The study site is a degraded raised bog within the Amtsvenn-Hündfelder Moor, formerly used for peat extraction. The study area comprises approximately 600 ha along the German–Dutch border between Gronau and Ahaus in North Rhine-Westphalia, Germany (52.1756° N, 6.95486° E; Fig. 1). The EC tower at the site (DE-Amv) is an ICOS (Integrated Carbon Observation System) associate station implementing high quality standardized greenhouse gas flux observations (Gharun and Behrens, 2025). The long-term (2006–2025) mean annual temperature and precipitation are 10.8 °C and 791 mm, respectively, based on data from the Ahaus weather station operated by the German Meteorological Service (Deutscher Wetterdienst, DWD; station ID 7374).

In this degraded raised bog, peat cutting stopped in 1979, but drainage ditches remained open. Peat thickness (predominantly Sphagnum peat) at the site is roughly 4.3 m, with the upper 10–25 cm of peat being affected from degradation due to drainage, as reflected in lower C to N ratios (30–40), elevated N (1.5 %–2.0 %) and P (500–800 mgkg-1) content and increased degree of peat decomposition compared to the underlying pristine peat (Lemmens et al., 2026). In the investigated area the peat body was only cut on the outermost margins. Water is continuously lost from the raised peat block to the adjacent lower parts of the bog where peat extraction took place and through an open drainage ditch to the south of the site. A vegetation survey on five sampling points within the footprint was conducted in July 2022. The vegetation within the footprint is relatively homogeneous across directions. It is dominated by an almost closed canopy of dwarf shrubs with a coverage of 50 % to 75 % common heather (Calluna vulgaris), 25 % to 50 % cross-leaved heather (Erica tetralix) and individuals of Vaccinium oxycoccos and bog cranberry (Andromeda polifolia). Irregularly interspersed are patches of graminoids, mostly purple moor-grass (Molinia caerulea), with a cover of up to 25 % and higher coverage at the outer margins of the investigated area. Some heterogeneity is introduced along overgrown drainage ditches where individuals of downy birch (Betula pubescens) and silver birch (Betula pendula) occur, with smaller individuals scattered in the footprint (Fig. 1). However, these ditches stretch across the footprint in several directions and provide no separation of the footprint into different sectors. Typical peatland (Sphagnum spp.) mosses are barely present anymore, only a few individuals can be found in drainage ditches and outside of the tower footprint in low lying ponds. To control the encroachment of the woody vegetation the site is managed with light sheep grazing.

Several ancillary climatic and edaphic variables were recorded at a frequency of one minute and aggregated to half-hourly means. TA and relative humidity (RH) were measured with a HMP155 temperature and humidity probe ((Vaisala Oyj, Vantaa, Finland). Precipitation (P) was measured with a non-heated rain gauge TR-525M (Texas Electronics Inc., Dallas, Texas, USA). Four radiation components–upward (SWOUT) and downward (SWIN) shortwave radiation, as well as upward (LWOUT) and downward (LWIN) longwave radiation–were measured using the CNR4 heated net radiometer (Kipp and Zonen B.V., Delft, the Netherlands). Soil variables were measured in two soil profiles at 5 cm depth. Soil water content (SWC) and soil temperature (TS) were measured with the Hydraprobe II (Stevens Water Monitoring Systems, Portland, Oregon, USA), secondary measurements of TS were conducted with LI-COR 7900–180 TS sensors (LI-COR Inc., Lincoln, Nebraska, USA). Soil heat flux (G) was measured with self-calibrating HFP01SC heat flux plates (Hukseflux Thermal Sensors B.V., Delft, The Netherlands).

All meteorological data was screened for faulty measurements. In a first step the resulting gaps were filled from a secondary meteorological tower 20 m from the EC tower, which provided measurements of TA, RH, P and SWC (with sensors in that order being: S-THB-M00x, S-RGB-M002, S-SMC-M005, all by Onset Hobo, Bourne, MA, USA). Remaining gaps in TA, RH and P were filled with data from the station Ahaus of the DWD (station ID: 7374), ca. 10 km from the station. Potential offsets between the data sources due to different sensors or measurement heights or depths were corrected using ordinary least squares regression (OLS). Global radiation was not available from any auxiliary tower, so final remaining gaps in SWIN were closed using machine learning, specifically with the XGBoost regression tree method (Chen and Guestrin, 2016), using the day-of-year, the hour of the day, TA, P and SWIN derived from the ERA5land (Muñoz-Sabater et al., 2021) dataset as predictors.

The water table was measured with a pressure transducer ca. 20 m from the tower (MX2001 connected to a RX3000 data logger, Onset Hobo, Bourne, MA, USA). A second WT logger (DWLR-PA logger, Prignitz Mikrosystemtechnik GmbH, Wittenberge, Germany) was installed ca. 50 m from the tower as part of the Moorbodenmonitoring (MoMoK) project by the Thünen Institute of Climate-Smart Agriculture (Frank et al., 2025). The MX2001 WT logger had several power outages and was moved in the beginning of 2025, resulting in data gaps and a non-continuous time series. The complementary DWLR-PA WT logger covered the period from 19 September 2023 until 5 March 2025. To create a full time series spanning from the beginning of 2023 until the end of 2025, the data measured with DWLR-PA was used as a reference, and the time spans before 19 September 2023 and after 5 March 2025 were inserted from the MX2001 data. To remove offsets due to the sensor positions, data from 2023 and 2025 from the MX2001 were individually fitted to the complementary WT data, with an ordinary least squares (OLS) regression (R2 = 0.91 for 2023 and R2 = 0.88 for 2025).

