The study site Chamau is a permanent grassland located in the canton of Zug, Switzerland (47°12^′36.8^′′ N and 8°24^′37.6^′′ E, 393 m a.s.l.). In 2005, an EC and meteorological station (part of FLUXNET as CH-Cha) was established, and CO₂ and H₂O vapour fluxes have been continuously measured since then. The average annual temperature was 9.9 °C, and the average annual precipitation sum was 1147 mm over the study period (2005–2024). The soil of the grassland parcels surrounding the flux station is classified as a Cambisol/Gleysol and has a silt loam soil texture, with a pH of 5.3 in the topsoil (Roth, 2006). This grassland has been intensively managed, with four to six mowing events per year, occasional grazing by sheep, and organic fertilizer application (mainly as slurry, normally applied once after each mowing event, on average 260 kg N ha⁻¹ yr⁻¹ from 2005 to 2024). In February 2012 and August 2021, the sward was renewed, i.e., the existing sward was terminated and ploughed, and a new sward was sown. The vegetation in the grassland includes English ryegrass (Lolium perenne), common meadowgrass (Poa pratensis), Italian ryegrass (Lolium multiflorum), white clover (Trifolium repens), red clover (Trifolium pratense), and dandelion (Taraxacum officinale) (Feigenwinter et al., 2023b).

The Chamau grassland site consists of two parcels A and B (Feigenwinter et al., 2023b), and detailed management information from this site has been continuously documented since 2005, including major activities like mowing, grazing, fertilization, and sward renewal. A nitrous oxide (N₂O) mitigation experiment was carried out from 2015 to 2020, where parcel A received standard fertilization, while parcel B was oversown with a higher legume proportion and received no fertilization (Feigenwinter et al., 2023a; Fuchs et al., 2018). Since parcel B dominated the footprint area in earlier years (2005–2013; Fig. B3 in Feigenwinter et al., 2023b), and the two parcels were under similar mowing and grazing scheme during the entire study period (Table B2 in Feigenwinter et al., 2023b; Tables A1, A2), we adopted the management scheme of parcel B to define regrowth periods in the current study.

We defined the period between two mowing/grazing/sward renewal events as a regrowth period (excluding the day(s) when management occurred). For grazing events, the start of any given grazing was defined as the end of the previous regrowth period, and the end of this grazing was considered the start of the next regrowth period. Each regrowth period consists of a minimum of 10 d. If several management activities took place only a few days apart within the entire footprint, we defined the latest activity as the start of the next regrowth period. For example, in earlier years, the parcels were occasionally first partially grazed and then partially mowed within a certain time window. We thus considered the start date of the grazing as the end of the previous regrowth period, and the mowing date as the beginning of the next regrowth period. A “DaySinceUse” variable was generated for each day in a regrowth period to indicate the number of days after the last mowing/grazing/sward renewal event. An average daily fertilizer nitrogen (N) input for each regrowth period (DailyN, in kg N ha⁻¹ d⁻¹) was calculated based on the total fertilizer N received during the period and the length of that period. If parcel A and B had different N inputs for certain regrowth periods, the average of the two parcels was used as the total N input. GPP and Reco during each regrowth period (in g C m⁻² d⁻¹) were calculated based on the cumulative CO₂ uptake or release over the regrowth period (in g C m⁻²) divided by the length of the regrowth period (i.e., number of days). If the center of regrowth periods was between April and September, i.e., the main growing season for agricultural grasslands in central Europe, they were classified as spring–summer regrowth periods, else as autumn–winter regrowth periods.

Since July 2005, molar densities of CO₂ and H₂O vapour were measured with open path infrared gas analysers (IRGAs; LI-7500; LI-COR Biosciences, USA) at 2.42 m measurement height and tilted 15° from the vertical, together with measurements of 3D wind speed and wind direction using ultrasonic anemometers (R3–50, Gill Instruments Ltd., UK) at 20 Hz time resolution. Meteorological variables, including air temperature (TA, in °C), relative humidity (RH, in %), atmospheric pressure (PA, in hPa), incoming and outgoing shortwave and longwave radiation (SW IN/OUT and LW IN/OUT, in W m⁻²), incoming and outgoing photosynthetic photon flux density (PPFD IN/OUT, in µmolm-2s-1), precipitation (PREC, in mm), soil temperature (TS, in °C), and volumetric soil water content (SWC, %), were also measured. For more details on instrumentation, see Feigenwinter et al. (2023b).

