the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 30 Jul 2025
Drivers of long-term grassland CO2 fluxes and regrowth: effects of management and meteorological conditions over time
Abstract. Grasslands serve a unique role in the global carbon (C) cycle and cover about 30 % of the European and about 70 % of the Swiss agricultural area. Carbon dioxide (CO2) fluxes of managed grasslands are substantially influenced by land management practices and meteorological conditions, but the temporal development of these drivers is still uncertain. With the eddy covariance (EC) technique, net ecosystem CO2 exchange (NEE) can be directly measured, and then partitioned into gross primary production (GPP; amount of CO2 fixed through photosynthesis) and ecosystem respiration (Reco; amount of CO2 released via plant and soil respiration). We used 20 years (2005–2024) of EC fluxes, meteorological data, and detailed management information collected from an intensively managed grassland site (Chamau) in Switzerland, and employed machine learning approaches, i.e., eXtreme Gradient Boosting (XGBoost) models in combination with SHapley Additive exPlanations (SHAP) analyses, to identify drivers and their temporal contributions over two decades. Our study aimed to (1) investigate intra- and inter-annual variations in grassland CO2 fluxes, (2) assess magnitude and drivers of GPP and Reco during the regrowth periods (i.e., after mowing, grazing, or reseeding), and (3) quantify driver contributions to GPP and Reco over time, with focus on management and extreme events. CO2 fluxes showed pronounced intra- and inter-annual variations, driven by both management activities as well as meteorological conditions. Despite significant increases in temperature and decreases in soil water content (SWC) during the two decades, GPP and Reco rates during regrowth periods remained stable, and no significant trend over time was detected, suggesting adapted, climate-smart decision making of the farmer. The most important drivers of GPP in the long-term were light, management, and temperature, while Reco was mainly driven by temperature, GPP, and management. However, during extreme drought periods in the peak growing season (June, July, August), SWC increased in importance and limited GPP. In contrast, the impact of nitrogen (N) fertilization was more differentiated, either acting in parallel with SWC, suggesting low N availability during drought periods, or increasing GPP in years after sward renewal despite low SWC. Overall, our study provided novel insights into relevant drivers of grassland CO2 fluxes and their complex temporal contributions in the short- and long-term. Our results suggest that even small climate-smart management adaptations could be promising solutions for stabilizing important grassland processes, such as grassland regrowth, under current and future climate.
- Preprint
(2700 KB) - Metadata XML
- BibTeX
- EndNote
Status: open (until 12 Sep 2025)
-
RC1: 'Comment on egusphere-2025-3562', Georg Wohlfahrt, 13 Aug 2025
reply
General comments:
This is a well-written paper that reports on 2 decades of CO2 exchange measurements at an intensively managed grassland in Switzerland with the aim of disentangling the influence of management amidst variable environmental conditions and ongoing climate change. The paper builds up on Feigenwinter et al. (2023) who analyzed the first 16 years, here the focus is more on the analysis of the drivers using a machine learning approach. I think the manuscript is ready for publication following some minor changes, detailed below.
There are however two terminology issues that I struggle with and ask the authors to consider:
- The authors analyze what they refer to as regrowth periods in between management events, especially harvesting. What I struggle with is the terminology “GPP/RECO regrowth rates” which the authors use to refer to the GPP/RECO during the regrowth periods. The terminology to me however suggests GPP/RECO “to regrow”, i.e. rebound, during these periods, which may not be the case. In fact, the negative SHAP values for days since last management and GPP suggest a negative relationship. I think the authors could simply say something like GPP/RECO during regrowth periods, which may be a little awkward at times, but less ambiguous.
