The authors present a manually labeled contrails dataset derived from Meteosat Second Generation SEVIRI imagery. The manuscript describes the labelers as following established best practices from the literature for their labeling campaign such as having multiple labelers annotate each scene and providing temporal context to the labelers. Additionally the authors set a nice precedent, being the first group to additionally collate and provide auxiliary data (CiPS and ProPs, Ash color scheme, reflectances/BTs, land-sea mask, surface altitude) for each scene that is useful for further scientific/ML analysis. The manuscript is well written, clear and comprehensive in its documentation of the dataset. The dataset is of high quality and fills an unmet need of contrail labels in geostationary imagery over the MSG domain, and represents a substantial amount of expert manual labor that would be nontrivial to reproduce. I had no trouble reading the article nor downloading and using the dataset based on it.

There are however two major technical corrections needed in the data that must be resolved before acceptance:

1. CiPS fields are missing from at least 8 scenes (~6% of scenes): `scene_025.nc`, `scene_026.nc`, `scene_027.nc`, `scene_041.nc`, `scene_096.nc`, `scene_097.nc`, `scene_098.nc`, `scene_099.nc`.
2. Rotation/flipping appears to be afflicting individual SEVIRI channels provided with the scenes, somewhat randomly but upon investigating a sampling it appears to be affecting what looks like possibly more than half of scenes, especially in the reflectances/NIR/WV_062. Perhaps these auxiliary data were generated by an ML dataset pipeline with rotation/flipping training augmentations accidentally left turned on?

Minor issues to also address please:

Data:

1. Providing the data in units of 'pixels' is convenient in some regards but can make it challenging for users of the dataset to report things like contrail lengths/widths/flux in standard SI units like kilometers or Watts/meter^2. Consider providing an auxiliary variable for each scene that provides the area of each pixel in the scenes, or code to calculate the same. For example Appendix B of https://egusphere.copernicus.org/preprints/2025/egusphere-2025-3739/ details a method to do so using an approximation of each pixel as a parallelogram.
2. In the scenes missing CiPS fields, the variables actually contain -32767.0 despite the NetCDF metadata declaring the _FillValue as NaN
3. Appendix A lists variable names as for example IR108 but files use for example IR_108

Manuscript:

1. Because the dataset spans 11 years it could theoretically be susceptible to long-term calibration inhomogeneities ('sensor drift'), which in scientific analysis might erroneously show trends over time that were in fact entirely due to sensor aging. I was not able to find any statistically significant evidence of sensor drift in the dataset. Nevertheless, authors are advised to please see section 2.1 for details on the inter-calibration of SEVIRI (https://essd.copernicus.org/articles/15/5153/2023/essd-15-5153-2023.html) and affirm explicitly in the manuscript whether the shortwave auxiliary variables were derived from inter-calibrated reprocessed data ("Climate Data Record") or from archived realtime-calibrated data. I believe this applies to just the reflectances, since CiPS and ProPs are longwave only and in SEVIRI longwave channels are generally less susceptible to sensor drift.
2. There are a few places discussing the subjectivity of labels based on labeler (dis)agreement, where the statements appear to apply to all contrail labeling from satellite imagery; please be sure to avoid misleading readers by scoping these statements to being about the specific imager. Consider reviewing https://amt.copernicus.org/articles/18/1115/2025/ which recently analyzed the effect of pixel size on detectability of contrails, and reports that detectable contrails contribute (70±2) % of the net radiative forcing in simulated 2 km resolution images.
3. Line 20: the cited studies are simulation studies, not "Observational studies"
4. Line 34: starting the sentence with "For instance" implies that Ng et al 2024 is not a publicly available dataset, but it is actually publicly available.
5. Line 88: is ECMWF really at 6 hour timesteps? I'm familiar with ERA5 that is either 1hr (nominal) or 3hr (ensemble); may help to cite the specific ECMWF product used
6. Line 130: citation (and ideally quantification) needed for the assertion that contrails are "mostly identified as thin ice clouds" - many geostationary satellite cloud products can misclassify thin cirrus as clear sky (or water cloud if they are over a lower cloud)
7. Table 1: are these pixel-area-weighted frequencies or just pixel count frequencies?
8. Line 164: "object-wise agreement" is first mentioned here but the algorithm for converting contrail pixel masks to linear objects is not described until section 4.2.1
9. Section 4.1: please explicitly mention that while other labeling campaigns (such as Ng et al 2024) have provided labelers with some rasterizations of advected flight trajectories to inform their labels, they were not provided to labelers here.
10. Conclusion: a comment would be welcome about the size of the dataset and if the authors expect it to be useful for training deep ML models (or more likely for simpler ML models or validation/qualitative insights only?)

This paper describes the development and characteristics of a dataset of manually labeled contrails in images captured by the MSG SEVIRI instrument.

The paper is extremely well written: I did not find any typographical errors. It reads well and I have very few line-by-line comments, which mostly pertain to the figures.

The study contributes to our understanding of the limitations in manually labeled datasets of contrails in satellite images, which is crucial given that these datasets are typically used to create and evaluate automated contrail detection techniques.

