the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
A new method for large scale snow depth estimates using Sentinel-1 and ICESat-2
Abstract. Knowledge about seasonal snow accumulation is important for managing water resources, but accurate estimates of snow depth at a high spatiotemporal resolution are sparse, especially in mountainous regions. In this paper, we outline a novel approach to estimate snow depths using Sentinel-1 C-band synthetic aperture radar (SAR) and ICESat-2 LiDAR observations. Specifically, we estimate snow depths at 500-meter spatial resolution by correlating increase in Sentinel-1 volume scattering with snow depths derived using ICESat-2. Sentinel-1’s vast spatial coverage and frequent 6/12-day revisit cycle makes it promising for monitoring seasonal snow accumulation, but capturing the volume scattering signal within a dry snowpack and relating it to snow depth remains challenging. Using ICESat-2, we retrieve thousands of high accuracy snow depth observations covering the Southern Norwegian Mountains. ICESat-2 has a low revisit time of three months, but by matching observations with the temporally nearest Sentinel-1 scene, we significantly enhance spatiotemporal resolution. Our results demonstrate that our ICESat-2 calibrated Sentinel-1 snow depths can estimate snowfall magnitudes in deep dry snow (>0.6 meters), achieving an accuracy of 0.5–0.7 meters, significantly improving estimates made by the SeNorge snow model in remote mountainous regions. This study highlights the potential of utilizing ICESat-2 to derive Sentinel-1 snow depths, improving snow monitoring capabilities in data-sparse regions.
- Preprint
(4379 KB) - Metadata XML
-
Supplement
(2513 KB) - BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on egusphere-2024-3850', Simon Gascoin, 02 Mar 2025
This article introduces a new method to calibrate spatially-continuous Sentinel-1 snow depth retrievals using sparse but presumably more accurate ICESat-2 snow depth retrievals. I think it is a great idea and to my best knowledge it has never been published elsewhere. The manuscript reads well and the results are clearly presented. The main issue in my opinion is that the authors did not compare the performance of their calibrated S1 snow depth to the standard product by Lievens et al. Therefore it is difficult to know if it is worth the hassle to process gigabytes of ICESat-2 data. Another issue in this study is that (if I understood correctly) a subset of ICESat-2 data was used to evaluate the S1 snow depth retrievals whereas the same data were also used for calibration. These ICESat-2 tracks should be excluded from the calibration data to draw robust conclusion.
Minor comments/suggestions
L39: snow models forced by numerical weather models can provide the SWE directly unlike remote sensing methods
L43: the three references refer to airborne/terrestrial lidar. I guess the sentence should read “Accurate snow depth estimations from ground-based lidar”.
L43: “laser altimetry (LiDAR)”: Lidar is another acronym. NB) « LiDAR », « lidar » to homogenize.
L46: I think that the literature review should mention previous attempts to retrieve snow depth from ICESat-2 after the study by Treichler and Kääb with ICESat 1.
L50 “ICESat-2 has similar limitations” in fact ICESat-2 is a huge step forward compared to ICESat!
L68 “Our Sentinel-1 methodology” : maybe “Our methodology” reads better
L74 « covering a one » delete « a »
L91 “The Sentinel-1 radar system transmits waves of a single polarization (horizontal or vertical) but receives both polarizations. Over most land areas, Sentinel-1 transmits vertically polarized waves, yielding co-polarized (VV) and cross-polarized (VH) images” a bit unclear what is transmitted
L119: Can you clarify why masking vegetated areas solves the issue of soil moisture changes?
L121 “> 0.3”
L124: Forests should already be masked by the previous NDVI thresholding.
L145: Can you describe a bit further the “customized surface height algorithm”? In particular, what are the advantages against other, higher level, products like ATL06. In addition, it would be interesting to show an example of ATL03 point cloud and the extraction of the surface elevation.
L151 “(Geonorge, 2024)” the reference is insufficient: indicate a web site, a documentation. Are the acquisition dates known? Is it sure that the ground was snow-free?
L161 “S1”: figure ?
L152: Can you estimate the fraction of the study area that was excluded after all these criteria? It seems very restrictive but maybe not so much… In addition it could be useful to mention in the abstract that the study focused on relatively flat open terrain. Also, I could be useful to provide somewhere the area of the study area.
L160: Specify which parameters exactly
L160: “were” why do you switch to the past tense?