Half-hourly CO₂ and H₂O vapour fluxes were calculated with EddyPro (v7.0.9) based on the covariance between the turbulent fluctuations of gas concentration and vertical wind speed. Widely used community guidelines (Aubinet et al., 2012; Pastorello et al., 2020, Sabbatini et al., 2018), including corrections for density fluctuations (Webb et al., 1980), double axis rotation (Wilczak et al., 2001), high-pass (Moncrieff et al., 2005) and low-pass filtering effects (Horst, 1997), were applied. The default self-heating correction for the open path LI-7500 (Burba et al., 2008) was not applied because it was developed specifically for vertically mounted IRGAs, and applying it to a 15°-tilted setup without a parallel reference flux for validation would risk introducing systematic over-corrections, larger than the random uncertainty of the measurement (Deventer et al., 2021). Time lags were detected using covariance maximization in a short time window (between 0.05 and 0.55 s), with a default time lag (typically 0.30 s) if no clear time lag was found. Post-processing for quality assessment and gap-filling was done in the Python library diive (Hörtnagl, 2025a), NEE flux partitioning into GPP and Reco based on the night-time method (Reichstein et al., 2005) was done with the R package REddyProc (v1.3.3, Wutzler et al., 2018). During quality assessment, a series of quality control flags were applied to classify data quality (after testing steady-state, data completeness, spectral correction factor, IRGA signal strength, outlier detection), and turbulence-based constant friction velocity (USTAR) thresholds were applied (0.05, 0.07, and 0.09 m s⁻¹ for the 16th, 50th, and 84th percentile of the USTAR threshold distribution, respectively). An overall Quality Control Flag (QCF) was generated, integrating all assessments to filter for high quality data. Resulting data gaps were then filled using random forest models utilizing a sliding three-year window for annual model training (Breiman, 2001; Hörtnagl, 2025a). For all further analyses, we used quality-controlled fluxes that were filtered with the 50th percentile of the USTAR threshold distribution, subsequently gap-filled using random forest models, and partitioned into GPP and Reco.

For meteorological data, a similar quality assessment procedure was applied with the diive library (Hörtnagl, 2025a), and all screened variables were aggregated into half-hourly resolution. Vapour pressure deficit (VPD) was calculated based on TA and RH. For TS (at 0.04, 0.15, and 0.4 m depths) and SWC (at 0.05, 0.15, and 0.75 m depths) measurements, averages across all depths were calculated and used in the final analysis to represent the overall soil conditions over the entire soil profile. All data used in this study were derived from the most recent PI dataset (Hörtnagl et al., 2025), with all processing steps openly documented in a GitHub repository (Hörtnagl, 2025b).

During the past 20 years (2005–2024), the Chamau grassland experienced a mild temperate climate (MAT 9.9 °C, MAP 1147 mm). At an annual scale, TA at this site increased significantly by 0.071 °C yr⁻¹ (Mann Kendall test p<0.01), mean TS and minimum TS also increased by 0.070 and 0.095 °C yr⁻¹, respectively (p<0.01). The years 2018 and 2022 were the hottest years on record, with mean annual TA of 11.0 and 11.1 °C, respectively, while the years 2010 and 2013 were the coldest years, with mean annual TA of 8.6 and 9.0 °C, respectively. Annual PREC did not show a significant trend (p=0.87) over the 20 years. Nevertheless, the years 2018 and 2022 were the driest years on record, with annual PREC of 906 and 884 mm, respectively, while the years 2016 and 2024 were the wettest years, with annual PREC of 1351 and 1378 mm, respectively.

