the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 26 Aug 2026
No significant trends in snow free days at three long-term monitoring sites in the high-, mid- and low- Arctic Greenland as seen by MODIS Terra from 2000–2025
Abstract. Snow cover is a key component of Arctic ecosystems, regulating both biotic and abiotic processes such as plant phenology, soil thermal regimes, greenhouse gas fluxes, and the timing of glacier melt. We present a 26-year (2000–2025) record of snow-free days derived from the MODIS Terra cloud-gap-filled snow cover product at three long-term monitoring sites in high, mid, and low-Arctic Greenland. The MODIS-derived counts correlated well with in situ observations, most strongly at the high-arctic site, the only location where a direct comparison to in situ observations was possible. The mean snow-free season length is 76, 78 and 103 days at high-, mid-, and low-Arctic respectively, with a standard deviation of approximately two weeks at each site despite their contrasting latitudinal positions. We find no statistically significant trends in the number of snow-free days at any site over the record, consistent with the absence of significant trends in Northern Hemisphere snow cover in the Greenland and North American sector over the same period. The dominant driver of interannual variability shifts along the latitudinal gradient: snowfall was the dominant predictor of snow free day counts at the high arctic site, positive degree-days become co-limiting alongside snowfall at the low-Arctic site and at the mid-Arctic site positive degree-days are the only significant predictor, allthought snowfall and positive-degree-days together explain little of the interannual variance there, which point to additional physical controls. The high interannual variability in snow-free season length underscores the necessity of monitoring records spanning multiple decades in order to detect climate-change signals in snow cover and the many Arctic ecosystem processes that depend on it.
- Preprint
(5355 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
- RC1: 'Comment on egusphere-2026-3962', Anonymous Referee #1, 22 Sep 2026
-
RC2: 'Comment on egusphere-2026-3962', Mathieu Gravey, 09 Oct 2026
General comments
First, I would like to thank the authors for clearly stating their main result directly in the title rather than presenting it as a question. I particularly appreciate their willingness to report the absence of statistically significant trends. While identifying a statistically significant trend is relatively straightforward when sufficient evidence is available, interpreting a non-significant result is more challenging, as it does not necessarily demonstrate the absence of a trend. I therefore appreciate the authors' decision to present and discuss this result, which I consider scientifically valuable.
Overall, I find the study interesting and relevant. My comments below primarily concern methodological clarifications and a few suggestions that could improve the presentation and interpretation of the results.
Specific commentsSection 2.2 – MODIS Terra CGF snow cover maps and snow-free days
Choice of MODIS Terra (lines 53–60)
Why was MODIS Aqua not considered alongside MODIS Terra? This is not a request to extend the analysis, as there may be a good reason for this choice, particularly regarding the consistency of the long-term record. However, a short sentence explaining why only Terra was used would be welcome.
NDSI and polar night (lines 78–80)
The authors write:
"Since the NDSI depends on the visual spectrum it is not possible to obtain NDSI from MODIS during polar night."
I would normally use "visible spectrum" rather than "visual spectrum". More fundamentally, I would explain this limitation in terms of passive versus active remote sensing. The issue is that MODIS relies on reflected solar radiation and therefore requires sunlight, unlike an active sensor. I think this would provide a clearer physical explanation.
NDSI expressed as a percentage (Section 2.2)
I wonder whether expressing NDSI thresholds as percentages is the most appropriate choice. Although the MODIS product scales NDSI values from 0 to 100, NDSI is fundamentally a normalized spectral index rather than a percentage. Expressing it in percent could give the impression that it represents a directly proportional physical quantity, which is not the case.
I would suggest clarifying this distinction or using dimensionless NDSI values throughout the manuscript.
Line 83 – Typographical error
"thresholds of of 10" should be corrected to "thresholds of 10".
Reprojection of MODIS data (lines 72–75)
Out of curiosity, is there a particular reason for reprojecting the MODIS data to local UTM coordinates rather than performing the analysis directly in the native MODIS sinusoidal projection? Given that much of the analysis is based on pixel-level statistics, I would have expected that working in the original projection might simplify the workflow. I do not expect this choice to significantly affect the results, but I would be interested to understand the motivation behind it.
