ClimAVA‑SWE: A High‑Resolution CMIP6‑Based Snow Water Equivalent Dataset for the Western United States

Khoshnood Motlagh, Sajad; Smith, Kayla; Zhang, Wei; Null, Sarah; Miller, Anna; Chikamoto, Yoshimitsu; de Lima Moraes, Andre Geraldo

doi:10.5194/essd-2026-303

Preprints

https://doi.org/10.5194/essd-2026-303

Preprints

30 Jun 2026

| 30 Jun 2026

Status: this preprint is currently under review for the journal ESSD.

ClimAVA‑SWE: A High‑Resolution CMIP6‑Based Snow Water Equivalent Dataset for the Western United States

Sajad Khoshnood Motlagh, Kayla Smith, Wei Zhang, Sarah Null, Anna Miller, Yoshimitsu Chikamoto, and Andre Geraldo de Lima Moraes

Abstract. The ClimAVA-SWE dataset provides bias-corrected daily snow water equivalent (SWE) estimates at approximately 4 km spatial resolution across the western United States, publicly available through the Harvard Dataverse and can be accessed via its official repository at https://doi.org/10.7910/DVN/SCD2VT (Khoshnood Motlagh, de Lima Moraes and Smith, 2026). The dataset is generated using the Spatial Interactions Downscaling for Snow Water Equivalent (SPID-SWE) method, a data-driven statistical downscaling framework that integrates high-resolution reference SWE data (NSIDC-0719) with daily outputs from an ensemble of 14 CMIP6 global climate models (GCMs). SPID-SWE employs a dual random forest modeling strategy that explicitly distinguishes snow accumulation and ablation phases, improving seasonal SWE representation relative to single-phase approaches. ClimAVA-SWE spans a historical period (1981–2014) and future projections (2015–2100) under three Shared Socioeconomic Pathways (SSP245, SSP370, and SSP585). The pixel-based design of SPID-SWE independently downscales each grid cell across space and time, preserving both spatial and temporal variability in the SWE signal. Although performance is conditioned on the quality of the driving input data, the dataset demonstrates strong accuracy and computational efficiency, providing a robust and scalable resource for high-resolution climate and hydrological impact applications.

Received: 21 Apr 2026 – Discussion started: 30 Jun 2026

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, 2432 KB)

Supplement (11329 KB)

Download & links

Preprint (2432 KB)
Metadata XML
Supplement (11329 KB)
BibTeX
EndNote

Sajad Khoshnood Motlagh, Kayla Smith, Wei Zhang, Sarah Null, Anna Miller, Yoshimitsu Chikamoto, and Andre Geraldo de Lima Moraes

Status: open (extended)

Post a comment Subscribe to comment alert

RC1:
'Comment on essd-2026-303', Anonymous Referee #1, 22 Sep 2026 reply
ESSD-2026-203
This paper extends the downscaling method of de Lima Moraes and Khoshnood Motlagh (2024), which was applied to daily precipitation min and max temperature, to SWE. The authors apply bias correction (Quantile Mapping) and statistical downscaling (Random Forest) to fourteen CMIP6 GCMs.
I found the results to be a bit circular while at the same time lacking specific insight; some aspects of the methods need additional explanation. I could have used a bit more detail on the SPID method because I found myself referring to the original publication to try and clarify aspects. I found the analysis of those results in the context of snow and the applied SPID method to be a bit thin. For example, the authors present results for two GCMs with two different spatial resolutions but don’t really discuss how and why their method might be sensitive to spatial resolution. Further, I would have liked better linkages between sections to show how one set of results informs or impacts another piece. For example, how do the residual errors of the bias corrected and downscaled GCM compare? This might give insight into how the RF model is operating and how it is translating from coarse to fine resolution.
The manuscript contains a substantial number of technical errors surrounding the referencing. In text citations tack dates and the reference list contains errors, incomplete citations and broken dois. Some of the statements in the results refer to information not presented in the figures/Tables or that cannot easily be interpreted from them.
Major comments
In extending to SWE the original method needed to be adapted; the authors divided the snow season into accumulation and ablation and trained separate models for each. Given that this is a key element of extending to SWE I would have liked to see separate evaluations of the accumulation and ablation periods in addition to evaluation of the full snow season. For example, is the linear accumulation period easier to downscale than the ablation period or no difference? I can see how this is sort of built into the snow metrics analyzed, however the authors don’t ever mention the dual-approach outside of the description of the SPID-SWE approach. I’d like to see the authors integrate the model design into the analysis of the results a bit more.

