Predicting Vessel Tracks in Waterways for Maritime Anomaly Detection

Many approaches to vessel track prediction and anomaly detection rely only on a vessel's positional data. This paper examines whether including tide and weather data into the track prediction model improves accuracy. We predict vessel tracks in waterways using a bi-directional Long Short -Term Memory (Bi-LSTM) approach and a transformer model. For this purpose, the boundaries of the Elbe and Weser river waterways are merged with vessel position data. Additionally, tide data, as well as weather information, will be used to train the model. To ascertain whether this additional data improves the accuracy, the models have been trained with and without tide and weather data and evaluated against each other. Furthermore, we have investigate whether the predictions can be used for detecting anomalous vessel behaviour. Our results show that the lowest average error and the best RMSE, MSE, and MAE values have been achieved with the Bi-LSTM, where no tide and weather data have been used for training. We have also found that the transformer model is more accurate than a linear prediction model, which is used as a baseline. In addition, we have shown that deviations between predicted and real tracks can be labelled as anomalous. The results have shown that including tide and weather data does not necessarily improve the predictions. Adding data with a low information content to train a machine learning model may introduce noise or bias into the model. We believe that this phenomenon explains our results. Thereby this paper shows that simply adding this data to train the track prediction model may not enhance the overall accuracy.