SnowGalileo: A Pre-trained Earth Observation Transformer for Daily, 100 m Fractional Snow Cover Mapping

Reil, Marlena; Kaltenborn, Julia; Allum, Donovan J. M.; Pelletier, Francis; Roessler, Sebastian; Shrestha, Samip; Lv, Zhibang; Truckenbrodt, John; Schwaizer, Gabriele; Nagler, Thomas; Pomeroy, John W.; Marsh, Christopher B.; Montpetit, Benoit; Jonas, Tobias; Tseng, Gabriel; Rolnick, David; Dietz, Andreas J.; Baumhoer, Celia A.

doi:10.5194/egusphere-2026-3833

Preprints

https://doi.org/10.5194/egusphere-2026-3833

Preprints

08 Sep 2026

| 08 Sep 2026

Status: this preprint is open for discussion and under review for The Cryosphere (TC).

SnowGalileo: A Pre-trained Earth Observation Transformer for Daily, 100 m Fractional Snow Cover Mapping

Marlena Reil, Julia Kaltenborn, Donovan J. M. Allum, Francis Pelletier, Sebastian Roessler, Samip Shrestha, Zhibang Lv, John Truckenbrodt, Gabriele Schwaizer, Thomas Nagler, John W. Pomeroy, Christopher B. Marsh, Benoit Montpetit, Tobias Jonas, Gabriel Tseng, David Rolnick, Andreas J. Dietz, and Celia A. Baumhoer

Abstract. Mountain snow is an important component of the cryosphere that directly affects downstream livelihoods. Accurate monitoring of mountain snow is crucial, yet remains challenging due to complex topography and frequent cloud cover. Although combining data from multiple Earth Observation (EO) satellites can improve spatial and temporal coverage, extracting Fractional Snow Cover (FSC) from sensors with different spatial and temporal resolutions remains difficult. AI-based Earth foundation models can process diverse sensor inputs and have been successfully applied to various remote sensing tasks; however, they often lack snow-specific design considerations. This paper presents SnowGalileo, a pre-trained transformer model designed to integrate EO data to map snow-covered areas. We evaluated SnowGalileo's potential for generating daily, gap-free FSC maps at 100m resolution in mountainous regions. SnowGalileo combines one week of satellite time-series data from multiple sensors, including Sentinel-1, Sentinel-2, Landsat, Sentinel-3, MODIS, and VIIRS, with topographic, land cover, and meteorological information to predict FSC for a given day. The model was pre-trained using masked autoencoding and fine-tuned using labeled data from various mountain ranges across the Northern Hemisphere. In addition to clear-sky conditions, SnowGalileo is evaluated under a wider variety of conditions than was previously possible, including cloud cover and lack of high-resolution (~10–30 m) satellite imagery. For the Canadian Rockies and the Swiss Alps, respectively, SnowGalileo achieves RMSEs of 0.099 and 0.124 on clear days, 0.144 and 0.209 on cloudy days, 0.185 and 0.262 on days without high-resolution imagery, and 0.200 and 0.299 on cloudy days without high-resolution imagery. SnowGalileo also consistently outperforms random forests, support vector regressors, and multi-layer perceptrons. While the current product is a proof of concept that has undergone limited validation across geographic regions, operates on 1km × 1km tiles rather than full maps, and has restricted capabilities in challenging conditions, this approach could ultimately enable the continuous, gap-free operational generation of FSC time series for mountain regions worldwide.

Received: 02 Jul 2026 – Discussion started: 08 Sep 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

Status: open (until 24 Oct 2026)

