Maneuver pattern classification is fundamental for understanding and predicting the dynamic behaviors of aerial vehicles operating in increasingly complex airspace environments. However, existing rule-based and data-driven approaches are constrained by the scarcity, imbalance, and limited maneuver diversity of real-world flight data, leading to a restricted generalization capability and a reduced robustness to noise. To address these challenges, we construct a standardized Maneuver Pattern Library, a curated dataset of simulated flight trajectories encompassing five representative maneuver primitives: climb, descent, left turn, right turn, and loiter. Trajectories are generated using the X-Plane 12 flight simulator under controlled conditions to ensure maneuver diversity and label consistency, refined through noise reduction and cubic spline interpolation, and rendered from synchronized top and side views with time-encoded color gradients to preserve temporal continuity. Building upon this dataset, we propose DualView-LiteNet, a lightweight Siamese convolutional network designed to jointly learn complementary spatial and temporal cues from dual-view trajectory representations through parameter sharing and feature fusion. In addition to comprehensive comparisons with multiple baseline models on the simulated benchmark, we further evaluate the trained model via direct inference on a real-world ADS-B dataset collected from ADS-B Exchange, without any fine-tuning. The consistent performance observed in this sim-to-real setting demonstrates the practical feasibility and generalization capability of the proposed approach. The experimental results show that DualView-LiteNet achieves an accuracy of 97.64%, with its precision, recall, and F1-score all reaching 0.98 on the benchmark dataset, validating its effectiveness for aerial maneuver pattern classification and establishing a reliable reference for future research.
Yang et al. (Sat,) studied this question.