Bias-adjusted projections of snow cover over eastern Canada using an ensemble of regional climate models

Bresson, Émilie; Dupuis, Éric; Bourgault, Pascal

doi:https://doi.org/10.5194/egusphere-2025-3979

Preprints

https://doi.org/10.5194/egusphere-2025-3979

Preprints

10 Sep 2025

| 10 Sep 2025

Bias-adjusted projections of snow cover over eastern Canada using an ensemble of regional climate models

Émilie Bresson, Éric Dupuis, and Pascal Bourgault

Abstract. In the context of climate change, stakeholders and decision makers need easily accessible bias-adjusted projections of snow cover and indices produced from those to develop adaptation plans. To meet this need, we produced an ensemble of regional climate projections statistically bias adjusted of snow water equivalent (SWE) in the province of Québec, Canada. This bias adjustment required some fine-tuning to operational methods, mainly due to the seasonality in the SWE. We calculated indices of interest for several sectors based on the bias-adjusted SWE. These indices included the maximum of SWE as well as the duration, start, and end of the snow season, and the days without snow cover. In eastern Canada, snow cover tended to persist for shorter periods as the climate warms, with symmetrical shrinkage at the beginning and the end of the snow season, with the exception of the Nunavik region. The maximum SWE was projected to decrease in the southern part of the domain and increase elsewhere. The snow season in the Côte-Nord, southern Québec and St-Lawrence River Valley regions would be increasingly interrupted by sequences of days without snow cover, whereas this would not be the case for the northern and central Québec regions.

Received: 15 Aug 2025 – Discussion started: 10 Sep 2025

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.

Download & links

Preprint (PDF, 3630 KB)

Supplement (5438 KB)

Download & links

Émilie Bresson, Éric Dupuis, and Pascal Bourgault

Status: final response (author comments only)

RC1:
'Comment on egusphere-2025-3979', Anonymous Referee #1, 09 Oct 2025
Review of the paper “Bias-adjusted projections of snow cover over eastern Canada using an ensemble of regional climate models” by Bresson et al.
In this paper, Bresson et al. present bias-adjusted projections of snow water equivalent (SWE) over Eastern Canada. They applied bias-adjustment methods from the climate community to snow outputs from an ensemble of regional climate models using the ERA-Land SWE as a reference gridded product for debiasing. Several indices were then derived from the bias-adjusted times series of SWE such maximal annual SWE, snow cover duration, date of snow cover onset, … The authors finally presented how these indices are projected to change in different subregions of Quebec during the 21^st century.
The topic of this paper is relevant for stakeholders and decision makers interested in the future of the snow cover in Eastern Canada. However, this paper could have reached a larger audience among the snow community if the authors had better explained their debiasing methodology and quantify its impact on the projections of the different SWE-related variables as detailed in my general comments below. This paper would have also benefited from an evaluation of the ERA5-Land SWE over Quebec using in-situ snow observations to better justify the choice of ERA-Land as the reference gridded snow product for this study. Therefore, at this stage, major revisions are required before this paper can be considered for publication in The Cryosphere. My main comments are listed below as general comments and are followed by specific and technical comments.
General comments
The introduction of the paper is rather short and does not present well the scientific context of the study. It should give more details about the bias-adjusting methodologies that are widely used in climate services. These methods are not necessarily well known by non-experts in climate projections. In addition, the introduction could give more detail about the challenges associated with the debiasing of snow variables that present a strong seasonality. The authors mention L 44 that the only debiasing method for SWE has been proposed by Michel et al. (2023). They could explain briefly what the characteristics of this method are and why does it work for snow variables. At the end of the introduction, the readers need to understand why bias-adjusted variables are crucial for climate services and the associated challenges for snow variables.

