km to 11 km and 6 km, respectively, by incorporating the probabilistic features and applying trigonometric transformations.
The model achieved an R2 score higher than 98% for cargo and tanker vessels, demonstrating its ability to make accurate path decisions.
Quotes
"The probabilistic features have an F1 Score of approximately 85% and 75% for each feature type, respectively, demonstrating their effectiveness in augmenting information to the neural network."
"Our model achieved a R2 score of over 98% in different techniques using varying features. The high R2 score is a natural outcome of the well-defined shipping lanes. However, our model stands out among other forecasting approaches as it demonstrates the capability of complex decision-making during turnings and path selection."