The Bernadouze peatland, located at an altitude of 1343 m in the eastern French Pyrenees (42.80273° N; 1.42361° E), covers roughly 4.7 ha and is part of a national biological reserve. Designated as a regional biological reserve since 1983 and a Natura 2000 site since 2007, it is one of the four sites of the French National Peatland Observatory Service SNO-Tourbieres. Characteristic of alpine and southwestern European mountainous peatlands, it is classified as a soligenous fen, continuously receiving water from precipitation and surface runoff. It has formed over the past 5000 years within a 1.4 km watershed with steep slopes averaging 50 %. A beech forest surrounds the fen, extending up to 1800 m. The peatland has an average peat depth of 2 m, reaching up to 6 m in some areas. The vegetation comprises species typical of both ombrotrophic areas, such as Sphagnum palustre and Sphagnum capillifolium, and minerotrophic areas, including Carex demissa and Equisetum fluviatile . This distribution is illustrated in Fig. 1 of , which shows the spatial arrangement of vegetation types across the study site.

Piezometer wells are 50 mm diameter PVC tubes, distributed across the peatland to cover the full spatial extent of the site (Fig. 1 from ), and their placement corresponds to the locations where chamber measurements were conducted. Water table depth (WTD) data were recorded at 1 h intervals, and the mean value from the nine piezometers is used throughout the study. As in precedent studies of the same peatland we used the S2M reanalysis, which combines the French weather service SAFRAN meteorological reanalysis (8 km resolution) with the SURFEX/ISBA-Crocus snow cover model . The S2M chain was applied in a local mode at the scale of the “Couserans” massif (a mountainous region in the central Pyrenees, southwestern France), assuming homogeneous weather conditions across the massif for equivalent altitudes but accounting for local topographic features in the calculation of shortwave radiation. The S2M model has a vertical resolution of 300 m and provides hourly outputs. The topographical features of the peatland, including altitude, slope, and aspect, were carefully accounted for to ensure that the atmospheric variables extracted for the site accurately represent local conditions.

Hourly validated time series of CO2 fluxes (GPP, ER, and NEE) spanning 2017 to 2022 are available, as detailed by . These time series were derived from statistical models based on monthly CO2 flux measurements using the static chamber technique. The use of the statistical model allows reconstruction of daily fluxes, enabling direct comparison with the ISBA model outputs at the same temporal resolution. The reconstructed fluxes are considered spatially representative of the peatland due to the coverage and replication of the chamber measurements. In the following sections, these reconstructed datasets are used as a reference for model validation, and readers should note that the validation is performed against the derived statistical model rather than the raw measurements.

In the present study, we utilize the ISBA model, which is part of the SURFEX land surface modeling platform (Version 9) and serves as the land surface component of the global Earth System Model CNRM-ESM . ISBA is widely employed to simulate surface processes, but it does not currently represent mosses or Sphagnum, which are essential components of ecosystems such as peatlands. To address this limitation, we introduce a new plant functional type (PFT) designed specifically for modeling mosses and Sphagnum within the ISBA framework. This work details the implementation of this novel PFT and the associated modifications to the model to better represent the dynamics of peatland ecosystems. The other PFTs, representing vegetation types such as herbaceous plants and trees, have been thoroughly described in the existing literature .

First, based on the ISBA outputs, we derived Eq. (), which relates, for each soil layer, the change in volumetric water content to the change in water equivalent height, assuming the layer is saturated. 16wj,sat(t)∂hj(t)∂t=dj∂wj,soil(t)∂t with wj,sat the volumetric water content of soil layer j at saturation, hj the height of water in the layer j, dj the vertical width of the layer j, wj,soil the volumetric water content of soil layer j.

Then, we derived the ISBA-diagnosed WTD by summing over the soil layers down to 2 m depth, i.e. those where variations in water content are significant. This calculation assumes that the variation of wsat with depth z is negligible, allowing it to be factored out of the summation (see Fig. A3a): 17WTDISBA=∑jΔhj=∑jΔwj,soil×djwj,sat where Δhj, and Δwj,soil represent the changes in water equivalent height and soil volumetric water content, respectively, calculated explicitly through the temporal discretization of Eq. ().

The dryness index is based on the work of (see Fig. A3b). We first define the daily soil water deficit dif(t): 18dif(t)=Tairnormalized(t)-WTDnormalized(t) We use the daily mean of Tair(t) and WTD(t) to calculate dif(t). Normalization of each variable (X) was done following Xnormalized=X−XminXmax−Xmin. A minimum WTD value of 1m was imposed to elevate the normalized WTD values, since the diagnosed WTD tends to produce water tables that do not drop sufficiently during droughts. This adjustment allows the normalized WTD to better discriminate drought periods, which would otherwise be overly smoothed by the index (see Fig. A3b).