L165: can you clarify if these statistics were obtained over the same area as the snow depth retrievals, i.e. after masking vegetated areas, lakes, slopes > 10° etc.
L182 “between January – April, 2019-2021” maybe clarify: from, to…
L192 Figure 4. It would be interesting to color the points according to their date to better understand how the temporal aggregation of multiple products influence the regression.
L220: The first paragraph of the Eesults reads like the presentation of the study area. I suggest to move it to Sect. 2
L323 “snow. (B “ delete the “.”
L325: additional evaluation studies were published recently, please check if you think they can be useful to your discussion :
Bulovic et al. 2025 WRR https://doi.org/10.1029/2024WR037766
Sourp et al. 2025 HESS https://hess.copernicus.org/articles/29/597/2025/
L332 : “snowfall” Do you mean snow depth?
L332: can you clarify “limitation persists in Sentinel-1’s ability to capture the snow variability”. Do you mean spatial, temporal variability?
L339: “if you only have” I think it should be rephrased (colloquial)
L341 “biased”
L355 it is rather 2 km resolution, and 500 m the ground sampling distance
Conclusion: do you think that a perspective could be to develop a temporally varying calibration instead of a single regression for all dates?
Citation: https://doi.org/10.5194/egusphere-2024-3850-RC1 -
RC2: 'Comment on egusphere-2024-3850', Désirée Treichler, 11 Mar 2025
Combining Sentinel-1 and ICESat-2 data is an interesting idea that has the potential to mitigate some of the shortages that these data sources have when used separately, and especially compared to combining Sentinel-1 with sparse weather station data only. This is absolutely worth exploring, and readers will be interested to learn whether this approach could be useful for their region of interest. My main comments are:
- How much improved/different are the presented results from the approach presented by Lievens et al. (2019)?
- The results part of the study focuses on validation and less so on an analysis of the generated snow depth maps themselves. The validation strongly relies snow amounts and the seNorge dataset, which is well known to (consistently) overestimate snow depths at high elevations and not a suitable/sufficient validation dataset in the way it is used here. The conclusion that seNorge data is wrong is a sidetrack, and the actual performance and limitations of the presented method/output snow depth maps remain unclear. Do the generated Sentinel-1 maps correctly reproduce spatial patterns and the evolution of the snow pack over time?
- ICESat-2 snow depths are a similarly new method for snow depths as Sentinel-1 but here introduced as reliable ground truth in a fairly uncritical way. I suggest a more critical consideration of the potential and limitations of both ICESat-2 and Sentinel-1 data for the purpose of snow depth estimates throughout the manuscript, and workflow.
- What about uncertainties? (Both in the input data and output snow depth maps?) Currently nowhere presented or discussed in the manuscript, but I found some figures in the Supplement. Uncertainties/errors should get higher priority in the workflow and discussion.
- Visualisations: I suggest replacing satellite imagery backgrounds with neutral backgrounds in all map figures where coloured foreground content is presented. The background noise interferes with the presented data/colours, which makes it unnecessarily hard to read and interpret the figures.
---
Abstract:
- L16-21: the method could be presented in a more logical order for increased readability, and some phrasings are unclear: “remains challenging” is confusing here (have you solved this or not?), what does “high accuracy” mean (and can you really claim that), and what is meant by “significantly enhance spatiotemporal resolution”?
- Readers will want to judge whether they can apply this method in their area and for their purpose: What is the spatial and temporal resolution and extent of the output data? Uncertainties and limitations of the method and resulting snow depth maps? Accuracy stated in L23: compared to what, how was this validated?
- L22, “estimate snowfall magnitudes”: misleading, do you mean volumes? I.e., positive snow volume changes (after a snowfall event)? This shows up several places in the manuscript.
- L23, “improving” sounds like you are actually improving/employing this model, please rephrase.
Introduction:
---Snow depth, volume, and SWE seem to be used interchangeably here, but they are not the same
L50: ICESat-2 is introduced in a rather uncritical way without mentioning that this satellite has been used for snow depth estimates. There are both validation studies and studies where this data is used in combination with models or other data. The spaceborne lidar snow depth review by Fair et al. (in review) and references therein may be helpful: https://doi.org/10.5194/egusphere-2024-3992
L57: to my knowledge, Lievens et al (2019) only studied C-band radar
L71ff: review - maybe less details are sufficient/less confusing to introduce the scope of the validation (move details to later) - here, locations like Hardangervidda and the meaning of ICESat-2 “tracks” are not yet introduced.
