The manuscript investigates aerosol retrievals performed on normalized sky radiances obtained from an upward looking all-sky camera. Retrievals are first simulated using synthetic data corresponding to a variety of aerosol types to estimate uncertainties under known conditions. The retrieval is then applied to real data obtained from a sky camera located in Valladolid, Spain and the resulting retrieved parameters are compared with their counterparts from AERONET inversions. It is found that the method has good sensitivity to AOD in scenes with low to moderate aerosol loading and it can characterize some aspects of the particle size distribution, especially in the fine mode.

The study demonstrates that upward looking cameras have the potential to retrieve useful aerosol information even without an absolute calibration. This approach has the potential to provide a low cost means of significantly expanding ground-based remote sensing coverage and the topic is certainly relevant for AMT. There are a few points that I feel need further clarification though, particularly some details regarding the preparation of the measurements used as input to the retrieval. Additionally, I would recommend another round of proof reading for English language usage.

––––––––––––––––––––

Detailed Comments

––––––––––––––––––––

1)    LN 60: Since the comparisons are with retrieved values the term "retrievals" is probably more appropriate than "real measurements".

2)    LN 91: "geometries" should be "scan" since each of the two scans contains observations at multiple geometries.

3)    LN 104: It would be helpful to specify the field of view of this fisheye lens.

4)    LN 106: Given that stock Bayer filters are typically quite broad, will there still be significant spectral crosstalk (i.e., band overlap) between the three channels? It would be good to provide some description of the magnitude of this effect as well as a definition of exactly how the "effective wavelengths" are calculated. I would also be interested to hear the authors thoughts on how this could potentially impact the retrievals (perhaps at a later point in the manuscript).

5)    LN 122: I believe the measurements made here cover to most of the sky. Presumably using observations at so many viewing angles would place a high computational burden on the retrieval but what was the motivation for subsampling these observations to the AERONET hybrid scan viewing geometry in particular?

6)    LN 126: This is a little ambiguous in my view. Is the spectral dependence accurately captured here as well, or just the angular dependence? It would be good to include an equation and/or make the description of the normalization procedure more precise.

7)    LN 128: I am struggling to track the remainder of this paragraph. Which uncertainties are used for the 5% screening criteria? If all observations with uncertainties greater than 5% are removed in QA, how is the uncertainty of the remaining quality screened 605nm data still above 5%?

8)    LN145: Were any corrections applied to account for gas absorption or was it assumed negligible?

9)    LN 150: How is the aerosol's vertical distribution modeled by GRASP in this study?

10)    LN 152: Could other points in the FOV be used to provide more measurements if <6 data points are available within the hybrid scan geometry?

11)    LN 156: A more precise explanation of "residual higher than the uncertainty" should be provided. The exact expression used to calculate the residual and the corresponding threshold value would be helpful.

12)    LN 165: Is it possible to provide some metric conveying the strength of these temporal smoothness constraints? (e.g., typical level of autocorrelation)

13)    LN 179: I'm not following what is meant by "using the Valladolid coordinates as a reference" and how the analysis went from seven to nine aerosol scenarios. Were two additional scenarios added corresponding to Valladolid aerosols? If so, how were these two scenarios derived?

14)    Figure 4/6: These figures might be a bit easier to follow if all panels on each row had the same y-axis scaling.

15)    LN 240: It is shown that the normalized sky radiances do not change significantly with AOD at high aerosol loads. Said another way, there are many possible aerosol concentrations that produce the observed normalized radiances so, while the retrieval lacks sensitivity to AOD in these cases, it still should be able to fit the measurements relatively well. Thus, I am wondering why the number of converged cases (N) decreases to near zero at the higher AODs?

16)    Figure 6: The authors might consider including a plot of normalized (instead of absolute) size distributions so that the retrieval errors are not dominated by uncertainties in the total concentration, especially at high AODs.

17)    LN 265: How exactly is the lognormal's σ defined here? I generally take σ to be a unitless quantity.