The authors use the SWE from ERA5-Land as the reference gridded product for debiasing. They only refer to performances reported in Kenda and Fletcher (2025) and Mudryk et al. (2024) to justify their choices. However, this justification remains vague. For example, they refer to the evaluation of ERA5-Land for different ranges of elevation detailed in Kenda and Fletcher (2025). However, this evaluation concerns Canada as a whole, including regions such as the Prairies or the Arctic where the snow conditions differ significantly from those found in Quebec. For this reason, I recommend the authors add an evaluation of the performances of ERA5-Land across Quebec using reference SWE measurements from the CanSWE dataset complemented by SWE data collected by the province of Quebec. Such rich dataset would allow a comprehensive evaluation of ERA5-Land in Quebec that would (i) strongly justify the choice of ERA5-Land as the reference product for debiasing and (ii) help identifying regions of lower performances for ERA5-Land where the result of the debiasing should be considered more carefully. In itself, a solid evaluation of ERA5-Land would be a very interesting outcome of this paper.

A key component of this paper is the method used for the bias adjustment of SWE from regional climate models. In its present form, the paper does not explain well and illustrate well why applying a bias adjustment is crucial before deriving relevant projections for different snow indices. Section 2.4 should be heavily revised to better explain the bias-adjustment method. The following points should be considered:

It is not clear how the method proposed by Michel et al (2023) addressed the issue associated with the seasonality of snow mentioned in the introduction as one of the main challenges for the bias adjustment of snow variables.

It seems that the bias adjustment method allows for a downscaling from the native resolutions of the RCM (which are not mentioned in the paper) to the resolution of ERA5-Land. Can the author explain how such downscaling is possible?

A pre-processing step is applied to the SWE time series during the melting season (P 6 L 135-142). It would be very valuable for the reader to have a figure that illustrates the impact of this pre-processing step. In addition, is this pre-processing applied to each of the grid cells?

I understand that the authors want to focus their analysis on the bias-adjusted projections as part of group that provides climate services to stakeholders. However, from a scientific point of view, I recommend the authors to quantify if the projections of the different snow indices would have been different without the bias adjustment. Such analysis would highlight the importances of carrying out such bias adjustment when considering snow variables. For climate service, it may be obvious for certain variables such as near surface temperature, but it seems that bias adjustment is less usual for snow variables due to methodological challenges. This analysis would be a solid contribution from this paper that would extend beyond stakeholders interested in the future of the snow cover in Quebec.

