Eurocontrol’s aviation sustainability unit (ASU) develops AI systems aimed at monitoring and mitigating non-CO2 emissions from European air traffic. Contrails, thin ice-crystal clouds formed behind aircraft, account for an estimated 30% of aviation’s non-CO2 climate impact. Improving the ability to detect and segment contrails from satellite observations may support future route-optimization strategies.

The ASU team, composed of three data scientists, primarily works with GOES and MTG ash-RGB imagery, a modality that differs substantially from natural-image RGB and presents domain-shift challenges for standard vision models.

Eurocontrol evaluated several off-the-shelf architectures, including YOLO, RF-DETR, and SAM to detect and segment contrails. Early experiments showed that performance varied significantly across models, with some struggling to generalize well on ash-RGB data.

Contributing factors included:

• Limited utility of standard pretrained weights due to modality mismatch
• A relatively small dataset (~20,000 images) with strong class imbalance (many samples contain no contrails)
• Sparse segmentation label distribution and a high proportion of negative samples

These constraints motivated the team to explore alternative strategies for representation learning, pretraining, and model distillation, to improve scalability and develop a reproducible engineering workflow.

ASU team tested a range of techniques to improve segmentation robustness, including:

1. Domain-specific pretraining on satellite imagery.
2. Use of unlabeled data to compensate for limited annotated samples.
3. Model distillation pipelines with LightlyTrain.

LightlyTrain was used to implement reproducible distillation workflows compatible with models such as YOLO and RF-DETR. According to the ASU team, the framework helped unify augmentation and training configurations and simplified debugging efforts across architectures.

They also highlighted that the documentation and implementation were straightforward. At the same time, the team continued evaluating other pretraining and segmentation strategies to determine the relative impact of different components in the pipeline.

Early experiments demonstrated improvements across segmentation models:

• Distillation improved performance for YOLO-based models, with increases in true positives and upward shifts in metrics such as mAP and DICE.
• Gains were observed even when pretraining and fine-tuning used the same dataset, though the magnitude of improvement varied.
• The ability to mix labeled and unlabeled data helped reduce earlier bottlenecks, but the degree to which this contributed relative to architectural changes is still being quantified.

ASU team is currently preparing more formal benchmarks comparing off-the-shelf models, their custom pretraining strategies, and the distillation workflows developed using LightlyTrain.

As Ana summarised, “We saw an increase in true positives. Once the numbers are ready, we expect the improvements achieved through distillation via LightlyTrain to be significant,” though the team notes that the final evaluation will depend on comprehensive comparisons.

Outlook

The project is ongoing. Future work includes:

• Releasing benchmark results across multiple architectures and training
strategies
• Assessing the generalization of distilled models on new ash-RGB data sources
• Quantifying the relative benefits of unlabeled-data usage, domain-specific pretraining, and distillation frameworks

While early findings indicate promising improvements, Eurocontrol emphasizes that the conclusions will rely on the forthcoming quantitative benchmarks rather than anecdotal trends.