Overview

This manuscript presented an approach to studying sea ice ridges from medium- and high-resolution SAR imagery from TerraSAR-X validated against Helicopter-born EM (HEM) datasets acquired over the Baltic sea. This study proposed a method to delineate linear ridge features from SAR images regarding the local density of bright SAR pixels over a certain percentage. This study found a linear relationship between SAR bright pixel percentage (BPP) and HEM ridge coverage. Although acknowledged, this study does not contribute towards quantifying sea ice roughness/ridges corresponding to SAR backscatter, which is a major gap in the sea ice literature. However, this study argues that the proposed method can aid safe navigation through ice-infested water.

We find this characterisation mostly correct. On the question of quantifying ridging from SAR we remark that the approach can describe the variation of  ridge rubble coverage in a relative fashion. Local measurements of ridging conditions,  IceSAT-2 profiles, or observations from an ice going ship can be used to locally remove or reduce the relative uncertainty and the result can be extrapolated to an extended area. The other aspect of ridging quantification from SAR, or ridge height, remains a gap as remarked. However, ridge sail block dimensions increase and the sail geometry scales up  on the average with ridge size as  larger ridges are created from thicker ice. If the bright returns are from favorable block arrangements the distances between neighboring bright returns should then increase with ridge size. We did not follow this clue further as we did not have matching surface data for the high resolution image.

Major comments:

The paper provides important details on the subject matter, works of literature, and proposed methodology, supported by necessary figures. My major comments are as follows:

I have a concern regarding the structure of the paper. The paper is comprehensive; however, sections 1-3 can be synthesized well to shorten the length. There are a few very short sentences in the manuscript, which can be added to the previous sentence. Similarly, a few very short paragraphs can be merged with the previous section to keep the flow consistent in the manuscript. Please check and correct this issue throughout the manuscript. At least, section 2 needs to be synthesized to have a better flow of the content, which I find scattered in the current version.

Both reviewers remark on the length and too short sentences. We clearly should and seek to  improve in both matters and will also have the final version of  manuscript chekced for language. This answer is the same for both reviewers.  

The methods and results are mixed in sections 4-7, making it difficult to follow. I think the authors should separate results and discussion.

We considered this carefully when writing the paper and also had  an earlier version with methods and results more clearly separated. This made the methods section long and difficult to follow as the statistical model, which is expected to be novel for most readers, must be expounded and validated both for SAR and for the ridge data. This was further complicated by the fact that the ridge data set is compared with non-concurrent high resolution image in terms of the statistical model while for the medium resolution image the ridge data is concurrent validation data.  We considered also the possibility to present the theoretical parts as an appendix. These parts contain however the main new methodological advances towards the quantification of ridging from SAR and  should in our opinion be presented in the actual paper. There are also several different and in part independent methodologies introducted in the paper. Thus we found the present arrangement, where each methodology is followed by results, a  more accessible alternative.

 In Section 5 we present the approach of counting BPN in pixel blocks with some practically oriented methods of presentation. This section ends to an introdcution to main statistical idea, or the considering how BPN values change in small changes of BPP. This section might better serve as an introdcution in Section 6 when combined with introductory parts of 6.1.  We will also  try to make more clear in Section 6 what is the basic approach, what are theoretical validations of the concepts,  and what are the applicable results.  The simulation exercise in 6.6. was intended the provide additional validation to the basic concepts but it also illustrates these in an easily understood way. We find ourselves the result quite striking but presenting it earlier in the paper might raise confusion on what is the subject matter. Section 7 is a detailed application/validation study that uses only the  primitive concepts  of the approach to compare SAR and surface data but seeks to be careful  in its use of validation data.

When TSX was acquired, the paper motioned the air temperature as -2.3 degrees C. Since the ice was first-year sea ice and the snow had brine, I wondered whether the snow was brine-wetted at the bottom had an impact on X-band backscatter. A sea ice study on C-band SAR imagery reported moist snow at -3.1 +/- 1.5 degrees to have a melt onset signature. Since air temperature was warmer in Baltic during TSX acquisition, how could this affect the SAR statistics presented here?