- The authors suggest, e.g. in the abstract but also elsewhere, that the fact that there was no trend in CO2 exchange over the two decades despite ongoing climate change shows that the farmers are using a climate-smart management. This statement to me implies that the management is deemed climate-smart as it prevented a decrease in the CO2 sink strength. This ignores the possibility that an alternative (truly climate-smart?) management could have profited from ongoing climate change and increased the sink strength. Neither option (a decrease or increase in sink strength was prevented by the actual management) can be answered with the present data that are conditional on the actual management. This would need a manipulative experiment (with alternative management like in Ammann et al. 2007) or the use of some model which represents management and the resulting consequences on CO2 fluxes (which would be an intriguing follow-up). I thus suggest to down-tune the climate-smart aspect and rather leave it with saying that the adaptive management that the farmer practiced in response to interannual and intra-seasonal variability in weather conditions apparently was able to keep CO2 exchange stable in the face of ongoing climate change during the two decades of observations.
Detailed comments:
- l. 8: the temporal development of management practices and meteorological conditions is uncertain? Aren’t the interactive effects of these on grassland CO2 fluxes uncertain?
- l. 9-10: this sentence could be removed in the abstract without loss of information
- l. 42: GPP and RECO are the essential part of C cycling of any ecosystem
- l. 71: in my view Wohlfahrt et al. (2008, 10.1029/2007JD009286) were one of the first grassland papers to look into the interactive effects of management and environmental drivers and in fact also analyzed data in periods stratified by management (harvesting) events
- l. 130: what about the self-heating correction of the Li-7500 – I guess at least during the early phase of the time series the used models required this correction? In addition, the early Li-7500 models had some intrinsic lag of the digital signals that could be increased on the software side to result in a lag that is some multiple of the sampling rate in order to be removed – what lag value was set – 0.3 s?
- l. 135: which approach for flux partitioning was used – day or nighttime?
- Fig. 1e: given the length of the time series I feel a bit overwhelmed with the day-to-day variability and thus I suggest showing CO2 fluxes on a monthly timescale, possibly as a stacked bar chart that might nicely visualize the interplay between GPP and RECO on NEE
- Fig. 1: would it possible to add an additional panel that shows the cutting events, grazing periods and re-sowing events?
- Table 1: is huge but conveys limited information and might thus go into the supplement?
- Fig. 3: I suggest adding Fig. A2 as a third panel here; overall the information content of this figure is limited - GPP/RECO is smaller during the off-season period with short days and larger during the warm period with long days
- l. 270: DaysSinceUse shows negative SHAP values for GPP – correct? Does that mean that GPP declines the more time has passed since the start of the regrowth period? If so, might be worthing spelling this out
- l. 311: these significant differences are not visible from Fig. 6b-c
- l. 372-377: these data should be first introduced in the Results section
- l. 471: … in C cycle model simulations …
Citation: https://doi.org/10.5194/egusphere-2025-3562-RC1 -
RC2: 'Comment on egusphere-2025-3562', Andreas Ibrom, 02 Sep 2025
reply
Overall assessment (do not require specific comments by the authors)
The study presents and analyses a rare, high-quality long-term data set of field measurements in an intensively managed ecosystem, led by the colleagues who run the site. The data includes CO2 exchange and management data. The scientific focus is on identifying drivers for interannual variability of gross-primary productivity (GPP) and ecosystem respiration (Reco) in ‘normal’ and extreme years. The manuscript clarifies nicely that management and especially the aboveground and the complete renewals of the grassland vegetation mark significant short and long-term disturbances putting additional constraints on comparability of periods and years. Weather and management variability and their interactions, which I believe can be expected by the concept of adaptive climate smart management, characterize the challenge, when investigating this ecosystem, compared to, e.g., natural vegetation.
The work creates order in the time series through some useful classification. A period of 20 years is long enough to identify significant environmental trends and weather extremes. The period includes two grassland renewal events and numerous aboveground canopy harvests that define regrowth periods (RPs), still of different length and located in differing seasons. The simplicity and clarity of these decisions to define the perspectives on how to look at the time series and analyse it, is one of the strengths of this work. One sub-period of 5 years, i.e. still ¼ of the investigated period, when part of the study area was subject to a comparative experiment, adds complexity and the way this has been dealt with, raises some questions.
The complexity of the data set requires a complex analysis approach with careful selection of drivers, which I will discuss below. Going for a machine learning approach (XGB) for analysis might be a good choice, as it puts the least constraints on the results compared to alternatives, such as alternative empirical mechanistic modelling approaches. But the different nature of the results, especially the results from the SHAP analysis, are yet a bit difficult to understand and I suggest more explanation and guidance for the reader.
