Automated Route Clustering for Air Traffic Modeling

Abstract

An approach for identifying and approximating well-traveled routes in a historical dataset of flight trajectories is presented. The intent is to use these routes to construct a network model of air traffic flow for use in strategic planning. The flight trajectories are clustered based on spatial and dynamic similarity measures. A coarse clustering process first groups trajectories using origin, destination, and average ground speed information; then a fine clustering process uses the Fréchet distance between pairs of trajectories in the coarse clusters as the more accurate measure of spatial similarity to further divide the trajectories into smaller clusters. The number of resulting clusters is automatically determined by employing a combination of three performance indices. This two-step clustering process has two benefits. It reduces the computational burden because the Fréchet distance computations are required for fewer trajectory pairs due to the initial coarse clustering step. Secondly no manual tuning is required to determine the final number of clusters. A method of detecting and categorizing outliers is presented which fits well with the clustering
process. The clustering and outlier processes are demonstrated on a historical dataset for a region composed of 6 Centers with 21 airports.