At the monthly scale, meteorological conditions showed pronounced seasonal variability (Fig. 1a–d). Overall, monthly TA reached maximum values in July (mean of 19.6 °C) and minimum values in January (mean of 1.0 °C). Monthly TA (Fig. 1a), PREC (Fig. 1b), and VPD (Fig. 1d) did not show significant trends over time (p=0.21 for TA, p=0.69 for PREC, p=0.51 for VPD), while SWC (Fig. 1d) decreased significantly over the 20 years (slope =-0.016 % per month, p<0.01). Looking at each specific month (i.e., across 20 years), TA and TS increased significantly in the winter months (December to March; Table A4). August was the only month to experience a decrease in PREC, which otherwise remained stable. SWC significantly decreased in late summer and autumn (August to October; Table A4). Based on the z-scores, four months of extreme air dryness, seven months of extreme soil dryness, and six months of compound extremes (i.e., high VPD and low SWC) were identified (Fig. 2). These extreme conditions mainly occurred between May and September, particularly during the peak growing season months (June, July, and August) in the years 2018, 2019, 2022, and 2023.

Daily NEE, GPP, and Reco showed strong inter-annual and seasonal variations over the 20 years (Fig. 1e, Table A5). Annual cumulative NEE ranged from 139 g C m⁻² yr⁻¹ (in 2012) to -503 g C m⁻² yr⁻¹ (in 2014), with an overall mean (± standard deviation) of -260 (±172) g C m⁻² yr⁻¹ (Fig. A1). In 2012 and 2021, two grassland renewal events (i.e., ploughing and reseeding of the grassland) took place, with very different effects on annual NEE: while the event in 2012 led to a strong net CO₂ loss (cumulative NEE of 139 g C m⁻² yr⁻¹; Fig. A1), the event in 2021 resulted only in a weaker-than-normal net CO₂ uptake (cumulative NEE of -163 g C m⁻² yr⁻¹; Fig. A1). All three flux components were strongly influenced by management practices (Fig. 3a, Tables A1, A2) such as mowing and grazing. Strongest net CO₂ uptake occurred in early spring (March–April), while the largest net CO₂ release happened in early winter (December–January).

The Chamau grassland was managed intensively during the past two decades, and experienced in total 94 mowing events, 319 d of grazing, and 102 fertilizer applications. Moreover, the grassland was renewed twice, in 2012 and 2021 (Fig. 3a). This resulted in 115 grassland regrowth periods during the 20 years of this study. The average (± standard deviation) length of the regrowth period was 58 (±47) d, with shorter regrowth periods in spring–summer seasons (39±20 d) and longer regrowth periods in autumn–winter seasons (100±60 d). During spring–summer seasons (i.e., April to September), about four mowing/grazing events took place (about two to three events in extreme dry years of 2018 and 2019). A slightly increasing trend (p=0.01) was found for the length of regrowth periods over the 20 years of the study: While the length of regrowth periods during the spring–summer seasons increased significantly (p<0.01), no significant trend was found during the autumn–winter seasons (p=0.87). Generally, longer regrowth periods corresponded to lower GPP and Reco. Overall, mean (± standard deviation) GPP during regrowth periods was 7.2 (±2.6) g C m⁻² d⁻¹ and Reco during regrowth periods was 6.6 (±2.5) g C m⁻² d⁻¹ (Fig. 3b–c). The average GPP and Reco during regrowth periods in spring–summer seasons were 8.6 (±1.4) and 7.9 (±1.6) g C m⁻² d⁻¹, respectively. In comparison, average GPP and Reco in autumn–winter seasons were 4.1 (±1.4) and 3.7 (±1.5) g C m⁻² d⁻¹, respectively. We did not observe any significant trend for GPP during regrowth periods over the 20 years (overall p=0.13), regardless of seasonality (spring–summer season: p=0.10; autumn–winter season: p=0.33). Similarly, Reco during regrowth periods did not change significantly over time (overall p=0.07), not during the spring–summer (p=0.08) nor the autumn–winter seasons (p=0.84).

