General Comments:

This work presents a remote sensing technique that utilizes multi-angle polarization information to retrieve cloud fraction and cloud optical depth. Numerical experiments using 1D radiative transfer are presented to demonstrate the technique along with an extremely welcome field validation. The paper is well suited to publication in AMT. The technique is an important contribution as it provides a means to mitigate the weakness of coarse resolution that upcoming space-borne multi-angle polarimeters will have to deal with.

There are, however, several aspects of the paper that should be expanded before publication to give a more complete accounting of the technique. These include a discussion of the applicability to ice phase clouds, the definition of cloud fraction, and the issues inherent to the application of the retrieval to clouds with 3D geometries. The 3D nature of the cloudy atmosphere is an inextricable component of multi-angle retrievals such as this one and should be addressed within this paper, even if only with a combination of qualitative and idealized quantitative arguments. I therefore recommend major revisions to this paper to address the specific comments listed below.

Specific Comments:

1. Exactly which of the instruments listed in the introduction is this technique applicable to? It seems to me that for general solar zenith angles, only hyperangular measurements will be able to provide the simultaneous scattering angles required, and HARP2 is focused on for what appears to be this reason. If this is the case then this should be made clear when the polarimeters are introduced in Lines 51 to 65. Will there be limitations in the regional coverage of this technique applied to HARP2 because of this due to scattering geometry requirements?
2. I also wonder whether the choice of wavelength is based on HARP2’s hyperangular channel or whether there would be some other more optimal wavelength for this technique.
3. The technique is stated to retrieve cloud fraction. I am left wondering exactly how this cloud fraction is defined. The definition will tend to depend on the purpose of the product, which is not made clear. If the purpose is to retrieve the geophysical variable of cloud fraction for model evaluation, then that is typically defined based on the projected area of clouds onto a space oblique plane. On the other hand, if the point is to identify the presence of clouds within pixels in an image, then that is a completely different target for the retrieval when non-nadir views are considered, as they are here. The definition of the target/truth will determine how the output of the retrieval is validated and also guide how a potential product is used and so this should be made clear. Perhaps the goal is to produce a resolution-independent estimate of cloud fraction due to the potential to provide a continuous output? This would certainly be an important contribution as most moderate-resolution satellite cloud masking products are not necessarily designed to explicitly estimate cloud fraction but rather to mask pixels for a variety of purposes.
4. A clearer target and application for the retrieval may help to focus the introduction which, in its present form, is very close to just a listing of different retrieval methods and instruments. Ideally, this section should present a problem to which the proposed technique is a solution. For example, how bad is the problem of unresolved clouds at 2-4 km resolution vs. 1 km? Why can’t we just tune threshold algorithms to retrieval pixel-by-pixel cloud fraction at coarse resolution? Just a few of many relevant references: (Stubenrauch et al., 2024; Dutta et al., 2020, Wielicki, B. A., and  L. Parker, 1992)
5. If the choice of target is a pixel-by-pixel cloud fraction, then which pixel is it? Is it the pixel observing at 140 degrees or the one observing at 90 degrees? For general cloud geometries, observations at these two scattering angles will not observe the same fractional coverage of their field of view (and both will be distinct from the vertically projected cloud fraction). This may seem relatively minor relative to the precision of the technique (which is evaluated using specMACS) but will introduce an instability of the algorithm (i.e., a systematic error) to the target cloud type and the solar zenith angle which will alias into regional and seasonal variability and may be quantitatively significant. This is a fundamental and unavoidable aspect of the technique (as it is proposed here) and should be discussed especially as to how it will affect the use of the product to measure its target.
6. The same issues apply to the retrieval of cloud optical depth. For a model, optical depth is precisely defined as a vertical integral. Exactly how is this retrieved quantity defined here?
7. Due to the required choice of scattering geometry the observation at 90 degrees or 110 degrees will tend to observe regions of the surface that are shadowed by the cloud (i.e. when compared to near-backscatter). The technique seems to rely on observations of polarized surface reflection to determine cloud fraction. The shadowed regions of the surface will not provide this strong polarization signal despite being clear. It seems that this will induce a systematic error in the technique that varies with the shadow fraction of the fields of view (which will also differ between the two views). Shadow fraction will vary with cloud geometry such as area, spacing and cloud-base height and aspect ratio. While I appreciate and support the stated intention to examine the retrieval using 3D radiative transfer with complex cloud fields derived from Large Eddy Simulations in a subsequent study, this does not preclude the need to present and explain these basic features of the retrieval within this paper, perhaps with simple idealized clouds such as cuboids. It would also be interesting to see whether there is any detectable signals of these effects in the specMACS data, though this would require development of a pixel-by-pixel shadow mask.
8. The authors state on Line 218 that the technique will deliver accurate cloud fraction and cloud optical thickness over the ocean. This statement seems overly strong in the case of cloud optical depth even in the highly simplified case of plane-parallel atmospheres. The estimation of the cloud optical depth from the cloud-bow degree of polarization will have a correlated error with the determination of the droplet effective radius and droplet effective variance. For example, the magnitude scaling parameter in the least squares fit for the shape of the cloudbow polarization pattern is implicitly sensitive to both cloud fraction and optical depth. It seems that to properly understand the error characteristics, all four parameters should be jointly retrieved.
9. There is no mention of ice within this paper. Some discussion of whether this technique works for ice phase clouds or mixed phase clouds should be included. Due to the coarse resolution, there will also be a further lack of uniqueness in the cloud phase in the actual data when compared to moderate resolution (1 km) imagers. This possible limitation should be discussed but even if the technique relies on liquid scattering signals, this technique will still be a valuable contribution because most small clouds (and hence partially cloudy pixels) originate from low-level liquid clouds.
10. The discussion of non-oceanic surfaces appears slightly incomplete. The authors mention that for brighter unpolarized surfaces the technique is impossible but do not come to a conclusive statement about dark unpolarized surfaces, simply stating that the retrieval will be more uncertain. For the case with the unpolarized surface, the retrieval is reliant on the molecular polarization signal and there will be an ambiguity about whether the polarization signal is due to cloud top height or cloud fraction. This sensitivity to cloud top height is not examined for a dark unpolarized surface. It would be helpful to reference the accuracy with which the DoLP can be measured so that the relative statements about uncertainty are translated to actual retrieval uncertainties.
11. I am also curious as to whether the retrieval concept presented is overly simplified and doesn’t fully exploit the observational information content. There are some dependences of the retrieval on wind speed and aerosol optical depth documented within the paper. For the purposes of reducing systematic errors and ensuring proper uncertainty propagation, the more variables that are explicitly retrieved (even if with strong priors), the better. Given the hyperangular observations from HARP2 and other multi-spectral observations, is it not possible to jointly retrieve wind speed/surface roughness and aerosol optical depth (e.g., Knobelspiesse et al., 2021) in combination with cloud fraction using this information and some prior information on aerosol composition? I think it would be valuable for the authors to discuss the feasibility of this as a possible extension to their work.
12. The validation of the technique against field data is an extremely valuable component of the paper. Some greater discussion of the threshold-based cloud mask that the technique is being compared against is warranted. The features used in the cloud mask are listed but the tuning of the thresholds is not. Is the cloud mask clear-conservative or cloud conservative? Or is it designed to optimally estimate cloud fraction over 2.5 km regions? This is important for understanding any agreement or lack thereof between the coarse resolution technique proposed here and the high-resolution threshold-based technique. As I understand it, the high-resolution reference mask also makes use of polarization features to separate sunglint and from non-sunglint. Perhaps this point, and the ambiguity of masking in sunglint regions with just intensity measurements should be more emphasized as a strength of the technique.
13. For the specMACS data, I would appreciate seeing retrievals that used the cloudbow reff/veff as input as this is the best guess for the retrieval and would reduce any possible error compensation in the validation.

