Enhancing air traffic complexity assessment through deep metric learning: A CNN-Based approach

Abstract

Air traffic complexity is related to the workload of air traffic control officers and pilots, subsequently leading to potential effects on flight safety and efficiency. However, assessing air traffic complexity accurately is still a question in the concept of Air Traffic Management (ATM). In this study, a model for air traffic complexity assessment is proposed based on deep metric learning. Specifically, the air traffic data collected by surveillance radars are adopted to generate the air traffic image set, from which air traffic features are extracted based on the Convolutional Neural Networks (CNN) model. After that, the deep metric learning method based on Asymmetric distance, Aggregation loss, and Edge hard loss, called AAEDM, is applied to address the problem of class imbalance in air traffic images. Finally, the traffic complexity assessment model is proposed based on AAEDM. The proposed model is validated through comprehensive experimentation using two established standard datasets. The results of these experiments indicate the outstanding proficiency of AAEDM, particularly in scenarios involving unbalanced data. The proposed model can extract deeper features of air traffic than traditional machine learning methods, outperforming other models for air traffic complexity assessment. This study can help assess air traffic complexity and improve the robustness of the ATM system.