Predicting Ship Tracks in Waterways Using Long Short-Term Memory

In recent years, several machine learning approaches focusing on the prediction of ship tracks based on ship position and navigation data have been proposed. This work builds on a previous work that predicts ship tracks using positional and weather data. The former include speed over ground (SOG), course over ground (COG), the distance between transition points (TPs) and starboard buoys as well as the turn at TPs. These TPs are the intersection points between the ship's tracks and the lines between port and starboard buoys. Since the best models were trained with bidirectional Long Short-Term Memory (LSTM) networks without weather data, this paper further investigates whether weather data and additional bathymetry data can improve the prediction with these networks. However, we changed the prediction in this paper by combining several models instead of using only one model for it so that each output feature is trained separately in a different model to improve their individual predictions. The models were evaluated by the following metrics: Root Mean Squared Error (RMSE), Mean Squared Error (MSE), and Mean Absolute Error (MAE). In addition, a Wilcoxon signed-rank test was applied between two models to measure the statistical significance of model differences. The results indicate that the prediction can be improved by separating the training for the output features. Furthermore, the bathymetry data improve the prediction of COG, TP's distance and angle, and the prediction of SOG can be improved by using bathymetry data with bidirectional dilated LSTM networks. The weather data provided only better predictions of the feature COG. The commercial ship track data were taken from the German part of the North Sea from 2020 for the training, validation and test data, and the bathymetry and weather data were interpolated for the TPs.