Segmentation optimization in trajectory-based ship classification

The paper presents an analysis over eleven trajectory segmentation techniques applied to the study and experimentation of ship classification problems based only on kinematic information. Using the experimental framework introduced in previous works, it cleans, smooths and extracts trajectories from real-world Automatic Identification System (AIS) data. It also applies three balancing solutions to address the lack of an equal distribution among classes. In total, 196 classification experiments have been carried out, which have been presented with a multi-objective analysis to consider the imbalance problem and conflicting metrics (total and minority class accuracies). The results show a Pareto front with different viable solutions for the classification problem, without a dominant one over the rest. The segments generated in the best experiments (Pareto front) are analysed using specific metrics to compare their impact in the classification problem.