An eddy covariance tower was installed in September 2022. The tower setup includes a LI-7200RS closed path infrared gas analyzer (LI-COR Inc., Lincoln, Nebraska, USA) and a Gill HS-50 anemometer (Gill Instruments Ltd., Lymington, Hampshire, UK). In December 2023 a LI-7700 open path CH4 analyser (LI-COR Inc., Lincoln, Nebraska, USA) was added to the setup. Gas concentrations and wind speeds were measured on a tripod at 2.77 m height and with a frequency of 10 Hz. The raw EC data was processed into half-hourly aggregates using the EddyPro software version 7.0.9 by LI-COR and adhering to best practices agreed on by the community (Sabbatini et al., 2018). Raw 10 Hz data were filtered for spikes, dropouts and absolute limits (Vickers and Mahrt, 1997). A filter was applied to remove CO2 measurements when the signal strength was below 80 % and CH4 measurements when the signal strength was below 10 % (McDermitt et al., 2011). Anemometer tilt was corrected using the double-rotation method (Wilczak et al., 2001). Time-lags between measurements of gas concentrations and wind components were accounted for using the automatic time lag optimization (Sabbatini et al., 2018). Spectral losses were corrected by deriving reference (co-)spectra for well-developed turbulent conditions and then correcting measurement spectra for high-pass and low-pass filtering effects (Fratini et al., 2012; Moncrieff et al., 2004). CH4 concentrations measured with the open-path analyser were corrected for density fluctuation using the WPL correction (Webb et al., 1980) and spectroscopic effects (McDermitt et al., 2011).

Since measurements were done at a low height (2.7 m) no profile measurements of gas fluxes were conducted. Storage fluxes were derived with the default single-point estimation implemented in EddyPro and added to the half-hourly CO2 flux to calculate the net ecosystem exchange (NEE). The fluxes were then filtered to remove outliers, periods with erratic behaviour potentially induced by faulty sensors and poorly developed turbulent conditions. As a first step all data points with a quality flag of 2 were removed from the dataset, following the quality flagging system with flags 0, 1 and 2 (Mauder and Foken, 2015). To remove outliers representing unrealistically high fluxes we determined the 99.9 % and 0.1 % quantiles of the highest quality fluxes (flag 0 based on the flags mentioned above) separately for day and night-time of CO2, CH4, the sensible heat flux H and the latent heat flux LE. Values above and below these thresholds were removed. Nightly conditions were determined using a threshold of less than 10 Wm-2 SWIN. After removing fluxes above the absolute limits, remaining spikes were determined with a conservative approach, removing data points that are four standard-deviations above or below the mean of a running window of 30 d. Finally, data measured under poorly developed turbulent conditions (u∗-filtering) was filtered out with the R-package REddyProc V. 1.3.2 (Wutzler et al., 2018) using R version 4.5.2 (R Core Team, 2025). To estimate the range of possible u∗ thresholds, we generated 100 bootstrap resamples and derived the empirical distribution of the threshold. From this distribution, we extracted 37 evenly spaced quantiles between the 0.05 and 0.95 quantiles (in increments of 0.025). The final derived thresholds across bootstrap-derived quantiles ranged between 0.09 and 0.19 while the central estimate of the threshold for the years 2023, 2024 and 2025 were 0.10, 0.13 and 0.13. All further processing steps (gap filling and partitioning) were repeated for all the potential u∗-thresholds to allow an estimation of the uncertainty associated with the choice of the central threshold.

To derive annual budgets of fluxes the gaps resulting from the above-described steps needed to be filled. Deep ensembles of neural networks have been shown to yield good predictive performance as well as reliable estimates of the respective model uncertainties (Vekuri et al., 2025). We trained an ensemble of five neural networks (Lakshminarayanan et al., 2017) for each u∗-threshold to fill half-hourly CO2 and CH4 fluxes. The predictors used were TA, VPD, SWIN, WT and fuzzy variables derived from the hour and the month turned into sine and cosine waves. We followed a model setup previously shown to yield good flux predictions (Vekuri et al., 2025) using the Tensorflow library (Abadi et al., 2016). The models consist of two hidden layers with 50 nodes each and ReLU (Rectified Linear Unit) activations. Since flux errors approximately follow a Laplace distribution (Hollinger and Richardson, 2005) the loss function of the model was the Laplace negative log likelihood. The models are trained to predict the mean and the logarithm of the scale parameter of a Laplace distribution, yielding a mean prediction as well as the predictive uncertainty. Following (Vekuri et al., 2025) the models were initialized with random weights (He et al., 2015) using the Adam optimizer (Kingma and Ba, 2017). A randomly sampled subset corresponding to 10 % of the data was used as a validation set for early stopping with a patience of 20 epochs. The final gap filling uncertainty consists of the epistemic uncertainty, the uncertainty across the predicted means from the five model iterations, and the aleatoric uncertainty, the mean of the predicted Laplace scale parameters, added in quadrature. To test the performance of the gapfilling models for data gaps of different lengths (12 h, 1 d, 10 d, 20 d, 30 d) we randomly sampled 20 % of the data randomly in contiguous blocks of the respective length, making sure that each block contains at least 50 % of non-missing data. This data was withheld during model training and later used to calculate the R2 and root mean square error (RMSE). This procedure was repeated 10 times for each gap length. For this test we used NEE filtered with the annual central u∗ estimate.