Specific Comments
P1 L 14: the terminology “the northern part of the northern hemisphere” is rather vague. Can the authors clarify? Maybe give a range of latitude.
P 1 L21: what do they authors mean by “best understood”?
P1 L22: “surface temperature”: is it the actual surface temperature (“skin temperature”) or the screen-level temperature (taken for example at 2 m above the ground)?
P1 L24: it would be highly relevant for the readers to add here references that illustrate how challenging it is to adjust the bias of variables with rare occurrences or strong seasonality.
P2 L 30-35: these sentences contain several statements that should be supported by appropriate references. For example, the statement “to better reproduce … the processes such as sublimation or ablation” is really vague and must be supported by references.
P 2 L35: Offline simulations with snowpack schemes are often carried out at continental or global scales (such as the ERA5-Land product used in this study or the Crocus-ERA5 dataset (Ramos Buarque et al., 2025).In this context, I recommend the authors to rephrase the sentence “Consequently, this method could be better adapted for specific purposes at a local scale”
P 2 L 43: the term “snow cover” used here is confusing since it is already widely used in the paper to refer to snow in general. Maybe use “snow cover fraction” since it is the variable of interest in the paper of Matiu and Hanzer (2022).
P 2L 48-49: can the authors explain briefly what are the problems that arise with snow simulations at high elevation?
P 2 L59: It would be interesting to know the mean elevation of the different subregions considered for the analysis. In particular, it would illustrate well the contrast between Southern Quebec and SLRV.
P3 L 71: what do the authors mean by “flexibility”? Would it be possible to reformulate to be clearer?
P 4 L 89: what do the authors mean by “mismatch”? Between which datasets? What was the nature of this “mismatch”?
P4 L 91-92: this sentence should be reformulated since Figure 3 in Kenda and Fletcher (2025) presents an evaluation of the SWE from ERA5-Land across Canada (including region below 50N). Kenda and Fletcher (2025) did not only evaluate ERA5-Land in northern Canada above 50N.
P 5 L 103: how many simulations were considered in this first ensemble?
P 5 L 117: Why are the authors using the argument about the availability of SWE observations to justify focusing on the region below 50 N in their selection criteria? Indeed, SWE observations are not used in this study to evaluate ERA5-Land (see my second general comment) and are not assimilated in ERA5-Land.
P 5 L 119: how many candidates were present in the initial ensemble? Such information is interesting to better understand how strict the selection criteria were.
P 5 L 119-120: it would be good to know what the selection criteria were in McCrary et al. (2022) to justify why it makes sense to compare the two ensembles.
P 6 L 147: does the value of 375 mm refer to ERA5-Land?
P 7 L 157-159: Has the bias correction already been applied when presenting the results for the two ensembles?
P 12 Figure 9: The dots on the different subplots seem to be rather noisy, especially for the northern and central domain. Is it because of very few events (even only one) over the 30-yr period are considered when computing the mean duration of noSCseq? Showing results aggregated longer time periods can potentially reduce the noise and make the figure easier to read.
Technical Comments
Text
P2 L 32: “CROCUS” is not acronym and can be written “Crocus”.
P 2 L 46: “water content in the snowpack” could be confusing. Maybe use “total water content of the snowpack” to make sure that it does not only refer to liquid water content in the snowpack.
P2 L 55: It could be worth mentioning the other Canadian provinces that are included in the simulation domain.
P 5 L 104: the year is missing for the reference to Mearns et al.
P 5 L105: Explain the meaning of the acronym “SM”. It should be changed throughout the document. I also recommend the authors to mention to which specific table or figure they are referring to in the Supplementary Material.
P 5 L 108: the term “melt” or “melting” is often preferred to “thaw” when referring to snow.
P 5 L 131: please double check to reference to (Themeßl et al., 2012). Is the family name written correctly?
P 6 L 144: this sentence can be included in the previous paragraph to avoid having a paragraph made of a single sentence.
P 6 L 148: explain that the names of the regions such as Charlevoix, … are shown on Fig. 2.
P 16 L 264: add the corresponding DOI.
Figures
Figure 10: the different levels of transparency are not visible in this figure.
Citation: https://doi.org/10.5194/egusphere-2025-3979-RC1
RC2:
'Comment on egusphere-2025-3979', Anonymous Referee #2, 16 Oct 2025
The authors present an interesting study of RCM-based snow simulations that have been bias-corrected. For variables bounded by zero, such as snow or precipitation, statistical bias-correction methods must account for the (ill-defined) mapping of zeros to many possible non-zero values -- a "snow creation" problem. The "small value insertion" methods sometimes used to bias-correct precipitation cannot be easily applied to snow because there is a strong temporal correlation, so sequential treatment of the snow time series is necessary. The authors reference the method described in Michel et al. (2024) for the method they apply. This involves replacing values at the end of the snow season with an exponential decay of SWE down to a minimum threshold before applying a standard (multiplicative) empirical quantile mapping. The simulations chosen for the study cover the province of Quebec, Canada and are comprised of several models, two forcing scenarios, and a variety of snow metrics that summarize the characteristics of the seasonal snow.

Overall, the model ensemble comparisons show that SWEmax, SSS, SSE, SSD, and noSC/noSCseq indices all change under climate projections. In general, the strength of the responses increase in time and with radiative forcing (SSP4.5 vs 8.5), but there are some regional differences in these responses. In some regions, the snowpack is projected to become much more sporadic and short-lived, while there appears to be a near-term increase in snow in northern regions.

