General comments

This is a well-written paper that provides an extensive introduction to and description of a large, novel observational dataset of contrails. I have mostly minor comments, related to small details in the writing and presentation of the results. These can be found under “line-by-line comments”.

There are some open questions that remain after having read the manuscript, that I think should be addressed before the paper is published.

• How were the to-be-labeled video sequences selected? Given a particular sequence start, how was its length determined? How do these choices affect the representability of the dataset, and the generalization performance of algorithms trained with it?
• Do you expect that the video-based models would perform better as the length of the clips used for training them is increased? If I interpret the paper correctly, the maximum length of video clips that are used is 5 frames (so 2.5 minutes). This seems to be quite short, especially when placed in comparison to the length of temporal contexts used when analyzing geostationary satellite imagery (e.g. Ng et al. 2023).
• It is found that the image-based models show better instance-segmentation performance than the video-based models, but that the latter also provides “tracking” output. What is the quality of these “tracking” outputs and is it sufficiently better than what could be obtained from the “better instance segmentation” output in combination with a simple tracking algorithm (e.g. based on overlap between image frames).
• Could more examples of how the image pre-processing technique helps identify contrails, be given? Currently, only 1 such result is included, but more examples that show the value of the re-projection and contrast enhancement would be helpful in my opinion.

Line-by-line comments

P1L6: “don’t” -> “do not”, “aspect” -> “aspects”

P1L21: “trap outgoing” -> “reduce outgoing”

P1L22: “as due to aviation CO2 emissions”

P2L30: “context-dependent and extremely difficult to model reliably” -> “highly variable and challenging to model.”

P2L33: APCEMM stands for “Aircraft Plume Chemistry, Emissions, and Microphysics Model”

P2L53: I think it would be nice for the authors to explain in more detail why the attribution of observed contrails to flights is difficult when using geostationary imagery, and how the introduced dataset does not suffer from this problem.

P2L56: I think I understand what this paragraph is trying to say, but I feel like it could be made clearer. Perhaps state this as, “existing datasets of contrails annotated in observational data such as the OpenContrails dataset, do not track individual contrails over time or provide information on the flights that formed them”?

P3L70: is there no reference for this Mask2Former approach?

P4L106: Although named after Schmidt and Appleman, the criterion in its current form was originally presented by (Schumann, 1996) so I would suggest also citing that paper here.

P4L109: “trap” the more physically correct term is “reduce”.

P4L111: “The precise relative impact” what exactly is meant here with “relative impact”?

P4L117: “Atmospheric imagery” what is meant with this term?

P5L131: “coverage” -> “spatial coverage”

P8L198: the labeling of these images was done at the “semantic segmentation” level, not at the instance segmentation level.

P8L205: “Thanks to the 10-minute temporal…” This sentence does not make much sense to me. Perhaps it could be left out, as the rest of the paragraph is clear.

P8L209: “A 2025 update” Technically, the instance-level labels were always there, but not released as part of the 2023 Kaggle competition that utilized the OpenContrails dataset. So I would suggest to leave out this sentence.

P8L210: The authors could also cite (Pertino et al., 2024) here.

P8L219: “of dataset” –> “of a dataset”. The authors should note that earlier studies have collocated contrails between different remote sensing instruments. Examples are (Iwabuchi et al., 2012; Mannstein et al., 2010; Vazquez-Navarro et al., 2010).

P8L220: “Altitude” appears twice in this sentence.

P9L244: “Non-data-driven image-analysis techniques” -> “Traditional image analysis techniques”

P9L254: Citation “Jarry et al.” missing a year. Happens elsewhere as well.

P10L266: “Leveraging as well on Hough-based line detection” This sentence doesn’t make sense to me.

P10L282: provide exact coordinates of camera location, if possible.

P12L315: “or reviewing” -> “of reviewing”

P14L371: “U-net” written differently than elsewhere in the paper

P14L387: “k-means” maybe write “k” in italic?

P23 Figure 5: Include whether time is “UTC” or not. Same goes for Figure 7. Additionally, “Raw image” is used for an image that has already undergone quite some processing, so perhaps a different terminology could be used here?

P24 Figure 6: I think it would be helpful to combine figures 5 and 6 to make it easier for the reader to perform the visual comparisons.

P24L594: “erroneously merges contrails 5 and 6 into a single prediction” I don’t see this in the results at all? Is this potentially a typo? Should it be contrails 3 and 6?

P24L595: “merges contrails 5 and 6” Again, I don’t see this.

P26L638: “we are deployed” -> “we are deploying” ?

P26L656: “Integrating these tasks …” This sentence doesn’t make sense to me.

References

Iwabuchi, H., Yang, P., Liou, K., Minnis, P., 2012. Physical and optical properties of persistent contrails: Climatology and interpretation. J. Geophys. Res. Atmospheres 117.

Mannstein, H., Brömser, A., Bugliaro, L., 2010. Ground-based observations for the validation of contrails and cirrus detection in satellite imagery. Atmospheric Meas. Tech. 3, 655–669. https://doi.org/10.5194/amt-3-655-2010

Pertino, P., Pavarino, L., Lomolino, S., Miotto, E., Cambrin, D.R., Garza, P., Ogliari, E., 2024. Ground-Based Contrail Detection by Means of Computer Vision Models: A Comparison Between Visible and Infrared Images, in: 2024 IEEE 8th Forum on Research and Technologies for Society and Industry Innovation (RTSI). IEEE, pp. 254–259.

Schumann, U., 1996. On conditions for contrail formation from aircraft exhausts. Meteorol. Z. 4–23.

Vazquez-Navarro, M., Mannstein, H., Mayer, B., 2010. An automatic contrail tracking algorithm. Atmospheric Meas. Tech. 3, 1089–1101. https://doi.org/10.5194/amt-3-1089-2010