12 Mar 2021
12 Mar 2021
Sentinel-1 snow depth retrieval at sub-kilometer resolution over the European Alps
- 1Department of Earth and Environmental Sciences, KU Leuven, Leuven, Belgium
- 2Department of Geosciences, Boise State University, Boise, ID, USA
- 3WSL - Institute for Snow and Avalanche Research SLF, Davos, Switzerland
- 4ZAMG - Zentralanstalt für Meteorologie und Geodynamik, Vienna, Austria
- 1Department of Earth and Environmental Sciences, KU Leuven, Leuven, Belgium
- 2Department of Geosciences, Boise State University, Boise, ID, USA
- 3WSL - Institute for Snow and Avalanche Research SLF, Davos, Switzerland
- 4ZAMG - Zentralanstalt für Meteorologie und Geodynamik, Vienna, Austria
Abstract. Seasonal snow in mountain regions is an essential water resource. However, the spatio-temporal variability in mountain snow depth or snow water equivalent (SWE) from regional to global scales is not well understood due to the lack of high-resolution satellite observations and robust retrieval algorithms. We demonstrate the ability of the Sentinel-1 mission to monitor weekly snow depth at sub-kilometer (100 m, 300 m and 1 km) resolutions over the European Alps, for 2017–2019. Sentinel-1 backscatter observations, especially for the cross-polarization channel, show a high correlation with regional model simulations of snow depth over Austria and Switzerland. The observed changes in radar backscatter with the accumulation or ablation of snow are used in a change detection algorithm to retrieve snow depth. The algorithm includes the detection of dry and wet snow conditions. For dry snow conditions, the 1 km Sentinel-1 retrievals have a spatio-temporal correlation (R) of 0.87 and mean absolute error (MAE) of 0.17 m compared to in situ measurements across 743 sites in the European Alps. A slight reduction in performance is observed for the retrievals at 300 m (R = 0.85 and MAE = 0.18 m) and 100 m (R = 0.79 and MAE = 0.21 m). The results demonstrate the ability of Sentinel-1 to provide regional snow estimates at an unprecedented resolution in mountainous regions, where satellite-based estimates of snow mass are currently lacking. The retrievals can improve our knowledge of seasonal snow mass in areas with complex topography and benefit a number of applications, such as water resources management, flood forecasting and numerical weather prediction.
Hans Lievens et al.
Status: open (until 07 May 2021)
-
RC1: 'Comment on tc-2021-74', Anonymous Referee #1, 09 Apr 2021
reply
This paper builds on the work of Lievens et al., 2019 to extract snow depth from S-1 data in the Alps. As mentioned by the editor, this work is of high relevance to the snow community but also to many other research areas such as water management, tourism, climate change and biodiversity. I appreciate the work that is done here but in its current state, I cannot recommend this paper for publication since I feel there are too many unknowns and too much processing done on the S-1 imagery to be able to retrieve some sort of good quality snow information and give a proper assessment of the results shown here. This is reflected in my comments below.
Contrary to what has been stated by the authors in their response to the editor's comments, I am not skeptical of the relationship between the C-band signal and thick alpine snowpacks. I do question the physics of the approach used in this study and am concerned about the multiple layer of data smoothing in order to get good correlations with modelled data.
If the authors are willing to provide more information on the imagery processing and modify it to make it more physically accurate, I strongly believe this work has great value to the scientific community.
General Comments:
As mentioned above, I do agree with the authors that the cross-pol channel of S-1 can be sensitive to a thick snowpack but I disagree with the physical explanation of the authors. The physical interaction of the microwave signal with the snowpack is very complex and is not solely related to surface/volume scattering and single/double bounce. With snow layer thicknesses close or smaller than the wavelength, you have many interference and coherence effects in the signal. Recent work has shown that volume scattering and depolarization of the SAR signal comes mostly for the snow anisotropy (Leins et al., 2016) and the vertical/horizontal structuring of the snowpack at C-band. This can be achieved by a stratified snowpack horizontally or with snow grains that are structure vertically/horizontally through metamorphic processes. I would agree that with a thicker snowpack, chances are you will get more anisotropy but this is not shown with in situ measurements, temporal analysis or snowpack stratigraphic information.
With all the processing done to the SAR imagery, it is impossible to assess the physical interactions of the SAR signal with the snowpack since the data has been smoothed multiple times and transformed radiometrically and geometrically. You have multi-looking (averaging 10x10 pixels), border noise removal, thermal noise removal, terrain correction and reprojection to the WGS84 projection. The multi-looking is especially concerning given the topographic complexity of the Alps. It is smoothing all the topographic information (which is crucial for snow retrievals) and emphasizing only the areas of significant snow (snow drifts) which is not representative of a 100m grid cell in the Alps. Then you add incidence angle correction using a DEM (30m) that is of lower resolution than the pixel spacing (10m) of the original image. A DEM with similar resolution should be used but also, the topographic information has already been altered from the multi-looking which is not representative of the local topography. Then there's temporal averaging (Eq.2) which alters the signal even further. Finally, outliers are replaced by a 12-day average to smooth the data once more.
