the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 15 Jun 2026
A hybrid physics and machine learning approach highlights and alleviates soil moisture downscaling challenges
Abstract. Soil moisture information is crucial for predicting natural hazards like droughts and landslides, improving climate and weather forecasts, understanding soil-atmosphere exchanges and feedback, and better managing water resources. Soil moisture networks are sparse, and remote sensing products lack necessary resolution. Downscaling techniques aim to enhance resolution, but face challenges because the ancillary data they rely on are often sparse and lack representativeness.
Here, a ”downscaling playground” is created over the Contiguous USA using physics-based hydrological simulations that allow downscaling techniques to be trained and tested without practical limitations. We show that existing biases in soil moisture station locations lead to spatial overfitting, resulting in poorer performance than, for example, if those stations had been randomly located. The biases lead also to unwarranted confidence in the improvement achieved with downscaling, as testing is carried out on stations in similar locations. Finally, we propose a new paradigm that fuses physics-based modelling (ParFlow-CLM) with machine learning (XGBoost) and downscales remote sensing observations (e.g., SMAP) to produce a 1km2 soil moisture product over CONUS. We test the approach thoroughly against in situ observations.
Competing interests: At least one of the (co-)authors is a member of the editorial board of Hydrology and Earth System Sciences.
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.- Preprint
(52487 KB) - Metadata XML
- BibTeX
- EndNote
Status: open (until 19 Aug 2026)
-
RC1: 'Comment on egusphere-2026-2956', Heye Bogena, 24 Jun 2026
reply
-
AC1: 'Reply on RC1', Elena Leonarduzzi, 08 Jul 2026
reply
We thank the reviewer for a thorough, constructive, and helpful review. We address the major comments in turn below; as this is the interactive discussion phase, we focus here on the substance of our response and will translate the agreed points into concrete manuscript changes (and a formal response-to-reviewers) afterward.
On the metrics themselves, we'd like to clarify our reasoning and hear whether it addresses the concern. Our choice to lead with the Crow et al. (2022) framework rather than conventional RMSD/ubRMSD/ME wasn't meant to sidestep a direct comparison against observations, it reflects what the framework is designed to isolate. Conventional metrics computed against point observations mix temporal and spatial performance together, and the temporal component tends to dominate the signal (which is also why, as you note, most downscaling studies struggle to show clear improvement against ground stations in these metrics, the target dynamics are already largely captured or completely missed by the coarse product, so a downscaled product has limited room to improve on them). The Crow et al. framework is built specifically to separate the two, so that any spatial (sub-footprint) information added by downscaling doesn't get swamped by temporal correlation that the coarse product already captures.
That said, in Figure 1, the scatter plots of RMSD, ubRMSD, R², and correlation for RF, XGBoost, and SMAP against in-situ observations, computed station-by-station. The three products are broadly comparable on these conventional statistics, with XGBoost/RF showing a somewhat wider spread but similar medians (slightly worse) to native SMAP. We still believe that using Criteria 1 of Crow et al. rather than these “more conventional” statistics is most informative and less impacted by scaling issues and the strongest spatial variability. Rather than asking "does the pixel value match this one point through time," it's structured to test whether the downscaled product preserves the temporal fidelity that's already present in the parent coarse product, without that comparison being corrupted by the fact that a station isn't a pixel-mean estimate. In other words, it's evaluating "did downscaling break the temporal information we started with" rather than "does this pixel match this point," which is a fundamentally noisier and more scale-mismatched question.
On whether the product "outperforms" SMAP, this is a fair and important challenge, and we will revise the summary and conclusions accordingly softening some of the conclusions. However, we would like to clarify where we believe the novelty of this work actually lies, since we think this reframes the rest of the discussion. It is not, in our view, in the specific downscaled product itself, or even in the specific choice of algorithm. One could in principle apply the same approach to a different satellite product, or swap XGBoost for another machine learning method, and possibly improve the downscaled approach. The novelty is in the framework and approach: using physics-based simulations as training data sidesteps two limitations that are otherwise very hard to avoid. First, it is not limited in space: the simulation provides a gridded estimate everywhere within the domain, not just at the sparse set of locations where stations happen to exist, so the training (and validation) data are not subject to the sampling biases we demonstrate in the first part of the paper. Second, and just as importantly, it avoids the scale mismatch problem entirely: we are training a model against a target that is already a grid-cell estimate, consistent in support with what we want to predict, rather than against a point measurement standing in for a pixel. This is, we think, a more fundamental contribution than the performance of any one downscaled product.