I also have some more general questions:

• The period over which satellite data is acquired is relatively long: are there any relevant issues related to sensor degradation over this time period, that could affect the creation of this dataset? And what is the rationale behind the decision to label images in the period 2013 – 2018, but then also between 2023 and 2024?
• Were the images projected in some way, to address image distortion at higher satellite viewing zenith angles? Were the locations of images that were labeled in some way limited by the viewing zenith angle?
• Why was the decision made to only consider images from the hour leading up to labeled image? And was this found to be sufficient, or would a longer period have been helpful? For example, (Meijer et al., 2022) uses 2 hours.
• Were the labelers trained in some way? The paper mentions a labeling guide: it would be very helpful to include this with the paper, like was done by (Meijer et al., 2022; Ng et al., 2024). Additionally, I believe Ng et al. (2024) implemented a “training process” that had labelers learn on a “gold-standard dataset” before “graduating” to labeling new images.
• How was labeler agreement affected by the proportion of land/sea pixels in an image? How was labeler agreement affected by the number of contrail pixels present in the image?
• How does the size of this dataset compare to previous labeled datasets (Meijer et al. 2022, Ng et al. 2024)? It would be interesting to compare the number of labeled pixels (accounting also for the lower resolution of the SEVIRI instrument) and the fraction of contrail pixels within these datasets. The distribution of contrail pixels among images might be another interesting feature to compare.

Another general comment is that for certain words, like “sun”, the capitalization is inconsistent across the paper. It might be worthwhile to do a check throughout the manuscript.

Once my questions and line-by-line comments are addressed, I consider this paper ready for publication.

Line-by-line comments

Figure 1: It might also be helpful to include the original satellite image (like figure 5a or 5e) next to this for interpretation.

Line 101: Why does this say: (e.g. 2021)?

Figure 2: The legend is cut-off. It might also be helpful to include the original satellite image (like figure 5a or 5e) next to this for interpretation.

Figure 6: The resolution of this figure is a bit too low to inspect the ash RGB images. I think it may also be helpful to combine the 3 different labels in a way that highlights their agreement and disagreement, as I now have to switch back and forth between the different labels to identify differences.

Line 178: “the own image assessment” what is meant with this?

Table 3: It is not clear to me how the different columns in this table are computed. There are 6 columns, but 14 labelers in total?

Figure 7: perhaps include horizontal axis labels for plots b) and c).

Figure 8c): abbreviation “PC” not defined.

Line 370: “as global attribute” -> “as a global attribute”

Table A1:

• “timestamp” is in UTC?
• row “native_coordinates”: “boundary box” -> “bounding box”

Table A2:

• First row: “Consensus” should this be capitalized?
• “acquisition_time”: First letter in “Description” not capitalized

References

Meijer, V.R., Kulik, L., Eastham, S.D., Allroggen, F., Speth, R.L., Karaman, S., Barrett, S.R.H., 2022. Contrail coverage over the United States before and during the COVID-19 pandemic. Environ. Res. Lett. 17, 034039. https://doi.org/10.1088/1748-9326/ac26f0

Ng, J.Y.-H., McCloskey, K., Cui, J., Meijer, V.R., Brand, E., Sarna, A., Goyal, N., Van Arsdale, C., Geraedts, S., 2024. Contrail Detection on GOES-16 ABI With the OpenContrails Dataset. IEEE Trans. Geosci. Remote Sens. 62, 1–14. https://doi.org/10.1109/TGRS.2023.3345226

Gabriel et al. present a manually labeled dataset of aircraft-induced contrails derived from MSG/SEVIRI imagery. Given the radiative importance of contrails, particularly in the context of quantifying aviation-related anthropogenic climate forcing, this dataset represents a valuable contribution. The provision of a carefully constructed ground truth is especially relevant for the development and evaluation of machine learning–based contrail detection algorithms.

The manuscript is clearly written, and the methodology for contrail identification and labeling is well described. The inclusion of auxiliary variables improve the usefulness of the dataset. The data structure and metadata are generally well documented in both the text and the accompanying tables. I have only a few minor comments that should be addressed before the manuscript can be considered for publication in ESSD:

Line 80: It may be worthwhile sharing why the considered dataset has a gap of 5 years between 2018 and 2023.

Line 122: The manuscript states that approximately 40% of scenes contain no visible contrails and 60% were selected to include contrails. Was this segregation intentional? The wording suggests that it was. If so, this should be clarified, including the rationale for this choice.

Line 124: The manuscript notes that the selected 140 images are not uniformly distributed in space and time. This suggests that the dataset is not intended to represent real-world contrail climatology. In addition, the relatively small sample size indicates that the primary application of the dataset is likely to be the development and validation of contrail detection algorithms (e.g., ML-based approaches), rather than for statistical analyses of contrail occurrence or detailed studies of meteorological drivers. I suggest clearly highlighting these points in the manuscript.

Table 3 is not discussed in the manuscript text. Also, it is not clear what the columns represent.

Figure 8 presents contrail length and width distributions in pixel units. Since pixel size varies with viewing geometry, it would be more physically meaningful to express these quantities in distance units (e.g., km).

Line 255: It is worth noting here that the auxiliary parameters related to cloud optical properties are only available during daytime.