X-VORTEX: Spatio-Temporal Contrastive Learning for Wake Vortex Trajectory Forecasting

TL;DR

X-VORTEX tackles the challenge of wake vortex analysis from sparse LiDAR data by learning physics-aware, spatio-temporal representations in a self-supervised manner. It introduces a two-view contrastive framework grounded in Augmentation Overlap Theory, combining temporal evolution with spatial sparsity through a time-distributed encoder and a temporal aggregator. The approach yields strong unsupervised representations, enables high-precision center localization with only $1\%$ of labels, and supports accurate short-horizon trajectory forecasting, outperforming heuristic, image-based, and fully supervised baselines. This workflow reduces labeling needs and provides a practical pathway toward real-time wake-vortex advisory capabilities with robustness to noise and decay, with potential applicability to other atmospheric flow phenomena captured by remote sensing.

Abstract

Wake vortices are strong, coherent air turbulences created by aircraft, and they pose a major safety and capacity challenge for air traffic management. Tracking how vortices move, weaken, and dissipate over time from LiDAR measurements is still difficult because scans are sparse, vortex signatures fade as the flow breaks down under atmospheric turbulence and instabilities, and point-wise annotation is prohibitively expensive. Existing approaches largely treat each scan as an independent, fully supervised segmentation problem, which overlooks temporal structure and does not scale to the vast unlabeled archives collected in practice. We present X-VORTEX, a spatio-temporal contrastive learning framework grounded in Augmentation Overlap Theory that learns physics-aware representations from unlabeled LiDAR point cloud sequences. X-VORTEX addresses two core challenges: sensor sparsity and time-varying vortex dynamics. It constructs paired inputs from the same underlying flight event by combining a weakly perturbed sequence with a strongly augmented counterpart produced via temporal subsampling and spatial masking, encouraging the model to align representations across missing frames and partial observations. Architecturally, a time-distributed geometric encoder extracts per-scan features and a sequential aggregator models the evolving vortex state across variable-length sequences. We evaluate on a real-world dataset of over one million LiDAR scans. X-VORTEX achieves superior vortex center localization while using only 1% of the labeled data required by supervised baselines, and the learned representations support accurate trajectory forecasting.

Figures (8)

• Figure 1: Measurement geometry of a ground-based LiDAR (L) scanning perpendicular to the runway. The scan captures the radial velocity of the Port and Starboard vortices trailing a landing aircraft wartha2022characterizingqu2025explainable.
• Figure 2: Large Eddy Simulation (LES) visualizing the roll-up process. Vorticity sheets shed from the wingtips and flaps merge to form the characteristic counter-rotating vortex pair.
• Figure 3: Overview of X-VORTEX. Top: self-supervised pre-training on unlabeled wake sequences using weak and strong augmentations of the same sequence, optimized with InfoNCE. Bottom: downstream adaptation on labeled sequences, where the pretrained spatial encoder and temporal aggregator are frozen to produce representations for (i) vortex center localization via a segmentation-based soft-center head and (ii) short-horizon trajectory forecasting via a lightweight prediction head.
• Figure 4: Qualitative Visualization. A sequence of five consecutive LiDAR scans ($t$ to $t{+}4$) showing the radial velocity field, where warm colors indicate motion toward the sensor and cool colors indicate motion away. The counter-rotating port and starboard vortices appear as paired regions of opposite velocity. Green and orange circles denote ground-truth vortex centers, while crosses indicate model predictions.
• Figure 5: Alignment and uniformity during self-supervised pre-training (PointNet backbone). Validation alignment and uniformity over the first 20 epochs are shown for clarity. Alignment improves rapidly as positive pairs are pulled together, while uniformity decreases steadily, indicating increasing dispersion of embeddings across the hypersphere. Both metrics stabilize early, suggesting stable contrastive training without representation collapse.