With that framing in mind, we return to the validation question. We are not claiming to have produced a product that is uniformly better than SMAP in the conventional sense: Criterion 1 (largely temporal) is indeed close to unchanged, and Criterion 2 (spatial) is mixed and sometimes negative, exactly as you point out. As noted above, this is consistent with the broader literature: downscaling studies validated against point stations with standard statistics hardly ever show significant, consistent improvement, for the scale-mismatch reasons we discuss throughout. What we think is genuinely novel is that Criterion 2 shows a clear, physically interpretable seasonal cycle in where and when the downscaling adds spatial value (and where it doesn't, e.g., at higher elevations), a pattern that is already visible in the results we present, and that, to our knowledge, no prior downscaling study has been able to isolate and demonstrate, because doing so requires separating spatial from temporal skill in the first place (as Crow et al., first suggested) and breaking it down into individual daily values. So, our claim is narrower than "better product" and more like "a framework that lets us diagnose when and where sub-footprint downscaling is adding real information" which is a more modest but, we think, more defensible and more useful contribution. We will restate this explicitly in the revised conclusions, so the claim matches the evidence.
Finally, we thank the reviewer for mentioning Montzka et al.'s work. The authors’ approach downscales using static soil-texture-derived sub-grid variability alone. We will make the case in the Discussion that this captures only the time-invariant component of sub-footprint heterogeneity, whereas the ParFlow-based approach could in principle also reflect dynamic drivers that modulate sub-footprint variability in time, which could also be one of the explanations for Criterion 2 signal being seasonal rather than static (although we agree that Criterion 1 shows no temporal improvement overall).
Thank you again for such a careful and engaged review, it has already sharpened how we plan to frame the paper's contribution.
Figure 1. Performances of the SMAP downscaled products (with Random Forest and XGBoost) or SMAP without downscaling against in situ observations. Scatters are colored by stations’ elevation. The red lines indicate the median of each product and statistical metric. Note: the limits of the y-axis for the Coefficient of Determination were manually adjusted removing some (ca. 5) of the very negative outliers.
Citation: https://doi.org/10.5194/egusphere-2026-2956-AC1
-
AC1: 'Reply on RC1', Elena Leonarduzzi, 08 Jul 2026
reply
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 7 | 11 | 2 | 20 | 0 | 1 |
- HTML: 7
- PDF: 11
- XML: 2
- Total: 20
- BibTeX: 0
- EndNote: 1
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
The authors introduce a "downscaling playground" based on ParFlow-CLM simulations and use it to investigate fundamental limitations of traditional downscaling approaches that rely on sparse in-situ observations. They show that the spatial distribution of monitoring stations can introduce strong biases and overoptimistic estimates of model performance. Building on this framework, they propose a hybrid approach that combines physics-based simulations with machine learning (XGBoost) to downscale SMAP soil moisture to 1 km over CONUS.
The manuscript addresses an important topic in hydrology and remote sensing and raises a valuable point regarding representativeness and validation. The idea of exploiting physically based simulations as a surrogate training dataset is very innovative and has the potential to influence future downscaling studies.
However, in its current form, the manuscript contains a few weaknesses that limit the strength of its conclusions, including aspects of methodological clarity and the interpretation of results; these are detailed in the major and minor comments below.
Major comments:
The central assumption of the study is that ParFlow simulations provide realistic high-resolution reference data, and this assumption underpins almost all experiments. Although the authors acknowledge its limitations (Sections 4.2 and 4.3), several conclusions appear stronger than warranted by the evidence presented. While the manuscript compares the downscaled soil moisture product with independent in-situ observations using the metrics proposed by Crow et al. (2022), the validation would benefit from a more direct assessment of how well the downscaled product reproduces the observations themselves using conventional statistics (e.g., RMSD, ubRMSE, ME).
