This paper describes the aerosol optical depth retrieval candidate algorithm for the EarthCARE mission. The method described in the paper shows quite some limitations, mostly on the limited number considered valid for retrieval purposes. The results presented should be further discussed and the algorithm should be better placed in a broader context, highlighting its unique features, pros and cons, compared to other existing algorithms. Overall, the paper could be strongly improved with more precise explanation and much more elaborated discussions.

General comments

• The verification of a retrieval algorithm against simulated data is expected to represent the best possible retrieval performances, as all the assumptions are the same both in forward and inverse modelling and as the true state of the scene is known, which is never the case from actual observations. Therefore I am a bit puzzled to see such low correlation against the simulated test data set. Either there is something wrong in the construction of this exercise, or there is something I am missing.
• The selection of valid pixels is mentioned in several part of the manuscript, yet it remains confusing. To summarise, the algorithm only processes open ocean water and dark vegetation, is this correct? If so, it should made it clear in the paper. Sometimes you refer to “relatively bright” or similar statement that are not very conclusive. Also, in Table 1 the values you chose for the surface reflectance in the NIR are quite low, I assume this is because dry vegetation is also excluded from the processing. If this is not the case, you should include larger value of surface reflectance. Also, in what quantity is the latter expressed? BRF, BHR, ecc.. Please clarify.
• State vector: it is not clear what does the algorithm actually retrieve: the AOT at 550 nm (as stated at L163) or the AOT at 670 nm and (over water) at 865 nm, as suggested from the abstract and introduction? I seem to understand that the retrieved quantity is the AOT at 550nm, it should therefore be made clear that the AOT at 670 and 865 nm is derived from 550 nm based on your (quite strict) aerosol composition assumptions).
• I think that there is a bit of confusion between absolute accuracy and RMSE, which is dependent on the magnitude of the quantity you’re measuring. The RMSE can largely increase with the AOT range, it should not be taken as a reference for absolute accuracy purposes.
• The impact of your assumptions should be better assessed, e.g. lambertian surfaces, fixed atmospheric composition, etc.

Minor comments

• L4 670 nm over ocean and valid land pixels, and at 865nm over ocean.
• L36 It has been applied, for instance, to MODIS
• L40 The algorithm from Luffarelli and Govaerts is not satellite dependent. In the very same paper you refer to it is applied to PROBA-V as well. Please correct.
• L52 Please update the reference to Govaerts and Luffarelli, 2018
• L55 “where possible” is very vague. Please be more clear: “on valid land pixels”, “on dark vegetated surfaces”
• L79 Add reference to the DEM model used.
• L134 Explain why the method to describe water bodies is not suitable for coastal water. I assume it is because a fix value of chlorophyll content is considered?
• L143 […] the AOT, the aerosol scattering phase function […]
• L170 make sure you use the same symbols as in Eq. 8
• L171 Therefore, the state vector […]
• L175 justify the value of 0.03 and 0.001. Do they come from a certain number of tests, from literature, …?
• L242 surface reflectance in terms of? Albedo?
• L254 You should mention the issue of the radiative coupling between surface and atmosphere
• L254 more complex than
• L267 “empirically found”. How? What is the accuracy?
• Eq. 21 make sure to define all symbols
• Figure 3: the scatterplots show values close to 0.15 which are not really visible in the images. Also there is a cloud of points in figure 3 f) and 3 g) where the M-AOT strongly underestimate AOT between 0.05 and 0.1. Do you have a possible explanation for this?
• L325 “Explained variance”??
• Table 3 Here you suddenly mention “all simulated scenes”. You should mention in the text that more scenes were simulated and refer to the annex.
• L357 Missing reference
• L376 “more sophisticated” More than what?
• L381 transferred = extrapolated?
• L387 The correlation for that scene is 0.97 and 0.99 over ocean and still reaches 0.87 […]
• L422 “relatively bright” - please avoid this kind of statement and be precise

The upcoming EarthCARE mission will include measurements of M-AOT, or column aerosol optical depth at 670 nm (over land and ocean) and 865 nm (over ocean). The study presents the algorithm for these retrievals and results using simulated EarthCARE data and the MODIS L1 reflectances, comparing this output to existing AOD measurements from AERONET and MAN and to the MODIS aerosol products. It is strange that for both the synthetic test scenes and the MSI algorithm applied to MODIS L1 reflectances, the performance is dramatically better over ocean than over land; this raises questions about its surface reflectance estimation. The comparisons to MODIS and AERONET data need clearer justification. My other specific comments are below:

Line 16. “Aerosols have a special role in the overall context” of what? Consider rephrasing.

Line 21. Another unclear use of “context”. Does this mean the lidar data needs to be combined with MSI to obtain geolocation?

Lines 82-83. What is the spatial resolution of one MSI pixel? Measurements of aerosols at cloud edges (and of aerosols in or near dust or smoke plumes thick enough to be mistaken for clouds) are valuable in themselves, and this seems like it would greatly reduce spatial coverage.

Lines 232-334. How often would the land cover type map be updated to account for land use change?

Fig. 3 and Fig. 4. If there’s a way to generate RGB images for these synthetic data test scenes, they would be helpful to orient the reader. Meanwhile, the skill seems very low for the land retrieval, but both scenes are mostly ocean; it’s hard to tell how much of this is limited sampling.

Line 357. A reference placeholder (?) has been left unfilled.

Lines 374-375. The references are for the MODIS Dark Target algorithm, but the MODIS L2 aerosol file includes retrievals from Dark Target, the Deep Blue algorithm, and a combination DTDB. Make sure to specify which is meant.

Lines 381-383. Unless the MODIS comparison is meant to use the Deep Blue or DTDB retrievals exclusively, the L2 product reports Dark Target AOD at 660 nm (land and ocean) and 860 nm (ocean) directly. Inferring the values from the Ångström exponent and the 550 nm retrieval instead risks losing precision.

Lines 383-384. The MODIS Ångström exponents are calculated for ocean only by the Dark Target retrieval and for land only by Deep Blue, so the underlying algorithms being compared here are not equivalent. Also, double check these wavelengths.

Fig. 6. The corresponding MODIS RGB image would be helpful here, too.

Fig. 8. Any idea why these scatterplots form branches at higher AOD values? It’s still odd that the correlation for land is so much lower than for ocean.

Line 402. 5 km is a tighter spatial radius than is commonly used for satellite/AERONET matches, and this does not necessarily improve representativeness. How were the match criteria chosen?

Line 414. Are there plans to report a 670 nm/865 nm Ångström exponent as part of the MSI product?