Comparison of the performance of the downscaling on Fine (FR) versus Coarse (CR) resolution GCM (Section 3.1 and 3.2). From what I can tell you are comparing the testing/validation results of two different GCMs that have spatial resolutions at the extremes of the ensemble. This means that there are elements of both spatial resolution and the model itself – i.e. snow amounts and distribution, that are unrelated to spatial resolution. It wasn’t completely clear whether this step was performed on the raw or bias corrected GCM data [see comment on Methods], but I assumed it was the bias corrected. If so, the differences stemming from meteorology and snow model should be reduced but not eliminated as presented in the authors’ own results of the bias correction (Sect. 3.3.2). I recommend the authors more clearly state the limitations of this comparison – i.e. that it includes aspects not limited to spatial resolution. One approach to really isolate the impact of spatial resolution could be to resample the FR to the resolution of the CR but this is probably beyond the scope of the study.

Methods – clarification required. A bit unclear or inconsistent between Figure 1 and the text – was the bias corrected data used to train the RF model or was the RF model trained on the raw output and then applied to the bias corrected data? The text seems to suggest bias correction happens first and then the RF model is trained on the bias corrected data but Figure 1 suggests the bias correction and RF are two independent steps that come together in the final downscaling piece. Along the same lines, are the results presented in 3.1 and 3.2 of the final bias corrected + downscaled data or is it spatial downscaling of the raw GCM data? Sorry for my confusion here.

I found the evaluation of the bias correction with the same reference used to bias correct the GCM to be a bit unsatisfying. I admit I’m not an expert in bias correction or QM so my comment might be completely off base. If so, please disregard. I understand that the QM approach corrects based on the distribution, so using the reference to validate might be appropriate. However, the language used to present the results that had me a bit puzzled. For example, is closer alignment with the reference really an ‘improvement’ or just closer alignment with the reference used to bias correct the data? Further, given that the approach corrects based on the distribution aren’t the results presented in Fig 6 expected? Do these figures add anything to the paper or simply confirm that the method is operating as intended?
Did the authors consider a split-validation approach or independent data for additional validation?

Were any other bias correction approaches (e.g. Quantile Delta Mapping) considered?

Additional validation – SNOTEL is not an independent reference as the authors state. Did the authors consider using other data? e.g. ASO (might provide additional insight regarding the spatial downscaling), snow surveys from California Water Resources and NCRS.

Organization – I found the organization of the results to be a bit counter intuitive. Some of the section headings could be more specific and descriptive (3, 3.1, 3.2 and 4). I’d suggest moving the discussion of the raw GCM and impact of bias correction before the results of the ‘Grid-cell Level Validation’ and ‘Watershed-Based Validation’. If the authors decide to keep the existing structure I suggest adding a ‘roadmap’ to the results and also more descriptive section headings. E.g. ‘First we present the results of the SPID-SWE (bias correction + spatial downscaling) for the historical period (split validation) at both the grid cell and basin level. The grid cell validation focuses on two products with spatial resolutions at the extremes of the ensemble. This is followed by an assessment of the applied bias correction. Finally, we present the full downscaled historical and future SWE estimates for all fourteen GCMs.’

Suggested structure:
3.0 Results
3.1 Evaluation of GCM SWE and Bias Correction
3.1.1 GCM SWE
3.1.2 Impact of Bias Correction
3.2 Validation of SPID-SWE
3.2.1 Grid-cell Level validation
3.2.2 Watershed-level validation