Further on the processing, I would avoid talking about sigma-nought when Eq. 1 converts the sigma-nought into a pseudo-gamma-nought multiplied by cos(40). I say pseudo here because the incidence angle used to convert sigma-nought is the 100m reprojected angle and not the gamma-nought values from the SAR imagery calibration.
If we accept the processing chain of the SAR imagery, it is still unclear that what the correlations are showing is linked to the snow depth. The errors obtained from the SAR retrievals (Figure 11) are most of the time larger than the precision of the reference data which is the model simulations. It is very difficult to determine that the correlations are statistically significant in this case and also looking at Figure 10, most of the comparison points are grouped around 0 which tends to falsely boost the correlation.
Given that modelled data is often smoothed and often have difficulty capturing extreme snow conditions and that the SAR data has been smoothed many times and outliers replaced by temporal means, I can’t say I am surprised to see a good empirical relationship.
Also, asking scientists to identify themselves in order to get access to the data used in this study does not comply with the open data policy.
Specific comments:
P.3L.5: I would disagree with the claim that an increase snow depth automatically causes an increase in volume scattering. If their is not sufficient anisotropy in the snowpack, there will not be any volume scattering in C-band. The theory will show that even if you increase the snow depth and keep all other snowpack parameters constant, you will not have a significant increase in volume scattering
P.3L.6: Again, this comment is highly dependent on the stratigraphy and anisotropy of the snowpack. This section needs to be supported by snowpit measurements of the studied area or referred to past work done in the area analyzing the snowpack properties.
P.3L.7: This comment is most likely true for the studied area but again, no reference or field measurement is provided to support this claim.
P.3L.9: Again here, I strongly disagree with this claim. The microstructure, anisotropy changes and stratigraphy, especially in the bottom layers of the snowpack will most likely drive the changes in sigma0.
P.3L.30: Even though this is common processing of SAR imagery, this is considerably altering the SAR signal, considerably smoothing it and making it very difficult to link to any ground snow properties.
P.3L.32: Multi-looking (or block averaging here) is a good way to reduce speckle noise in flat terrain. Here though, the topography is very complex (as mentioned by the authors) and it is emphasizing on the geometric distortions and the areas of significant snow (snow drifts) which is often not representative of a 100m grid cell in alpine areas.
P.4L.10: Using "local" incidence angle correction on a multi-looked image is not an accurate method. A DEM with similar resolution as the raw image should be used to correct for local incidence angle before multi-looking.
P.4L.15: This relationship was developed for areas of flat terrain and is not representative of the studied area. Proper analysis of the backscattered signal as a function of local incidence angle needs to be conducted in alpine areas in order to find the proper normalization relationship. A before and after image should show that this is not normalizing the image properly. Also, this is exactly taking sigma-nought and converting it to gamma-nought and then multiplying it by cos(40).
P.4Eq.2: Here again, temporal smoothing of the data. There's no way of linking the spatio-temporal snow properties of the original SAR imagery.
P.4L.27: Excluding March to July is very subjective here. First, it is removing a lot of snow properties variability which can occur in March. Anisotropy and stratigraphy is stronger in the later winter season. Second, with climate change, we know that wet snow is detected outside of this period.
P.4 L.30: This is not rigorous. Removing outliers is another method to smooth out the data and get better correlation with modelled data. But here they are not only removed, they are replaced by a smoothed average.
P.5 Eq.5: Is A applied to the ratio or only the cross-pol channel?
P.6L.17: I appreciate this approach where the index varies in time but I feel like the threshold is still limiting. I would see a temporal analysis of the SAR signal through multiple years to try and identify the proper threshold.
P.6L.25: Again, the February start is very subjective as wet snow conditions can be detected earlier and the September-November period is most likely to be the period where you have the highest backscatter and all the values that are 3dB below might be because of small surface moisture or percolating water which is not uncommon in Alpine snow.
P.11L.7: There is no mention of layering and anisotropy which is most likely the main reason of signal backscattering of dry snowpacks.
P.11L.11-13: These comparisons do not really apply to the current studies. As was mentioned by the authors in the response to the editor these studies were conducted in shallow snow conditions in tundra/taiga landscapes.
P.11L.20: This is a strong assumption since in alpine regions you can have strong surface roughness that will depolarize your signal.
