General comments

The manuscript provides a novel sub-millimeter ice cloud retrieval algorithm that uses radiometer along-track scanning simulation data and a hybrid Bayesian MCI and OEM optimization approach to reconstruct two-dimensional ice cloud scenes. The manuscript is well-written and rich in information. I believe the manuscript deserves to be published in AMT. I only have a few minor issues that need to be explained by the authors.

Specific comments

1. Line 79. Is the current CoSSIR not capable of along-track scanning?
2. Line 99. Since CloudSat is not scanning, meaning that an orbit of 2C-ICE is striped, how to understand the 3D scene here?
3. Line 101. How to understand the IBA technique here, whether it is by averaging 1D simulated observations within the field of view or by extracting multiple atmospheric profiles at different locations on the path to synthesize a scene?
4. Figure 5. What spatial extent of cloud parameters are reconstructed in one along-track scan, is it the entire scan? How do you handle the front and back scans where there is a crossover?
5. Line 190. If the BMCI performs successfully, does the OEM not execute, why the minimum database case is 25?
6. Line 197. Does the local Gaussian a priori constrain mean the a prior IWC and covariance matrix used for OEM?
7. Line 197. That is, even if the BMCI fails, at the start of the OEM, it still has to find 25 cases in the database to update the a posteriori PDF. I'm a little confused here.
8. Line 237. How to understand auto-adjustable BMCI? Is it any different than the BMCI in 1D?
9. Line 265. Computational resources and speed are really an issue with this algorithm, roughly how much computational resources and computation time are needed at the moment?

This is a well written manuscript that illustrates the benefits of performing 2D tomographic retrievals as a way of improving on historical 1D retrievals.  While the results between 1D-nadir, 1D-tomographic and 2D-tomographic retrievals could have been anticipated, the value of this paper is that it clearly outlines viable approaches for the 1-D and 2-D tomographic inversions. The treatment of uncertainties and correlated uncertainties, while perhaps not the only way to deal with the problem, is sound and pushes our treatment of correlations length in OEM retrievals beyond our current methods.  I think this paper is definitely worth publishing since it not only explores a new technique, but also opens up the possibility of occasionally flying cross-track scanning satellites sideways in order to obtain better microphysical insight as suggested here.

Specific issues the authors should address

Line ~110 – The results presented here depend to some extent on the details of the particles that were assumed to make up the liquid and ice clouds, and how supercooled water was handled.  To make the results more reproducible, I would encourage the authors to make their PSD assumptions explicit. Because the profiles are artificially simple, the authors should also add a caveat to the conclusions.  It may be that changes in PSD properties or the inclusion of super-cooled water in the cloud may be another reason to perform 2D tomography but that has not been demonstrated.

Line ~140 – the suggestion is that Tb are computed for each CloudSat profile at multiple angles needs a little more explanation. Because slant path computations cut through multiple horizontally adjacent profiles, this leads to uneven layering of the slant path profiles. Exactly how the authors handled this was not clear from the description. 

Line ~150 – If Tbs are constructed from single profiles but simply for different view angles, then the vertical correlations are assumed in the prior data.  Some discussion here about how this is ultimately handled in the retrieval would be appropriate here.

Line 220 – Simply averaging the same voxels retrieved from each view angle in the 1-D scheme seems to artificially simplistic.  Given that the authors have the goodness of fit from the OE, should this not be a weighted average?