The water salinity in the Bay of Bothnia is from 3 to 4 permille. Ice bulk salinity typically less than 1 permille for level ice and ridge blocks can be almost fresh; we add the following reference ( A.J., Weeks, W.F., Kosloff, P. and Carsey, S., 1992. Petrographic and salinity characteristics of brackish water ice in the Bay of Bothnia. CRREL Report 92-13). ). The field measurements of ice samples reported by  Hallikainen (1992) (will be added as reference) had ice salinity range from 0.2 to 2 permille depending on the location, time, and weather history.

Brine expulsion on the  surface may occur for thin level ice after rapid freezing but is expected to have drained from thick midwinter ice in the Bay of Bothnia that also does not become flooded by snow load any more. Thus we do not expect that any consequential flooding  occurred in the SAR covered areas. In the case of the flooding the salinity of brine is low and brine would effect on the backscattering level mainly through the increasing snow wetness. So the  role of salinity is small. Some moisture in the snow cover in the 2011 TSX data probably have slightly decreased the contrast between level and ridged ice. However, a large fraction of the ridges are still visible as seen  in Fig.13. The assumption that the bright returns are predominantly from favorable ridge block arrangements is valid in these sea ice conditions.

We can address this in the discussion section.

Minor comments:

Since the manuscript focuses on TSX images, the title should reflect the frequency used in this study. Please include ‘X-band’ in the title.

The methods are not restricted to X-band and we believe that the results would have been similar if the SAR images available to us would have had different band(s), that is, the assumption that brightest returns are predominantly from ridges would have been valid. The soundness of this generally known property has been demonstrated for the X-, C- and L-band  in the Baltic Sea, e.g., in studies Mäkynen (2004) and Dierking (2010). There are some sea ice conditions outside ridging which generate strong radar response, brash ice zones likely being the most common. Hence the identification of open water and brash ice areas must be done before the application of the proposed method.  However, the functionality of the method needs to be systematically studied and conditions where the assumption does not hold can be conceived. We will address this in the discussion section and would like to retain the title as it is. 

If section 7.1 can be considered validation, the title should show that information so that the reader can refer to the section title to find necessary information without going into the details of the text.  discriminative

We work towards improvement here, e.g. in terms of subsections.

Was both the imagery acquired in ascending or descending mode? Please confirm. A mixture of modes can seriously impact SAR backscatter from ridge sail direction.

In our case the time interval between the fine and the medium resolution images  is five years and the target area for the images is hence not the same. We believe that even if the time difference would have been short and imaging from different directions would have applied, the proposed method should have yielded similar statistics. The method refers the statistics of ensembles of pixels (pixel blocks), not single  pixels.  For each pixel in the block that contains ridge rubble it depends on the favorable arrangement of the ice  blocks whether the pixel has a bright return (i.e. one above the assigned intensity threshold) or not. As the orientation of ice blocks is basically random the same applies to the occurrence of favorable arrangement and bright return. If the modes or angles are different, different pixels in the pixel block will appear bright and also the number of bright pixels is likely to change. Due to randomness the same statistics is expected to emerge. 

However, we would like to remark on possible effect of the incidence angle in each SAR image (e.g. Mäkynen and Karvonen 2017) . This may affect also the statistics although ridge sails as sloped features of randomly oriented scatterers are less sensitive to incidence angle effects. In the high-resolution image used to study the BPN statistics  no effects could be discerned and  the change of incidence angle across the 8.7 km wide study area  may be neglected. For the medium resolution image study area extending about 60 km laterally this situation may be different but is obscured by the gradient of the ridge density towards the Finnish coast. We will address this issue in the discussion.  

(Mäkynen, Marko, and Juha Karvonen. "Incidence angle dependence of first-year sea ice backscattering coefficient in Sentinel-1 SAR imagery over the Kara Sea." IEEE Transactions on Geoscience and Remote Sensing 55, no. 11 (2017): 6170-6181)

We also corrected the issues in the remaining comments below:

In-text references are not included correctly. For example, page line 17 should be ‘As verified by Dierking (1999)…”. Please correct similar issues throughout the manuscript.

Page 17, line 16

?? should be replaced with an equation number.

 Page 25 line 9

What does ‘?’ denote?

Page 26 line 1

Check and correct the section title

Page 23

Section number needs to be updated after 7.1. Currently, the result section shows 7.1.1. One section in between does not have a section number.

