This paper presents a novel seismic approach aiming to temporally monitor the formation of sea ice through the deployment of a dense nodal seismic network. Thorough processing of ambient noise through beamformed cross-correlations resulted in the recovery of waveguided modes that were subsquently used as the basis for a Bayesian inverse scheme.

The paper is well written and constructed, as are most coming out of this group, and is suitable for publication upon very minor revisions. My few questions/comments are as follow:

1) On paragraph 165: consider describing a bit better in math the procedure related to the SVD decomposition of the FK transform.

2) Could additional dispersion information have been retrieved by simply picking the maximum of the beamformer at every frequency?

3) Why pose as an MCMC instead of a full grid search? In your case, your forward model is purely analytical unless I am mistaken, which means that a quick parallel implementation of a grid search should be quite feasible.

4) I’m curious as to why things seem to be generally insensitive to density, since this is an important stated objective of your study. I would recommend you explaining what you mean by “The inversions give a value that is very stable, at around 910±82 kg.m−3 for the EW direction, and 908 ± 80 kg.m−3 for the NS direction” on 320, and a second (perhaps unintentionally repeated) time: “Nevertheless, the actual uncertainty is probably much smaller, as all of the inversions give a very stable value” on 325, and again . The posterior density is more or less flat, and thus the mean value here, regardless of its consistency across lines, should probably not be treated as a constrained parameter.

Perhaps an added statement as to why you think the inversion fails to robustly constrain density might be helpful, or in what ways it could be accounted for.

The authors present an innovative passive seismic study that monitors the evolution of sea ice thickness and mechanical properties using passive seismic noise. The authors combine both array processing with ambient noise cross-correlation to produce dispersion curves for guided waves within sea ice. The results of these dispersion measurements are used to invert for ice thickness, Young's modulus, Poisson's ratio and ice density using a Bayesian framework. 

The manuscript is interesting and merits publication in The Cryosphere. It is concise and well written but I think improvements could be made to the clarity of the paper. I have a number of minor comments and suggestions which are detailed below.

- Section 2.1: Seismic array would benefit from a paragraph describing the deployment of the seismic instruments, with details on how the geophones were installed i.e were they mounted prior to deployment, burial depth if at all and any challenges or difficulties encountered. 

- Section 2.2.1 would benefit from a few sentences on the theoretical foundations on the retrieval of Green's functions from ambient noise. This section should also include a description of the processing steps, parameters used and the justification for their selection e.g., why and how was a 5 minute window selected for the noise correlation length. Also which stations were used and if correlations were made across stations components e.g., N with E. 

- Section 2.2.2: In this section, it is unclear as to which of the seismic instruments (1C or 3C) is used in the beamforming and subsequent calculation of the noise correlation function. I assume that the authors beamform seismic noise using the 1C stations before using noise directed within 10 degrees of the 2 3C lines to compute the correction function using the 3C stations.

- Section 2.2: A flow chart summarising the beamforming and noise correlation function workflow should be included to improve the clarity of this section.

- Section 2.3: The authors should expand upon the SVD methodology used to compute the dispersion curves. 

- Section 2.4.2: In the inversion, why haven't the author's used uncertainties from their measurements for the estimate of sigma in equation 6? It is unclear from the text why this isn't possible. I'm also unsure as to why the Simulated Annealing method is used instead of the burn-in phase of the MCMC. Could this not affect the sampling of the posterior distribution?

- Results and Discussion: This section would benefit from comparisons between the retrieved parameters and previous estimates in past studies e.g., How does the Young's modulus compare with other estimates? Likewise, how does the estimated value of density compare with meteoric ice and sea ice? Additionally, the inversion appears to be insensitive to density. Why do you think this is the case? How could the inversion be improved to account for the insensitivity to density?

- Figure 1: A map of Svalbard and the location of Lake Vallunen highlighted should be included.