The XGBoost models for daily GPP and Reco performed well (Fig. A2), with a RMSE for GPP of 1.2 g C m⁻² d⁻¹ and an R2 of 0.91 on the testing set (Fig. A2a). The performance of the Reco models was better, with a RMSE of 0.6 g C m⁻² d⁻¹ and R2 of 0.96 on the testing set (Fig. A2c). The three most important drivers for daily GPP were PPFD, TS, and DaySinceUse (Fig. A2b), while the three most important drivers for daily Reco were TS, GPP, and DailyN (Fig. A2d).

For both GPP and Reco, the drivers showed strong seasonality over the 20 years of the study (Fig. 4; SHAP analysis 1). PPFD and temperature normally increased GPP in summer compared to the overall mean, while low light and low temperature conditions in winter decreased GPP compared to the mean (Fig. 4a). At daily scale, higher PPFD (Fig. A3a) and higher TS (Fig. A3b) had more positive effects on GPP than lower PPFD and lower TS (i.e., SHAP values of both drivers generally increased (and then saturated for TS) with increasing magnitude of the driver values). When averaged over the entire regrowth period, management events (i.e., mowing and grazing, represented by DaySinceUse) typically had a negative effect (i.e., negative SHAP values) on GPP compared to the overall mean GPP (Fig. 4a). However, at daily scale (Fig. A3c), SHAP values for DaySinceUse were first negative, but then steadily increased towards positive before becoming stable after around 20 d, indicating that GPP increased the more time had passed since the last management event. Compared to the main drivers of GPP (light, temperature, and management), water-related drivers (i.e., SWC, VPD, and PREC) had relatively smaller contributions and showed a less strong seasonality. However, during the extreme summers (i.e., 2018, 2019, 2022, and 2023), SWC increasingly limited GPP (negative SHAP values compared to normal years). The main drivers of Reco, i.e., TS and GPP, showed a similar seasonality, with higher TS and higher GPP in summer generally promoting respiration, while in winter, these drivers limited Reco (Fig. 4b). At daily scale, the effects of these two drivers on Reco became more positive with higher TS (Fig. A3d) and higher GPP (Fig. A3f). Compared to the drivers of GPP, fertilizer N inputs were more important for Reco than DaySinceUse, while water-related drivers had relatively low importance and smaller effects on Reco. During the extreme summers when low SWC occurred, negative effects of SWC on Reco were more obvious in 2022 and 2023 compared to earlier years (e.g., 2018 and 2019).

Focusing on the years with exceptional management during the 20 years of this study, i.e., the renewal years 2012 and 2021, we tested the functional relationships between PPFD and GPP for the two years before and after the renewal years vs. all other normal years (Fig. 5). During all periods, GPP increased with mean daily PPFD, while maximum GPP (GPPmax, i.e., the maximum of the curves in Fig. 5) was reached at a PPFD of about 30 mol m⁻² d⁻¹ (for years before the renewals) or at about 45 mol m⁻² d⁻¹ (for all other normal years and years after the renewals). Moreover, spring–summer seasons in all years showed rather similar average GPP during regrowth periods: 8.2 g C m⁻² d⁻¹ before the renewal years, 8.6 g C m⁻² d⁻¹ after the renewal years, and 8.9 g C m⁻² d⁻¹ for normal years. However, GPPmax was lower before the renewal years (8.5 g C m⁻² d⁻¹) compared to normal years (9.7 g C m⁻² d⁻¹) and years after the renewals (9.3 g C m⁻² d⁻¹), despite the fact that the two years (2022 and 2023) after the second renewal (2021) experienced more extreme weather conditions during summer than the normal years (Fig. 2).