Section 2.3 – In situ data
Line 98 – Albedo threshold
A brief justification for the albedo threshold of 0.4 used to identify snow-free conditions would be welcome. I assume that this threshold could be calibrated or at least evaluated against the available observations. Is there a specific reason for choosing this value?
Figure 5 – Elevation distributions
I wonder whether cumulative elevation distributions would be easier to interpret than the current representation, as they would directly indicate the fraction of the area located below a given elevation.
However, this is mainly a matter of presentation, and I have no strong objection to keeping the current figure if the authors consider it more appropriate.
Section 2.4 – Reanalysis data
Line 104 – Gap-free records
The authors state that they require "gap-free records" for the analysis. I understand the need for a consistent dataset covering the full study period, particularly because the in situ records do not extend back to 2000 at all sites. However, I wonder whether "gap-free" is really a requirement, since most statistical methods can accommodate missing observations. Perhaps the authors could clarify that the main motivation is to ensure a consistent temporal coverage across all sites rather than a strict statistical requirement.
Lines 114–116 – Elevation-related bias in CARRA
The authors acknowledge that the absence of elevation downscaling may introduce a bias in absolute precipitation amounts but argue that this should not substantially affect the analysis of interannual variability.
I fully agree with this reasoning in the context of the study's conclusions. Such a bias could potentially introduce an artificial trend, whereas it seems much less likely to systematically eliminate an existing trend. Given that the main result is the absence of statistically significant trends, I consider this approach reasonable and do not request any modification.
Section 2.5 – Statistical analysis
Lines 138–139 – Software packages
I am not convinced that listing the Python packages used for the analysis is particularly informative from a scientific perspective, as equivalent implementations of the same statistical methods should produce comparable results.
If the intention is to ensure reproducibility, it would be more useful to provide the package versions through the shared code environment or dependency files in the GitHub repository.
Section 3 – ResultsGeneral comment on the statistical approach
I do not intend to discuss every statistical choice or hypothesis adopted in the manuscript, including the selection of statistical tests, null hypotheses, or significance criteria.
There are many defensible approaches to such analyses, and different researchers might reasonably make different choices. I fully understand that methodological decisions must be made to move the analysis forward.
My comments therefore focus on aspects for which clarification is necessary to understand or assess the reported results.
Choice of the 40% NDSI threshold (lines 140–145)
The authors justify the selection of the 40% NDSI threshold based on the apparent convergence of snow-free day counts at lower thresholds and its correspondence to an SCF of approximately 50%.
To me, this reads somewhat like an a posteriori justification of the selected threshold.
Why were thresholds evaluated only in increments of 0.10? Given that in situ observations are available, I would expect that an optimal threshold could potentially be identified through calibration, or at least that the sensitivity to threshold selection could be explored more systematically.
Alternatively, a finer sensitivity analysis (e.g., increments of 0.01 or smaller) would be computationally inexpensive and could help assess the robustness of the results. While the authors already report that no significant trends were found at the tested thresholds, it would also be interesting to know whether the identified relationships with climatic drivers are similarly robust to the threshold selection.
I am not necessarily requesting a full optimization procedure, but I think the authors should provide a clearer justification for the threshold selection and discuss its influence on the results.
Table 3 – Interannual variability in snow-free days
The definition of "Interannual variability in snow-free days (5 yr)" in Table 3 is unclear. Does this refer to a rolling standard deviation computed over overlapping 5-year windows, variability calculated from five consecutive annual observations, or simply a trend calculated over a 5-year period?
Please clarify how this metric is computed.
If overlapping windows were used, temporal autocorrelation would be introduced by construction, even if the original annual observations can be considered independent. This could affect the significance of the Mann–Kendall test. In that case, please explain how this autocorrelation was accounted for in the statistical analysis.
Furthermore, Table 3 reports a statistically significant trend in this variability metric at Disko (p = 0.0063), whereas the manuscript emphasizes the absence of significant trends in snow-free days. Although these results are not necessarily contradictory, it would be useful to explain this distinction explicitly and discuss the physical interpretation of the reported increase in variability.