Technical Comments:

Line 93: I was a little confused on a first read by the statement that all simulations had a liquid cloud layer located from 2-3 km and same droplet effective radius. It would be helpful to note that this restricted setup is just to illustrate the main sensitivity to cloud fraction and optical depth and that more sensitivity tests are performed later.

Line 304: Sun-glint

Relevant References:

Wielicki, B. A., and  L. Parker (1992),  On the determination of cloud cover from satellite sensors: The effect of sensor spatial resolution, J. Geophys. Res.,  97(D12),  12799–12823, doi:10.1029/92JD01061.

Stubenrauch, C.J., Kinne, S., Mandorli, G. et al. Lessons Learned from the Updated GEWEX Cloud Assessment Database. Surv Geophys (2024). https://doi.org/10.1007/s10712-024-09824-0

Dutta, S.,  Di Girolamo, L.,  Dey, S.,  Zhan, Y.,  Moroney, C. M., &  Zhao, G. (2020).  The reduction in near-global cloud cover after correcting for biases caused by finite resolution measurements. Geophysical Research Letters,  47, e2020GL090313. https://doi.org/10.1029/2020GL090313

Knobelspiesse, K., Ibrahim, A., Franz, B., Bailey, S., Levy, R., Ahmad, Z., Gales, J., Gao, M., Garay, M., Anderson, S., and Kalashnikova, O.: Analysis of simultaneous aerosol and ocean glint retrieval using multi-angle observations, Atmos. Meas. Tech., 14, 3233–3252, https://doi.org/10.5194/amt-14-3233-2021, 2021.