P.11L.33: This is normal since most of the volume scattering and depolarization will come from the forest cover. For this study, I would have masked out the forested areas because this adds unnecessary complexity to a study that is already complex. Masking the forested areas would allow to focus on the snow retrieval without getting confused in multiple empirical relationships and heavy data processing.
-
AC1: 'Reply on RC1', Hans Lievens, 15 Apr 2021
reply
We thank the reviewer for the detailed feedback on our manuscript. We will address all the comments in detail during the revision. However, there are number of comments, especially on the S-1 processing, on which we disagree and which we address here to stimulate the open discussion.
- Reviewer comment: “With all the processing done to the SAR imagery, it is impossible to assess the physical interactions of the SAR signal with the snowpack since the data has been smoothed multiple times and transformed radiometrically and geometrically. You have multi-looking (averaging 10x10 pixels), border noise removal, thermal noise removal, terrain correction and reprojection to the WGS84 projection. The multi-looking is especially concerning given the topographic complexity of the Alps. It is smoothing all the topographic information (which is crucial for snow retrievals) and emphasizing only the areas of significant snow (snow drifts) which is not representative of a 100m grid cell in the Alps. Then you add incidence angle correction using a DEM (30m) that is of lower resolution than the pixel spacing (10m) of the original image. A DEM with similar resolution should be used but also, the topographic information has already been altered from the multi-looking which is not representative of the local topography. Then there's temporal averaging (Eq.2) which alters the signal even further. Finally, outliers are replaced by a 12-day average to smooth the data once more.”
Author response: We strongly argue for the opposite: Useful snow information can only be obtained if the processing of the S-1 data is adequate, and our processing is conforming the state-of-the-art. There are several steps involved in the S-1 data processing, but none of these steps involves ‘smoothing’:
- Border noise removal and thermal noise removal are very basic and standard procedures that are recommended by any literature source or handbook, and for any application that uses S-1 backscatter data.
- The data is corrected radiometrically for the local incidence angle impact, similar to the way gamma0 is calculated. This appropriately reduces the impact of the local incidence angle and therefore will better reveal the relationship between backscatter and snow depth. Note this is a rescaling rather than a smoothing operation.
- The data was geometrically corrected by range-Doppler terrain correction, which is also a standard processing step, especially in terrain with complex topography, that improves the geo-location of the radar measurements.
- We believe there is a misinterpretation of Eq. 2. This equation explains the bias-correction of the backscatter data by the rescaling of the mean and standard deviation. This is again not a smoothing but a rescaling step. We moreover strongly recommend such rescaling for any application that aims at combining measurements from different relative orbits of S-1.
- In summary, we are strongly convinced the above-mentioned processing steps are fully conform with the state-of-the-art.
The reviewer also mentions that the multi-looking to 100 m is especially concerning. We are surprised by this statement. The multi-looking effectively reduces the pixel spacing of the backscatter measurements from 10 m in the original S-1 data (which is below the ~20-m spatial resolution) to 100 m in the multi-looked data. The result is thus similar as if one would have an instrument that measures backscatter at a native pixel size of 100 m, but with reduced noise (e.g., speckle).
- If the 100-m scale is problematic to retrieve snow depth according to the reviewer, what is then the take on novel satellite mission concepts, such as dual-Ku band SAR, that propose resolutions up to 500 m?
- The multi-looking is not only applied to reduce speckle noise, but also to keep the computation time for the processing and the data storage feasible. Our intention is to perform a consistent processing also at the larger scale, including other mountain regions and the full S-1 archive. Such processing would no longer be feasible using the high-performance computer that we have access to at a further reduced pixel spacing.
- Reviewer comment: “The errors obtained from the SAR retrievals (Figure 11) are most of the time larger than the precision of the reference data which is the model simulations. It is very difficult to determine that the correlations are statistically significant in this case and also looking at Figure 10, most of the comparison points are grouped around 0 which tends to falsely boost the correlation.”
Author response: Figure 11 does not show the accuracy of the model simulations, but the accuracy of the S-1 retrievals with respect to the in situ snow depth measurements. We did not show the validation of the model simulations in this study, in order to focus on the validation of the S-1 retrievals. Furthermore, the model simulations of OSHD are including the assimilation of in situ measurements, and can therefore not be independently validated using these same measurements. We are surprised that the reviewer questions the statistical significance of the time series correlations shown in Figure 11, which are mostly higher than 0.8 for sites reaching snow depths above 1 m. We agree that the inclusion of zero snow depths can slightly increase the correlations. Therefore, Figure 5 shows time series correlations (against model simulations) both with and without the exclusion of zero snow depths. Even though more data are clustered around low snow depths in Figure 10, the density plots clearly demonstrate the overall agreement between the S-1 retrievals and the in situ measurements also for the high snow depths, especially for the coarser 300 m and 1 km retrievals.