While I believe this paper to be a valuable contribution and a good fit for publication in The Cryosphere, I have several comments that should be addressed. Overall, I find that key methodological information should be clarified and/or added. My comments below will provide more information.

Major comments:

Simulation selection: In the Data section (2.3), the authors describe the process of excluding some model realizations from their ensembles. My understanding is that this was done before the main analysis of the study because of the bias-correction step. If this is the case, I think section (4.2) needs to be moved and combined with (2.3). I would also suggest a supplementary plot showing the selection variables compared to ERA5-Land for each model/realization considered (before exclusion). This would support the claim that a "first scan of the simulated SWE showed large discrepancies".
Additional comment related to this: the stippling in figures represents 80% agreement among models, but there appear to be 4 RCP4.5 simulations chosen and 7 RCP8.5 simulations chosen (Table S1). 80% therefore corresponds to consensus in the RCP4.5 simulations but agreement in either 6 or 7 out of 7 RCP8.5 simulations. I request that the authors clarify this point and discuss the implications.

Total number of years: Given that ERA5-Land has a long back-extension (to 1950s), would there be value to using a longer training period for the statistical bias correction?

Method description: There are several aspects of this study that need to be more clearly outlined within the paper.
Was SWE data from the ERA5-Land reanalysis used for bias-correction? Is the model output hourly or daily? If different, how is this reconciled?

Was the full Michel et al. (2024) algorithm used? If it was, what data was used for the clustering variables (slope, aspect, altitude...)? If not, how was the EQM applied (how many quantiles were estimated, is there sufficient data to separately do this at each grid cell)?

Clarify the SWE replacement method (L138-143). This is the algorithm I imagine from reading: find the end of the snow season using a 1 mm threshold (by finding the last >1 mm value before 14 consecutive days below 1 mm); replace all values after that date by half the previous day's value.

If half of 1 mm is replaced in the dataset up to a 0.001 mm cutoff, that yields roughly 6 more days of
nonzero snow, but in the model selection stage, the authors look for models for which SSE can differ by up to 15 days. What impact could this have on bias correction?

The authors mention in L230 that varying the threshold for snow cover is outside the scope of the study. I wonder, though, what the difference would be if another threshold (e.g. 2mm instead of 1mm) were used for the snow insertion before bias-correcting. Were there any tests done for this?

Could snow be added to the snow start season (before bias correction)?

I echo the first reviewer's call to explain the details of the bias correction process.

How was Quantile Mapping applied to the projections (2041-2070 and 2070-2099)? In these cases, there is no corresponding "true" distribution to match (e.g. ERA5-Land). Does the method adequately account for the climate change signal? If a quantile delta mapping was applied, please explicitly state this and how it works for a variable like SWE.

How does the bias correction potentially affect an index like noSCseq? How are zero SWE values within the snow season season corrected? This factor might impact snow projections where zero values become more frequent during the snow season in some regions.

Editorial comment: I would suggest to revise the paper for tense consistency and some other minor English errors (L25 "precipitations" instead of "precipitation", L117 "experimented" instead of "experienced").

Figures and tables:

Section (2.1) Are there any lakes or water bodies resolved at the 0.1 degree scale for any of the RCMs? I see that lakes are masked in most figures.

Table 1: should include noSCseq, defined as the average duration of snow-free conditions starting on each day of the year (excluding 1 day and 14+ day snow-free conditions). Table 2: include resolution, otherwise unclear why MPI-ESM-LR-CRCM5 would be different when produced by a different institution.

Figure 2: the second and third columns have colours which are hard to see, might suggest cutoff of +/- 15% instead of 20%. Lakes are not masked in panel (g). Figure 5 is mentioned before Figure 4, rearrange order. Figure 8: It appears that coastal regions gain a lot more snow-free days than continental regions, especially so in the 2070s compared to historical. This is not mentioned in the discussion. Lakes are not masked in panel (a).