Minor comments
L88: ‘the existing research’ about what? Please be more specific.
L94-97: Inaccurate description of GlobSnow product for two reasons. (1) GlobSnow is a combined passive microwave and in situ product – it assimilates in situ snow depths (2) GloSnow does not provide SWE estimates over complex terrain.
L98-101: Did you consider comparing your downscaled output to ASO data?
L104-105: Not completely accurate or complete. MERRA2 is 0.5°x0.652° which is closer to 50km. There’s also ERA5 (31 km/0.25°, Hersbach et al. 2020) JRA-3Q (0.375°, Kosaka et al. 2024).
Hersbach, H., Bell, B., Berrisford, P., Hirahara, S., Horányi, A., Muñoz-Sabater, J., Nicolas, J., Peubey, C., Radu, R., Schepers, D., Simmons, A., Soci, C., Abdalla, S., Abellan, X., Balsamo, G., Bechtold, P., Biavati, G., Bidlot, J., Bonavita, M., De Chiara, G., Dahlgren, P., Dee, D., Diamantakis, M., Dragani, R., Flemming, J., Forbes, R., Fuentes, M., Geer, A., Haimberger, L., Healy, S., Hogan, R.J., Hólm, E., Janisková, M., Keeley, S., Laloyaux, P., Lopez, P., Lupu, C., Radnoti, G., de Rosnay, P., Rozum, I., Vamborg, F., Villaume, S., Thépaut, J.-N., 2020. The ERA5 global reanalysis. Quarterly Journal of the Royal Meteorological Society 146, 1999–2049. https://doi.org/10.1002/qj.3803
Kosaka, Y., Kobayashi, S., Harada, Y., Kobayashi, C., Naoe, H., Yoshimoto, K., Harada, M., Goto, N., Chiba, J., Miyaoka, K., Sekiguchi, R., Deushi, M., Kamahori, H., Nakaegawa, T., Tanaka, T.Y., Tokuhiro, T., Sato, Y., Matsushita, Y., Onogi, K., 2024. The JRA-3Q Reanalysis. Journal of the Meteorological Society of Japan 102, 49–109. https://doi.org/10.2151/jmsj.2024-004
L115-116: references for this statement?
L115-117: Isolated sentence. Does this go with the preceding or following paragraph?
L118-119: same as above (L115-117)
L120-121: Would be good to note, at least in parentheses, what some of these HR SWE products you are referring to are.
L124: long-term analysis of past and future snowpack dynamics
Output 4km grid – is this the same grid as the UA SWE data? Please clarify the output grid specs somewhere.
L171: How did you treat cases where partial pixels overlap?
L195-198: could do with a bit more details on SPID method (see general comment)
L206-207: Could this be suggesting that the coarse resolution can't capture patchy snow and partial snow cover within the coarse pixel that is common during the ablation period?
L226-228: Why not just use the original reference data in its original resolution?
L270-272: ‘results’ of what? It would help the reader if this text was a bit more specific. Perhaps something along the lines of ‘Downscaling results for two example GCMs bounding the range of spatial resolutions in our GCM ensemble are presented below. Results include both pixel-based and watershed validation spanning 42 years (1982-2023)’.’
L281: Is this (metric vs swe magnitude) shown anywhere? When you say 'accuracy' are you only referring to correlation here? Areas with higher SWE seem to have larger RMSE, although I find it a bit hard to infer from the maps.
L284: How ‘small’ and is there any relation to SWE magnitude? Or, is it related to snow cover duration or something else?
L286: What are the SWE classes you are referring to? Where is the evidence (Figure/table/etc.) for this? Do you have results aggregated by SWE magnitude or 'SWE class' you can refer the reader to? One idea could be to show the mean or peak SWE values in another map.
L290-293: From the information provided it seems that it’s more stronger correlations rather than necessarily higher performance overall? Perhaps the figures aren’t telling the full story.
L306-307: Can you think of why the FR might have a less negative bias than the CR? Second, why might you expect differences in performance at different spatial resolutions. You decide to present the extremes but need to relate the results to the input data and downscaling setup. For example, with the CR the model searches much farther away from the pixel in question than it does for the FR – how might this impact the downscaling? There will be more 4km pixels with the same overlapping and adjoining GCM-resolution grids at the CR than the FR. Does this impact the results? If so, how? I’d like the authors to pick some of these elements to discuss the results in terms of the input data and model setup.
L318-319: Are you able to show these basins on one of the figures somehow? This would help the reader tie what is shown in Fig 2 and 3 with Table 2.
Section 3.2 – see major comment #2 and L306-307.
L367-374: This statement contradicts L139-141. SNOTEL is not independent of the reference dataset. Your reference dataset assimilates SNOTEL observations so it is expected that there will be a strong correlation between the two. (see also major comment #5). Second, Fig 4 shows that the downscaled data differs from the reference and SNOTEL, I would have liked to have seen the authors touch on this if only briefly.
Sect 3.3.2 – The existing text should be condensed.
L385-386: Logically you'd think that a bias correction to a given reference would improve agreement with that same reference or am I missing something here? Maybe clarify how and why the QM approach might not lead to improvement. (see also major comment #4). Again, I apologize if I’m completely off-base here. If so, please disregard.
L410: remove ‘excluding snow-free area’ (is redundant)
L442-442: Isn’t this expected? Or am I missing something?
L445: It’s interesting to me that the range is larger after correction. Any idea as to why this is?
L450-451: Any idea why certain models show greater improvement than others and why some might show worsening?
L473: Could this be due, at least in part, to the fact that if the GCM underestimates (which your results seem to suggest) that there’s a good chance the snow season (later onset, earlier snow-free) will also be shorter if the model ties snow cover to SWE. Since your QM application can’t add snow, it won’t be able to adjust these dates and associated time metrics unless they occur too early (onset) or too late (snow-free). Might this suggest that the timing metrics, at least onset and melt, aren’t appropriate metrics given the limitations of the QM you described? I’m curious why you think there was little change in the date of peak.
L524: consistency of the ClimAVA-SWE or of the bias corrected GCM data?
L547: suggest ‘Most downscaled GCMs’ for higher specificity
L549-554: Are these the same patterns of over/under estimation you saw after bias correction? i.e. Does this mirror the residual biases in the GCMs after bias correction or is it additional (or less) error as a consequence of the downscaling?
L650-660 – Data Availability. Please add somewhere that there is one file per 15yr period for a total of 280 individual netcdf files. (2 per GCM for historical and 6x3(scenarios) per GCM = 20 per GCM x 14 GCMs = 280). See also comment Data.
Tables and figures
Figure 2
MPE color scheme not helpful. Suggest restricting the range.