Post a comment Subscribe to comment alert

RC1:
'Comment on egusphere-2026-3833', Anonymous Referee #1, 29 Sep 2026 reply
The manuscript presents a novel methodology that adapts and fine-tunes the Galileo Earth Foundation Model (Tseng et al., 2025) for daily FSCA mapping at 100m. The model integrates multi-sensor satellite observations, climate reanalysis products, and other ancillary variables. The authors report that their model achieves much better performance compared to traditional ML models when evaluated in the Canadian Rockies and the Swiss Alps under various conditions (e.g., cloudy or cloud-free, with or without high resolution images, various canopy density).
The method shows scientific novelty, like the use of EFM, considering the temporal sequence to deal with the cloud contamination, the fusion of multi-sensor high resolution images with moderate high spatial resolution sensors for mapping of daily FSCA at finer spatial resolution, etc., and the results indicate value to the cryospheric remote sensing community. However, the manuscript in its current form is not suitable for publication.
I have a couple major concerns:
1) A major claim of this paper is its ability to overcome cloud contamination, which is one of the primary challenges in optical snow mapping. However, rather than evaluating the model on images contaminated with real clouds, the authors used an open-source cloud generator tool (Czerkawski et al., 2023) to simulate synthetic clouds. It is scientifically invalid to claim a methodology solves cloud contamination without validating it against satellite data with real cloud cover.
2) Additionally, the model is validated across a mere 16 scenes. For an EFM claiming broad generalizability across complex mountain terrains, this sample size is insufficient to prove scientific robustness. Extra validation datasets, such as lidar-based high-resolution snow cover or other high-resolution FSCA products across other mountain regions, are necessary to truly evaluate the transferability and generalization of this model.
3) Lastly, the uncertainty of the FSCA in the high-resolution Landsat and Sentinel datasets is not well discussed or evaluated. How do these uncertainties propagate into the EFM, and how do they ultimately impact the final FSCA accuracy? Furthermore, the labeled target FSCA appears to be based on an equation designed to convert MODIS NDSI to FSCA, which is not evaluated using a Landsat dataset. The authors must clarify how the labeled FSCA data was generated (Section 3.2.2). It appears the target data does not include images with clouds. I am curious: how does the model learn to correctly predict FSCA under clouds if it is never trained on target labels that exist under cloudy conditions?
The organization of the manuscript is highly circular and confusing, mixing methodology with results. Descriptions of the datasets used, the model introduction, the pre-training, and the fine-tuning processes are scattered across different sections, making it very difficult to understand what was actually done in this study versus what was already done in previous studies. Additionally, the Results section contains lengthy methodological explanations that do not belong there.
Some examples:
“For example, clear-HR results were obtained using models trained on the clear-HR setup, while cloudy-HR results were retrieved using models trained on the cloudy-HR setup. In practice, a model should be selected that performs well across all input data setups.”

“However, under the cloudy-no HR setup, the clear-no HR SnowGalileo model outperforms the cloudy-no HR SnowGalileo model. This may be due to insufficient model runs (only one model run is reported here), or because the cloudy-no HR data does not provide enough information...” Relying on a single model run and speculating on why it failed undermines the scientific rigor of the paper.

I strongly encourage the authors to completely revision the manuscript to provide a rigorous academic context, and to redesign the validation experiment using real (not simulated) cloud-contaminated data.
Specific comments:
Abstract
Please explicitly indicate what reference (ground truth) dataset was used for model training and which dataset was used for validation.

Introduction & Background
Overall, the Introduction fails to provide a comprehensive, logical review of recent FSCA mapping techniques. Paragraphs 1 through 3 are overly lengthy and consist of broad, generic statements regarding the importance of mountain snowpack that appear heavily reliant on AI generation. The entire background section is a list of citations with little critical insight into how this study addresses existing knowledge gaps. I highly suggest rewriting the Introduction and Background sections. The authors must put more effort into demonstrating a genuine understanding of prior work.

Some examples:
Physically incorrect statement: “FSC is important to hydrology because it indexes the redistribution of snow by wind, vegetation, and gravity (Pomeroy and Gray, 1995) and also the effect of spatially inhomogeneous ablation on the areal snowpack (Pomeroy et al., 1998).” FSC does not index redistribution. Please correct this for physical accuracy.

Poor citation: “Multiple algorithms and methods have been developed to derive FSC mainly from multispectral sensors such as Landsat (Missions and Tools), MODIS (Painter et al., 2009), or AVHRR (Zhao et al., 2024).” "Missions and Tools" is not a valid academic citation.

