A deep learning framework for predicting aircraft trajectories from sparse satellite observations.

Area of Science:

• Aerospace Engineering
• Artificial Intelligence
• Data Science

Background:

• Satellite-borne optical sensors offer global air-traffic monitoring potential but face challenges due to fragmented observations and low temporal resolution.
• Accurate trajectory forecasting is crucial for air-traffic management but is hindered by data limitations from space-based sensors.

Purpose of the Study:

• To introduce HiFormer, a novel deep learning framework for direct multi-step trajectory prediction using sparse, space-based observations.
• To develop a robust method capable of capturing diverse motion patterns and dependencies from fragmented flight data.

Main Methods:

• HiFormer integrates convolutional, recurrent, and attention mechanisms in a unified architecture for comprehensive sequence modeling.
• A large-scale synthetic dataset of 12,000 trajectories and a real-world dataset of 1000 fragmented ADS-B flight segments were created and utilized.
• The framework processes short-term maneuvers, medium-range trends, and long-range dependencies in a single forward pass.

Main Results:

• HiFormer achieved up to a 30% reduction in multi-step prediction errors on synthetic data.
• The framework demonstrated a 10% improvement in prediction accuracy on real-world ADS-B flight tracks compared to existing methods.
• Experiments confirmed HiFormer's effectiveness in handling sparse and irregular observational data.

Conclusions:

• HiFormer provides a robust and effective framework for trajectory prediction with sparse, space-based observations, significantly advancing air-traffic monitoring capabilities.
• The developed model shows promise for improving forecasting in various domains that rely on incomplete or irregularly sampled data.
• This work addresses critical data challenges in space-based surveillance and paves the way for more reliable global air-traffic management systems.