the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 22 Jul 2026
Review article: A decadal review (2015–2025) of machine learning models applied for satellite-based snow depth retrieval
Abstract. Ongoing climate warming is impacting the frequency and magnitude of extreme weather. The high sensitivity of snow to changes in temperature and precipitation makes it a primary indicator of climate change. Previous studies have proven that the snow cover extent has decreased with rapid warming. Nevertheless, this remains controversial, and no solid conclusion has been reached regarding snow depth changes. Numerous remote sensing‐based approaches have been used to derive spatially continuous snow depth. However, challenges remain in capturing and understanding the spatial variability of snow depth because of the non-linear and so-called ‘ill-posed’ problems associated with inversion framework. Machine learning (ML) techniques (including deep learning) are beginning to play important roles in advancing snow depth retrieval with microwave remote sensing, owing to their strong ability to fit nonlinear, nonexplicit functional relationships between snow depth and massive amounts of geoscience data. However, a systematic review of ML applications in snow depth retrieval with remote sensing is notably absent from the literature, and the trajectory for future advancements remains ambiguous. This review comprehensively summarizes the implementation and progress of snow depth research using microwave remote sensing over the last decade (2015–2025), and analyzes current research directions and areas where further developments are needed. An analysis of the literature reveals that the number of ML-related articles has increased over the past 10 years, rising from 3 to 33. By first-author affiliation, China and the United States lead in terms of contributions, accounting for almost 70 % of papers. We also found that western countries are actively engaged in high-resolution snow depth retrieval (ranging from meter to hundreds of metres) at regional or catchment scales (especially over mountains), which is attributed to their dense and comprehensive ground-based and airborne field campaigns (e.g., SnowEx, NoSREx, and ASO Lidar etc.). While China focuses on snow depth retrieval at the global scale or regional scales, typically at a coarse spatial resolution (10 or 25 km) or spatial downscaling (1 km or 500 m). Our decadal review concludes with five existing paradigms, namely, the coupling of ML and snow physical model (snow electromagnetic model or process model); developing snow electromagnetic models for simulating microwave signals in assimilation or iteration algorithms; optimizing snow electromagnetic models by providing key inputs or accelerating operational efficiency; improving existing gridded snow depth products by data fusion, bias correction or assembly techniques; and downscaling coarse snow depth products to a fine-scale resolution. However, some challenges and unresolved issues still exist. Our future efforts should aim to bridge the disparity in model–observation mismatch, integrate fundamental physical laws into ML structures, enhance the quality of ML training samples, and improve snow depth estimates under complex conditions (e.g., in mountainous and polar regions and during the snowmelt season). This paper provides a comprehensive review of the applications of ML techniques in snow depth remote sensing, focusing on current paradigms, existing challenges, and potential future research directions. We believe that ML techniques hold significant potential for addressing the challenges associated with the quantitative inversion of snow depth and deepening our understanding of the spatial variability of the snowpack globally.
- Preprint
(3247 KB) - Metadata XML
- BibTeX
- EndNote
Status: open (until 06 Sep 2026)
- RC1: 'Comment on egusphere-2026-2859', Anonymous Referee #1, 22 Aug 2026 reply
-
RC2: 'Comment on egusphere-2026-2859', Anonymous Referee #2, 01 Sep 2026
reply
Thank you for the invitation to review this manuscript. The paper is a review of machine learning (ML) applications in satellite-based snow depth retrieval. This topic is highly relevant to the journal and represents an important area of development for the community. The authors have compiled a large number of papers that contribute to this field.
My primary concern is that the manuscript frequently mischaracterizes the methodology and findings of the cited literature, and provides more of a superficial review of the source material. In numerous examples, the claims made in the text are entirely unrelated to, or unsupported by, the cited papers. Furthermore, the manuscript lacks logical flow and synthesis; rather than providing deep insights into the evolution of ML-based snow-depth retrieval algorithms, it reads like an uncritical compilation of information lacking careful findings about the main knowledge gaps, contributions of the paper, and future directions.
Below, I have detailed my major concerns and specific line-by-line corrections.
Major Concerns:
- Mischaracterization of cited literature:
Throughout the manuscript, papers are cited incorrectly. A major flaw is the cross-use and conflation of snow cover, snow properties, snow depth, and SWE. These are distinct parameters. The papers included in this review cover all of these types but are not well separated. The manuscript needs to explicitly focus on snow depth (as the title suggests) and properly distinguish it from SWE and snow cover.
- Lack of proper synthesis and organization:
The papers are not organized or grouped to reflect the evolution of ML techniques or retrieval algorithms. Strangely, the authors summarize findings based on the country of publication (Lines 270-285) rather than treating the topic as a unified community effort or organizing by methodological advancements.