I wonder if a DEM underlay or outlines of the mountain regions would help somehow? It's difficult to pick out what's happening for the correlation especially.

What are the white areas in this figure and in Fig 8a and why aren’t they also masked in Fig 3 and 7 and in the downscaled dataset? I assumed these are areas with no snow-covered days bit not clear.

Figure 6 and associate text: Since QM adjusts the distribution isn't an almost identical CDF expected? How much does this section add to the paper? See major comment #4.
Technical corrections
In-text citations contain errors – missing dates and inconsistent format
Reference list contains errors: no date, journal missing, not all dois link correctly. Please carefully check all in text citations and full reference list.
Examples:
Livneh et al. (no date) should be:
Livneh, B., E. A. Rosenberg, C. Lin, B. Nijssen, V. Mishra, K. M. Andreadis, E. P. Maurer, and D. P. Lettenmaier, 2013: A Long-Term Hydrologically Based Dataset of Land Surface Fluxes and States for the Conterminous United States: Update and Extensions. J. Climate, 26, 9384–9392, https://doi.org/10.1175/JCLI-D-12-00508.1.
Repeated de Lima Moraes 2024 reference: L831-836
L957-958: Incorrect Daymet reference
(No date) ‘Daymet: Daily Surface Weather Data on a 1-km Grid for North America, Version 4 R1 Get Data’. Available at: https://doi.org/10.3334/ORNLDAAC/2129.
Some isolated sentences that should be combined into proper paragraphs – e.g. L115-119.
L56: relevant instead of true
L537 – ‘Table. 2, examining basin-scale snowpack behavior and inter-model variability.’ Random sentence fragment. Remove.
Data
I found the number of files to be a bit cumbersome. Are there any upload/download restrictions preventing merging of multiple netcdfs from a single model? It would be nice to at least have one file for the historical period. Ideally there would be one file per GCM for the historical and one per scenario for predictions.
Areas and times where the bias correction and/or downscaling is not applied should be masked and flagged accordingly.

Reply
Citation: https://doi.org/10.5194/essd-2026-303-RC1

Sajad Khoshnood Motlagh, Kayla Smith, Wei Zhang, Sarah Null, Anna Miller, Yoshimitsu Chikamoto, and Andre Geraldo de Lima Moraes

Supplement

https://doi.org/10.5194/essd-2026-303-supplement

Data sets

ClimAVA‑SWE: A High‑Resolution CMIP6‑Based Snow Water Equivalent Dataset for the Western United States Sajad Khoshnood Motlagh, Andre Geraldo de Lima Moraes, and Kayla Smith https://doi.org/10.7910/DVN/SCD2VT

Sajad Khoshnood Motlagh, Kayla Smith, Wei Zhang, Sarah Null, Anna Miller, Yoshimitsu Chikamoto, and Andre Geraldo de Lima Moraes

Viewed

Total article views: 360 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	Supplement	BibTeX	EndNote
185	78	97	360	77	122	89

HTML: 185
PDF: 78
XML: 97
Total: 360
Supplement: 77
BibTeX: 122
EndNote: 89

Views and downloads (calculated since 30 Jun 2026)

Month	HTML	PDF	XML	Total
Jul 2026	52	10	6	68
Aug 2026	32	6	3	41
Sep 2026	66	34	68	168
Oct 2026	35	28	20	83

Cumulative views and downloads (calculated since 30 Jun 2026)

Month	HTML	PDF	XML	Total
Jul 2026	52	10	6	68
Aug 2026	32	6	3	41
Sep 2026	66	34	68	168
Oct 2026	35	28	20	83

Viewed (geographical distribution)

Total article views: 325 (including HTML, PDF, and XML) Thereof 325 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 14 Oct 2026

Download

Preprint (2432 KB)
Metadata XML

Short summary

Snow stored in mountains is a critical water source for the western United States. We developed a high-resolution dataset of daily snow amounts from 1981 to 2100 under three future greenhouse gas emission scenarios. Using statistical methods applied to fourteen global climate models, we produced estimates across the region at fine spatial detail. The dataset reliably captures peak snowpack, geographic patterns, and long-term changes. It supports water planning and climate impact studies.


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