In general, I see this work as a model for such long-term empirical studies in managed ecosystems, for sure a strong scientific contribution to quantitative Biogeosciences.
Some general critical comments are as follows (please comment and take action, where applicable):
- G1: Devising extreme months with a clear and simple Z-score approach on soil physical and atmospheric drivers for drought makes good sense. The classifying variables are well chosen, because a capacitive variable (SWC) and an atmospheric state variable (VPD) are combined to represent the accumulated (SWC deficit) and actual stress (VPD). For the latter, potential evaporation might possibly be an even stronger variable.
- G2: Machine learning methods are relatively novel, and I believe the interpretation of the results is still a challenge. From my own experience with this text, there is a large risk for a reader, not yet familiar with the SHAP analysis, of miss-interpreting the results from the SHAP analysis, e.g., a “negative effect” as “negative relationship” between a driver (D) and a response variable (R). A negative SHAP value (“negative effect”) shows rather only that the contribution from D has made R lower than the reference. This is irrespective of the sign of a relationship: a positive relationship (dR/dD > 0) makes R small at low values of D a negative relationship (dR/dD < 0) makes R small at high values of D. To avoid misinterpretation by the readers, please consider explaining this possible trap for understanding the text correctly.
- G3: However, I claim, for scientific understanding, relationships are more relevant than just effects. By examining the effects further, e.g. by looking into the relationship between the effect and the magnitude of D, you might be able to say something more about the nature of the relationship (see, e.g., D29) and possible interactions between drivers.
- G4: Although the text introduces it correctly, I first falsely assumed that SHAP analysis 1 was based on RP, rather than on days, while SHAP 2 was on a daily basis. Just to confirm, is it correct that in both cases the SHAP analysis was performed on a daily basis but the daily results from SHAP 1 were then presented as RP averages in Fig. 4? For the next points, I presume that this is correctly understood.
- G5: I was surprised by the apparent lack of coherence between SHAP 1 and SHAP 2 analysis. Doesn’t this show how sensitive the results are to the choice of the baseline (and of course the variables)? This should be mentioned when interpreting the results from such analysis.
- G6: Possible a priori relationship between response and driver variables: Defining RPs that can be compared across seasons and years seems a very appropriate approach, however, the RPs are of different lengths (see e.g. Fig. 6, where one RP spans over three months in 2018, while others span only over one month in 2022 and 2023). I wonder whether common relationships between driver variables and the RP length cause some artificial interdependence (circular logic) among some driver variables and, and even more worrisome, among some drivers and the response variable GPP.
I mean especially the two drivers DaySinceUse and DailyN, which represent management that both depend RP length. From a farmers perspective, i.e. planning the RP length to reach a certain goal, the RP length is inversely related to productivity and thus response variable daily GPP level.
The particular relationship between DaySinceUse and RP length is that only in cases when RP length is high, DaySinceUse can reach high values. In these cases, high DaySinceUse values coincide with low daily GPP and high RP length.
The particular relationship between DailyN and RP length is a mathematical consequence of the definition of DailyN , i.e. the DailyN value per unit fertilized N will inversely decrease with increasing RP length.
Please comment on the possible effects on the results of the analyses from these interdependencies or consider amelioration by different driver definitions.
- G7: Critical reflection of using time as a driver: The variable DaySinceUse increases linearly with time until reaching RP length, i.e. it simply represents time as such or canopy age. The analysed relationship is thus the timeline for the development of GPP during an RP.
I wonder, what is the rationale behind using time as a driver? Short development time coincides with high productivity, but the drivers for production are not time, but rather the growth conditions. It will be obvious that the XGB-SHAP analysis will ‘turn’ time into a driver owing to the a priori decision of the user defining time as a driver.
Please explain the usefulness of time as a driver.
- G8: The above problem raises some fundamental questions about the general meaning of driver and response variables. The term adaptive management suggests that management, e.g. defining the time for harvest, i.e. the RP length, can both be a driver and response – a clear distinction may even be impossible.