Overview

The work presents a new method to derive sea ice ridge height through bright pixels and ridge sail number in one high resolution and one medium resolution TerraSAR-X image. The paper uses a statistically derived method to connect the SAR images to the ridge density.

 We find this characterisation correct.

Major comment:

The text would benefit from being shortened, it’s presently very long, 30 pages excl. references, and it deters from making the work shine. The text is in places composed of short sentences making it a bit difficult to read, and some of these sentences are not grammatically correct. This unfortunately lowers the quality of an otherwise nice manuscript build on solid physics. It would therefore be nice to see the work going through a thorough language check.

Both reviewers remark on the length and too short sentences. We clearly should and seek to  improve in both matters and will also have the final version of  manuscript checked for language.  This answer is the same for both reviewers. 

It is stated that the work can be used to aid safe navigation, but it could perhaps be explained a bit more thorough how that is planned.

 In the example figures we delineate the 30% of most deformed ice as divided into three 10% classes. Although the conversion factor to actual rubble coverage is lacking and the height characteristics is not known a ship entering the area may provide these by its own observations and then distinguish easily navigable areas from those it should avoid. The spatial distribution of the  classes can be also  complement ice charts that estimate the total deformation form different sources. A third possibility, possible in the frequently navigated Baltic, is to determine from AIS data the ship response to the classes and use this to interpret them in terms of difficulty of navigation. We will add some sentences in this theme.

This study is conducted using X-band images, which possibly due to the high resolution offered by TerraSAR-X, is renowned for being very good for ridge detection. Would it be possible to transfer this study to the operational used C-band SAR images? Perhaps the authors could speculate if this would be possible and what implications it would have. As the study is done by authors at FMI is it perhaps possible to also find overlapping C-band images for the time and area of the X-band images used in this study?

 As we commented also to the other reviewer we expect the method to apply to other bands if the brightest returns are predominantly from ridge rubble. We sought to support this by operations that obscured the ridging signature in the image but not in the contextual image. Conditions where the wavelenght matters can conceived however, as remarked to the other reviewer.

By inspecting the results shown in Mäkynen (2004) for the scatterometer data at X- and C-band, we notice that the discriminative power of sigma0 between different ridging categories is roughly the same at X- and C-band. This is applies also to the difference between level and ridged ice areas. This study suggests that the proposed method works also for C-band SAR imagery. Further studies are expected to clarify these issues, and IceSAT-2 can provide a source of validation data.

The air temperatures were quite high during the time of the SAR image acquisitions, up to -1C. How will this affect the analysis? As ridges often trap more snow than the level sea ice, would the method presented here be more sensitive to a temperature change than a level sea ice area.

Shortly yes. The temperature controls the amount of wetness in the snow cover. With increasing snow depth and/or snow wetness the backscattering decreases. Due to aeolian transfer the snow tends to accumulate in ridge sails. If average snow thickness is small smooth level ice areas may even be snow free and all snow is accumulated against sail slopes. Our earlier results (Haapala et al. 2013)  indicate that for thick snow cover (average ~0.45 m) the snow thickness in ridge sail  rubble areas may be almost twice the average and snow can cover completely individual sails with height  up to three times snow thickness. Thus if the snow is moist the effect is significant. The upper parts of  higher sails usually remain ice free in windy conditions however. Thus the contrast between ridges and level ice may decrease on the average but bright returns from upper sail parts are expected to remain. For an extended area with homogenous snow and sail height conditions the density of bright returns is then  intergral sum over the returns from sail height classes for which the returns are conditioned by snow thickness. However, for each class the returns can be assumed to follow the same randomness that obtains for snow free conditions, and  the hypothesis behind the statistical model can be assumed to  still hold.  Hence the method is sensitive to temperature and snow thickness but  is nevertheless able to detect variations in ridge rubble coverage. The conclusion is that our basic hypothesis that the brightest returns come from ridge rubble remain valid in most cases except possibly  in very wet and thick snow conditions over shallow ridges  when the contrast between level and ridged ice disappears and the backscattering level overall is low.

