Abstract

Traffic Management systems generate a huge amount of track data daily. Flight trajectories can be clustered to extract main air traffic flows by means of unsupervised machine learning techniques. A well-known methodology for unsupervised extraction of air traffic flows conducts a two-step process. The first step reduces the dimensionality of the track data, whereas the second step clusters the data based on a density-based algorithm, DBSCAN. This paper explores advancements in density-based clustering such as OPTICS or HDBSCAN*. This assessment is based on quantitative and qualitative evaluations of the clustering solutions offered by these algorithms. In addition, the paper proposes a hierarchical clustering algorithm for handling noise in this methodology. This algorithm is based on a recursive application of DBSCAN* (RDBSCAN*). The paper demonstrates the sensitivity of these algorithms to different hyper-parameters, recommending a specific setting for the main one, which is common for all methods. RDBSCAN* outperforms the other algorithms in terms of the density-based internal validity metric. Finally, the outcome of the clustering shows that the algorithm extracts main clusters of the dataset effectively, connecting outliers to these main clusters. Document type: Conference object