The most important practical outcome of the study is the downscaling application of the SMAP observations. However, validation against observational data does not convincingly demonstrate that the proposed product consistently outperforms the original SMAP product. Criterion 1 remains essentially unchanged, whilst Criterion 2 is close to zero or even negative (e.g. −0.211 for XGBoost), with improvements occurring only during certain time periods and, above all, at stations at low altitudes. The results therefore provide evidence of the feasibility of the proposed approach rather than clear proof of superior performance. In this context, some statements in the summary and conclusions appear exaggerated, particularly those suggesting that the approach ‘mitigates the challenges of downscaling soil moisture’ or ‘enables the production of high-resolution soil moisture estimates’. I recommend toning down these claims and instead emphasising the methodological potential of the hybrid framework, whilst acknowledging that its advantages over the native SMAP have only been partially demonstrated by validation against observational data.
The manuscript is generally well written but would benefit from tightening, particularly in the Introduction, improved consistency in terminology, and a more focused narrative that reduces redundancy and enhances readability. In addition, the experiments could be presented more clearly, for example by introducing a dedicated subsection in the Methods and by improving the clarity and informativeness of Fig. 1. This subsection should also explain the assumptions made, for example, why only 20 days were used for the testing and how these were selected.
The Discussion would benefit from a clearer comparison with existing downscaling approaches, as it currently lacks sufficient context within the broader literature. In particular, Montzka et al. (2018) proposed a physically based method using high-resolution soil properties to downscale SMAP to 1 km without requiring hydrological modelling. Given the relative simplicity of that approach, the manuscript should better clarify the added value of the more complex framework presented here and whether it leads to improved performance or broader applicability.
In my opinion, the Conclusions would benefit from the inclusion of a brief outlook outlining future research directions, potential pathways toward operational application, and remaining challenges for transferring the proposed framework beyond the current setup.
Specific comments:
L5: The term "downscaling playground" sounds somewhat informal. Consider replacing it with a term such as "simulation-based testbed", "controlled experimental framework", or "benchmark framework", which would better convey the methodological rigor and purpose of the approach.
L6: At this stage it is unclear what “existing biases in soil moisture station locations” means.
L11: CONUS needs to be introduced.
L11: The final sentence of the abstract is not effective. Abstracts should conclude with the principal findings and/or their broader significance rather than with a description of a methodological step.
L13: “soil moisture estimates”
L17: Rather old citation.
L22: I assume that "soil moisture" is meant here. Please use a consistent name for the variable throughout the manuscript and avoid switching between different terms unless a distinct quantity is intended.
L50: Which data sources are you referring to.
L80: Unclear which components are meant here.
L103-104: This should be reformulated into clearer way.
L104: Why did you choose 10 cm? SMAP’s effective penetration depth is generally considered about the top 0–5 cm of soil (e.g. Brown et al., 2013)
L123-125: This statement understates the representativeness problem. Even at 1 km² resolution, the support volume is still many orders of magnitude larger than that of typical in-situ sensors. This issue remains central to the interpretation of validation results and should be more explicitly emphasized.
L220-225 and L244-248: These sections should be moved to a separate section within the Methods section, where the experiments are described.
L242: The meaning of “classic downscaling” should be explained.
L377: Consider closing with an outlook.
Figure 1: The clarity of the schematic could be improved. The caption is also rather brief and could be made more informative by explicitly describing the main steps of the workflow.
Literature
Brown, M. E., Escobar, V., Moran, S., Entekhabi, D., O'Neill, P. E., Njoku, E. G., ... & Entin, J. K. (2013). NASA's soil moisture active passive (SMAP) mission and opportunities for applications users. Bulletin of the American Meteorological Society, 94(8), 1125-1128.
Crow, W. T., Chen, F., & Colliander, A. (2022). Benchmarking downscaled satellite-based soil moisture products using sparse, point-scale ground observations. Remote Sensing of Environment, 283, 113300.
Montzka, C., K. Rötzer, H.R. Bogena and H. Vereecken (2018): A new soil moisture downscaling approach for SMAP, SMOS and ASCAT by predicting sub-grid variability. Remote Sensing 10(3): 427. DOI: 10.3390/rs10030427