The paper is interested by the reflection of ice slab covering snowpack in the near-infrared domain. It aims at retrieving two parameters (namely ice slab thickness and snow “grain size”) from spectral BDRF measurements. To this end, direct modeling based on a radiative transfer model and inverse method related to Bayesian inference was used.
Although the general structure of the paper is clear, the paper is hard to follow because of imprecise writing. Many different and sometimes fuzzy terms are used for the same meaning. The overall objective of the paper is not clearly stated. The first sentence of the abstract, the title and the objective stated P2L31 do not match. The overarching goal of the paper needs to be clearly stated and used to conduct the introduction. It would also be made more attractive and adequate for TC audience by specifying the geophysical problem the authors want to solve (in the introduction with details, not a one-sentence at the end of the conclusion). On which planets is such ice over snow configuration found ? Which are the expected values for the parameters (ice thickness and snow “grain size”) found there. The results of the paper should then be interpreted in terms of the geophysical problem(s) in the discussion (which is not a discussion otherwise).
Regarding the practicle application to remote sensing, it would be interesting to test how the degradation of the BDRF sampling (e.g. MODIS takes a very limited number of “angels” in a day) change the value and uncertainty. Application to the sensor/mission targeted by the authors would be most relevant.
The inversion method developed seems inadequate regarding the speed criteria. Reference to this criteria should be removed or a variational method should be used to offer a benchmark, such methods are so common that any scientific language has a few of them (see details below).
At last to avoid misunderstanding, it has been suggested that “grain size” is reserved to the “maximal extent” measured with a hand lens (Fierz 2009). Optical radius or diameter is the precise and relevant term for the present study. I recommend it should be changed throughout the paper.
My recommendation is to improve the readiness and interest for TC audience before acceptation.
Detailed comments.
* English editing is required. Remove future in the introduction and in most of the paper.
* remove excessive words like “ultra” “golden era” “kill”
* L6P3 is not clear (what is spectral spot?).
* L16P3 is not clear.
* In general “inversion” is unclear with the model and parameter to be estimated precisely defined. Use retrieval, estimate, estimation, …
* L24P3. It's difficult to follow with these parameters not properly defined.
* L1P4. Not clear. Remove this information if you mean something that is not in this paper.
* the model is half described which is acceptable owing to the earlier publication but this is not reader-friendly. Consider to improve this section.
* L10P4. What means “from IPAG” ?
* L12P4. How large ?
* end of P4. A picture of the experiment and slab would be helpful.
* L5P5. Figure order must be ascending.
* Section 3.2.3 should be in the results, at least the interpretation.
* L1P6. Why ? Is there any reason to use the older reference at wavelength larger than 1 microns ? Computation of the uncertainty associated with this issue would be interesting.
* L15P6. Are “m” and “d” vectors ? If yes, use bold face. Considering only two parameters are used here, I suggest to explicitly write the formula with the parameters instead of “m” or “d”. This will be easier to follow.
* L17P6. “this problem” is not clearly stated. Please define what “close” means. This vocabulary does not sound Bayesian.
* L17P6. “each quantity as a probability density function (PDF)” seems to be a leap. Do you mean each quantity is considered as a random variable ?
* L18P6. What means “In non-linear direct problems” ? Non-linear, direct and the plurial of problems are odd.
* L22P6. Do you mean Multivariate normal distribution ?
* L22P6. I don't understand what is ri compared to d ?
* L25P6. “with σ i being the standard deviations of each measurement. “ There is a conceptual problem here. A single measurement has potentially error but not a standard deviations. Random variables and sets of samples have standard deviations.
* L5P7. Which also supposes that the covariance matrix is fully known. In a typical Bayesian application it would be included in the parameter spaces.
* L3P8. How fine ? Why not to use a Monte-Carlo Markov Chain sampler which is incomparably more efficient and has not the problem of chosing a step ?
* L14P8 and L17P8. Reference
* L1P9: “The best match” please rephrase in proper Bayesian terms. Or remove any reference to Bayesian method and likelihood, and use fit, cost function, etc...
* L11P9: Is it really justified considering that the ice absorption dataset has a resolution ~10nm. Give a rational or remove.
* L13P10: What is a “stack” of pdf ?
* L20P10: not sure to follow what is “posterior uncertainty”. Is it the standard deviation ? I'm also uncomfortable with relative uncertainty larger than 100%. Do you mean relative error ?
* L21P11: What means compatible ?
* L31P11: The sentence starting by “It would require...” is not clear.
* L5P12: Why roughness and tilt of the sample would have the same results in the BDRF ? The description of the model roughness is clearly needed in this paper, even if it was already published elsewhere.
* L17P12: “mismatch between the best fits” these terms do not sound Bayesian.
* L24P12, L12P13: even if snow has been evolving, it would be interesting to show the grain size measurements ? How are obtained the “independent measurements” ?
* L17P13: The sampler used (lookup table) is very inefficient compared to many MCMC samples. Moreover, considering that for most applications the uncertainty will be of little use, using a variational optimization algorithm which returns the Hessian should be sufficient. Variational optimization is probably another order of magnitude faster than MCMC. This must be made clear because the text indirectly suggests that the presented method is efficient. The model seems indeed fast, but since the number of calculations is huge, a much slower model with a good inversion method would perform as well.
* L28P13: The first sentence suggests that by using Bayesian inversion, modeling error was accounted for. This is not the case (as written in the paper). The method only accountd for uncertainties resulting from observation errors (and considering the errord are known while it could be estimated).
* What is proposed for the data availability (cf TC data policy) ?
* Figures 4, 5, 6: increase the size of the axes and tick labels.
* Figure 3: is the phase angle the relevant parameter ? I mean is the BDRF only dependent on the difference of the incident and emergent angles ? Or I miss the point.
* Figure 4: unit on y-label (should be m^-1)
* Figure 5: uppercase are missing.
* Figure 7 and 8 unit for the reflectance (is it percent?). Same in Figure 10 and 13 ?
* Figure 8. Remove “.0” everywhere. Adjust colorscale on the right plots to avoid decimal numbers.
* Figure 10: “The thicknesses indicated in the captions”. Captions/Titles are missing ?
* Figure 13: Why means “R” on this graph ? |