18)    LN 295: It is a little surprising to me that more real retrievals pass the convergence test (>80%) than simulated retrievals (<65%, even for the lowest AODs). Do the authors have any thoughts as to why this might be?

19)    LN 302: My understanding is that GRASP radiative transfer assumes plane-parallel geometry. Are errors in GRASP's forward modeling a significant concern at such high SZAs?

20)    LN 368: Since the sentence is referring to the bias (not the spread) of the points "Most of the differences" should probably be replaced with a term that means the center of the distribution (i.e., "mode").

21)    LN 451: I found the first few sentences of this paragraph to be very confusing. I recommend rewording.

22)    LN 518: Why is AERONET able to retrieve absorption with hybrid geometry but GRASP-CAM has no sensitivity? Is it due the lack of absolute calibration in the GRASP-CAM data? If the authors have an intuitive explanation of this phenomenon, it would be helpful to include it.

––––––––––––––––––––

Technical Comments

––––––––––––––––––––

LN 25: "makes" should be "means"

LN 32: "using all" should read "all using"

LN 34: It would be clearer to say "...capable of measuring..."

LN 74: While the acronym is common, "m.a.s.l" should be defined within the text.

LN 144: "AWRONET" needs correcting.

LN 389: I think this should reference Figure 13, not 12.

LN 413: Should say "...than the ones obtained..."

General comment

The manuscript covers aerosol retrievals using normalized radiances from an all-sky camera (simulated and measured) and the GRASP algorithm. The study has been very thoroughly performed and described, and presents interesting and relevant results regarding generalizing all-sky cameras (or normalized radiances) towards aerosol retrievals. The subjects also fits the scope of the journal and I recommend its publication, with some very minor comments. 

Specific comments

P4, L117. Is there a reference, where the hybrid geometry is described in more detail?

P7, L192pp. I think the explanation of NSR sensitivity to AOD with dominance of aerosol scattering (Rayleigh vs. aerosol) is not very convincing (wouldn’t this AOD sensitivity be the same in an atmosphere without Rayleigh?). Maybe it is because the normalization factor (sum over all radiances) increases with AOD and reduces the relative differences in NSR.

Fig 2. What is the reason for the “kinks” in the radiances at around 40° scattering angle? If I am not mistaken, these features do not appear in almucantar scans (radiance plotted against azimuth not scattering angle).

P9, L218. So the noise in the synthetic data is the same as stated on p5, L128? Also, what is distribution of the random noise? I would guess normal, but it should better be stated.

P11, L258. It works surprisingly well for SZA=30°, considering AERONET limit of 50° (is that correct?) with almucantar measurements only. So the hybrid scan seems effective at small SZA.

Would this mean that really most of the information is in the small scatting angles, <100°?

Fig. 5. I am wondering if the higher deviations of the retrievals towards higher AOD, could also be due to (or at least affected by) the lower number of successful retrievals, i.e. a statistical effect. Of course, the retrieval success decreases due to decreasing sensitivity to AOD.

Since the STD is really interesting in relation with the Md (e.g. is ΔAOD=0 within the errorbar also for high AOD?), I believe it would be a good alternative to combine right and left columns of the plots and plot STD as error bars (or uncertainty bands) for Md. Maybe with a second uncertainty of the mean MD/sqrt(N), N successful retrievals. This will be too much for all scenarios in one plot, so this would need a different grid of plots. Might be worth a try.

Fig. 6 As a general question: If red is the median of noise retrievals, I would expect red and blue line to be very similar (as the retrieval input, the median of noised radiances should be the radiance without noise). This is almost always true except for the coarse mode of MALD. Any idea why?

Fig.7. Again, as for Fig.5, I think it would better if Md and STD are plotted together.

Fig. A33. There is actually quite little one can read from this plot of the complete time series. On this time scale, it would be easier to visualize aggregate values, e.g. monthly means.

Finally, the manuscript can in general be improved with respect to English grammar / language. For example, missing articles: P1, L1. the GRASP code. P1,L8. As a result.