Sentinel-1 cross-polarization ratio as a proxy for surface mass balance across east Antarctic ice rises

Thore Kausch, Stef Lhermitte1, Marie G.P. Cavitte, Eric Keenan, and Shashwat Shukla

Overview

This paper seeks to explain the relationship between the Antarctic ice sheet surface mass balance (SMB) and the cross-polarization ratio from Sentinel 1 C-band SAR data. The study aims to build upon previous work by the author using the RACMO2 model to estimate snowfall on the ice sheet. I found the study interesting as it attempts to examine whether an empirical ratio method (HH/HV cross polarization ratio at C-band) is sensitive to SMB. The authors use SMRT and SNOWPACK as part of their analysis to constrain the snowpack properties and attempt to explain the processes driving the cross-polarization response.

General comments

The authors should be commended on their study which attempts to shed some light on the cross-polarization ratio approach. Attempting to find a physical explanation for the Sentinel-1 response is an important task, especially in the estimation of ice-sheet SMB for which this work could be highly applicable.

The paper is moderately well written in some places but in many other places it lacks specificity and clarity, or is confusing, or is sometimes unintentionally misleading. This makes it quite difficult to follow in places. For example, on line 25 (Page 2), the authors state that the Lievens et al (2019) approach can map snow across the northern hemisphere. This is incorrect since it has been applied to deep snow in mountain domains only. On line 30 of Page 2, the authors claim that Lievens et al (2019) "highlights that the radar wave interactions within the snowpack and the volume scattering response of the snowpack are what is driving cross-polarization ratio variability" when in reality, the paper states in their discussion that "future modelling efforts to unravel the different mechanisms that cause the demonstrated sensitivity to deep snow at C-band" should be encouraged. In other words, the Lievens study finds an empirical relation between the VV/VH cross polarization ratio for deep mountain snow but is unable to provide an explanation of the physical processes driving the observed outcomes. So they speculate (perhaps plausibly) that it is a volume scattering response, although no evidence can be applied to explain this relationship. The authors of this paper take a similar approach and have identified an HH/HV cross-polarization ratio relationship with the SMB. But again the full explanation needs further thought. I suspect that it is present in the microstructure information in the SMP data but this detailed information is lost during aggregation in the analysis. 

The experimental method is simple but potentially effective: a) to compare GPR measurements across 3 ice rises with S1 data, and b) to attempt to explain the patterns observed. Comparing the cross-pol ratio with SMRT model estimates, forced by SNOWPACK, is a useful approach. However, the results are not easy to interpret in the format provided and the description of the results is very hard to follow, often because the writing is not precise enough. For example, in the Section 4.2 descriptions of the results in Figure 7, the text should precisely refer to the specific figure panel and its attribute. I found that without a specific identification of leeward and windward side in the figures, the reader is left having to figure it out.

Specific Comments

P2. Line 16, The authors claim that SAR is sensitive to snow microstructure and is independent of cloud cover. In fact, studies show that cloud cover can impact the radar response and that the SAR the sensitivity to snow microstructure is frequency-dependent. I would encourage the authors to be more precise in their writing.

P2 line 33. Using the co-pol HH channel will have less sensitivity to volume scattering-dominant processes that are evident in the VV co-pol channel, as used by Lievens et al (2018). Likely this may impact the results so the authors should explain more what the impact might be. 

P3 Line 15. The locations of the LIR, HIR and DIR are not labelled in Fig 1 making this quite difficult to assess and understand what leeward/windward actually means. 

P4 Line 11. Since ERA5 was used to gap fill in 2018, the authors should provide an assessment of the uncertainty of this gap-filling since wind speed has a significant impact on the SNOWPACK estimates.

P4 eq 1. Please define all units used. 

P4 Line 27, Equation (1) and P5 lines 1-5 and throughout the paper. The authors should be more specific in their language referring to "grain size". This variable is a critical parameter in the SMRT estimation process and there are several emerging terms regarding what is meant by grain size (effective grain size, measured or observed grain size, optical grain size). Also, the concept of a "real grain size" is somewhat misleading. 

P5 Fig 1. Does the scale bar for B (HH pol) apply to the cross-pol in C? This should be stated. Also, the GPR tracks are not clear - the authors provide a more detailed map of these tracks at the 3 locations.

P6 lines 14-15. 50 m spatial resolution is much finer resolution than the Lievens (2019) approach. They have noted in their paper that the smoothed backscatter data was posted to 1 km. Why did you select 50 m?

P6 Section 2.5. Did the authors include speckle filtering in their workflow? Even for EW data, speckle noise may have an impact and when the data are averaged, the speckle (multiplication noise) could have an impact on the averaging process of the S1 data. How do they know that this doe not have an effect?

P9 Line 7. Can the authors explain what they mean by "looping the 2018 input AWS for 20 years"?

P10 Line 3-6. I disagree that there is "good agreement between SNOWPACK grain radius (?) and the SMP snow grain radius (?). There is much more variability in the SMP data than observed by the model indicating a lack of model sensitivity. Can the authors explain what this might be caused by and the importance of this?