Uncertainty of the annual CO2 balance was computed by adding in quadrature the aggregated uncertainties arising from random measurement errors (Finkelstein and Sims, 2001), u∗- threshold selection, quantified as the standard deviation of annual balances derived from gap-filled flux time series generated using alternative u∗ thresholds (Pastorello et al., 2020), and aleatoric and epistemic gapfilling model uncertainty (Vekuri et al., 2025).

Finally half-hourly filled CO2 was partitioned into the two components of gross primary production (GPP) and ecosystem respiration (Reco) using the night-time partitioning approach (Reichstein et al., 2005). Daily NEE, GPP and Reco were derived from the half-hourly data as daily averages. As a quality control we tested the energy balance closure (EBC) of the flux data. The energy balance is calculated according to Eq. (1). 1LE+H=(LWIN-LWOUT)+(SWIN-SWOUT)-G It describes whether the measured fluxes of sensible (H) and latent (LE) heat flux together sum up to the total energy input into the ecosystem measured by the instrumentation minus the soil heat flux (G). The linear slope between the left and right side of Eq. (1) resembles the EBC and is used as a quality criterion of the fluxes (Foken, 2008). The footprint of the tower was estimated using a two-dimensional flux footprint model (Kljun et al., 2015).

Seasonality is often derived from digital imagery or satellite products (Richardson et al., 2018). However, the camera installed at the site provided no reliable data in the first year due to a misconfiguration. We additionally tested the use of the MODIS normalized difference vegetation index (NDVI) product for deriving seasonality metrics (Didan, 2021). However, the analysis was subject to considerable uncertainty due to the 16 d temporal resolution and missing observations caused by cloud cover. We therefore derived the seasonality and dates of key phenological changes (start of season, SOS; peak of season, POS; end of season, EOS; length of season, LOS) based on the annual course of the GPP partitioned from NEE The change dates were detected using a fitted double logistic sigmoid function (Gonsamo et al., 2013): 2GPP(t)=L11+exp⁡(-k1⋅(t-t0))+L21+exp⁡(-k2⋅(t-t1)) where t represents the numeric index of the time series, L1 and L2 are the asymptotic amplitude, k1 and k2 are the parameters controlling the steepness (curvature) of the transitions and t0 and t1 are the inflection points to the early- and late-summer transitions in GPP. The change dates were determined by calculating the third derivative of Eq. (2). The date of the first local maximum and the last local minimum were taken as the start and end dates of the season. The peak of the season was determined as the date of the maximum of the fitted function. The seasons were then defined such that the non-growing season (NGS) was before and after the start and end of the growing season, early growing season (EGS) is between the start and the peak, and late growing season (LGS) between the peak and the end of the growing season. To comprehensibly visualize the derivatives together with the GPP we normalized both the mean daily GPP and double sigmoid as well as the derivatives to a range of zero to using Eq. (3) 3xi′=xi-min⁡(x)max⁡(x)-min⁡(x) where x is the whole time series, xi is the specific value at index i and xi′ is the normalized value. Uncertainty in phenological transition dates was estimated by bootstrapping model residuals. Residuals from the fitted double sigmoid model were resampled and added to the modelled seasonal cycle to generate 300 realizations of the GPP time series. For each realization, the model was refitted and SOS, POS, and EOS were recomputed. Confidence intervals were then derived from percentiles of the resulting distributions.

To compare the climate and fluxes between the 3 years, anomalies of daily mean fluxes and climate were calculated using a block-averaging method. Daily anomalies were calculated as the deviation of each daily data point from the mean of the daily data within a ± 3 d-of-year window across all 3 years. Such anomalies were calculated for NEE, GPP, Reco, TA, SWIN, WT and VPD and are in the following denoted as z with the respective variable as subscript, e.g. zNEE.

The drivers of gas fluxes at a daily timescale were analysed using forward stepwise multiple linear regression models (MLR) between the anomalies of daily fluxes and anomalies of environmental drivers. Few days have no data gaps, e.g. due to filtering of low turbulence conditions at night. Thus, when working on the daily timescale, some gapfilled data needs to be included to avoid excessive data loss and skewed filtering (skewed towards nighttime when turbulence is lower). To minimize the effect of gapfilling on the driver analysis we used a threshold of maximum 50 % gapfilled data per day. We tested the influence of the percentage by repeating the analysis for thresholds from 10 % to 80 %. The results are presented in Sect. S2 in the Supplement). As the potential drivers of the fluxes zSWIN, zTA, zVPD and zWT were included. Additionally, we added multiplicative interaction terms between all drivers as predictor candidates. Interaction terms are in the following denoted with a colon between two variables (e.g. zTA:zSWIN). We only included WT and not SWC as a predictor because there were several gaps in the time series of SWC due to sensor failures and WT and SWC are closely correlated. Similarly, the time series of TA was more complete than that of TA and TA and TS are correlated (Pearson R = 0.83), thus we only included TA as a predictor. To compare the relative influence of the drivers, all predictors were standardized to the range of zero to one before fitting the model. Separate models were trained for the targets zNEE, zReco and zGPP. Predictors were sequentially added to the model. First the predictor was checked for collinearity. If the variance inflation factor (VIF) was larger than five, the model did not improve in its Akaike Information Criterion (AIC) and the adjusted R2 did not improve at least by 0.05 the predictor was rejected. Since driver strengths are likely to change between the non-growing and through the growing season, e.g. due to phenological development, separate models were created for the different seasons (see Sect. 2.4). All models were built in Python version 3.7. MLR models were built using statsmodels version 0.14.6 (Seabold and Perktold, 2010). Preprocessing of the data made use of the scikit-learn package version 1.8 (Pedregosa et al., 2011). Each fitting was repeated on 1000 bootstrap-resampled datasets to derive confidence intervals of the coefficients.