Focusing on years with extreme summer months (i.e., July and August of 2018, July of 2019, July of 2022, and June of 2023; Fig. 2), we found normal daily PPFD (Fig. 6a), but significantly higher-than-normal daily TA, TS and VPD, accompanied by lower-than-normal SWC (Fig. 6b, c). Mean monthly TA in July 2018, 2019, 2022 was 1.2, 1.2, and 1.1 °C above the 20-year monthly average (19.6 °C), respectively, and 1.8 °C higher in August 2018 compared to 20-year monthly average (18.6 °C). All extreme summer months had SWC more than 10 % lower than the 20-year monthly averages. Daily GPP during the peak growing seasons (June, July, August) in 2018 and 2019 (i.e., years before the second renewal) was smaller than the mean predicted GPP of non-extreme months during the peak growing seasons of normal years (9.7 g C m⁻² d⁻¹), while GPP in 2022 and 2023 (i.e., years after the second renewal) was more dynamic, ranging from 1.3 to 15.5 g C m⁻² d⁻¹ (Fig. 6d). Mowing and fertilization events were also more frequent in the recent years (2022, 2023) compared to earlier years (2018, 2019), accompanied by higher amounts of N input. In contrast to the findings over all 20 years (Fig. 4), management-related drivers, i.e., DailyN and DaySinceUse, gained higher importance during the extreme summer months, while PPFD lost importance (Fig. 6d). In addition, SWC surpassed TS in importance, and its limiting effect on GPP was clearly seen in all four extreme summers, stronger in 2018 and 2019 than in 2022 and 2023. However, effects of fertilizer N inputs on GPP differed during the four summers: In 2018 and 2019 (i.e., years before the renewal in 2021), N fertilization effects acted similarly to SWC (both drivers having negative SHAP values), suggesting low N availability during low SWC periods, while in 2022 and 2023 (i.e., years after the renewal in 2021), fertilizer N inputs increased GPP (positive SHAP values) despite low SWC values. The relationship between SWC and its SHAP values for GPP during those four extreme summers showed a clear trend with limitation of GPP at low SWC (negative SHAP values; Fig. A4). This limitation of GPP by SWC was stronger in the years 2018 and 2019 (before the renewal; lowest SHAP values) compared to the years 2022 and 2023 (after the renewal in 2021).

Overall, CO₂ fluxes (NEE, GPP, and Reco) at the Chamau grassland site showed significant intra- and inter-annual variations (Figs. 1e, A1), with sometimes higher inter-annual variation (-260±172, 2193±263, 1934±244 g C m⁻² yr⁻¹ for NEE, GPP and Reco, respectively) than reported earlier for managed grasslands (e.g., Ammann et al., 2020; Heimsch et al., 2021, 2024; Peichl et al., 2011; Rogger et al., 2022). Reasons for these differences are manifold, considering that our study spanned two decades of measurements (2005 to 2024). Besides short-term changes introduced from meteorological conditions and frequent management events, we also captured the impacts of multiple extreme periods, particularly dry and hot summer months in 2018, 2019, 2022, and 2023 (Fig. 2), a six-year N₂O mitigation experiment with reduced organic fertilization (2015 to 2020; Feigenwinter et al., 2023a; Fuchs et al., 2018), as well as infrequent management practices such as two complete sward renewals (in 2012 and 2021; Feigenwinter et al., 2026; Merbold et al., 2014), all contributing to the highly dynamic nature of ecosystem CO₂ fluxes observed at our site.

