DAMOT: A Novel Algorithm for High-Speed Target Sparse Flight Trajectory Interpolation

High-speed target flight trajectories typically involve extensive temporal and spatial ranges, with data frequently missing in both discrete and continuous segments. Missing data lead to overall trajectory sparsity, and hindering the development of downstream applications. Most existing algorithms primarily address discrete missing data, lacking the capability to interpolate trajectories through continuous missing segments. To address this issue, we propose a Transformer-based decoupled adaptive learning multi-headed output algorithm (DAMOT) for sparse flight trajectory interpolation. DAMOT leverages the self-attention mechanism to learn inferential relationships among individual trajectory points and incorporates observed data to predict missing states. Using a multi-head attention framework with multi-head outputs, it decouples multi–dimensional attributes—such as latitude, longitude, altitude, and speed—to reduce systematic errors in attribute–specific interpolations. A boundary prediction layer further enhances spatial constraints for missing spans by leveraging boundary point information. Meanwhile, an adaptive loss function dynamically adjusts the model's focus during different training stages, improving convergence. Experimental results demonstrate that DAMOT's span–masking training method offers greater robustness to continuous missing data, and that its architecture and loss function achieve higher interpolation accuracy compared to existing approaches.