the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
Grassland yield estimations – potentials and limitations of remote sensing, process-based modelling and field measurements
Abstract. Grasslands make up the majority of agricultural land and provide fodder for livestock. Information on grassland yield is very limited as the fodder is directly used at the farms. Data on grassland yields would be needed, however, to inform politics and stakeholders on grassland ecosystem services and inter-annual variations. Grassland yield patterns are often varying on small scales in Germany and estimations are further complicated by missing information on grassland management. Here, we present three different approaches to estimate annual grassland yield for a study region in southern Germany. We apply (i) a model derived from field samples, satellite data and mowing information (RS), (ii) the biogeochemical process-based model LandscapeDNDC (LDNDC) and (iii) a rule-set approach based on field measurements and spatial information on grassland productivity (RVA) to derive grassland yields per parcel for the Ammer catchment area in 2019. All three approaches reach plausible results of annual yields of around 4–9 t/ha and show overlapping as well as diverging spatial patterns. For example, direct comparisons show that higher yields were derived with LDNDC compared to RS and RVA, in particular related to the first cut and for grasslands mown only one or two times per year. The mowing frequency was found to be the most important influencing factor for grassland yields of all three approaches. There were no significant differences found in the effect of abiotic influencing factors, such as climate or elevation, on grassland yields derived from the different approaches. The potentials and limitations of the three approaches are analysed and discussed in depth, such as the level of detail of required input data, or the capability of regional and inter-annual yield estimations. For the first time, three different approaches to estimate grassland yields were compared in depth resulting in new insights in their potentials and limitations. Grassland productivity maps provide the basis for long-term analyses of climate and management impacts and comprehensive studies of the functions of grassland ecosystems.
- Preprint
(3193 KB) - Metadata XML
- BibTeX
- EndNote
Status: open (until 07 Apr 2025)
-
RC1: 'Comment on egusphere-2024-4087', Anonymous Referee #1, 05 Mar 2025
reply
The paper describes and compares three different approaches for grassland yield estimation. Given the limited availability of spatially explicit data on grassland yields and the pressure of global change on agricultural ecosystems, this paper addresses a relevant topic and emphasises the potential of fine-scale monitoring using high-resolution spatio-temporal earth observation and environmental data.
The main novelty of this paper seems to me to be the presentation of a novel remote sensing-based approach that combines modelled AGB time series and mowing events, all derived from free and globally available satellite data. This approach is compared with an established process-based model (LDNDC) and a simple rule-based approach. The paper is well written and clear.
I recommend it for publication but encourage the authors to make some improvements to clarify and increase its contribution to a larger audience. My main criticism is expressed in the first comment, which suggests strengthening the focus and improving the results section.
General comments
1) I think the relevance of this paper could be increased with a stronger focus on the novel RS-based approach and its ability to capture small-scale spatial and temporal variations, with management information being a prerequisite, also for LDNDC. The importance of spatial/temporal variation is highlighted in the abstract and introduction, and its potential to be captured by the RS-based approach is also highlighted in the conclusion under L593/594, but in my opinion, this is not adequately represented in the results. I recommend a more detailed analysis of this.
A stronger focus on the RS-based approach would make it clearer what the novelty of this paper is, especially as the other two approaches are being currently under review in a separate paper, if I understood correctly. In this regard, it might be worth shortening the description of LDNDC in the methods and the detailed and rather descriptive comparison with the two other approaches (results and discussion section, e.g. L376-L407), which I think are interchangeable to some extent. This could be complemented with some in-depth analysis of the novel RS-based approach. For instance, I think the currently made statements such as “All three approaches reach plausible results of annual yields of around 4-9 t/ha and show overlapping as well as diverging spatial patterns” are not very meaningful. I'm also not sure how relevant this is for people who don't work with LDNDC, or if it allows general conclusions to be drawn for process-based models?
Can the authors show (map, stats) that small-scale variations are better captured with the RS-based approach (e.g. looking at adjacent in-situ measurements / within and across parcel variations)? Same for the temporal variation, e.g. which approach better captures AGB of the first cut (see statement L26/L470; it’s not clear to me, was it compared with in-situ measurements?)? I think these points are rather well covered in the discussion but could be improved in the results section. Also consider changing the title to give more weight to the RS-based approach and clearly mention the focus in the introduction.