With climate change, GPP of terrestrial vegetation has been observed and predicted to increase (Keenan et al., 2014; Knauer et al., 2023; Myneni et al., 1997; Piao et al., 2007) due to CO₂ fertilization, increasing temperature, and longer growing seasons favouring photosynthesis, unless other resources become limiting, such as water or nutrients (He et al., 2017; Keenan et al., 2023; Madani et al., 2020; Terrer et al., 2019; Xu et al., 2019). Often, studies focused on natural unmanaged ecosystems or forests at regional and global scales (Anav et al., 2015; Cai and Prentice, 2020; Davi et al., 2006; Norby et al., 2010), but studies on grasslands also observed increasing trends in GPP and Reco with higher temperature in the past (Shi et al., 2022; Wang et al., 2019). These studies were often based on experiments in less intensively managed or even unmanaged grasslands, consequently with less intervention potential for farmers or land managers. In contrast, despite significant increases in mean annual TA of around 1.4 °C over the past 20 years and decreases in SWC during late summer and autumn months (Table A4), we did not observe significant trends in GPP nor Reco of our intensively managed grassland (Fig. 3b–c). This clearly indicates that despite observed climate change, the farmer managed to maintain GPP during regrowth periods and forage production over the past two decades, also seen in the harvest exports reported by Feigenwinter et al. (2023b) for the same site (335±42 g C m⁻² yr⁻¹ between 2005 and 2020). The local farmer adapted management practices, e.g., skipped individual harvests in extreme years (e.g., 2018), thereby extending the regrowth periods (Fig. 3a). These detailed adjustments are typically difficult to capture by standard large-scale models and coarse-resolution management datasets, demonstrating that management information at high spatio-temporal resolution is crucial to reliably model GPP of agricultural systems. However, management data at high temporal resolution is difficult to obtain, therefore large-scale modelling studies often use harmonized datasets at annual time scales (Winkler et al., 2021), potentially introducing substantial bias in modelled GPP.

In addition, infrequent management events such as sward renewals, carried out every 5–15 years depending on vegetation composition and soil texture (Schils et al., 2007; SUPER-G, 2018), are typically not considered in C cycle models, partially due to lack of information about their occurrence but also simply due to lack of data for such events. The observed differences in annual NEE during the two renewal events at the Chamau grassland (139 vs. -163 g C m⁻² yr⁻¹ for 2012 and 2021, respectively; Fig. A1) might be explained by differences in seasons (February 2012 versus August 2021) and soil disturbance intensities (ploughed at 20 cm in 2012 and 3–4 cm in 2021), which subsequently influenced the establishment and regrowth of the new sward. Other studies in managed grasslands have also reported different magnitudes of CO₂ fluxes after renewal, influenced by timing, intensity, and environmental conditions of the renewal events (Ammann et al., 2020; Drewer et al., 2017; Li et al., 2023b). Such observations, albeit rare, are impactful in terms of C dynamics and underscore the necessity to include such infrequent destructive management practices in long-term flux studies as well as modelling frameworks.

NEE represents a combination of two large opposing fluxes, i.e., GPP and Reco, based on two physiological processes, i.e., photosynthesis and respiration, which respond differently to changing environmental conditions (Feigenwinter et al., 2023b; Peichl et al., 2011; Schwalm et al., 2010; Wang et al., 2019). Thus, using two separate machine learning models for GPP and Reco (XGBoost models with high performance, R2>0.9 on testing datasets), we identified their main drivers as well as the temporal development of daily driver contributions (SHAP analysis 1). Confounding effects from overlapping predictor variables were reduced by using a broader set of variables for the driver analysis compared to the gap-filling model. Daily aggregated data for the driver analysis (instead of half hourly as for gap-filling) furthermore reduced potential correlation among predictor variables. In contrast to other studies, we not only included high resolution meteorological data but also information about management events in our analyses at daily time scales. Consistent with previous studies (Gilmanov et al., 2007; Wohlfahrt et al., 2008), PPFD and temperature were the main drivers of GPP and Reco, respectively (Figs. 4, A2). However, management-related variables such as the number of days since the last management event and fertilizer N inputs were always among the top four to five important drivers for daily GPP and Reco, respectively. This highlights their importance for a comprehensive understanding of C dynamics and greenhouse gas fluxes in managed grasslands (Hörtnagl et al., 2018). By using time since the last management event (i.e., DaySinceUse) as a driver, we incorporated the management information into the models in a relatively straightforward way. However, the simplification of our approach should also be recognised: Rather than acting as a direct ecophysiological driver, DaySinceUse serves as a temporal index that tracks the time elapsed since the most recent management intervention (e.g., mowing or grazing). By establishing this temporal relation for each data record, the model can account for the recovery dynamics and residual effects of management that are otherwise omitted from standard meteorological drivers. The machine learning analyses also picked up the smaller effects of fertilizer N inputs as a driver for GPP during the N₂O mitigation experiment (2015 to 2020), when organic fertilization – and thus N inputs – was replaced with higher fractions of legumes for half of the area (Feigenwinter et al., 2023a; Fuchs et al., 2018). Since the majority of temperate grasslands in Europe are actively managed (Eurostat, 2023), long-term and detailed management information, combined with more field observational measurements such as EC fluxes, can support improved management strategies to enhance grassland ecosystem services under changing climate (Wang et al., 2025). Although detailed management information has been used successfully in our XGBoost models, we also acknowledge the limitation from our simplified representation of N fertilization on the daily scale. We represented N availability after fertilization by distributing the total N applied evenly across the regrowth period (see Sect. 2.2), although the actual temporal dynamics of N availability most likely follow a more complex pattern, involving different processes and drivers such as soil properties, precipitation, soil moisture and temperature but also microbial activity. However, such data are very rare, and model assumptions can thus often not be validated due to insufficient evidence in the existing literature. While we employed a simplified approach, we clearly recognize the need for remotely sensed or in-situ quantified daily N availability to further improve modelling of CO₂ fluxes in managed grasslands.