Final remarks
Overall, I appreciate the study and particularly its emphasis on reporting the absence of statistically significant trends. I believe this is an important result that deserves to be communicated. To be honest, given the substantial year-to-year variability already highlighted by the authors, I would not have expected to detect a statistically significant trend over the study period.
I hope these comments will help the authors
Citation: https://doi.org/10.5194/egusphere-2026-3962-RC2
Data sets
Snow cover fraction on land from MODIS at the three GEM sites (NDSI) Signe Hillerup Larsen https://doi.org/10.22008/FK2/HG6HY4
Snow free days from MODIS at the three GEM sites Signe Hillerup Larsen https://doi.org/10.22008/FK2/537JIP
Model code and software
GEM_snow_cover_from_MODIS Signe Hillerup Larsen https://github.com/GEUS-Glaciology-and-Climate/GEM_snow_cover_from_MODIS
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 190 | 113 | 66 | 369 | 63 | 67 |
- HTML: 190
- PDF: 113
- XML: 66
- Total: 369
- BibTeX: 63
- EndNote: 67
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
This manuscript has used a 26 year time series of MODIS data to analyse trends and variability in the number of snow-free days at three sites across Greenland. The authors have also attempted to determine which climatic drivers are responsible for the interannual variability in snow-free days at the different sites by fitting an ordinary least squares model. They conclude that there is large variability in number of snow free days but no significant trends over the time series analysed, but that the sites exhibit different climatic drivers of the variability. Overall it is an interesting paper that uses similar datasets/approaches to earlier work, making the results comparable and complementary to existing knowledge on how snow cover is changing in the Arctic.
General comments
While I don't think it was necessary to assess the number of snow-free days using NDSI thresholds > 50%, the majority of my comments relate to the analysis of the climate drivers. I also think that rather than only analysing the MODIS data at specific sites/areas across Greenland, it would also have been interesting to map the derived variables and their trends for Greenland as a whole so that any changes and trends (or lack of) can be assessed in the wider context. Moreover, the authors also emphasise in the last part of their discussion, the importance of monitoring trends in terms of ecological impacts – and mapping these variables over a larger spatial area would therefore allow potentially vulnerable zones to be identified.
Specific comments
L82-83: I am a bit confused about why it was necessary to test thresholds >50% for computing the number of snow-free days. In my mind it would be hard to argue that a grid cell can be considered snow-free if more than half of it is snow-covered. Indeed the authors also note that the higher thresholds are more relevant to identifying melt onset, making the results somewhat irrelevant for snow-free conditions.
L114: Snowfall is summed over a hydrological year but PDD is summed over a calendar year? Would this potentially impact the OLS model results if the time series are acquired from summing over different time periods?
Also I might expect correlations between snow-free days and snowfall/PDD to be different depending on the time of year. The rate of snowmelt would likely control the variability in number of snow-free days in the spring through timing of snow disappearance, and this depends on how much snow there is to melt and also the temperature, whereas in the autumn I might expect other factors to control the number of snow-free days through variability in snow onset date, which is maybe more dependent on temperature because this determines whether the precipitation actually falls as snow or not.
Have the authors considered fitting the OLS model for two parts of the year 1. the start of the hydrological year through to snowmelt, and 2. following snowmelt until snow onset (which can also be determined from MODIS)? One might expect that the roles of snowfall/PDD would differ during different periods of the year eg. PDD may steer the number of snow-free days more strongly when the ground is already snow-free (greater PDDs = lower likelihood of snowfall onset = more snowfree days) whereas during the snow-covered season perhaps both variables play more equal roles?
L131. Snowfall and temperature/PDD are two factors that influence the number of snow-free days but this view seems a little simplistic. What about precipitation as rain falling on an existing snowpack? ROS events are known to accelerate melt processes more than temperature alone so I would expect rain could be an important explanatory variable as well as PDD and snowfall. Moreover earlier studies mentioned in the results (Liu et al. 2024) have identified wind as a factor since it can redistribute snow – why not download this dataset as well since it is also available from CARRA? Perhaps the variables identified as the strongest drivers of snow-free days no longer remain the strongest if accounting for other meteorological variables such as wind and rainfall?