Monthly changes in amounts of carbon fluxes may be caused by seasonal climatic changes while the underlying response functions remain the same. To assess whether the environmental response of GPP or Reco fundamentally changed across the years, we established monthly temperature- and light-response functions. From those we extracted the characteristic parameters temperature dependency of respiration Q10, the light-use efficiency α and the maximum CO2 uptake rate Pmax⁡. Note that these functions were not used to derive partitioned data but only to detect changes in the ecosystem responses to light or temperature.

We derived Q10 by fitting an exponential Lloyd–Taylor model (Lloyd and Taylor, 1994) to night-time CO2 fluxes (NEEnight) with Eq. (4). NEEnight was used instead of the partitioned Reco because Reco was already modelled based on a TA response curve. 4NEEnight(T)=Rref⋅exp⁡E0⋅1Tref-T0-1T-T0 Rref is the baseline respiration rate at reference temperature, T0 the baseline temperature set to -46 °C, Tref the reference temperature (set to 10 °C) and E0 the activation energy parameter that determines the temperature dependency of the model. After the determination of E0, the temperature response parameter Q10, depicting the change of respiration with a 10 °C change in temperature, was calculated with Eq. (5). 5Q10(T)=exp⁡E0⋅1T-T0-1T+10-T0

To determine the light response of the ecosystem fluxes we fitted a Michaelis–Menten rectangular hyperbolic light-response model (Bao et al., 2019; Falge et al., 2001) to daytime net ecosystem productivity (NEP) using Eq. (6). NEP is the inverse of NEE, which places net CO2 uptake as positive values, making the positive relationship with increasing radiation more intuitive to interpret. Daytime periods were defined as those with incoming solar radiation of more than 10 Wm-2. 6NEP(SWIN)=α⋅SWIN⋅Pmax⁡α⋅SWIN+Pmax⁡-Rd α is the light-use efficiency (µmolJ-1), the initial slope of the curve representing the efficiency of carbon taken up per quantum of light, respectively. Pmax⁡ is the maximum rate of CO2 uptake (µmolm-2s-1) and Rd (µmolm-2s-1) represents daytime ecosystem respiration.

Detecting the effects of TA and VPD on carbon fluxes is complicated by their mutual correlation and co-variation with other drivers such as SWIN. We therefore employed a sequential filtering approach to isolate the effect of each variable while controlling for the others (Fig. S3 in the Supplement). Measured (non-gapfilled) half-hourly NEE and the respective partitioned data were first filtered for daytime conditions (SWIN > 10 Wm-2) and then for near light-saturated GPP using the 80th percentile of SWIN (570 Wm-2). Optimum conditions (VPDopt and TAopt) were determined using boundary line regression, in which GPP is grouped into 1 °C bins, 1 hPa bins, respectively, and separate regression lines are fit to percentiles from those bins (0.55–0.95, steps of 0.1), capturing the maximum GPP response envelope. VPDopt and TAopt were defined as the bin where the highest three consecutive percentile regressions reached their maximum (tolerance: ± 1 bin). Data were then filtered to retain only conditions above these optima, as these are the ranges where a potential impact of TA or VPD can be expected.

To evaluate the effect of high TA while controlling for VPD (and vice versa), data were filtered to a narrow window of the controlled variable, with the window size selected as the largest window showing no detectable correlation (p > 0.05) of the controlled variable with the flux (tested from 0.5 to 6.5 °C, hPa respectively; Fig. S4 in the Supplement). For the TA analysis, the VPD window was centred on the mean VPD of data with TA above the 50th percentile of TA > TAopt (window: ± 2.1 °C). For the VPD evaluation, TA was centred on the mean TA of data with VPD above the 90th percentile of VPD > VPDopt (window: ± 5 hPa; Fig. S4). The different percentiles were required to capture high conditions of the evaluated variable and enable the removal of the correlation with the controlled variable while capturing enough data for a regression analysis. Finally, OLS regressions (statsmodels v0.14.6; Seabold and Perktold, 2010) were used to quantify the effect of TA or VPD on Reco, GPP and NEP, with separate models for the early and late growing seasons (see Sect. 2.4).

The additional meteorological data used to supplement the data measured at the EC tower completed the data with a high level of agreement. The R2 of the linear models between TA and RH from the secondary tower and and the data measured at the EC tower was more than 0.90. For TA from the nearby DWD station R2 was 0.96 and for RH 0.80. The R2 on test data of the model used to fill SWIN was 0.97.