To consider only periods of positive water deficit, we define the function f(t) as: 19f(t)=dif(t),if dif(t)>00,otherwise The Dryness Index (DI) is then calculated by integrating f(t) over the summer period: 20DI=∫julyaugustf(t)dt This formulation ensures that only days with a positive soil water deficit contribute to the DI, providing a simple and physically meaningful measure of summer dryness.

To compute Pearson correlation coefficients (r2), root mean square errors (RMSE) trends and associated p values, we used the statsmodels Python library. In particular, the Ordinary Least Squares (OLS) (https://www.statsmodels.org/dev/generated/statsmodels.regression.linear_model.OLS.html, last access: 11 May 2026) regression function was employed to fit a linear model and perform an F-test to assess the statistical significance of the relationship.

To assess the relative contribution of each season to interannual variability in net ecosystem exchange (NEE) (Fig. 6), we applied SHAP (SHapley Additive exPlanations), a game-theoretic approach widely used for interpreting machine learning models. Specifically, we trained an Ordinary least squares Linear Regression model from sklearn python library (https://scikit-learn.org/stable/modules/generated/sklearn.linear_model.LinearRegression.html, last access: 11 May 2026) to predict annual NEE from seasonal values (summer, autumn, spring, and winter). We then used SHAP values to estimate how much each seasonal predictor contributed to the model output. Rather than relying on model coefficients or explained variance, we computed the mean absolute SHAP values across all observations to quantify each season's average influence. These contributions were normalized to express their relative importance as percentages. This approach provides a robust and interpretable measure of feature relevance, even for collinear predictors. The SHAP methodology is described in detail at: https://shap.readthedocs.io/en/latest/index.html (last access: 11 May 2026).

A simple linear model using R2 or an F-test tends to attribute explanatory power to the first variables entered into the model. When predictors are correlated as is the case here with seasonal NEE components (e.g., summer NEE , autumn NEE) this can lead to shared variance being unfairly credited to one variable over another. In contrast, SHAP values, based on Shapley values from cooperative game theory, average the contribution of each variable across all possible combinations of input features. This results in a fair and consistent allocation of importance, even when variables are highly interdependent or collinear.

At the Bernadouze site, we obtained a 64-year meteorological data series (1959–2022) from the S2M reanalysis, which was used as input for the ISBA land surface model. To establish a contemporary carbon balance for the various carbon compartments, we simulated a 7000-year spin-up to account for the peatland's age, repeating the 64 year atmospheric forcing. This approach ensured a realistic accumulation of carbon in the soil reservoirs. The model was run in a one-dimensional configuration, at a single grid point corresponding to the location of the Bernadouze peatland.

For model validation (Sect. 3), we constrast simulations with 100 % Sphagnum cover and 100 % herbaceous cover. Herbaceous cover was simulated with the use of the ISBA PFT boreal grassland (BOGD). For the remainder of the study (Sect. 4), the vegetation distribution is assumed to consist of 70 % herbaceous plants and 30 % Sphagnum mosses, based on the cartography provided by . Accordingly, the GPP, ER, and NEE values presented correspond to this vegetation mix. A subsection is dedicated to analyzing the sensitivity of carbon fluxes to vegetation composition.

Figure 1 compares three time series from 2017 to 2023, illustrating Gross Primary Productivity (GPP), Ecosystem Respiration (ER) and Net Ecosystem Exchange (NEE) modeled by a statistical model (grey) and the ISBA model for Sphagnum (orange) and Herbaceous (blue). The top plot (a) shows GPP with r2 values of 0.6 and 0.62 (RMSE of 1.9 and 1.6 µmol-2s-1, see Fig. A4a, b), indicating moderate agreement between the models. The middle plot (b) shows ER with r2 values of 0.82 and 0.44 (RMSE of 0.9 and 1.6 µmol-2s-1, Fig. A4d, e), revealing stronger agreement for ER, particularly for Sphagnum respiration. The bottom plot, Fig. A4c, f, shows NEE with r2 values of 0.1 and 0.2 (RMSE of 1.6 and 1.3 µmol.-2s-1). Additional scatter plots for the mixed vegetation configuration (Fig. A4g–i) highlight contrasted performances across fluxes. For GPP, the mixed simulation shows a level of agreement comparable to that of herbaceous vegetation alone (r2=0.62, RMSE = 1.58 µmol-2s-1). For ER, the mixed configuration exhibits an intermediate performance, but with a substantial improvement compared to herbaceous vegetation alone (r2=0.60, RMSE = 1.32 µmol-2s-1). For NEE, the mixed simulation yields higher skill scores (r2=0.24, RMSE = 1.17 µmol-2s-1) than either vegetation type considered separately, although overall correlations remain low.