What do these uncertainties mean for the interpretation of the analysis?
Is using the term “driver” together with a statistical analysis that is not able to detect cause-effect relationships (just effects) at all appropriate?
- G9: General reflection of the usefulness of DailyN as a driver for daily GPP: If I am right, DailyN is the only driver variable that includes averaging over the RP length. For a SHAP analysis that is based on daily values, the definition of DailyN is counterintuitive and the naming does not reflect what is actually going on (the fertilization is not daily). The authors will agree that N-availability might be the more relevant factor for GPP. N-availability will be larger right after the fertilization event (after reaching the rooting zone) and will decrease maybe not with time but with growth (and leaching, emissions etc.) over RP. Do you see a possibility to define an alternative daily variable “available N” (AN) that parameterizes the decrease from the amount of fertilized N over the length of the RP or even scaling it negatively with GPP (as proxy for growth)?
- G10: In general, please explain the term adaptive climate smart decisions / management. This is important for two reasons, i) it is used in the interpretation and the conclusions and ii) the definition might help to better understand the nature of the RP, i.e. as depending on certain a priory rules and expectations/ observations on productivity.
- G11: The study concludes (L474-L476) adaptive climate smart management as a factor for homogeneous production despite weather trends (likely climate change induced). Is this ‘just’ a plausible speculation or did your study show this? Maybe I overlooked it, I did not find clear evidence in the presented results for this statement that comes up in the discussion, the conclusions and is highlighted in the abstract. A quantitative analysis would examine the interaction between adaptive climate smart management and production, probably in contrast to a plausible BAU scenario.
I wonder whether XGB generated predictions could be used for scenario calculations or whether mechanistic models would be needed to substantiate such speculation. I deem this worth to be clarified in the discussion. I do not suggest such study to be included here. The study is rich enough, but its limitations need careful consideration, i.e. what can be concluded from its results.
General recommendation: I deem the overall quality of the manuscript to be very high and inspiring and maybe its clarity is the reason why it provokes some critical thoughts. I do not claim that this review from reading the manuscript a couple of times, can be assumed to be exhaustive and accurate, as I lack particular knowledge that the Authors probably have. I expect though clarifying responses and look forward to the answers by the authors. It depends very much on these answers, whether minor or major revisions will be necessary.
Detailed comments (please comment and take action, where applicable)
- D1: The title includes the word regrowth, which implies biomass production while it is used here as re-establishment or recovery of GPP and Reco. I suggest using the more neutral “grassland CO2 exchange:” instead of “grassland CO2 fluxes and regrowth:”
- D2: L 16: consider starting a new paragraph before “CO2”.
- D3: L 20 and L 25: make a decision on whether the study showed or suggested a relationship between CO2 exchange and “adapted, climate smart decision making” (see also G11).
- D4: L53-54: Define “atmospheric dryness” – explain, why does it not include “reduced precipitation”.
- D5: L56-L57: While promote productivity makes sense “promote CO2 fluxes in general” does not– consider rewording.
- D6: L64: is the word ‘buffer’ appropriate here? ‘mitigate impact of … on …’?
- D7: L73 – L77: Please clarify, do you mean the nonlinear, interactive, and highly dynamic “nature of drivers” or rather the nonlinear, interactive, and highly dynamic “nature of responses”? Please consider the difference between “dynamic” and, e.g. “variable”? What would fit better here?
- D8: L84: (objective 1) Do you deem the investigation of something as a scientific objective?
- D9: L96 – what do you mean with “destroyed”? was just ploughed, or extracted and removed?
- D10: L103-L107: If I am right, this is an important decision on how to use and interpret ¼ of the time series. I wonder how this decision has influenced the results. I would like you to discuss the alternative(s), e.g. separating fluxes between parcels and using only the comparable one, parcel B, for this study.
- D11: L109-112: Please specify “This” in “This allows” – the logic between the two sentences is not clear (to me). Did you merge shorter periods into one RP? If this was the case, did you check, whether the results in these RP differed from the others?
