the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 28 Jul 2026
Development of an automated contrail-to-flight attribution using Meteosat Second Generation satellite data
Abstract. Persistent contrails are a major contributor to the effective radiative forcing from aviation. Optimizing flight trajectories to avoid regions prone to the formation of warming contrails has therefore been proposed as a mitigation strategy to achieve the international climate targets for the aviation sector. However, the precise attribution of observed contrails to individual flights remains a challenge for both the evaluation of avoidance measures and the validation of contrail models. This study presents a fully automated evaluation framework designed for matching contrails to flights using observations from the Spinning Enhanced Visible and Infra-Red Imager (SEVIRI) instrument aboard the geostationary Meteosat Second Generation (MSG) satellite and flight trajectories. The method utilizes an automated contrail detection algorithm. Although such automated detections designed for MSG/SEVIRI often face a trade-off between high detection efficiency and low false alarm rate, the proposed matching process substantially improves detection reliability through a multi-step verification: contrails are preselected based on their spatial and temporal occurrence as well as their spatial orientation with respect to the flight trajectories, followed by temporal tracking. A new aspect is the development of a life cycle-based confidence score to derive a quantitative matching score. In contrast to previous approaches, the proposed method is entirely observation-driven, requires no model input, and enables rapid, large-scale application. The framework’s performance is demonstrated through two specific case studies. Future applications include the assessment of contrail avoidance trials across the Europe-African airspace and adaptation to other satellite configurations.
- Preprint
(12113 KB) - Metadata XML
- BibTeX
- EndNote
Status: open (until 01 Nov 2026)
- RC1: 'Comment on egusphere-2026-3046', Anonymous Referee #1, 18 Aug 2026 reply
-
RC2: 'Comment on egusphere-2026-3046', Anonymous Referee #2, 14 Oct 2026
reply
Remarks to the Author:
The authors present a fully automated, completely observation-based, multi-step framework that matches contrails to individual flights. The precise attribution of contrails to individual flights is relevant because it allows evaluation of climate-optimized flight trajectories. Such re-routed flights aim to avoid regions prone to the formation of (persistent) contrails, one strategy - among others - to achieve climate-neutral aviation by 2050. The framework uses rapid scans (5-minute temporal resolution) from MSG-SEVIRI which cover a latitude range from about 15° to 70°N at 3 km by 3 km spatial resolution at nadir. The fully automated framework is new and in comparison to existing algorithms it does not rely on external wind data to determine contrail advection. The authors aim to prevent the two primary sources of uncertainty (i) inherent inaccuracies in modeled wind fields, and (ii) the altitude-dependent advection errors caused by contrail settling. Instead, the assignment of contrails to flights is based on geometric criteria as well as several metrics that are combined into an overall value for the assignment.
The framework is built on different steps (i) collocation of flight segments to rapid scans, (ii) contrail detection and filtering, (iii) contrail-to-flight matching, (iv) contrail tracking, and (v) calculation of a matching score based on multiple criteria.
In MSG-SEVIRI rapid scans, contrails are detected using COCOS. The algorithm is briefly introduced, sufficient detail about its underlying machine learning model and training is missing. It may also be of interest to the reader how the COCOS algorithm performs in comparison to existing literature, such as Meijer et al. (2022) or Ortiz et al. (2024, 2025). It is also known that brightness-temperature differences are sensitive to rivers such as the Nile, leading to false detections; therefore, it would be interesting to learn how COCOS handles this issue.Detected contrails need to be selected as 'potential matches with the flight'. To account for horizontal advection, the authors assume a maximum wind speed at flight altitude (50 km/h) that defines an oval-shaped region with a radius that is defined as the product of the assumed wind speed and the elapsed time between flight passage and contrail detection.
Contrail tracking, then, relies on an existing algorithm, ACTA, which was translated from IDL to Python and for improvement slightly modified.