The ISBA model captures the interannual variability of GPP, ER and NEE effectively (Fig. A5), Consistently showing slightly higher values at the beginning of the growing season, much higher values at the end of the growing season, and stronger winter ecosystem respiration (ER) compared to the statistical model. Herbaceous vegetation is more sensitive to summer droughts than Sphagnum, leading to reduced GPP and ER in summer, especially during the 2022 drought, which caused a sharp decline in herbaceous GPP and ER.

Figure 3 consists of three panels, labeled (a), (b), and (c), illustrating the evolution of mean and maximum annual temperature, annual cumulative precipitation, and annual mean water table depth (diagnosed from ISBA) from 1959 to 2022, along with their respective trends and p values. Annual mean temperature (a) shows a significant (p value <0.05) increasing trend over the years. The annual mean temperature increases from approximately 7 °C in the early years to about 9 °C by 2022. Annual maximum temperature shows a similar evolution with an increase up to 8 °C from 1959 to 2022.

Annual cumulative precipitation (b) ranges from approximately 1000 to 2250 mm, while the Water Table Depth (WTD) (c) varies between -0.16 and -0.29 m over the whole period. A steady increase in the level of the water table is observed from 1967 to 1983, followed by a sharp decline, reaching its minimum in 1989. After this, the WTD rises again and stabilizes, although notable interannual fluctuations persist until the end of the period. Interestingly, the fluctuations in annual precipitation closely mirror those of the mean annual WTD, suggesting a strong correlation between these two variables. The trends of annual cumulative precipitation and annual mean WTD move in opposite directions, with precipitation increasing and the level of the water table decreasing. However, both trends are not significant (p values >0.05).

This section analyzes long term NEE dynamics and their drivers using ISBA simulations forced by the S2M reanalysis. Unless stated otherwise, all simulations and analyses presented in this section use the site vegetation distribution derived from , i.e., 70 % herbaceous plants and 30 % Sphagnum mosses.

Panel (a) of Fig. 7 shows the evolution of cumulative summer NEE (red) compared to cumulative annual NEE (blue) from 1959 to 2022. The cumulative summer NEE tends to “drive” the cumulative annual NEE almost always sharing the same sign, except in years near equilibrium (cumulative annual NEE = 0). This relationship is supported by a strong correlation between summer and annual cumulative NEE (r2=0.71, Fig. A7) and by the fact that summer NEE contributes approximately 40 % of the interannual variability of cumulative NEE (Fig. 6).

Panel (b) illustrates the evolution of the dryness index from 1959 to 2022. The trend shows an increasing pattern, with p=0.054, which is slightly above the conventional significance threshold of 0.05. As the dryness index increases, the cumulative summer NEE also increases, showing a good correlation (r2=0.6, Fig. A8). The dryness index exceeded 6 in only one year from 1959 to 1980, in five years from 1980 to 2001, and in six years from 2001 to 2022. This suggests an increase in the frequency of high dryness index episodes, indicating a rising frequency of summer droughts. Four years also stand out with particularly high dryness index values: 1989, 1994, 2003, and 2022, all of which occurred after the 1980s.

Figure 8 highlights the changes in the distribution of the annual cumulated Net Ecosystem Exchange (NEE) over 1959–2022. The black curve represents the probability density of NEE for a vegetation mix of 70 % herbaceous and 30 % Sphagnum, with shaded regions representing the 95 % confidence intervals accounting for variations in vegetation composition.

We observe that while the overall shape of the annual cumulative NEE probability density remains largely unchanged, significant differences arise depending on the vegetation type, particularly around the peak of the distributions. Sphagnum mosses amplify NEE extremes, either enhancing CO2 absorption by vegetation or increasing CO2 release to the atmosphere, resulting in a flatter distribution compared to herbaceous vegetation. In contrast, herbaceous plants have a buffering effect, with a distribution more concentrated around the main peak and a secondary, smaller peak slightly below +100 gCm-2yr-1, resulting in a bimodal pattern. These findings highlight the importance of considering different vegetation types, as they respond differently to changing environmental conditions. For completeness, the corresponding figures for GPP and ER are provided in the Appendix (Fig. A10).

The primary objective of this study was to evaluate the newly implemented Sphagnum PFT in the ISBA land surface model. Sphagnum photosynthetic activity is linked to water content in the top 10 cm of soil. While the model reproduces site scale carbon fluxes reasonably well given observational constraints, a more detailed validation of the water cycle would require eddy covariance data and multi site evaluations to assess parameter transferability. The aim was not to optimize parameters, but to test whether realistic behavior could be reproduced at a well instrumented site using literature derived values.