Fig. 1: I suggest you
- increase font sizes (flow chart, map legend and scale bar)- Panel a): it might be helpful for the reader if you indicate snow depth (distance between snow top and ground). Green line on top of the snowpack: misleading? ICESat-2 doesn’t measure continuously, and three beams are indicated. Sentinel-1: I suggest to limit the volume scattering indicators to the area under the sensor - not all snow is contributing to the yellow arrows.
- Panel b): you could distinguish between the linear regression from a subset of points, and its subsequent application to all of Sentinel-1 data
- Panel c), small inlet map: consider zooming out to show sufficient spatial context for non-Scandinavian readers
- Caption: What do the orbits stand for - imagery every 6th day each (they are overlapping)? ICESat-2 RGTs are misleading: only one theoretical path rather than six profiles, not considering off-pointing (all repeat paths were shifted in space) and data gaps (clouds, especially in the west). I suggest showing the actual data coverage/density.
Data and Methods
---I am missing a general description of the study area including its snow climatology, spatial patterns, magnitudes etc. L220ff would fit here and can be extended.
L87: I suggest you avoid abbreviations in section titles
L88ff: outdated (2-satellite again, now with 1C)
L94ff: introduce the concept/mechanism of volume scattering (rather than “information about surface properties”), and make sure you clearly distinguish volume scattering, ratios and changes thereof. Why is VH more sensitive to volume scattering from snow (=snow volume changes) than VV?
L104f: clarify that this is on the provider’s side, not your pre-processing
L110ff: individually for each (hydrological) year?
L114ff: the first day of snowfall varies throughout the area and with elevation (and hydrological year). Is this per pixel? What time period(s) do you end up with?
122ff: wet snow (and water) don’t fit the context of soil moisture/vegetation, explain
130ff: are data from orbits mixed, or all analyses (and CRreference) computed separately?
Figure 2: the map inlet in the right panel is maybe not needed. Colours will be better visible with a neutral background (e.g., hillshade or shoreline contours). Increase the legend size.
Caption: I suggest you refer to land and snow cover products rather than optical satellite imagery (suggesting you classified them)
L144: Introduce ICESat-2 and its data (track geometry, revisit time)
L145ff: “customized surface height algorithm”: please explain, is there any reference/code for that? Why do you use ATL03 data rather than ATL06/ATL08 or sliderule, as used by other studies? It seems strange when readily spatially averaged products at similar spatial resolution are available, and you later average to 2km segments (or some places to 500 m, e.g. Fig. S1). How are photons filtered (noise)? Co-registration? Clarify for the reader how you derive snow depths (differencing, how do you sample the DEM) and how your final snow depth sample looks like (one point every 10 m along-track..?). How accurate are these samples, do you filter out unreasonable values (large deviations visible in Fig. S1 in the supplement)? Are you using the same land mask as for Sentinel-1 data?
L157: the references you cite rather suggest that ICESat-2 snow depths need to be averaged at catchment-scale, with decimeter- to meter-scale bias/uncertainty in sloping terrain - after spatial co-registration.
L163ff: Is this bias consistent in space, with elevation, slope, and over time? Did you remove it?
L167: distinguish between tracks (RGT?), overpasses (may be a better word than tracks) and profiles (six per overpass). “Snow depth point measurements”: at what resolution - are these segments rather than points? Consider the terminology throughout the manuscript.
Figure 3: Mixing all years and months in one plot is misleading and introduces unnecessary noise. Snow depths are expected to change over time, and the reader is made to believe that all of this data is available for all points in time. I suggest you add at least one more panel where you show data from one single month.
2.3 Linear regression/ Fig. 4: -
Is this relationship consistent over time and in space? The study area exhibits a strong East-West snow depth gradient and the snow pack changes over time. Are temporal data pairs forming visual clusters in Fig. 4b?- This is not a 1:1 relationship - how do you deal with uncertainties?
- Temporally nearest: How many days at maximum? Do the examples in Fig. 4a) really only have one overpass each that fits? The ICESat-2 sampling pattern yields parallel tracks/profiles that gradually move East/West across the study area over the course of a few days, and for this large area there will be several overpasses that are close in time to the Sentinel-1 acquisition time - but in different locations.
- Fig 4a): ICESat-2 snow depths (overlapping circles) are not distinguishable. Maybe two distinct colour scales could work, or only show track locations.