Focusing on GPP and its most important driver PPFD, we further examined the physiological consequences of the two renewals in 2012 and 2021. Although the sward had shown a lower performance before the renewals (i.e., lower GPPmax, orange line, Fig. 5), compared to the following years with higher GPPmax (black and blue lines), the slopes of the light-response curves were very similar. Differences in GPPmax resulted in high net CO₂ uptake rates after both renewals (average cumulative NEE of -465 g C m⁻² yr⁻¹ for 2013, 2014, 2022, and 2023) and also higher yields than before the renewals (Tables A1, A2). Thus, the farmer's decision to renew the grassland with a multi-species mixture was substantiated. Furthermore, frequent oversowing (Tables B1 and B2 in Feigenwinter et al., 2023b) might further help to maintain the preferred species composition and a stable yield. However, even though oversowing more diverse mixtures (Schaub et al., 2020) or increasing legume proportions (Fuchs et al., 2018) can improve yield quality and are less destructive management interventions than sward renewals, they can be costly (Schaub et al., 2021), and positive results are not always guaranteed. Therefore, adopting a combination of these observed management practices is a crucial first step towards more climate-smart farming practices, helping to ensure stable production and to sustain ecosystem services over the long term.

Besides consistent positive effects of light and temperature on GPP (Fig. 4a; red and orange bars), fertilizer N inputs (Fig. 4a; moss green bars) also promoted GPP during spring–summer seasons. Moreover, our analysis demonstrated the complex interactions between N inputs and soil water availability: the positive effects of N on GPP diminished in drier years (e.g., 2018, 2019) or even became negative, indicating lower soil N availability and thus lower N uptake for GPP when soil moisture was low (Bloor and Bardgett, 2012; Hartmann et al., 2013). For Reco, negative effects of N inputs were more obvious during the spring–summer seasons 2015 to 2020 (Fig. 4b), i.e., when the N₂O mitigation experiment was ongoing, and the entire site received around half of fertilizer N input compared to other years (Feigenwinter et al., 2023a; Fuchs et al., 2018). Without N inputs via organic fertilizer (mostly slurry), the organic C inputs and the C sink strength were reduced as well (Feigenwinter et al., 2023b). This resulted in less C available for microbial soil respiration, which was also reflected in lower Reco during regrowth periods in spring–summer seasons during the N₂O mitigation experiment (7.4±1.3 g C m⁻² d⁻¹) compared to Reco in years with normal management (8.2±1.7 g C m⁻² d⁻¹).

Overall, being purely data-driven and highly site-dependent, the combination of machine learning models and SHAP analysis confirmed the highly dynamic nature of the drivers influencing GPP and Reco and also revealed the complexity of their individual contributions across temporal scales. Compared to traditional linear models, these new approaches not only captured the non-linear and combined effects among drivers at high temporal resolution (here daily), but also provided explanatory outputs such as driver importance on a daily basis, which enhanced interpretability of the machine learning models and further improved our understanding of these complex grassland systems.