Potential matches are analyzed using clusters and the contrail life cycle is evaluated based on five matching criteria. A cluster is a group of contrails in consecutive rapid scans that correspond to the same physical contrail (based on position, orientation, and length).
The combination of the five matching-criteria scores gives a matching score that assesses the consistency of the cluster with the flight segment.
Finally, the study presents 4 case studies that illustrate the framework's performance.Advanced cloud-tracking algorithms already exist and those reliably track clouds in geostationary satellite data, cf. TOBAC (Heikenfeld et al., 2019) or LIMOTTo (Seelig et al., 2025). These object-based tracking algorithms apply cross-correlation to calculate a displacement map - due to horizontal wind and/or advection - which is then used to forecast the position of cloud objects in the next time step.
The development of a method using a different Ansatz in detecting clouds (in the present case contrails) in MSG-SEVIRI data and tracking them in time, and additionally assigning a matching score for contrail-to-flight attribution is a valuable contribution.
However, because there is no 'ground-truth' contrail-to-flight attribution available, the reader would have at least expected a set of a few hundred or thousand cases, especially while the authors state that 'through batch processing multiple flights can be evaluated simultaneously'. This would allow for a thorough and robust verification of their framework. Furthermore, aggregate statistics can be calculated - i.e., contrail age, formation time, visibility, young/old contrail coverage etc - that can be compared to existing literature.
The framework could also be tested with the synthetic model-based reference dataset for GOES (Sarna et al., 2025).The manuscript provides sufficient detail in the methods section. However, sections could be improved.
For example Section 2.1, the authors introduce all the MSG satellites, which is not necessary, only name the one(s) used for the study. Precisely describe the region and the period used for the study. Next, the authors introduce which channels and combinations are relevant, and that the Ash RGB composite "is also used to highlight contrails". However, information is missing regarding exactly which channels and combinations of brightness-temperature differences are used. The data level (1, 1.5, ...) is also important to mention. The discussion about how complicated it is to observe contrails in geostationary satellite data could be dispensed with.
The reader might be interested in why Figure 1 shows 3 different channel combinations and which connection the subfigures have to the methodology. Lines 111 - 118 belong to Section 4.
Section 2.2 is a discussion about the contrail-appearance period in geostationary satellite data and could be shortened and moved to Section 3.
Section 2.4 is a mixture of introduction, discussion, and methodology. The authors might sort out the parts into the corresponding sections.
Surprisingly, the description of ACTA functionality and its modifications is outsourced to Appendix A, although it is an important part of the methodology. Lines 473 - 474 belong to Section 2.1. The thresholds could be mentioned explicitly in line 499.In Section 1, the authors describe how the application of wind-advection methods significantly degrades the attribution accuracy. Therefore, it is somewhat surprising to read in Section 3.2 that at cruising altitude a fixed wind speed of 50 km/h is assumed for horizontal advection. In dense air-traffic areas with contrail outbreaks and assuming a lower horizontal wind speed than 50 km/h this method would overestimate contrails near flight. And assuming higher wind speed at cruising altitude, the method misses contrails near flight. Can the authors comment on that?
Section 3.3 line 244-245: Clustering efficiently captures a part of the contrail life cycle, which is "complemented by ACTA ... particularly during early formation stages when they are not yet ... detected." Can the authors comment on that?
Section 3.3 line 257: Surprisingly, the description of the contrail-formation parameters is outsourced to Appendix B. Lines 530 to 545 belong to Section 4.
Section 3.4 line 277: Start the sentence on a new line - "The Proximity Score ..."
Section 3.4 line 287: Start the sentence on a new line - "The Advection Consistency Score ..."
Figure 4: It is understandable that an example with and without matching contrails is shown. However, the example with matching contrails makes the music and it is quite busy; why do the authors not neglect the example without contrail matches and, instead, show the two regions where the contrail matches appear as separate subfigures? It would also be interesting to see wind barbs at cruising altitude - perhaps ERA5 winds. This illustrates that ACTA chose the right direction for contrail advection.