P10 Line 21. Penetration depth in microwave research is defined as 1/e. Is this what the authors mean or do they mean the maximum depth beyond which no further response is observed?

P10 Section 4.1. I know this is pedantic but the authors seem to conflate Correlation R with coefficient of determination (R^2) which is the measure of the fit of a linear regression. Perhaps they can be consistent in their use of such standard terms.

Figure 5 should include a legend of the colours for improved clarity. 

P10 Section 3.5. The authors state that they use a stickiness value of 0.15 for all runs. How was this value selected and how sensitive are the results to it?

P10 Section 3.5. Why did the authors select the IBA and not, for example the DMRT approach. It would be helpful for the reader to provide this justification. Furthermore, what was the substrate condition used in the model - was it an infinite background somehow? A more comprehensive explanation of the model set-up would certainly help the reader follow the logic here. 

P10 The authors should include standard error metrics of the regression lines (the slope coefficient). What is the variability of the regression coefficients calculated? And how is this calculated?

P11/12. The role of Figure 6 is unclear. I understand it shows the SMB variations with cross-pol ratio but the patterns cannot be explained easily, despite the authors asserting that correspondence between SMD and cross-pol ratio is "clear". I can see that there is correspondence between the SMB and the cross-pol ratio for the HIR but for the LIR it is somewhat related but the DIR has only a moderate correspondence. It is unfortunate that in situ data are not available for the DIR and especially the HIR location where there is indeed the strongest agreement. The authors conduct an analysis of LIR based on the SMP, SNOWPACK, SMRT and cross-pol data. But no similar analysis can be undertaken of DIR and HIR because no microstructure data are available. This should be highlighted more clearly.

P12 Figure 6 is also confusing and needs clarification. First, what are the wind directions (guessing the black lines ?) and how do they represent wind direction? I assume that the P and P' labels mark the start and end of the transects? And the authors should mark all relevant figures including this one, with windward and leeward sides. Also, the axes text is too small.

P12. Lines 3-8. The authors claim that the density of snow might decrease with a constant addition of new snow, which might be reasonable leading on from Lienss et al 2020 in which the snowpack was located in a forest clearing in Finland where blowing snow is minimal. However, in reality would the windward side of an ice rise not be subjected to the development of a slab layer which would likely result in an increased snow density ? Furthermore, would blowing snow not be more likely to redistribute the snow from the windward to the leeward side of the rise? I understand that these processes are not included in the model/analysis but they are strong controlling factors of a snowpack state when non-flat terrain dominates. 

P13 Figure 7 and its description on p12-14. Why did the authors simply arithmetically average the microstructure information? A weighted average would be more appropriate given potential variations in each thickness and microstructure. For example, two equally thick layers with very different SSAs will give very different backscatter responses. I would have thought that weighted averages by layer thickness would be far more instructive. Plus it would be instructive to provide the reader with standard deviation of variation of the microstructure. The panel figures are too compressed - more should be made of them to provide better insight into the explanation of the cross-pol ratio data.

P14 Section 4.3. This section is not precise and needs to be written with more clarity. For example, line 14 is not necessarily the case because the averaging of all layer information in Figure 7 masks out the variability of potentially underlying processes that influence the grain radius and/or density values. Simply picking high/low SMB and correlating them with HV/HH and explaining by aggregated grain radius, density is perhaps rather too simplistic. 

Figure 8 and 9. What is the difference between the depolarization ratio and the cross-polarization ratio? The authors should be consistent.

P14 line 13. Do the authors mean R^2 value or R correlation? Also, for all correlations, the significance level must be included. 

P15 Figure 8C. How are the dotted lines estimated? The authors should explain. 

P14 Line 31. Suggest use "vice versa" rather than "the other way around" which is confusing.

P15 Figure 9. Why did the authors choose a 4 point running mean and a 100 point mean for the snow microstructure and cross-pol ratio respectively? 100 pixel running mean gives an averaging distance of 50x100 = 5km. Why did you not apply the same to the running average to the microstructure data?

P17. Lines 1-12. The question of anisotropy as an explainer is an interesting one. However, two problems emerge. The first is that the authors relate this to fresh snow which could indeed be the case for higher radar frequencies but for C-band, it is unlikely to have an impact at that wavelength - the Lievens et al C-band study (2019) is for deep snow only and is not sensitive to snow less than about 2 m (this is why it is applicable to mountain snow). And the Leinss (2020) study refers to X-Ku band - I would not expect it to be applicable at C-band (S1). 

P17 lines 13-18. Did the authors experiment by inserting rough layers in the SMRT which I believe is possible ? This might help to formally discount that that possibility.

P17 lines 24-29. This paragraph is confusing as it refers to the windward side only but with contradictory arguments. Also, based on the points above, it is conjecture and inconclusive.

P17 line 30-P18 line 2. The explanatory discussion can only really come from the analysis of the LIR data since there are no simulations of the other ice rises. This echos the point above about the role of Figure 6 which introduces a tantalizing relationship between SMB and the S1 cross-pol ratio for the HIR data. The only simulation data available are for the LIR for which the explanation is speculative from the analysis. Given that the LIR is the only place to have any explanatory power, this should be made clear at the outset and be clear in the discussion.