Using non-extreme summer months as the background dataset, SHAP analysis 2 focused on drivers of GPP only during extreme summer months, which minimized the confounding effects of strong seasonality in temperature and light conditions. This approach further improved the interpretability of our results, by isolating and focusing on the influence of extreme events-related factors such as SWC and VPD. Across all extreme periods, SWC strongly reduced GPP (SHAP analysis 2; Fig. 6d). In contrast, the effect of VPD was small (Fig. 6d), even on days with exceptionally high VPD values (Fig. 6c). This finding supported our long-term driver analysis (SHAP analysis 1; Fig. 4): here, SWC had a larger effect on GPP than VPD (mean absolute SHAP value of 0.6 for SWC and 0.2 for VPD during the peak growing seasons in 2018, 2019, 2022, and 2023), underscoring that soil water availability was more relevant for GPP than atmospheric water demand. This observation aligns well with previous studies reporting more significant negative effects of soil droughts (i.e., low SWC) on grassland CO₂ fluxes compared to those of heatwaves (i.e., high VPD; Li et al., 2016, 2021). Our finding is also consistent with the behaviour seen during the 2003 summer drought (Teuling et al., 2010) and the 2018 summer drought (Gharun et al., 2020), during which grasslands sustained evapotranspiration (ET) and thus GPP, as long as water was available in the soil, much in contrast to forests which rapidly reduced canopy conductance and ET in response to high VPD.

The effects of fertilizer N inputs were more differentiated during these extreme periods compared to the overall 20-year time series. While N inputs showed mostly negative effects on GPP during the years 2018 and 2019 (i.e., during the N₂O mitigation experiment and before the second renewal in 2021), effects of N inputs were typically positive during the years 2022 and 2023 (Fig. 6d). This suggests that after the second sward renewal in 2021, a resumed organic fertilization regime with a higher amount of fertilizer N input could further support recovery and increase regrowth rates after extreme periods (Hofer et al., 2017). Additionally, in response to the persistent droughts in 2018 and 2019 before the second renewal, the farmer delayed or skipped some mowing events (which typically occurred monthly during the peak growing season, Figs. 6d, 3a). In contrast, we observed even more frequent mowing events (three times between June and August) during 2022 and 2023 after the second renewal, which resulted in more regrowth periods and more variation in GPP compared to earlier years. While the first renewal in 2012 also resulted in increased productivity and forage yields (Feigenwinter et al., 2023b), our study clearly indicated that also after the second renewal in 2021 the new sward performed very well despite frequent and more intense extreme conditions in the two succeeding years, sustaining forage production also during such extreme summers.

Over the past two decades, Switzerland has experienced its hottest years on record (2022, 2023) and several extreme summers (e.g., 2018, 2022) (MeteoSwiss, 2023, 2024; Scherrer et al., 2022). In our study, we also observed consistent trends of rising temperature and frequent extreme summer months in recent years. The CO₂ fluxes of this intensively managed grassland responded very dynamically to changes in meteorology and to the already climate-adjusted management practices implemented by the farmer. However, it is unclear whether such small adjustments will suffice when extreme events intensify in the future (CH2025, 2025; IPCC, 2023). Some evidence suggests that less intensive management may buffer grasslands against extreme impact (Winck et al., 2025), which would require a different sward composition, lower livestock density, and further adjusted management practices. Adopting a less intensive management scheme for a single grassland site in a short time period is challenging. Therefore, site-specific adaptation of management strategies for both short- and long-term is needed to mitigate current and future climate risks (Gilgen and Buchmann, 2009; Woo et al., 2022). Further management options, such as multi-species mixtures (Craven et al., 2016; Finn et al., 2018; Isbell et al., 2015; Wang et al., 2025) and spatially adjusted precision-farming (Finger et al., 2019), can offer additional highly site-specific solutions to maintain selected ecosystem services and support the resilience of grasslands against extreme events.