Aircraft Trajectory Estimation in Distributed Sensor Networks via End-to-End Deep Learning

In the domain of air traffic control, the Automatic Dependent Surveillance-Broadcast (ADS-B) system has been widely adopted to provide real-time aircraft coordinates to ground stations. However, when the ADS-B system is unreliable due to attacks or interference, accurately predicting the aircraft coordinates solely based on a distributed ground sensor network becomes a major challenge. To address the limitation of existing prediction methods that fail to bidirectionally model both sensor sequences and trajectory sequences, we propose MAGE (Multisensor Aircraft Trajectory Prediction with GRU-based Estimation)-an end-to-end deep learning framework that integrates data from distributed ground sensor networks to achieve high-precision aircraft trajectory prediction. MAGE employs Bidirectional Gated Recurrent Units (Bi-GRU) to construct both a sensor encoder and a trajectory decoder, extracting the spatiotemporal features of multi-source heterogeneous sensor data and recursively decoding multi-step trajectory sequences from the learned high-dimensional representations. In the empirical study, we trained and evaluated MAGE using 3,000 trajectories from the OpenSky dataset. Taking longitude as an example, MAGE reduces RMSE and MAE to 2.82° and 2.2147°, respectively, while lowering MRE to $3.61 \%$, achieving improvements of $15.20 \%, 6.36 \%$, and $0.24 \%$ over the second-best model. These results indicate that MAGE can reliably estimate aircraft trajectories using only sensor data when ADS-B messages are unreliable, consistently outperforming other models across multiple error evaluation metrics. By providing accurate trajectory reconstruction, MAGE enhances the security and stability of air traffic communication networks.