Over the 3 years the mean annual TA stayed between 11.6 and 11.0 °C (Fig. 2a). Annual mean radiation was also in a close range between a maximum of 124 Wm-2 in 2025 and a minimum in 2024 with 116 Wm-2 (Fig. 2c). Both in 2023 and 2024 the site experienced exceptionally high amounts of precipitation with 1006 and 913 mm total rainfall (Fig. 2b), respectively, compared to a long-term average of 791 mm. Mean annual WT tracked annual precipitation, averaging -25 cm in the two wet years (2023 and 2024) and reaching its deepest level at -37 cm in the driest year, 2025 (Fig. 2d).

Comparing the individual months between the years revealed more striking meteorological differences. In 2024 the average TA in February, March and May was 7.7, 8.8 and 16.1 °C, making each of these months the warmest among the 3 years by a margin of at least 2 °C (Fig. 2a). In 2023 on the other hand the end of the growing season was the warmest among the 3 years by more than 1.5 °C at a mean of 17.3 and 12.9 °C in September and October. June 2023 was the warmest month of the 3 years with a mean monthly TA of 19.3 °C, 3.1 °C above the long-term average of 16.2 °C (Fig. 2a).

While annual cumulative rainfall was above average in both 2023 and 2024 (139 % and 126 % of the 20-year average; Fig. 2b), monthly distribution differed notably. In 2023, May and June were drier than usual (74 % and 51 % of average, while in 2024 March was the driest month (47 %) and July the wettest (171 mm; 220 % of average; Fig. 2b). The annual average WT in 2024 was similar to that in 2023. The lack of precipitation in May and June 2023 correlates with a sudden drop from -13.4 cm depth in April to -44 cm in June (Fig. 3d and e). In 2025 the year started with a precipitation deficit in March with only 3 mm, representing a 95 % reduction compared to the long-term average (Fig. 2b). The year 2025 was by far the driest among the 3 years, with March, May, June, and August all experiencing less than 50 % of the long-term average rainfall. Accordingly, the water table dropped earlier in 2025 and fell below -60 cm (Fig. 3e), resulting in a lower mean annual WT of -37 cm.

The EBC expressed as the slope between the sum of energy fluxes and the radiation budget minus the soil heat flux was 0.83 (Fig. S5 in the Supplement) for the 2023–2025 study period, which is slightly higher than the average closure of 0.8 found for the FLUXNET2015 dataset (Mauder et al., 2024). Due to filtering steps and gaps in the original data in total 29 % of CO2 flux datapoints were filtered out, 53 % of the night-time and 15 % of the day-time fluxes, respectively. Two events discontinued gas flux measurements in mid-summer, a lightning strike near the station in 9 July 2023 that broke the datalogger, as well as substantial damage caused by rodents on 30 June 2024 that led to a short-circuit, damaging controllers, gas analysers and soil sensors. The two incidents led to gaps of 19 and 29 d in 2023 and 2024, respectively. The ensemble neural networks filled gaps of lengths up to 30 d with an average R2 of at least 0.8 (Table 1) across ten repeated validation data samplings.

For CH4 fluxes 58 % of datapoints were filtered out. Fluxes remained very low during the study period (Fig. S2 in the Supplement) with the 25th to 75th quantile ranging from -9.2 to 5.01 nmolm-2s-1. The fluxes barely exceeded the uncertainty of the measurements, with the interquartile range of random measurement uncertainty alone ranging from 2.2 to 6.2 nmolm-2s-1. Because the CH4 fluxes were so low and irregular no correlations with environmental or meteorological variables were found, preventing a reliable gapfilling (R2 was near zero for any tested method) and thus the calculation of an annual CH4 budget. Therefore, CH4 fluxes are not included in any mention of the sites carbon budget, and no driver analysis was conducted.

The identification of the growing season revealed notable differences in timing of the start and end among the 3 years (Fig. 4). In 2024 the growing season started earliest, already on 25 March or day-of-year (DOY) 84 (Fig. 4b). In 2023 in comparison the growing season onset was detected 23 d later, on 17 April (Fig. 4a). The start of the growing season 2025 fell in between the other 2 years, 9 d after the onset in 2024 on DOY 94 (Fig. 4c). For each year the other seasonal transitions shifted along with the onset of the growing season. 2023 had the latest start of the growing season but also the latest peak at day 201 (20 July), 6 d later than in 2025, and the latest end of the growing season at DOY 311 (7 November), one week after the second latest in 2025 at DOY 306 (Fig. 4a). In 2024 the growing season started earliest, and it also ended earliest on DOY 301 on 27 October (Fig. 4b). Overall, growing season length was similar in 2024 and 2025 (217 and 212 d, respectively) and shorter in 2023 (204 d). The derived seasonal transitions were stable based on the residual bootstrapped confidence intervals. The maximum confidence interval ranged 8 d. Therefore, most transition points were significantly different across the years (all SOS and POS), only the confidence intervals of the EOS of 2023 and 2024 both overlapped with those of 2025 (Fig. 4).

In the three measurement years the CO2 budgets switched from a net source in 2023 with 112.6 ± 14.5 gCm-2 to a weak CO2 sink in 2024 with -24.8 ± 15.1 gCm-2 (Table 2). In 2025 the site was a net CO2 source again with emissions of 47.6 ± 27.8 gCm-2 (Table 2). In 2024, where the ecosystem was a weak carbon sink, Reco was ca. 101 gCm-2 higher than in the previous year 2023. GPP however increased by 239 gCm-2 from 1337.9 ± 45.2 to 1576.3 ± 18.2 gCm-2, surpassing Reco. In 2025 both Reco and GPP were lower than in the previous years with 1270.0 ± 14.6 and 1222.4 ± 11.1 gCm-2.