- Figure 6: The histograms should be normalised such that the area under each graph is 1 in order to show the PDF for each of the parameters. It is difficult to see the blue line indicating the maximum of each distribution and this should be changed to black or the shading of the histograms should be changed to grey or white. 

- Lines 78 - 81: A schematic diagram illustrating the displacements of the different modes of the waveguides would be beneficial. 

- Lines 280 - 281: Please include more detail on how the distribution is fitted to the posterior distribution. 

- Lines 284 - 285: Since the distributions are not Gaussian, I'm unsure how relevant it is quoting uncertainties for the results using the standard deviation. The interquartile range would be a more relevant measure of uncertainty.

This is a very interesting analysis of a uniquely dense sea ice deployment of seismographs using ambient noise guided mode dispersion curves to invert for bulk ice properties (in two propagation directions).  With a bit more detail (suggestions below) it’s suitable for publication in its present form as an initial analysis of a data set that I strongly suspect has a lot more to offer in more detailed analysis.  Will the data be publically available somewhere (not apparently in the Acknowledgments).

The authors have done a good job in fitting analytical expressions to (beamform-selected) autocorrelations, with modes decomposed via a SVD method, to extract dispersion curves in EW and NS directions.  The size and density of the array used (247 seismographs with interstation spacings as close as 1 m) is unprecedented to my knowledge for this type of application.  Additional deployment and experimental details would be appreciated.  Were the instruments buried, for example? Were ice cores taken for ice-truth measurements during of the experiment?  How long did it take to deploy these instruments at this scale (important to know if such experiments might be repeated in other locations).  Could similar (or at least practically useful) results be obtained with a smaller deployment (this can be tested by decimating the data set of NCFs).  This would be valuable to address given the motivations expressed in the abstract and paper.

Most of my comments below address where I suggest that methodological and other points might be usefully expanded (including, perhaps, some supplemental figures that could be in a appendix).

Complicating methodological or physical factors that may explain the observed variance in the parameters were underexplored (e.g., can the dispersion curves be better resolved or meaningfully smoothed to improve the data as shown in example in Figure 4?).  It was unclear (to me) exactly which stations were being correlated for use in constraining the NCFs (e.g., the linear array stations and azimuthally selected stations in the grid (which would mean that only stations intersecting a 10-degree cone from the linear array stations were actually utilized, but no stations from the corners, for example (?)).  It would seem that a great many potential crosscorrelations within the experiment were not used.  That’s fine for this initial study, but some additional detail and quantification of this geometry and data usage would be appreciated (with 247 stations there are potentially around 30,000 potential NCFs that could be calculated, of course; how many total NCFs went into the inversions here out of these potential ~30,000 station pairs?).  I’d suspect that a more comprehensive subarray strategy might yield improved results, and the ability to assess spatial variations (the authors hint that this will be a next step, along with icequake analysis in future work at the end of the paper). Structural variation across the study region may explain some of this four-parameter estimation experiment (but again it's not clear to me but it would be nice to assess this effect more thoroughly; it’s alluded to for example in line 270).

The authors employ a simulated annealing strategy to obtain a starting model.  It was unclear to me, at least, how this leads to new information regarding determination of the measurement errors, unless this was simply evaluated to be consistent with the chi-square value, which is simply consistent).  For this reason. I suggest that it is more accurate to indicate that the fit (e.g., line 280) between model and data is statistically “consistent”, rather than “excellent”.  They subsequently use MCMC to obtain a posterior PDF with a uniform (limited) range prior distribution for each parameter, but do not show the covariances of the parameters.  A posterior illustration of parameter covariances would also be helpful to illustrate the tradeoff space between the four parameters.  It is interesting that the inferred ice thicknesses differ systematically in the NS and EW directions in Figure 7.  Can the authors elaborate on this (in line 270 they allude to special variations in the physics of the problem, but I believe what they meant to indicate was something like “spatial variations in ice model parameters”.

All things considered, this is a significant work with a unique data set that could use a bit more polishing to realize its potential, and will no doubt be valuable for further analysis.