General comments

This study addresses a technical approach that potentially renders star-photometry cheaper to manufacture and to operate unattended. It implies a comprehensive and complex analysis, and the authors should be commended for completing it. The results are meaningful and encouraging. The paper is well written, and the figures are of good quality. It represents an important development, and I recommend it for publication!

Specific comments

Being the first such development properly analysed, and given the scarcity of operational star-photometers, the retried optical depth had to be compared with that of moon-photometers. Their accuracy remains however questionable, partly due to the ROLO model accuracy, and being prone to forward scattering in the presence of cirrus clouds or PSCs, that are very difficult to screen out with standard algorithms. I would advise therefore pursuing the analysis by also comparing the results with the available star-photometers in Lindenberg and Ny-Alesund. They can identify such clouds with a spectral screening (O’Neill 2016, doi: 10.5194/acp-16-12753-2016).

Also for a future analysis, instead of using effective wavelength and effective optical depth, one may take advantage of a modified The Beer-Lambert-Bouguer law, specifically developed for wide filters, as in Rufener 1986 (http://adsabs. harvard.edu/abs/1986A&A...165..275R), as it’s the case for most of the all-sky cameras used here.

The several subsequent screen algorithms to remove outliers may remove legitimate data points and underestimate the measurement uncertainty. For example, using several sigma as a criteria to remove outliers may not be appropriate in a non-gaussian, or non-symmetrical distribution. In this sense, particularly problematic may be the 1% constraint in the Langley fit. You may want to comment on this.

From my calculations, even with longer integration times, this camera should feature about 10 times more scintillation noise than the Lindenberg star-photometer. Therefore, the variability of the individual measurements in your Fig 8 may be due to this effect. Your filter may then simply select stars near zenith, that are less affected by scintillation. You may want to check and comment on this.

Technical questions 

The questions start with the number of the line where is referring to!

16 - If the precision uncertainty is 0.03-0.04, the 0.02 accuracy uncertainty may not be that relevant. In addition, I think the moon photometer accuracy may have larger uncertainties than 0.02 anyway, as it’s not a perfect reference to compare to anyway.

36 – “This methodology is followed by AERONET” – I would rather say “used” instead of “followed”, as the latter suggests that AERONET follows a method invented by (Toledano et al., 2018), which it’s not.

136 – “without demosaicing neither white-balance correction” - I don’t understand what this means.

165 – “long time exposures, star scintillation caused by atmospheric turbulence” cannot spread the light on several pixels. “Star scintillation” only varies the amplitude of the star irradiance, while the “atmospheric turbulence” can only spread the light to ~<20 arcsec, i.e. way less than the camera pixels (5.4 arcmin ~= 324 arcsec). Therefore, a “long time exposure” should not spread the light over several pixels, unless other effects may contribute, like instrument vibrations etc. The main reason should be “the camera’s point spread function”.

323 – Using “log” symbol for natural logarithm may be used only when the use of natural logarithm is implicit. Since for star magnitudes is usually used log in base 10, the natural base is therefore not implicit. I would recommend using “ln” for natural logarithm to avoid confusion. 

In addition, the “monochromatic” equation (3) may be used borderline to the camera OMEA-3C-TF, as the filters have usually smaller bandwidth than 40 nm (as specified in section 6.1 of Ivanescu et al, 2021, https://doi.org/10.5194/amt-14-6561- 2021). However, for the OMEA-3C, one certainly needs to consider a wide band-equation, like the equation (2) of Rufener 1986 (http://adsabs. harvard.edu/abs/1986A&A...165..275R). This was especially developed for such filters. Your “effective wavelength” solution seems an oversimplification. While it is not necessary to change your study to accommodate this formula, I think it’s necessary however to add a comment concerning this aspect. 

357 – “Table 2” – sigma/sqrt(N) gives uncertainties that propagates into OD of about 0.02 to 0.04. Since this is for the entire ~3 year period, for one year should lead (multiplied by sqrt(3))) to 0.03-0.07 uncertainty only due to log(USI⁰). Why doesn’t one see this in the final claimed 0.02 accuracy or 0.02-0.04 precision?

397 – “longer values”?

446 – The climate values for ozone and especially water vapor may be much further away from the actual values, than the value of the aerosol optical depth, leading to high uncertainties. How to you account for such uncertainties? On the other hand, TWVC is not necessary for OMEA-3C-TF, as all the water absorption bands are outside of its filters.

466 – “attributed to inaccuracies in the log(USI⁰)” – not sure why mentioning only this one in particular.

565 – “calibration values log(USI⁰) […] may not be optimal for all stations”. This should not depend on site location or local environmental conditions. USI⁰ is characteristic to a star and a camera. Some stars are variable stars. Also, the camera contribution may very in time due to optical and electronic throughput changes.

576 – “precisionranges” must be “precision ranges”

577 – “The lowest SD values are found for camera C013, which is the only one with a triband filter”, again, the best result obtained with this camera (including smaller y0 and b~=1) may be linked to its quasi-monochromatic filters.

584 – “The standard uncertainty is defined as the sum of the errors of the camera and photometer” – this may be true only for bias. For random errors one should add them quadratically.