The NGS was a CO2 source in all years, with relatively low interannual variability in NEE (max. difference of 46 gCm-2 between 2023 and 2025; Table 2). The growing season (GS) was a net CO2 sink in all years, with substantially higher interannual variability: NEE ranged from -152.4 ± 14.5 gCm-2 in 2024 to -35.7 ± 13.7 gCm-2 in 2023, a difference of 117 gCm-2. Reco and GPP were expectedly both higher in the GS than the NGS across all years. GS Reco ranged from 1086.5 ± 8.9 to 1324 ± 25.8 gCm-2, GPP ranged from 1141.1 ± 8.5 to 1476.5 ± 22.46 gCm-2, with both peaking in 2024. Interannual variability in GS carbon fluxes was higher than in the NGS, with Reco differing by up to 238 gCm-2 and GPP by up to 335 gCm-2 between 2024 and 2025. NGS Reco varied less, ranging from 183.6 ± 1.9 to 228.4 ± 8.8 gCm-2. Partitioning of Reco and GPP into seasonal sums showed low sensitivity to uncertainty in phenological transitions. Total sums for both the non-growing and growing season varied by less than 3 % between minimum and maximum growing season lengths based on the confidence intervals of the transitions (Fig. 4). NEE was more sensitive to uncertainty in growing season length. In 2023, the GS NEE increased by up to 17 % when integrating over the longest plausible growing season (DOY 103–313) compared to the best estimate (DOY 107–311), reflecting the inclusion of additional days with net CO2 emissions at the seasonal boundaries.

The monthly flux budgets reveal the decisive periods for the differences in the annual carbon budgets (Fig. 5). In 2024 CO2 uptake started earlier and was stronger compared to the other 2 years (Fig. 5a and b).

Both Reco and GPP were strongly increased in April to June 2024 compared to the other years (Fig. 5b and c), with GPP rising more than Reco. This led to reduced net emissions in April 2024 and a marked increase in net CO2 uptake in May and June 2024 (Fig. 5a).

The high total emissions of 2023 were accompanied by higher net emissions in March and April and lower net uptake in August than in the other years (Fig. 5a). From February to August, GPP and Reco were both largest in 2024, apart from July, where Reco and GPP were higher in 2023.

The multiple linear regression model performance varied considerably among seasons and target flux anomalies. In the non-growing season, the models for zReco and zNEE explained a large share of the variance (R2 = 0.67 and 0.82), while zGPP was poorly captured (R2 = 0.32), however this is expected as GPP is mostly absent in the NGS. The early growing season showed the lowest model performance across all seasons, with R2 ranging from 0.25 for zGPP to 0.41 for zNEE, thus daily climate anomalies alone explained a smaller share of flux variability during this period. In the LGS zNEE and zGPP were well explained (R2 = 0.76 and 0.63) and models for zReco showed moderate performance (R2 = 0.59). The chosen predictors, their signs and magnitudes of the effect sizes proved to be stable for a wide range of different thresholds of maximum allowed gapfilled data per day (see Sect. S2 and Fig. S1 in the Supplement for more details).

The selected predictor variables varied for each flux between the seasons (Fig. 6). For zNEE the strongest driver of increasing emissions was zTA in the NGS and LGS, thus higher TA increased net CO2 emissions (Fig. 6a). Notably the effect of zTA was reduced in the EGS. During the EGS, the interaction of anomalies in VPD and SWIN instead emerged as the strongest driver of increased NEE (thus increased emissions or decreased CO2 uptake). In contrast, anomalies in incoming radiation (zSWIN) were the strongest driver of increased CO2 uptake across all seasons. Concurrent conditions of warm TA and high light availability (thus the interaction zTA:zSWIN) exerted a small uptake-enhancing effect in the NGS and LGS. zVPD and zWT had similarly small effects, with zVPD and zWT decreasing NEE in the LGS and NGS, respectively. Note that WT becomes more negative as the water table drops. A negative effect of zWT on NEE therefore implies that lower (deeper) water tables are associated with higher (more positive) NEE values. Similarly, negative effects of zWT on zReco and zGPP translate to an increase in Reco or GPP with dropping WT.

zReco was primarily controlled by zTA across all seasons, with warmer conditions consistently driving enhanced Reco (Fig. 6b). A secondary effect of zWT indicated that drier conditions also increased Reco, particularly during the growing season. Conversely, positive VPD anomalies suppressed Reco during the EGS. Remaining significant effects were very small across all seasons (coefficients <0.1).

Increased GPP was associated with increased SWIN, TA, and lower WT, alongside the interaction of zTA:zSWIN in the NGS (Fig. 6c). Increased incoming solar radiation (zSWIN) had the strongest positive effect on GPP in the LGS and a weaker effect in the EGS. The influence of zTA was weakest in the late growing season. The interaction of temperature and radiation zTA:zSWIN had a positive effect on GPP during during the NGS. Across all seasons, the interaction zVPD:zSWIN exerted a negative effect on GPP, with a stronger impact in the EGS than the LGS, suggesting a negative impact of VPD on CO2 uptake under light-saturated conditions.

