the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 26 Jun 2026
Satellite-derived management indicators improve modeling of water and greenhouse gas fluxes in Swiss agroecosystems
Abstract. Agroecosystems regulate carbon, water, and nitrogen cycles, yet robust modeling of water and greenhouse gas (GHG) fluxes remains limited by incomplete or inaccessible information on field management practices. Although high-resolution remote sensing (RS) observations can detect management events such as mowing or harvest, their use for representing management intensity and associated impacts on ecosystem flux dynamics remains limited in existing models. Here, we developed an RS-assisted modeling framework to estimate daily latent heat flux (LE), net ecosystem CO2 exchange (NEE), nitrous oxide (N2O), and methane (CH4) fluxes across six Swiss FluxNet sites (two croplands and four grasslands) between 2016 and 2025. Sentinel-2 time series were used to derive leaf area index and RS-based field management indices (RS-FMIs), detecting mowing events, quantifying defoliation intensity, and identifying crop rotation and bare soil periods. These indicators were combined with meteorological drivers to train XGBoost models for each ecosystem type and target variable separately, and driver contributions were evaluated using SHapley Additive exPlanations (SHAP) analysis.
The RS-FMIs effectively captured in situ recorded management events and enabled improved reconstruction of daily flux variability. Model performances were strong for LE (R2 ≈ 0.89–0.90) and NEE (R2 ≈ 0.59–0.71), whereas N2O and CH4 fluxes were reproduced with moderate accuracy (R2 ≈ 0.37–0.55). Models using RS-FMIs performed similarly to those using well-compiled in situ management records, supporting the ability of RS-derived vegetation and management indicators to represent management effects. LE variability was primarily energy-driven and dominated by meteorological conditions, whereas vegetation dynamics and RS-FMIs played stronger roles in shaping NEE, N2O, and CH4 variability. These results demonstrate that RS-FMIs offer new opportunities to reconstruct management information and improve the representation of management effects in agroecosystem flux modeling.
- Preprint
(15635 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
- RC1: 'Comment on egusphere-2026-3522', Anonymous Referee #1, 22 Jul 2026
-
RC2: 'Comment on egusphere-2026-3522', Anonymous Referee #2, 07 Aug 2026
The manuscript “Satellite-derived management indicators improve modeling of water and greenhouse gas fluxes in Swiss agroecosystems” estimates latent heat and greenhouse gas (GHG) fluxes using a machine learning (ML) framework that applies meteorological (METEO) and other remotely sensed vegetation (RS-VPI) and field management indices (RS-FMI) as predictor variables. The authors show that by including RS based predictors in their ML framework, they improve flux estimation, especially the GHG fluxes. Overall, the study shows the potential offered by high resolution RS at detecting poorly documented land management events/practices, which in turn contribute to GHG emissions. The paper is properly written and structured, and is well suited for this journal. However, the modeling framework was only applied over a few (six) Swiss (2) crop- and (4) grass-sites, with its spatial generalization potential remaining untested. Remarks:
- The limitations regarding transferability and generalizability of the ML models developed in this study (which appear to only be applicable over their pre-selected Swiss sites) need to be addressed/clarified. First, the authors present two frameworks for the different land covers in their study sites. Why not train one model (i.e., one for each flux) that could be applied for both cover types—or even one that could be applied on other land covers such as forests? Second, GHG gas emissions occur globally, and it has generally been understood that such emissions contribute to climate change at global scales. GHG modeling frameworks that would be transferable or applicable at larger scales should therefore be preferred in place of localized ones. The authors need to discuss these limitations and clarify whether (and/or how) they intend to evaluate a more generalized framework over diverse cover types and broader spatial scales.
- This study proposes an interestingly simple yet visibly effective method to infer defoliation events from remotely sensed variables, specifically EVI. The authors show that the DII index is capable of timing defoliation events relatively well. However, the DII evaluations/sensitivity analyses performed are insufficient. E.g.: selection of window size(s) of surrounding background(s) appears somewhat arbitrary; why did the authors prefer EVI to some other variable like NDVI, which requires relatively less spectral information to estimate?; …etc.
To make the DII index applicable at larger (even global) scales across diverse land management practices and land-cover types, further analyses are required.
- The manuscript is sometimes too wordy. To improve flow and readability, rephrase/break long sentences, especially in the introduction and methods section.
L38-...: “Long-term eddy-covariance … Merbold et al., 2014)” For example, is too long and can be broken into at least 2 parts to make it easier to follow.
Specific comments:
L5-6: “… we developed an RS- assisted modeling framework …” – the authors should clearly indicate that the modeling performed here is based on applying machine learning modeling (i.e. statistical learning algorithms) and not physically based process modeling.
L13-15: “… LE (R2 « 0.89-0.90)” – for completeness, maybe also include the respective rmse or bias values here.