-
CC1: 'Reply on AC1 regarding', Joshua King, 19 Apr 2021
reply
The following comment was brought to my attention and I would like to respond as a community member involved with Ku-band SAR:
‘If the 100-m scale is problematic to retrieve snow depth according to the reviewer, what is then the take on novel satellite mission concepts, such as dual-Ku band SAR, that propose resolutions up to 500 m?’
The Terrestrial Snow Mass Mission (TSMM; a proposed dual-Ku bad SAR) in Phase 0 identified the need for higher resolution sensing in Alpine environments. A 50-m mode was proposed, trading swath width for improved resolution where canopy closure was high, or topography was complex. Course resolution 250 and 500-m modes were reserved for lower-complexity domains, generally exclusive of Alpine watersheds, addressing the need for high temporal revisit in downstream applications. Discussions about resolution will be critical moving forward, but the complexities involved with standard modes of TSMM as applied to Alpine environments have been acknowledge in Phase 0.
-
AC1: 'Reply on RC1', Hans Lievens, 15 Apr 2021
reply
-
RC2: 'Comment on tc-2021-74', Anonymous Referee #2, 20 Apr 2021
reply
The authors present an application of a change-detection algorithm to estimate SWE in the Alps using Sentinel-1 C-band SAR. They explore the effect of spatial resolution on their retrievals. This is an important and timely contribution, and should be of great interest to the community. The paper is well-written so I have very few minor comments. Instead, I’ll focus on a really key point which is that I think there is a great chance for readers to misunderstand the maturity level of the algorithm, based on how the paper is presented. This review is five related major comments that unpack this idea.
Major Comments
First, I do not think that the paper adequately reflects the fact that we still do not understand why this method works, even at a basic level. The manuscript instead makes it sound clear that the mechanisms are understood: e.g. in the introduction, page 2, lines 32-page 3, line 2. Taking their points one by one: to their first point (page 2 line 33), no reference was given, and no reason why having lower ground backscatter would change sensitivity to depth; to their second point (page 2, line 33), Chang et al. 2014 do not make this point, that I could see. Readers will assume after reading the introduction that it is obvious why the C-band cross-pol is correlated with snow depth, which is not true. In fact, the authors of this study only introduce the idea that the “physical mechanisms that cause this increase are still uncertain” in the Results & Discussion section (page 10, line 13). Please, bring this critical point into the abstract, introduction and conclusion!
Second, I think it is critical to communicate more clearly throughout that this is an empirical algorithm with calibration parameters that require known SWE data over the domain. The word “empirical” needs to appear in the abstract, in my opinion. Please somehow get this idea into the introduction, abstract, and conclusion.
Third, the authors need to point out that the algorithm only works well if you have accurate SWE data to calibrate against. Indeed, they need to just note explicitly that the accuracy of the approach they are using here is limited to the accuracy of their training data. I think this needs to be presented explicitly in the abstract and conclusions, to avoid reader misunderstanding.
Fourth, the authors should point out that in this study, they are calibrating here against very accurate model results. Here, they are applying the algorithm in this study over a domain where (in my opinion) the most accurate model results are available anywhere in the world. There is no other mountain range, to my knowledge, with the density of observations available in the Alps. Further, globally available model results in mountain ranges are inadequate for most applications, in terms of their spatial resolution and accuracy. See e.g. Mortimer et al. 2020. I think this needs to be mentioned in the conclusions.
Fifth, the authors need to acknowledge explicitly that the first four points mean that you could not use this approach globally, calibrated to models, and achieve the kind of results shown here; this point almost certainly will be lost on readers of the abstract alone. This is a major issue with the manuscript that needs to be addressed in the abstract and conclusions.
I hope the authors do not misinterpret any of these comments: they have done an amazing job uncovering this important new dataset. It has very important possible applications. Reworking the way the paper is presented should help the community get on board with this new dataset as quickly as possible.
Mortimer, C., Mudryk, L., Derksen, C., Luojus, K., Brown, R., Kelly, R., & Tedesco, M. (2020). Evaluation of long-term Northern Hemisphere snow water equivalent products. The Cryosphere, 14(5), 1579–1594. https://doi.org/10.5194/tc-14-1579-2020
Hans Lievens et al.
Hans Lievens et al.
Viewed
HTML | XML | Total | BibTeX | EndNote | |
---|---|---|---|---|---|
423 | 181 | 6 | 610 | 1 | 1 |
- HTML: 423
- PDF: 181
- XML: 6
- Total: 610
- BibTeX: 1
- EndNote: 1
Viewed (geographical distribution)
Country | # | Views | % |
---|
Total: | 0 |
HTML: | 0 |
PDF: | 0 |
XML: | 0 |
- 1