If the authors choose to keep their focus on comparing the three approaches, more in-depth analysis is needed, e.g. to clearly understand why one model over/underestimates yield and under what conditions. This should go beyond “simply” describing differences at hexagon-level and frequency distributions in the result section.
In the first chapter of the results, please provide more details about the RS-based model, e.g. prediction vs. observed plot.
2) More information regarding the in-situ measurements is needed (section biomass field data). In my opinion it is not sufficient to simply provide a reference, at least an overview of the sampling design is needed, addressing the nested design, the spatial and temporal distribution, and the number of samples per mowing frequency.
3) The results focus on showing the differences between the three approaches (map, histogram). However, these are difficult to interpret in terms of which approach performs better. A more detailed comparison/analysis with in-situ measurements would give more substance to this paper. However, I’m not sure if the in-situ data allows such further analysis? If so, this could help to address questions raised under comment #1, is linked to specific comments bellow and the potentially missing in-situ symbols in figure 3.
4) The comparison of the RS-based model with LDNDC is a bit difficult to interpret, as both approaches rely on Sentinel-2-based mowing events. Please clarify what this means for the interpretation, e.g. what would be the performance of LDNDC without spatially explicit mowing information (or is there another data source for this)?
As both models are highly dependent on the RS-based mowing dates, the finding that the spatial patterns or box plots (Figure 8) between the two approaches are similar and strongly related to the number of mowing events is somewhat redundant/obvious (e.g. L418, L467). I don't know how this could be resolved, perhaps by running separate models with/without mowing events?
5) Please review figure and table captions to be more self-explaining and consistent (e.g. Figure 4 vs. Figure 5 vs. Figure 6; rather mention the different modelling approach like in Fig. 4) and use consistent terms in the whole document (e.g. LandscapeDNDC vs. LDNDC).
Specific comments
L21: Change “present” with “compare”, or clarify which approach is new (RS-based) and which already exist / will be published separately?
L140 /Figure 1: Change to “CLCplus Backbone 2021”, consider citation, remove not-shown legend items. Make clear that the study area is the Ammer catchment and add catchment border to sub-figures.
L155: Was the validation done separately for this study or is the F1-score derived from Reinermann et al.? More details or citation is needed. Please check that the uncertainty of mowing detection is adequately addressed in the discussion section, as this is the most important parameter.
L166: Add plot locations to the map in figure 1 (maybe top-left)?
L170/171: More information regarding the sampling design is needed, I cannot quite follow how the n=111 samples are reached
L178 (Figure 2): Add validation procedure?
L180: Title not clear, consider “Remote sensing-based approach”
L195: Please check EVI formula for correctness (position of factor 2.5 and apostrophe after 1)
L205: Please cite the HR-VPP (https://doi.org/10.2909/c1c46cb2-b02b-4013-aae5-a54a8c018b1e)
L220: It would be helpful to get some information about the distribution of these data pairs, e.g. a figure with y=field plots, x=dates S2&field, colour=S2/field/pairs(train/test), table (supplementary) or at least providing some more information in the text (how many field plots etc.)?
L244: Introduce r2 instead of L247; use PRMSE like in L248
L282: It sounds like the RVA is based on an official table, but then it gets more complicated with the tables and the link to Kaim et al. (under review). It is a bit confusing, please check/revise.
L286: Do you mean “Table A1”?
L311: Improve caption. Maybe move table to appendix, to make clear that it belongs to Table A1.
L335 (Figure 3): I cannot find the in-situ measurements! Provide annual values (which are a bit hidden) as stacked-bar adding AGB from the individual events?
L342: 0.97 seems to indicate overfitting and might need to be discussed later?
L389 (Figure 6): I suggest adding r2 and variance to the plots. Same for Figure 7.