L72-74: “While spatial upscaling to unobserved regions remains outside the scope of this work, this …” – again, transferability/generalization of your ML models should be discussed a bit more than this. Could a similarly trained RS-driven ML model be easily applied in croplands and/or grasslands in other regions with similar climate or even globally?
L95: “No surface energy balance closure correction was applied to LE flux.” - Why? Is the energy imbalance at all 6 sites too little to warrant any corrections?
L116-119: “Due to the limited availability of gridded soil moisture and soil temperature data, these variables were not included in the analysis. Precipitation- based temporal features partially capture moisture variability.” - what about soil temperature? Satellite surface temperature (an important variable for inferring surface turbulent flux exchanges) can be used as a proxy for soil temperature..
Section 2.1.2: Restructure section for coherence. Too wordy.
L203: “… is the corresponding spatial standard deviation.” – ‘corresponding’ is not very clear here. spatial standard deviation of ? add …of EVI.
L213-…: “Sensitivity analysis indicated that a 1 km window best captured management-related disturbance signals in grasslands, whereas a 2 km window was more appropriate for croplands (Appendix B) “ – You come to this conclusion after basing your analysis on R^2, which is maybe not the best measure. Did you test with multiple objective functions, e.g. RMSE, bias …?
L226-228: “A field was classified as bare when the bare soil fraction exceeded 0.5” – 0.5 bare soil fraction would mean 50% vegetation cover fraction …is 0.5 fraction of baresoil not too low to consider the surface totally bare?
L249-…: “Grassland and cropland systems were modeled separately to enable a comparison of their sensitivities to environmental and management drivers within a unified modeling framework.” - the authors need to better clarify why separate ML frameworks (as per AppendixC: Table C1) were needed for the 2 cover types. Why not have a generalized ML model (per target flux) that could be applied over both land cover types?
L271: “of individual variables ” – since you also call LE and GHG fluxes ‘”target” variables’, you need to clarify what these ‘individual variables’ are – ‘…individual predictor variables (i.e. METEO, RS-VPI, RS-FMIs)’?
L295-…: “… The remotely-sensed DII was evaluated against recorded management events at representative long-term grassland and crop¬land sites before being used as an RS-FMI in the flux modeling framework. At the CH-Cha grassland …” - what was the consideration here for selecting the two sites as the representative sites. Indeed, DII peaks appear to generally coincide well with the real defoliation events over the grassland but not so well over the selected crop-site. Is this the case over the other 4 sites as well?
L329..., Figure5, Figure6: Can you also provide the evaluation performance metrics per site separately, in addition to the combined analyses (maybe in the appendix or supplements).
L331: “the NEE range, with underestimation of both strong negative and high positive fluxes” - from the scatterplots, strong negative NEE fluxes appear to be higher than the observations so not sure how that is an ‘underestimation’
L381-383: “… The grouped SHAP results showed increased contributions from RS-derived vegetation and management indicators in croplands compared to grasslands (Figure 9), indicating a stronger role of vegetation structural dynamics.” – the authors should clarify that this statement generally only applies to the LE and NEE fluxes (see figure 9)
L407: “… A LOSO ” - LOSO acronym needs to be described prior to use.
Appendix D: …the analyses here (Figure D1-noninterpolated vs Figure7-interpolated) only consider one [grass] site (CH-Cha) in 2021. Can the same conclusions be reached when all sites and/or other periods are considered?
Citation: https://doi.org/10.5194/egusphere-2026-3522-RC2 - The limitations regarding transferability and generalizability of the ML models developed in this study (which appear to only be applicable over their pre-selected Swiss sites) need to be addressed/clarified. First, the authors present two frameworks for the different land covers in their study sites. Why not train one model (i.e., one for each flux) that could be applied for both cover types—or even one that could be applied on other land covers such as forests? Second, GHG gas emissions occur globally, and it has generally been understood that such emissions contribute to climate change at global scales. GHG modeling frameworks that would be transferable or applicable at larger scales should therefore be preferred in place of localized ones. The authors need to discuss these limitations and clarify whether (and/or how) they intend to evaluate a more generalized framework over diverse cover types and broader spatial scales.
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 84 | 47 | 9 | 140 | 7 | 8 |
- HTML: 84
- PDF: 47
- XML: 9
- Total: 140
- BibTeX: 7
- EndNote: 8
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
The manuscript (egusphere-2026-3522)” Satellite-derived management indicators improve modeling of water and greenhouse gas fl uxes in Swiss agroecosystems” describes the development of an RS-assisted modeling framework to estimate daily latent heat flux (LE), net ecosystem CO2 exchange (NEE), nitrous oxide (N 2 O) and methane (CH4 ) fluxes across six Swiss FluxNet sites (two croplands and four grasslands) between 2016 and 2025.