We will seek to communicate, withour extending the text too much,  the essential content of the above argument on the effect of snow, temperature and other ambient conditions. In short, the generative hypothesis behind the statistics is that a small increase in bright pixel percentage changes the bright pixel numbers in way that depends linearly on BPN; and there are reasons to expect  that changes in ambient conditions follow the same pattern and therefore do not change the statistical model but only its parameters.

(Haapala, J., Lensu, M., Dumont, M., Renner, A.H., Granskog, M.A. and Gerland, S., 2013. Small-scale horizontal variability of snow, sea-ice thickness and freeboard in the first-year ice region north of Svalbard. Annals of Glaciology, 54(62), pp.261-266.)

HEM thickness measurements tend to underestimate underestimate thick and deformed ice (see, e.g., Haas et al., 2009; Mahoney et al., 2015), how will this affect the results presented here? What is the resolution for the laser data?

HEM data may be used to construct that are used to convert surface elevation characteristics to thickness charateristics. Due to the footprint (a few meters)  HEM blunts the maximum keel depths and the porous ridge keel may have nonzero conductivity that results to underestimation of average thickness if not accounted for in the HEM inversion model. On the other hand,  HEM instruments are  usually able to penetrate trough thickest Baltic ice types. In any case, the  conversion of surface characteristics to thickness is  not among the topics addressed in the paper and is rather an independent problem  shared by all approaches to thickness using surface data only. The laser measurement interval is about 0.5 m.

Minor comments:

It is fair to say that the ground truth data from 2011 was used for both datasets. Perhaps this could be rephrased slightly as it may come across as using 5-year-old data as in-situ could unfortunately question the study in the abstract.

 We will rephrase  ‘...acquired in 2016 and 2011, and ice surface profiles measured by a 2011 Baltic campaign. The images...’    The use of the surface data in the context of the two images is described later in the abstract, but we add ‘2016’ or ‘high resolution’  and ‘2011’ to make clear that the statistical model is formulated for RSN data that is not concurrent with the high resolution BPN data.

P4R20. Please provide reference/s for this.  - We will refer here to  Gegiuc et al., 2018  listed in the same section

P6R32. “…is used a” -> “a” and move “is used” to later in the sentence. - Corrected

P7R20. Is 1/km km^-1.  - Yes,  1/km is commonly used in the Baltic context for ridge density

P8R4. There is a reference to annual maximum extent, could this perhaps also be indicated in Figure 1, also as it is referred to this figure in the end of the sentence. - This is possible, alternatively we provide both extents in km^2 as only part of the Baltic is visible.

P8R13. Where is this temperature sensor located?  - We will add that the drifting station was RV Aranda  and that the data is from its weather station.

P8R26. How were these observations derived? From FMI ice charts? Ice observations? - FMI ice chart values based on observations and measurements by icebreakers and coastal stations.

Figure 4 is referred to before Figure 3. - Corrected

P17R16. Equation number is missing.  - Corrected

P23R18. Could the rubble field also be referred to as brash ice?

No, these are different. In rubble field the ice ridges sails are arranged so close to each other that the level ice fraction is minimal or nonexistent. It is a continuous field of ridge blocks, while brash is  floating melange of ice pieces of different sizes. Consolidated brash has rough surface with  bright but rather uniform SAR return in contrast to the randomly scattered bright returns from rubble fields.

P25R9 (?)  Corrected

P28R12-16. How were the scaling factor derived? Was there a statistically based fitting?

The scaling factor is from the slope of the quantile plot. The HEM RSN data followed gamma distribution in the continuous limit. Ridge density is special case of RSN statics  for 1 km segment length. The  HEM RRC of eq  (6) inherits the gamma as it depends linearly on the ridge density. The  SAR RRC is then also gamma as the quantile plot is linear. So we did not make statistical fitting here  but the the statistics is derived from the validation of the full scale system (1-3) for RSN in Sec 6.4.  We try to make this chain more explicit.

P29R8. Which frequency would be suitable for ridge detection?

According to the literature the L-band is best suited for ridge detection, see e.g. Dierking (2010)

 (Dierking, W. Mapping of Different Sea Ice Regimes Using Images From Sentinel-1 and ALOS Synthetic Aperture Radar, IEEE T. Geosci. Remot  S.,  48(3):1045 – 1058 https://doi.org/10.1109/TGRS.2009.2031806, 2010)