Lack of synthesis: “However, both NDSI-based and mixture-based methods remain sensitive to canopy cover... (Dietz et al., 2012; Rittger et al., 2013; Lv and Pomeroy, 2019; Bertoncini et al., 2022; Gascoin et al., 2024).” This fails to synthesize the specific findings of these papers.

Undefined terminology: The authors write, "...pixel-based approaches such as random forests, support vector regressors, and artificial neural networks (ANNs) are widely applied in the field..." It is unclear what is meant by "pixel-based approaches" in this context, as opposed to spatial or contextual approaches. Please define and clarify.

Data

The summary of the dataset used for model training is logically unclear, repetitive, and rambling. Please restructure for clarity.

Results

Section 5.1: This section is redundant and needs rewriting. For example: “Under the clear setup with HR coverage, SnowGalileo achieves an RMSE below 0.1 (see Table 2). SnowGalileo is a promising FSC mapper, showing improved FSC skill in mountain regions under various conditions, compared with SOTA models.”

Section 5.1: Please move this text to the Methods section: “SnowGalileo’s inference module can be used to produce daily roll-outs, i.e. all available data of the past seven days (and the current day) is used to predict an FSC tile, effectively performing a sensor fusion task... SnowGalileo is not simply interpolating between two known maps, but instead relies on satellite and meteorological information from the last week.”

Figure 5: Please provide a legend for sub-figures (a) and (c). Indicate which model setting is used (e.g., Clear-HR?). What are the white lines between the tiles in (b)? Furthermore, it is not scientifically rigorous to visually compare a 100 m FSCA product directly with a 500 m global operational product in (c). Finally, remove the concluding description from the caption (“In summary, SnowGalileo can bridge HR gaps in time-series, stitch predictions, and provide FSC at a higher resolution than current SOTA.”).

Section 5.3.1 (Figure 7): Were the snow-free areas removed before calculating the RMSE? The text states, “Performance appears to be higher during the summer months and the high winter season, and lower during snow onset and snow melt,” but visually it looks like the RMSE is highest in January and December. Please clarify.

Section 5.3.1: “Setting aside slight differences in regional performance scores, SnowGalileo clearly shows that it can skillfully perform FSC mapping on unseen mountain ranges, demonstrating its ability to generalize spatially.” The figures and analysis do not robustly support this conclusion. More rigorous validation is needed across diverse spatial domains.

Section 5.3.2: Lines 705-715 attempt to explain why the RMSE in January is high, but the reasoning is not robust. The claim that “January snowpacks are low-density, i.e. they are more strongly affected by wind drift... which could lead to rapid changes in FSC that are neither captured in the satellite nor ERA5 data” is highly speculative.

Section 5.4.1 (Performance under Forest Cover): This discussion is inappropriate. While SnowGalileo outperforms other models under forest cover, the labeled target data used to train/validate the models does not explicitly or accurately correct for the impact of the canopy on FSCA estimation in the first place.

Reply
Citation: https://doi.org/10.5194/egusphere-2026-3833-RC1

Viewed

Total article views: 635 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
342	231	62	635	58	54

HTML: 342
PDF: 231
XML: 62
Total: 635
BibTeX: 58
EndNote: 54

Views and downloads (calculated since 08 Sep 2026)

Month	HTML	PDF	XML	Total
Sep 2026	253	174	30	457
Oct 2026	89	57	32	178

Cumulative views and downloads (calculated since 08 Sep 2026)

Month	HTML	PDF	XML	Total
Sep 2026	253	174	30	457
Oct 2026	89	57	32	178

Viewed (geographical distribution)

Total article views: 635 (including HTML, PDF, and XML) Thereof 635 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 18 Oct 2026

Short summary

Many communities rely on snowmelt from the mountains for their water supply. However, monitoring snow from space is challenging because clouds and other factors can obscure satellite observations. We present SnowGalileo, an AI model that combines data from multiple satellites to produce snow maps. The model can produce accurate maps even when clouds are present and high-resolution data are missing. This helps to improve our understanding of snow conditions and their impact on water resources.


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