The framework is based on Sentinel-2 time series to derive leaf area index and RS-based field management indices (RS-FMIs), allowing to detect mowing events, quantifying defoliation intensity, and identifying crop rotation and bare soil periods. Four scenarios where tested to predict LE, NEE and GHG ; 1) meteo only, 2) Meteo and field management, 3) meteo and RS +vegetation performance index (VPI), and 4) Meteo, RS+VPI and field management indices (FMI)
Several indicators were combined with meteorological drivers to train XGBoost models for each ecosystem type (crop and grass) and target variable separately, and driver contributions were evaluated using SHapley Additive exPlanations (SHAP) analysis in order to assess the relative importance.
Analyses showed that RS-FMIs effectively captured in situ recorded management events and enabled improved rebuilding of daily flux variability. Model performances were strong for LE (R 2 ≈ 0.89–0.90) and NEE (R 2 ≈ 0.59–0.71), whereas N2O and CH4 fluxes were reproduced with moderate accuracy (R 2 ≈ 0.37–0.55). Models using RS-FMIs performed similarly to those using well-compiled in situ management records, supporting the ability of RS-derived vegetation and management indicators to represent management effects. Likewise the study showed that the incorporation of satellite-derived vegetation dynamics and disturbance signals, have a broader potential of remote sensing-informed management representations for monitoring agricultural greenhouse gas fluxes.
I read the manuscript with interest. It presents a novel methodology for predicting fluxes and analysing their drivers in agroecosystems. The manuscript is well written and well illustrated. Although the methodology involves complex modelling and data processing, the manuscript is generally easy to follow. In my opinion, the manuscript is suitable for publication after minor revisions
General comments .
Introduction: I suggest adding a few lines discussing the temporal heterogeneity of the different measurements.
Using EO images for modelling is not straightforward because of several limitations that are not mentioned, such as cloud cover, satellite revisit frequency, pixel size and image quality (see also L123ff).
Furthermore, fluxes are measured every 30 minutes (aggregated to daily values), whereas management events occur at much lower temporal frequency. Matching these different temporal scales is therefore an important challenge and deserves some Intro and discussion.
Materials and Methods: More clarification is needed regarding the EO products. It is not always clear where the products originate from (Copernicus, Sentinel processing chain, published datasets, etc.), how they were generated, whether they were produced by the authors, and whether existing products or previously published methods were used. Please clarify this around L121ff and provide appropriate references (DOIs where available).
The section describing the Sentinel-2 products contains a lot of information and is sometimes difficult to follow. I wonder whether a few italic subheadings would help distinguish between existing methods/products and the developments introduced by the authors. Alternatively, a small summary table could be useful, listing existing products, processing steps, improvements, training, and outputs.
Regarding Sections 2.2.2 to 2.2.4 onwards, I may simply have misunderstood the workflow, but this could indicate that L165–173 would benefit from additional explanation. Personally, I would move the model framework section after the descriptions of remotely sensed defoliation event detection, defoliation intensity estimation, and bare soil period detection, which would make the methodological flow easier to follow.
Section on sentinel-2 products, is a lot information and reader has difficulties to follow. I wonder if some sub-titles (in italic) may help to identify what was taken over from existing work/methods and what was added here I wonder if a small table may be useful... eg 3x collumns existing products and work/improvement/choices/ training... ect by authors
Sections 2.2.2 to 2.2.2 , might that I did not well understand , which might also show that L165-173 would need more detail (which arrive in L244), personally I would move the model framework after the Remotely-Sensed Defoliation Event Detection , Defoliation Intensity Estimation and Bare Soil Period.
discussion and Conclusion: In the last part of the Discussion (around L494 and L504) and in the Conclusions, the manuscript ends without a very clear take-home message. I encourage the authors to be more explicit about the readiness of the proposed framework. Is this mainly a proof-of-concept or first step, or is it already a mature and operational product? What are the main remaining limitations? What further developments would be needed before applying the framework more broadly or operationally?
Specific comments.
-It would also be useful here to briefly acknowledge some of the practical limitations of EO-based modelling (cloud cover, revisit time, spatial resolution, etc.).
-EO images in modelling is not so easy task due to many points not mentioned here, clouds, satellite passage, pixel size and quality (see L123ff)
-fluxes being measured on 30min (daily) and management events on monthly so to find the matching moment is a challenge
Please also explain the rationale behind selecting the four modelling scenarios. At present, the section is very brief and the progression from Scenario 1 to Scenario 4 is not entirely obvious. There is more detail in L244 but this is much later. Please think about moving sections ect
At present, it is difficult to interpret the figure because each driver contains both red and blue points of varying density. For example, for grassland CH₄, the "day of defoliation" variable contains both red and blue points. Does this indicate that high values can have both positive and negative effects depending on the context? A few additional sentences in the legend would greatly help readers interpret these plots correctly.