L433: Remove one bracket
L518: You could mention the potential to tackle cloud coverage with SAR
L550: In this study LDNDC also depends on cloud-free observations for the mowing dates
Citation: https://doi.org/10.5194/egusphere-2024-4087-RC1 -
RC2: 'Comment on egusphere-2024-4087', Anonymous Referee #2, 07 Mar 2025
reply
Grasslands make up most agricultural land and provide fodder for livestock, yet data on grassland yields is limited due to direct use on farms. Accurate yield information is crucial for informing policy and understanding ecosystem services and inter-annual variations. In this study, we estimate annual grassland yields in the Ammer catchment area of southern Germany for 2019 using three approaches: (i) a model combining field samples, satellite data, and mowing information (RS), (ii) the biogeochemical process-based model LandscapeDNDC (LDNDC), and (iii) a rule-based approach based on field measurements and spatial productivity data (RVA). All approaches yield similar results, estimating yields of 4-9 t/ha, with some spatial variations. LDNDC generally produces higher yields, especially for the first cut and grasslands mown one or two times per year. Mowing frequency was the most significant factor influencing yields, with no major differences in the impact of abiotic factors (e.g., climate or elevation) across approaches. This comparison offers new insights into the strengths and limitations of each method and highlights the importance of grassland productivity maps for long-term studies on climate and management impacts.
This study is a pioneering effort to compare entirely different approaches for estimating grassland biomass and provides valuable insights for future land use estimations. I recommend it for publication, with minor to moderate revisions focusing on enhancing clarity and broadening its relevance for a wider audience.General comments:#1 The key findings highlight the spatial and temporal differences between the three approaches. However, these differences are difficult to interpret in terms of the underlying reasons why each model or approach yields distinct results, and it is unclear which approach performs better overall. The results section presents only the spatial and temporal disparities without delving into potential explanations. Factors such as climate forcing uncertainties in the LDNDC model input, the quality of data, methods used for remote sensing photo analysis, and limitations of survey data could all influence the outcomes.To enhance the clarity and depth of the analysis, I recommend the following:
- A more comprehensive comparison, either in a table or in the discussion section, outlining the pros and cons of each method.
- An in-depth analysis that investigates the potential root causes behind the observed differences in results, including an exploration of the most significant factors affecting each approach (beyond mowing).
- A discussion on the future implementation of these approaches, specifying under which conditions each model is more suitable than the others.
#2 Each method, due to its underlying model mechanism and input data, introduces uncertainties. However, the study presents only the estimated annual AGB from the three approaches without providing an uncertainty range for each estimation. To draw a more reliable conclusion about the similarity of the three results, it is essential to include upper and lower bounds for each estimate. I recommend incorporating the potential uncertainties associated with the input data for each method and providing confidence intervals (upper and lower bounds) for the annual AGB estimations.#3 The results suggest that all approaches are highly sensitive to mowing dates, with the estimated annual grassland yields showing a strong correlation to the number of mowing events. To assess the stability and effectiveness of each model, I recommend conducting a sensitivity test by removing the number of mowing events from all three models. This would allow for an evaluation of how each model performs with the remaining input variables#4 The annual grassland biomass estimation is based on a particular region in this study. It would be better to give some insights in the discussion section to show the possibility and potential use case to apply those three approaches to a bigger region, what the limitations (such as data availability) would be, and what method(s) will be most likely to serve better.Specific comments:#1 Figure 3. Related to my comment #2, try to add an uncertainty range for the annual AGB#2 Figure 5. Try to reverse the color bar by using red as positive and blue as negative#3 Line 430 - 440 & Figure 9. Is the correlation between annual yields and temperature, precipitation, and elevation obtained without dropping dominant mowing events? If so consider eliminating mowing events from the model and test the sensitivityCitation: https://doi.org/10.5194/egusphere-2024-4087-RC2
Viewed
HTML | XML | Total | BibTeX | EndNote | |
---|---|---|---|---|---|
131 | 31 | 5 | 167 | 3 | 3 |
- HTML: 131
- PDF: 31
- XML: 5
- Total: 167
- BibTeX: 3
- EndNote: 3
Viewed (geographical distribution)
Country | # | Views | % |
---|---|---|---|
United States of America | 1 | 44 | 30 |
Germany | 2 | 30 | 20 |
Switzerland | 3 | 16 | 11 |
China | 4 | 11 | 7 |
Uruguay | 5 | 5 | 3 |
Total: | 0 |
HTML: | 0 |
PDF: | 0 |
XML: | 0 |
- 1
- 44