Figure 9: It's interesting that the period of the year with full snow cover (no fragmentation) can be seen from this figure. In Northern Quebec, full cover appears to last between early November and early May. It lasts a little longer in 2041-2070 (mid-Nov to Jun), and then shortens by 2070-2099 (Dec to mid-May). In the other regions, the full snow season shrinks uniformly over time or disappears completely (as in southern Quebec, SLRV). On the other hand, we do not see the duration of noSCseq get longer over time, indicating that there is still regular (short-lived) snowfall occurring throughout the winter months in all regions even at the end of century in these simulations.

Minor points or suggestions for wording:

L4: Replace wording with "...an ensemble of statistically bias-corrected regional climate projections of snow water equivalent..."

L19: Remove "etc." and replace with "and more".

L20: Could add some comments on the value added by statistical bias correction of multi-model output.

L24: I think "datasets" here should be "methods".

L24: Replace ";" with "." or adjust capitalization.

L25: Clarify why these SWE is difficult to handle. Perhaps use L110-112.

L32: Could add how coarse the GCMs and RCMs tend to be.

L32: What "biases" need to be "addressed"? How is this typically done, perhaps for other variables?

L44: Add some brief description of the method (e.g. "We propose using this method to insert nonzero snow near the snow end season before applying EQM to bias-adjust the SWE data.")

L85-99: Highlight the dataset that was chosen (ERA5-Land), including its strengths and weaknesses, rather than first mentioning others that were not selected. Please also mention the data assimilation in the ERA5- Land model and perhaps the known biases (which are now in Section (4.2)).

L90: Mudryk et al. (2015) does not cover ERA5, ERA5-Land, or B-TIM datasets, so might not be directly relevant here. Please review this reference.

L106: The authors should clarify if they mean discrepancies between models, scenarios (RCP4.5 vs 8.5), or realizations. I suggest the use of the phrasing "forcing scenario ensemble" instead of saying "RCP ensemble" throughout the paper.

L127: Could say SSE instead of "end of its snow season"

L175: Is it possible that this asymmetry in the response for Northern Quebec could be related to inserting nonzero snow values only in the snow melt season?

L206: unclear what it means to individually analyze "simulations" as opposed to "RCP ensembles".
Citation: https://doi.org/10.5194/egusphere-2025-3979-RC2

Émilie Bresson, Éric Dupuis, and Pascal Bourgault

Supplement

https://doi.org/10.5194/egusphere-2025-3979-supplement

Émilie Bresson, Éric Dupuis, and Pascal Bourgault

Viewed

Total article views: 1,362 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	Supplement	BibTeX	EndNote
1,309	43	10	1,362	25	22	22

HTML: 1,309
PDF: 43
XML: 10
Total: 1,362
Supplement: 25
BibTeX: 22
EndNote: 22

Views and downloads (calculated since 10 Sep 2025)

Month	HTML	PDF	XML	Total
Sep 2025	1,204	15	5	1,224
Oct 2025	105	28	5	138
Nov 2025	0

Cumulative views and downloads (calculated since 10 Sep 2025)

Month	HTML	PDF	XML	Total
Sep 2025	1,204	15	5	1,224
Oct 2025	105	28	5	138
Nov 2025	0

Viewed (geographical distribution)

Total article views: 1,459 (including HTML, PDF, and XML) Thereof 1,459 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 02 Nov 2025

Short summary

As climate changes, the need for easily accessible high resolution and unbiased projections of snow cover indices to develop adaptation plans increases. We produced such dataset for eastern Canada and made it publicly available. The analysis of the snow cover projection in Quebec up to the year 2100 revealed the following changes: a shortening of the snow season; a decrease of the maximum snow amount in the south and an increase in the north; more interruptions of the snow season in the south.


Total:	0
HTML:	0
PDF:	0
XML:	0