- Misunderstanding of scientific and technical challenges:
The authors’ assessment of the challenges facing ML in snow depth retrieval (e.g., Line 130) fundamentally misrepresents the state of the science. Stating that ML struggles because snow is not a "stationary spatial element" and that it "accumulates and melts" is scientifically inaccurate. This is basic seasonality. The actual, well-documented challenges in the PMW/ML community like domain transferability (models failing outside their training geography), a severe lack of high-quality spatial training data in complex terrain, deep-snow signal saturation, and the physical transition to wet snow, which drastically alters microwave emissivity. This section must be entirely rewritten to discuss true technical bottlenecks.
- Lots of unrelated material:
The introduction is poorly structured and fails to provide a clear view of the evolution of ML models in snow depth retrieval. Instead, it contains extensive unrelated material about snow cover, SWE, general Large Language Models (which have not been used in snow remote sensing). Other information like how to use ML to solve the VIE, ML applications in RTMs, generational challenging in snow remote sensing (i.e. the whole 4.4 session), which all show little connection to satellite-based snow depth.
- Excessive reproduction of figures:
The manuscript includes far too many figures adopted directly from original papers (Figures 7 through 17). A review paper should synthesize concepts and data into original summary figures, rather than simply adopting a dozen figures from existing literature.
- Inconsistent terminology:
Please use consistent language throughout the manuscript. For example, choose one standard term among "Physics-informed," "Physics-based," "Physics-guided," or "physics-embedded," and apply it consistently.
Specific Comments:
(Note: because the original line numbers appear incorrect in places, these comments refer to the corresponding line sessions).
Line 50: Please rewrite these two sentences to clearly define the differences between snow cover, snow depth, and SWE, and explicitly state the main focus of this review paper.
Lines 60-65: Please double-check the papers cited here. Several of these focus on SWE retrieval from Passive Microwave data, not snow depth.
Line 70: Consider reducing the historical content regarding radiative transfer models, as this is not the core focus of the paper.
Line 95: Spell out “DMRT”.
Line 105: Evora et al., 2008 is about SWE retrieval, not snow depth.
Line 110: Please verify this statement. Both Gao et al., 2009 and Tong et al., 2010 focused on SWE retrievals. Furthermore, only Tong et al., 2010 compared ANN models trained with AMSR-E and SSM/I.
Line 115: Do you mean “data assimilation”? It is unclear what is meant by "observation operators for radiance assimilation framework" and how ML models are used in this circumstance. Please clarify.
Lines 115-125: A large number of ML-related papers are listed here, but very few relate to the actual focus of this paper (snow depth retrieval). Please remove unrelated papers to keep the context concise.
Line 130 - 135: As noted in my major concerns, comparing the seasonal changes of a snowpack to "stationary spatial elements" (like soil types) as a limitation of ML is scientifically inaccurate. The challenges section must reflect actual ML limitations. Furthermore, Line 135 shifts to snow cover, which is off-topic.
Line 170: While LLMs are a recent advancement, they have not yet been meaningfully applied to snow remote sensing. This information is irrelevant to the paper's scope.
Line 255: The authors need to review all papers cited here, as many are unrelated to ML snow depth retrieval. Consequently, the accuracy of the main Figure 4, which is based on this text, is highly questionable.
Line 295 (Bair et al., 2018): The information provided regarding Bair et al., 2018 is incorrect. This paper is about using ML for SWE estimation, not attempting "to combine a snowmelt process model (ParBal) with an RF for real-time estimates in the watersheds in Afghanistan." Furthermore, ParBal is an energy balance model, not strictly a snowmelt process model.
Line 295 (Broxton et al., 2019): This paper is about SWE estimation using ML, not "using an ANN to estimate snow density."
Line 330: This sentence is logically flawed. Please reword for clarity.
Lines 375-405: This entire section focuses on using ML to solve the VIE, which has little practical connection to the core topic of satellite-based snow depth retrieval.
Lines 484-510: The challenges and future prospects section lacks deep scientific insight and reads as a disconnected compilation of recent ML developments that do not link back to the specific limitations of snow depth retrieval.
Comments on Figures:
Figure 1: Change the right y-axis label to ‘Percent of total publications (%)’.
Figure 2: The authors list Bair et al., 2018 as the milestone for stage two. As noted above, this paper is about SWE retrieval, not specifically snow depth during the melt season in mountains.
Reproduced Figures: For all figures adopted from original papers (e.g., Fig 9), it is strictly necessary to include the figure number of the original paper and properly cite the source in the caption.
Figure 18: The proposed future direction is highly speculative and merely compiles current research workflows. It is unconvincing without an explicit, physically grounded explanation of why and how this workflow would succeed.
Figure 19: Should this be "Loss Function Embedding (LFE)"? Please correct the caption and the main text where this is mentioned (e.g., Line 525).
Citation: https://doi.org/10.5194/egusphere-2026-2859-RC2
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 178 | 58 | 11 | 247 | 17 | 18 |
- HTML: 178
- PDF: 58
- XML: 11
- Total: 247
- BibTeX: 17
- EndNote: 18
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
Thank you very much for providing the chance to review “Review article: A decadal review (2015–2025) of machine learning models applied for satellite-based snow depth retrieval.” The authors present the last decade of machine learning research on snow depth from spaceborne sensors. While I believe this topic is certainly worthwhile, I found the methods to be poorly justified and the topics to drift substantially throughout the paper.