Is Figure 5 based on the flight segment in Figure 4 (left)? If so, the authors should describe it. Insets A to E appear, but they are not mentioned in the main text or in the caption. Clustering - a colorbar would increase readability. Matching - the authors chose a continuous colorbar, however, a discrete colorbar (same color for bin 0 - 0.2, same color for bin 0.2 -0.4, and so on ...) would greatly increase visualization of the matching score.
Section 5 - The authors state in lines 440 - 442 that their framework eliminates the source of meteorological uncertainty. However, they should fairly mention that their approach introduces uncertainty, i.e., due to (i) digital image processing, (ii) assuming a fixed maximum wind speed, and/or (iii) relying on geometrical parameters for connecting contrails within time steps. Line 448, lines 451 - 452, and line 461 ("... inherently scalable ...") - these statements still need to be demonstrated. Lines 91 - 92 belong to this section.
Check acronyms and abbreviations for first appearance/redundancy. For example,
line 22: write out COVID; otherwise introduce COVID if it is used several times in the following text,
line 59: NWP not yet defined,
lines 80 - 81: MSG-SEVIRI already defined in the abstract,
line 99: SEVIRI already defined,
line 137: The acronym COCOS should be introduced here,
line 335: see line 22, etc pp.The authors have not yet made their code publicly available, which would allow interested users to reproduce the study and apply the toolbox.
Section References:
Check titles for capitalization, i.e., lines 576 - 577, 582, 587, 589, 598, etc pp.Recommendation:
With the improvements mentioned above, I recommend publication.References:
Meijer, V. R. et al., (2022) "Contrail coverage over the United States before and during the COVID-19 pandemic" Environmental Research Letters, 17(3):034039, DOI 10.1088/1748-9326/ac26f0Ortiz, I. et al., (2024) "Satellite-based quantification of contrail radiative forcing over Europe: A two-Week analysis of aviation-induced climate effects", Submitted to SESAR Innovation Days 2024, pre-print DOI: 10.48550/arXiv.2409.10166
Ortiz, I. et al., (2025) "Robust evaluation of neural networks trained on the OpenContrails dataset", IEEE Transactions on Geoscience and Remote Sensing, DOI: 10.1109/TGRS.2025.3629628
Heikenfeld, M. et al., (2019) "tobac 1.2: towards a flexible framework for tracking and analysis of clouds in diverse datasets", Geoscientific Model Development, 12 (11), 4551–4570, DOI: 10.5194/gmd-12-4551-2019
Seelig, T., Müller, F., and Tesche, M., (2025) "LIMOTTo (Leipzig Institute for Meteorology Object Tracking TOolbox)", Zenodo. DOI: 10.5281/ZENODO2347.15076642
Citation: https://doi.org/10.5194/egusphere-2026-3046-RC2
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 266 | 125 | 43 | 434 | 37 | 40 |
- HTML: 266
- PDF: 125
- XML: 43
- Total: 434
- BibTeX: 37
- EndNote: 40
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
This paper describes an automated contrail-to-flight attribution algorithm. A few such algorithms exist in the literature (Chevalier 2023, Geraedts 2024, Sarna 2025). The main innovation of this paper is to apply its algorithm to the MSG satellite, which introduces new challenges due to its relatively low resolution. This is valuable since Europe is one of the most contrail-dense regions in the world and prior to the availability of MTG in late 2024 MSG was the only way to detect contrails in this region by geostationary satellite. Automated contrail-to-flight attribution systems are challenging to develop, so developing one and reporting on its performance is a valuable contribution. The paper therefore should be published after some revisions.
