Spatiotemporal Image-Based Flight Trajectory Clustering Model with Deep Convolutional Autoencoder Network

Abstract

Recent studies in four-dimensional flight trajectories attempted to identify the impacts of various flight trajectories and maneuver parameters on air traffic management efficiency and aviation safety. The previous studies attempted to cluster trajectories based on spatial scales. However, these might require converting the flight trajectories to equal lengths for sequence-based clustering. This paper proposes a novel trajectory three-channel image representation and Gaussian mixture model clustering based on several image-processing methodologies. The aircraft’s latitude, longitude, flight level, and ground speed are represented as corresponding pixel information of the image followed by image-based flight trajectory representation and clustering methods (including deep convolutional autoencoder (DCAE), principal component analysis (PCA) image dimensionality reduction, and image feature points extraction) using a half-year of automatic dependent surveillance-broadcast flight trajectory data in the Hong Kong flight information region. The computational results indicate that the image-based trajectory representation produces more insights for trajectory processing, such as the application of convolutional neural networks and image-processing algorithms. In addition, the DCAE model has better performance and robustness for trajectory feature extraction and similarity analysis than PCA, which will provide ideas for multiparameter trajectory similarity analysis and prediction.