- D12: L112: Be aware of the impact that averaging over differently long RP has on the meaning of the variable. Is there a negative relationship between the length of the RP and the average GPP or Reco value? Or has the definition of a minimum RP length of 10 days alleviated this relationship. From my distant perspective, if such relationship existed, it would explain relationships between effects from drivers that are (Day[s]SinceUse, dailyN) or are not (PAR, TA, TS, VPD) related to the length of the RP (see also G6).
- D13: L 109: Please add a short explanation on how was the position and length of the main “growing season” defined, and how this RP classification affected the analysis. Consider replacing ‘middle date’ by ‘center’
- D14: L125: consider “atmospheric” instead of “air”
- D15: L128: specify, probably, ‘volumetric’ SWC
- D16: L130: replace “community guidelines” with scientific references
- D17: L138: explain why using these percentiles instead of the usual ones 5 %, 50 % ,and 95 %?
- D18: L147-L149: Please explain this averaging choice considering the vertical distributions of roots and SOM.
- D19: L154 – L155: If only the GPP Reco averages have been tested, please add information, on whether the lengths of the RP showed a trend. Please specify: in the trend analysis did you exclude extreme months?
- D20: L165: Please clarify, what do you mean with “GPP regrowth rate” do you mean dGPP/dt or, in accordance what was explained above, “average GPP rates over RPs”? Then please explain the rationale for choice of a 2nd order polynomial as regression model for the analysis of light response function of GPP.
- D21: L168-L172: Specify when adding GPP as driver variable for Reco, why did you still include PPFD and VPD as drivers? Is it correct to say that that all other effects are then residual effects, i.e. effects of a variable on top of its effect on GPP?
- D22: L197 – The sentence does not make sense in the way that in both set-ups the same months (JJA) were selected without further distinction. Is in the sentence describing the second set after “for only the peak growing season” a reference to extreme years missing? Or did I misunderstand anything here?
- D23: L181 - L184: Clarify here that an effect is different from a relationship (see G2)
- D24: Section 3.1: Would the length of the vegetation period (VP) and the meteorological conditions in the VP - both more relevant for bioclimatological characterization- give a different picture? Consider moving Table 1 in the appendix and focusing only on significant trends here.
- D25: Figure 2: Nice and clear presentation and reasoning.
- D26: Section 3.2: can you provide information about interannual and seasonal variation of the length of RP (e.g. horizontal range bars or replace circle by rounded rectangles in Figure 3). Alternatively you might consider presenting the sums of GPP and Reco over RPs as alternative to the average in the same manner in a second figure. It might show, if I am right, the effects of adaptive management.
- D27: Section 3.3: Fig. 4: I find the choice of the transparent colors confusing because they do not match with the colors of the legend very well. Consider an alternative to show the different season categories if at all necessary. Mark in the figure when the grassland renewal has taken place and which were the months with extreme weather. Then explain that the x-axis variable is not time but number of RP from start of the investigated period.
- D28: Section 3.4: Make sure to mention that the difference between the canopy photosynthesis saturation level and the maximum of a polynomial GPP= f(PAR) have different meanings (see also D20).
- D29: L277-279: ‘negative effects of SWC on Reco’ is a very good example, how effects may sound counterintuitive. I think it would be good mentioning, recalling that SWC is low during droughts, low values of SWC have caused the low predicted Reco as indicated by the negative SHAP effect values (see also G2).
- D30: Section 3.5: the heading focuses on drivers, but, I believe, the main relevant results are the effects on GPP and Reco.
Citation: https://doi.org/10.5194/egusphere-2025-3562-RC2
Data sets
Dataset including 20 years of daily CO2 fluxes, meteorological data, and management information during regrowth periods from the intensively managed grassland site Chamau Yi Wang, Iris Feigenwinter, Lukas Hörtnagl, Anna K. Gilgen, Nina Buchmann https://www.research-collection.ethz.ch/handle/20.500.11850/745429
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 709 | 28 | 13 | 750 | 15 | 12 |
- HTML: 709
- PDF: 28
- XML: 13
- Total: 750
- BibTeX: 15
- EndNote: 12
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1