I have one major criticism of this paper which is that they do comparably little analysis of how their algorithm actually performs. They make the statement “The method demonstrates high potential for automated, observation-based verification of contrail mitigation strategies in practice” but I don’t think they’ve done enough validation to justify this statement. Producing validation as thorough as that in Sarna 2025 might not be feasible, but the authors should do better than just the 4 case studies presented in this work. Even increasing the number of case studies to 10 would be an improvement. Another approach, similar to Chevalier/Geraedts, would be to run their system on a few hundred flights and check if aggregate statistics of things like contrail age, diurnal/seasonal patterns, fraction of flights making contrails, are plausible. Ideally upon resubmission the authors should include some more efforts towards validating their methods. Failing that they should not stress much more strongly that they do not have a good estimate of the model's performance.
I have a number of minor comments listed below:
From what little validation the authors do, their method doesn’t appear to do very well. When compared to manual labels, their method only agrees on one out of four contrails, implying a precision and recall of 0.25. This is far less than what is achieved by Chevalier 2023, Geraedts 2024 or Sarna 2025 (as evaluated in Sarna 2025) [it's a small sample size and not an apples-to-apples comparison, but still I think this is a bad sign]. This does not mean that the work does not merit publication. Even if the performance does not match the best algorithms already in the literature, sharing how different design decisions affect performance is useful to inform future algorithm development. Furthermore the poor performance could in part be because of the poor resolution of SEVIRI. But I think the authors should at least call out this disparity, the statement ““The method demonstrates high potential for automated, observation-based verification of contrail mitigation strategies in practice” does not seem justified. It would be good if people reading this paper who might be interested in using the algorithm should at least be cautioned that it might not give the same results as some other methods in the literature. (For example, the authors plan to use this algorithm to validate a trial. If such a study gives a negative result, I would not necessarily say this implies the trial did not work, it might just be that the algorithm is not good enough to see the impact)
The authors claim that their method is scalable to large numbers of flights. This claim is somewhat undercut by the fact that the authors ran the algorithm on a total of 4 flights. (c.f. Geraedts 2024, which also claims a scalable algorithm and runs on >250 000 flights). The authors should prove that the algorithm is scalable by running on at least 100 flights. If this is too big a burden, then the algorithm is not actually scalable.
The authors made a design decision to limit the maximum advection speed to 50 km/h. The authors claim this is a fine decision, but I don’t agree. They point out that 50 km/h is higher than the mean, which may be true, but that’s not the right metric. The authors should instead compute what fraction of the time the wind is higher than 50km/h. The authors later try a larger wind error. They find in total 26 matching contrails instead of 17. They claim this is a small change, I don’t agree. I think ideally the authors would recompute all their main results with a larger wind threshold. Failing that they should at least change their claim that 50 km/h is sufficient. I would like to see increasing this threshold as a possible direction for future work.
Another design decision the authors make is to not advect the flights. They claim this is a strength of the paper, I think instead it is a weakness that likely contributes to the relatively poor performance of their method compared to existing literature. The authors correctly point out that wind data is uncertain. But having no data at all about the wind is more uncertain! For example the authors consider all contrails within 50 km/h of the flight (which I claim is not big enough). The wind uncertainty is only 20km/h, so if they used the advected flight instead they would need to consider a lot less flights. They have a similar issue with angle. In the introduction they authors claim that sedimentation + wind shear is a reason not to use advection, but then in their algorithm they assume a constant advection speed, which would not be a reasonable assumption of sedimentation + wind shear was really a problem. I think the authors' claims that not advecting is a virtue need to be toned down, and I think they should say that adding advection to their algorithm is a possible direction for future work.
The authors method does not consider multiple flights that might match the same contrail, and the authors correctly highlight this as a drawback of their work as well as discuss why it is hard. A reader might be left with the impression that such consideration is therefore infeasible, but in fact Chevalier, Geraedts and Sarna all do it. The authors should inform the reader of this.
In some of the case study figures I find the label ‘no decision’ confusing. I think the authors mean that these are flights that were within the max velocity of the flight but got a score of 0. Please be more explicit about these.