Major Comments
- I found the writing of this paper to be disorganized at times, with several paragraphs drifting in direction. For example, in the Abstract, the authors begin with extreme weather, then switch to snow depth as an indicator for climate change, and then switch to snow cover extent. A stronger beginning would simply be to discuss the importance of snow and the current uncertainties in measuring. Throughout the paper their structure is somewhat unclear. They present a synthesis on snow depth (but also at times SWE), and then proceed to focus mostly on microwave-focused retrieval algorithms, and then they end with future directions. The ending is the strongest part, but the authors do not report specific accuracy metrics across studies so it is difficult to know which challenges are the most important.
-Additionally, the authors focus on snow depth for the entirety of the paper, but they include SWE in the search term without providing justification for why/why not this might not be related and whether SWE was later screened out or not. These two properties are different and require different both ML models and algorithms, and it would be good for the authors to justify this methods decision.
Critically, I think the authors need to spend more time digging substantially through the papers to provide quantitative comparisons (not just one is superior to the other, but by how much) and repeat their search to either definitively include SWE or not. Finally, I think some stronger organization would help the reader to understand the overall flow, and I think the title does not reflect what the paper is about. The current paper provides an overview of the literature, but it has a strong focus on microwave, that I think this needs to be better reflected.
Abstract
Introduction
The introduction currently reads as a comprehensive literature for the progress made in microwave remote sensing for snow depth, evolving from various microwave models. They also switch between multiple snow properties including snow cover, snow depth, and SWE. I believe their title on decadal review of ML for snow depth is misleading. It seems a better title would be to clarify that the bulk of their review focuses on the evolution of spaceborne remote sensing for snow depth, touching on both physical and ML methods. As it reads right now ML isn’t introduced until paragraph 5.
Overview
2.2 Main countries and journals of published articles from 2015–2025
2.3 Spatial resolution and source of target predictors in snow depth retrieval
- Line 230-265: This is summary information, and in order to assess the progress and evolution of the field, it is necessary for the authors to dig more into these papers and examine how well each of these papers did for snow. Does the number of publications reflect the utility and where the future direction should go? Given that the authors claim that this is one of the primary goals for this paper, I believe it is necessary for this whole section to be expanded upon and for the authors to compare accuracy metrics from each of these papers.
2.4 Top 10 most cited articles in the past 10 years
- Line 293 – 297: Where was the citation information pulled from? Web of science? This needs to be defined in the methods with the search term. The analyses presented, as it stands now, appear to be somewhat ad hoc, without much justification for why they looked at certain criteria. The authors need to outline this at the beginning and how this relates to their overall goal of understanding progression and future direction.
- Line 297-300: “To improve SWE estimates in a semiarid interior region, Broxton et al. (2019) used an ANN to estimate snow density by combining snow surveys and Lidar measurements. Yang et al. (2020) first explored the potential of a RF approach in snow depth retrieval and demonstrated that its spatial transferability is poor due to the dynamic snow characteristics.” This is the first time the authors bring up SWE in their literature (besides mentioning it in the search term). It’s unclear how this relates to snow depth, and why it is included.
- Figure 6. How many citations were these top 10 papers? What do the authors think about these papers besides the fact that they were cited the most? It’s difficult to know how to interpret this information and what will be useful for the snow community based off this information.
- Line 313: “ML contributes to accurate snow depth retrieval with remote sensing techniques. Here, we summarize five main directions for improving snow depth estimation.” The authors say they will talk about five main directions but it is unclear what those directions are. The five main directions need to be stated at the beginning.
3.1 Coupling ML technique with snow physical model
- line 316-319: typo for examples. Also what are the physics-based snow models? How much was the error? This is all quite superficial information, and for this literature review it would be useful for the authors to include the current baseline of these methods.
3.2 Applying RTM for simulating microwave signals
- This is another place where the authors introduce an analysis that is separate from the literature (generating a database) without justification. Please bring this up in the beginning and explain how this fits with your overall view.
3.3 Optimizing snow RTM model
3.3.1 Providing key parameters for snow electromagnetic model
- Line 363: I like this section, but I am afraid you’ve lost this reader for how this relates to the initial research gap that was presented. This paper feels like it is discussing more the evolution of RTM models/ other types of microwave models for snow depth, using ML, and I believe this needs to be better stated in the beginning.
4.1 Bridging the disparity of model–observation mismatch
Line 482: “has been has been” please fix this typo
Line 495: “Previous studies have demonstrated that a multi-layer RTM is superior to a one-layer model (Zhu and Tan et al., 2018; King et al., 2018, 2020; Rutter et al., 2019; Pan et al., 2024).” By how much?
4.3 Developing high-spatial-resolution snow depth products
- Line 600: Here and throughout, I found that the Figures were dropped in without much description, and I found myself having to jump back and forth between the text and the figure to determine what the authors were trying to say. Can the authors be sure to incorporate what the main takeaway points are from each figure into the text? This will help with readability.