601 – “Andoya is a less suitable site for Langley calibration due to its higher latitude” – this is true only for some stars. Andoya should not be that far North (like Ny-Alesund is) in order to have air mass coverage issue for most the stars. Even for Ny-Alesund, probably half of the stars still cover the 2-5 airmass range in one night. Beyond this, the filtering algorithm discards those stars that don’t cover the required range. Then, why the higher latitude may not be good for Langley calibration? The unstable atmosphere? Why should a higher latitude have an unstable atmosphere? Do you have a reference on that?

604 – “estimating AOD appears to depend on the location where they are installed” – Why? Couldn’t be because of differences in available data?

630 – “the camera AOD at 466 nm shows lower values compared to the photometer”- this may be explained by the forward scattering error due to the larger camera FOV (see section 6.3 of Ivanescu et al, 2021, https://doi.org/10.5194/amt-14-6561- 2021). The forward scattering brings more light into the camera and the OD appears smaller. This should be more evident in the blue/UV, where the aerosol scattering is higher.

Global comment:

Strengths:

• Innovation: The study proposes all-sky cameras for star photometry, addressing the nighttime AOD data gap.
• Cost-Effectiveness: Using commercial imaging devices could make aerosol monitoring more accessible if proven robust.

Analysis:

• Assumptions: Star intensity stability and uniform aerosol distribution are supposed, but intrinsic variability and atmospheric inhomogeneity could introduce biases.
• Comparative Advantage: The study lacks discussion on trade-offs in spectral response, sensitivity, and precision compared to dedicated photometers.

Methodological and Calibration Challenges

Strengths:

• Data Extraction: Extracting starlight signals from wide-field images is technically innovative and promising for remote sensing.

Analysis:

• Calibration: All-sky cameras are not designed for absolute radiometry, requiring robust calibration strategies.
• Data Processing: Background subtraction, optical corrections, and noise propagation need clearer methodological details.
• Spectral Considerations: The study should address how it handles spectral mismatches affecting AOD retrieval.

Data Validation & Comparative Analysis

Strengths:

• Preliminary Validation: Initial consistency with existing measurements suggests potential feasibility.

Analysis:

• Systematic Biases: Possible AOD overestimation needs deeper investigation.
• Comparative Studies: A broader validation campaign against reference instruments across various aerosol regimes is necessary.
• Temporal & Spatial Variability: The method's ability to resolve short-term and spatial variations requires further evaluation.

Radiative Transfer Modeling & Algorithmic Considerations

Strengths:

• Model Integration: Radiative transfer models provide a strong theoretical basis.

Analysis:

• Model Assumptions: Question : Should the impact of multiple scattering and horizon effects be analyzed?
• Algorithm Robustness: Star identification and retrieval parameter sensitivity require systematic validation.

Operational Considerations & Generalization

Strengths:

• Scalability: The approach could enhance nighttime aerosol monitoring at low cost!

Analysis:

• Environmental Dependence: The influence of clouds, light pollution, and geographic variability should be better addressed.
• Hardware Variability: Standardized calibration across different camera models and long-term stability assessments are necessary.

Future Directions & Recommendations

• Improve calibration for sensor aging and optical distortions.
• Quantify uncertainties using advanced statistical methods.
• Enhance star detection and background subtraction algorithms.
• Participate or organize more validation campaigns under diverse conditions.

Conclusion

This study presents a very promising, cost-effective method for nighttime aerosol monitoring.

Refining the calibration, the validation, and maybe the error analysis will be crucial to ensuring its reliability and broader adoption in atmospheric research.

General comment:

The manuscript is about a relevant topic in aerosol remote sensing, specifically exploring the method of star photometry with all-sky cameras. The method yields interesting results, the text is well written, and so from my point of view, it is suitable for publication in AMT. Below, I have added some comments, and I leave it to the authors to decide if/ how to consider them for the manuscript. 

Specific comments:

L272, Eqn. 2. It's probably pretty standard, so it could be omitted in case the manuscript needed to be streamlined.

L382pp. While you can of course use a RT model, for the direct irradiance it would be a bit of an overkill, when you can use a simple equation (as in the Beer-Lambert law)?. 

By the way, also try to avoid the common confusion with libRadtran: it is not a “model” but a software package that includes multiple models (or solvers) like DISORT and MYSTIC (Monte Carlo code).

L417pp. Again, is it probably not necessary to use a RT model here? Basically it's about the cross sections that you use, no? Also, while Fig. 7 is of course relevant in general, it is maybe more appropriate in a textbook about remote sensing. To me it seems a bit distracting in a paper focusing on a specific implementation of star photometry with a camera. 

Fig. 9. I could think that there are more interesting or representative cases to be shown?

Fig. 10. I found myself a couple of times trying to relate camera number and location (for simplicity sometimes only number or location is used in the discussion). I know, the figure is quite busy already with text and numbers, but maybe there is a way to include the locations or generally have a more efficient labeling?

L563pp. So how exactly would an error in the Langley show in the correlation scatter plot? Considering the airmass dependence of the AOD error it would cause some scatter, but only in one direction (above or below the 1:1 line)? If yes, would it be radiometric instability or sth else that causes the low correlation?