Light-use efficiency α and Pmax⁡ showed a typical seasonal cycle with Pmax⁡ reaching up to 37 µmolm-2s-1 in June and July (Fig. 7). In April, May and June 2024 Pmax⁡ was strongly increased with 13, 26 and 38 µmolm-2s-1 (Fig. 7a) compared to an average of 5, 14 and 25 µmolm-2s-1 in the other 2 years. Further, the light-use efficiency in May 2024 was higher than in the other years with 0.11 µmolJ-1 compared to 0.04 and 0.07 µmolJ-1 in 2023 and 2025, respectively (Fig. 7a). From June to August 2024, the Q10 of respiration was higher than in the other 2 years, averaging 1.7 over the 3 months compared with 1.4, and reaching a maximum of 1.9 in June 2024 (Fig. 7c). In August 2025 both Pmax⁡ and the Q10 of respiration are notably lower than in the other 2 years with a Q10 of 1 compared to 1.2 and 1.6 and a Pmax⁡ of 26 µmolm-2s-1 compared to 34 µmolm-2s-1 in the other 2 years.

The filtering successfully isolated the effects of TA and VPD. Correlations between fluxes and controlled variables did not exceed a Pearson correlation of 0.13 (Fig. S6 in the Supplement). Notably, the correlation of fluxes with SWC was negative (-0.18 to -0.42) both in the data controlled for VPD and the data controlled for TA (Fig. S6), depicting increasing carbon fluxes under reducing soil moisture in the filtered data. We found no correlation between SWC and VPD (see S1 in the Supplement).

Throughout the growing season, even under highest encountered TA conditions warming was positively correlated with both GPP and Reco (Fig. 8d and f), but increased NEP in the EGS and decreased NEP in the LGS (Fig. 8b). In the EGS, GPP responded stronger to high TA than Reco (Fig. 8d and f), resulting in increasing net CO2 uptake (Fig. 8b). In contrast, in the LGS Reco increased stronger with TA than GPP (slopes of 0.65 and 0.42, Fig. 8d and f), leading to a decreasing net CO2 uptake (Fig. 8b).

Contrary to the TA response, VPD increasingly suppressed all fluxes in the EGS (Fig. 8a, c, and e). The decrease of GPP at high VPD during the EGS was almost twice as strong as the reduction in Reco with a slope of -0.98 (Fig. 8e) compared to -0.52 (Fig. 8c). In the LGS the responses changed markedly. Reco was not significantly correlated with VPD while GPP was still reduced but the effect was weaker than in the EGS with a slope of -0.33 (Fig. 8e). Together the reduction of GPP and the absent reduction of Reco in the LGS also resulted in a decrease in NEP similar to the EGS (slope of -0.37 compared to -0.46, Fig. 8a).

Annual CO2 budgets were variable, ranging from a net source of 112.6 ± 14 gCm-2 to a weak sink of -24.8 ± 15.1 gCm-2 (Table 2). These values fall between IPCC CO2 emission factors for rewetted temperate (-64 to 18 gCm-2) and extraction peatlands (110 to 420 gCm-2) (IPCC, 2014). According to the German CO2 emission inventory for wetlands, our study site with a WT persistently deeper than -20 cm and shrub-dominated vegetation, falls in the category of drained unutilized land, where emissions are estimated within the range of 70 to 1080 gCm-2 (Tiemeyer et al., 2020). With mean annual emissions of 45 gCm-2 the measured annual budgets confirmed a moderate but real contribution to global warming.

Past studies found that either the variability in Reco (Aslan-Sungur et al., 2016; Drollinger et al., 2019; Ueyama et al., 2014) or in GPP (Chivers et al., 2009; Li et al., 2021; Lund et al., 2010) controlled the annual variability of CO2 budgets. In the three measurement years the variability of NEE in the growing season was substantially higher than in the NGS (117 and 48 gCm-2 maximum difference; Table 2) and the differences in GPP were greater than those in Reco (335 and 238 gCm-2 maximum difference, respectively; Table 2). This highlights that the growing season vegetation activity played a pivotal role in shaping the interannual variability in CO2 fluxes.

Ranges of annual Reco and GPP at Amtsvenn (1270.0 ± 14.6 to 1551.5 ± 22.7 and 1222.4 ± 11.1 to 1576.3 ± 18.2 gCm-2, respectively) are high compared to other reported flux budgets from semi-natural bogs with shrubs in the footprint and similarly fluctuating water tables, where annual sums of GPP and Reco typically range between 600 and 800 gCm-2 (Humphreys et al., 2014; Hurkuck et al., 2016). The high GPP is likely attributable to the dense shrub canopy, as shrubs can sustain carbon uptake rates exceeding those of natural peatlands or even grasslands (Gyimah et al., 2020; Quin et al., 2015). Elevated Reco is consistent with the WT persistently below -15 cm, which promotes peat aeration and microbial decomposition (Ma et al., 2022; Swails et al., 2022). Additionally, Reco may be enhanced through vegetation activity. Root exudates released by ericaceous shrubs can increase CO2 emissions by increasing the microbial potential for cellulose and hemicellulose decomposition (Cai et al., 2024) and autotrophic respiration, which can make up more than 50 % of the total Reco in shrub dominated ecosystems (Rankin et al., 2023), increases with GPP (Waring et al., 1998). Finally, both Reco and GPP may be enhanced by the elevated nutrient availability in the upper peat layers (Lund et al., 2009).