- How about uncertainties?
L203ff: Which version of the seNorge model are you using? As far as I recall there are V1, V2, and seNorge_2018.
L211ff: Why these two overpasses? Were these data also used to establish the regression?
L214: refer to Fig. 8
Results
---Figures 4 and 5 suggest a far coarser spatial resolution than 500m
L235: comparison with modelled data, not the model itself
ff: The fact that seNorge overestimates snow depths is well known and has been shown by many other studies, including validations done by NVE (the seNorge data provider) as you cite, or from the extremely coarse ICESat data (predecessor of ICESat-2, Treichler and Kääb, 2017), showing that it is really not difficult to do better on absolute snow magnitudes in some areas. The area picked in Figure 7 shows extreme deviations. Could it be that in other parts with lower snow depths the differences would be much smaller - and Sentinel-1 might do worse…? This comparison therfore seems hand-picked and seNorge absolute snow depth values are obviously not reliable ground truth. Instead, wut what about relative patterns in space, time, elevation, for different slopes and aspects, how does your data do there?
Fig. 7: I suggest you show snow depths (both datasets) rather than deviations from seNorge data
L261ff, Table 1: I would expect to see time series plots or temporal analyses here. How comes that the bias changes each year -is it consistent within one year? And across the entire area? (seNorge is overestimating, yes, but apparently consistently?) Which orbit/regression model is used, and how do the other two datasets/regression models compare? Are there differences in space / for different weather stations? It is hard to understand how the measures are computed (over time? in space? "averaged"/"aggregated" how (table caption)?) and what they stand for. What does the "p-value" mean?
L305ff: fieldwork methods description, move to methods section
Figure 8:
- panel a: scales?
- panels b/c: show variability/uncertainties - the variability of the field measurements should provide an approximation (many measurements in one cell), and for Sentinel-1 data you need to come up with an uncertainty estimate
- panel d: show the result grid or size/location of the corresponding output map cell(s), for reference
- Sentinel 1 data seems to show a steeper decline in snow depth than seNorge/weather station data (panel c). That's interesting, how comes? Is this a consistent pattern across all the weather stations?
Section 3.4: I suggest you introduce a separate discussion section to make room for more critical reflection of your method and data.
L327f: how do you derive this 60 cm threshold? (If you only use data from January through April for areas above the three line, how much data with <60 cm snow depths is there..? The scatter plot y axis suggests barely any)
L333f: I don't understand how you capture variability (or the lack thereof) from Fig. 7 and your results in general
L348ff: This is not convincing, I believe the transferability of this method is far lower than what is suggested here. Southern Norway hosts an old mountain range that has comparably many flat areas above the tree line, with extended periods of dry, deep snow, and a far denser ICESat-2 pattern than mountain areas further south. Likely, both Sentinel-1 and ICESat-2 data are prone to more bias and uncertainties for steeper and warmer mid-latitude mountains. Also, despite repeat coverage (+/- tens of meters shift of ground tracks) in polar areas a DEM is still needed for areas that are not completely flat.
Conclusion:
---L355: 500m: is this the actual spatial resolution, or just the grid size (oversampled)?
The conclusion contains several measures/values that I can't clearly locate in the manuscript and that rather oversell the method/results.
LL366ff: Which inherent limitations? Timely SWE (rather than snow depth) estimates are indeed critical for, among others, hydropower management, but are these data really useful for this purpose given considerable bias (presumably inconsistent over time, and space)?
References:
---L467, NVE / seNorge: unclear reference, what is this for? A specific dataset?
Citation: https://doi.org/10.5194/egusphere-2024-3850-RC2
Viewed
HTML | XML | Total | Supplement | BibTeX | EndNote | |
---|---|---|---|---|---|---|
227 | 105 | 8 | 340 | 24 | 8 | 8 |
- HTML: 227
- PDF: 105
- XML: 8
- Total: 340
- Supplement: 24
- BibTeX: 8
- EndNote: 8
Viewed (geographical distribution)
Country | # | Views | % |
---|---|---|---|
United States of America | 1 | 114 | 34 |
Norway | 2 | 32 | 9 |
China | 3 | 30 | 9 |
France | 4 | 22 | 6 |
Denmark | 5 | 21 | 6 |
Total: | 0 |
HTML: | 0 |
PDF: | 0 |
XML: | 0 |
- 1
- 114