Methane fluxes were low and did not exceed measurement uncertainty. CH4 emissions require anoxic, waterlogged soils (Evans et al., 2021; Frolking et al., 2011; Zhang et al., 2022) and approach zero when annual average WT is below 20 cm (Calabrese et al., 2021; Evans et al., 2021), which was almost always the case in the three measured years (Fig. 3e). Repeated drying-rewetting cycles can further suppress CH4 formation through regeneration of electron acceptors (Gao et al., 2019; Knorr et al., 2009). Although Molinia caerulea, which is present with up to 25 % cover, could act as a methane conductor (van den Berg et al., 2020; Buzacott et al., 2024; Leroy et al., 2017), shrub dominance and persistently deep water tables likely explain the absence of detectable CH4 emissions. In Amtsvenn, the peat below 30 cm depth contains low concentrations of inorganic alternative terminal electron acceptors (Lemmens et al., 2026), yet ombrotrophic peat typically contains redox active organic moieties, contributing considerable electron accepting capacities (Guth et al., 2023). Thus, a suppression of methane formation by alternative electron acceptors during the warm summer period may also occur. Although nitrogen and phosphorus concentrations are elevated in the upper 10–15 cm of the peat, the WT rarely reaches these layers. Moreover, when it does, low TA likely co-limits methane production.

Notably, for a full carbon balance the export of dissolved organic carbon (DOC) through the drainage ditches may present an additional pathway of C loss (Rosset et al., 2022). DOC measurements were not conducted during the three measurement years and are thus not included in the annual budgets.

Studies have produced varying results on whether climate warming increases or decreases net CO2 uptake in northern peatlands, suggesting both increasing emissions of CO2 and CH4 into the atmosphere with warming (Hanson et al., 2020; Helbig et al., 2022; Qiu et al., 2022) or enhanced net CO2 uptake, especially in boreal Sphagnum-dominated peatlands (Zhao et al., 2026). In this study annual budgets of both GPP and Reco were highest in the warmest year 2024 (Table 2, Fig. 2a). A warm spring in 2024 (∼ 2 °C above other years from March to May) triggered an earlier growing season onset (DOY 84 in 2024 and 107 in 2023, Fig. 4a and b) and a rapid GPP increase (Fig. 5b) despite lower radiation in 2024 (Fig. 3c), driving the site to become a weak CO2 sink. The higher GPP was accompanied by an increased Pmax⁡ of up to 85 % between April and June relative to the other years (Fig. 7a). Enhanced CO2 uptake following pre-growing season warming was previously found in both fens and bogs (Adkinson et al., 2011; Heiskanen et al., 2021; Helfter et al., 2015; Peichl et al., 2014). In contrast to a boreal fen where Pmax⁡ increased with a lag in mid-summer following a warm spring (Peichl et al., 2014), Pmax⁡ at Amtsvenn was immediately elevated at the onset of the growing season in April and May (Fig. 7a). This may reflect the capacity of the evergreen shrub canopy to respond quickly to warm conditions without the constraints of leaf development. This dynamic shift in the response of CO2 fluxes to climate likely contributed to the low performance of the MLR models for this period.

In contrast to the effect of the warm spring, during the warm autumn in September and October 2023 (Fig. 3a) the net CO2 budget was not significantly different from the other years (Fig. 5a) because both Reco and GPP were similarly enhanced (Fig. 5b and c). The partitioned influence of TA further confirmed the different responses of CO2 fluxes to high TA in the EGS and LGS. In spring warmer conditions enhanced net CO2 uptake by stimulating GPP more than Reco, while autumn warming reduced net CO2 uptake by increasing Reco more than GPP (Fig. 8b, d, and f). The CO2 flux response to high TA thus depended critically on seasonal timing, consistent with findings on the large scale across northern ecosystems (Barichivich et al., 2013; Helbig et al., 2022).

The impact of high VPD conditions also changed seasonally. Under rising VPD net CO2 uptake was similarly suppressed throughout the NGS and LGS (Fig. 8a), but the individual responses of GPP and Reco differed drastically between the seasons. GPP was suppressed almost three times stronger (slope of -0.98 compared to -0.33) in the EGS than in the LGS (Fig. 8e). Reco on the other hand decreased with rising VPD in the EGS (slope of -0.52, Fig. 8c), which is also resembled in the impact of daily VPD on daily Reco (Fig. 6c). VPD and was not significantly correlated to VPD in the LGS (Fig. 8c). The decrease in Reco with rising VPD in the EGS thus reduced the impact of decreasing GPP on the overall net CO2 balance and led to the similar response as in the LGS. To the best of our knowledge, there is no study showing such a divergent response of CO2 fluxes to high VPD conditions between early and late growing season periods controlled for concurrent changes in SWIN or TA. While seasonally varying TA impacts have received some attention (Bubier et al., 1998; Heiskanen et al., 2021; Helbig et al., 2022; Helfter et al., 2015) the seasonal trajectory of VPD impacts has not been systematically evaluated. A detailed analysis of the leaf- or plant-level mechanisms behind this divergence is beyond the scope of this study. It may however represent a process counteracting increased spring CO2 uptake under future warming – particularly as extreme VPD and TA events are projected to intensify (Intergovernmental Panel on Climate Change (IPCC), 2023; Shekhar et al., 2024) and rising VPD is expected to place greater pressure on plant water regulation (Li et al., 2025; Novick et al., 2016).