The mixed representation, combining Sphagnum and herbaceous PFTs, accounts for contrasting responses to soil moisture. We removed the influence of soil moisture on respiration (θ(z)) for Sphagnum, allowing the moss layer to maintain microbial activity under dry conditions, while it was retained for herbaceous layers to preserve the soil moisture sensitivity of heterotrophic respiration. Although respiration from the herbaceous component alone is not improved, retaining θ(z) is consistent with previous validations of the ISBA model for herbaceous vegetation, ensuring the parameterization remains grounded in established formulations . Importantly, the combination of PFTs captures contrasting responses to soil moisture, introducing functional diversity that likely increases the robustness of ecosystem carbon fluxes. This mechanism is reflected in the observed modest improvement of NEE on the mixed vegetation dataset, even if GPP and respiration alone do not always show large gains. These results also highlight the broader uncertainty in representing heterotrophic respiration as a function of soil moisture: classical formulations derived from mineral soils may not adequately capture responses in organic soils, as noted in other peatland modeling studies , emphasizing the need for further research on moisture/respiration parameterizations.

By combining PFTs with contrasting functional responses, the model captures compensatory dynamics across vegetation types: herbaceous layers respond strongly to moisture deficits, while Sphagnum maintains near surface moisture and microbial activity. This functional diversity improves site scale carbon flux estimates and suggests increased model robustness under variable hydrological conditions, which could be further enhanced by including interactive dynamics between Sphagnum mosses and herbaceous following the work of but also competition and coupled carbon/water processes (Lippmann et al., 2023; Heijmans et al., 2008; Wu and Blodau, 2013a; Gong et al., 2020).

The methodology developed in this study aimed to investigate the variability and evolution of carbon fluxes in the Bernadouze peatland from 1959 to 2022 using a CSM validated for the present period (2017–2022). This approach provides insight into the long-term functioning of the peatland on a century-scale timescale, which has been scarcely explored in the literature due to the lack of suitable tools. This novel methodology provides access to an unprecedented temporal scale, enabling current observations to be interpreted within a broader historical perspective.

Over the past 64 years, the Bernadouze peatland has shown marked variability in net ecosystem exchange (NEE), while overall maintaining its role as a carbon sink. This variability is strongly influenced by climatic and hydrological conditions, particularly precipitation, water table dynamics, and air temperature, as highlighted in previous research . The reconstruction of water table height, together with the development of a dryness index that integrates both air temperature and water table depth, offers a robust explanation for the observed fluctuations in carbon fluxes, as also supported by other studies . Vapor Pressure Deficit (VPD) is generally an important factor to consider, particularly for vegetation development, as it influences both GPP and plant transpiration . However, at the Bernadouze site, VPD is low and exhibits little variation (Figure A9 (a)), suggesting it has a limited effect on NEE. Furthermore, VPD is strongly correlated with temperature, which captures much of its potential influence. A recent study in Northern Hemisphere peatlands also indicates that VPD has a neutral effect on vegetation and does not necessarily induce stomatal closure in vascular plants. The humid conditions at the site, along with the presence of bryophytes, help satisfy atmospheric water demand. Overall, air temperature and water table depth remain the primary drivers explaining NEE variability.

It is also observed that, in Bernadouze, summer NEE is the dominant contributor to the annual carbon balance, though the relative influence of other seasons varies across time periods. In recent decades, transitional seasons such as spring and autumn have become increasingly significant compared to earlier years. Understanding how climate change influences NEE seasonality offers key insight into the complex dynamics of carbon fluxes and the shifting balance between source and sink processes throughout the year, as also emphasized by . The dryness index developed in this study also appears to be a good proxy for summer NEE and, consequently, for annual NEE (Figs. A8 and A7). On other peatland sites where carbon flux measurements are not available, this index could potentially serve as a preliminary source of information to estimate the carbon balance of peatlands.

Over the past 64 years, the Bernadouze peatland has experienced an increasing frequency of severe droughts, as indicated by the calculated dryness index (Figs. 7b and A3b). These events have contributed to the destabilization of the NEE balance, particularly in 2022. Despite some differences in seasonal representation, both ISBA and the statistical model by agree that from July to November 2022, conditions fell completely outside the range of interannual variability. Similar dry summers have occurred in the past, notably in 1989, 1994, and 2003, and have consistently led to significant CO2 emissions into the atmosphere. The years 1989 and 2003 are recognized as having experienced different types of drought conditions in France, ranging from multi-year precipitation deficits (1989–1990) to short, hot, and dry periods (2003) . Similarly, 1994 is also identified as a year with a hot summer, preceded by a winter precipitation deficit in Southern Europe . The dryness index effectively captures these hot and dry summers, which impact vegetation, its development, and consequently, the NEE flux.

As the growing season lengthens and GPP increases due to rising air temperatures, a compensatory effect appears to be at play. The peatland's greening and higher summer GPP fluxes currently help mitigate the impact of droughts, allowing it to remain a carbon sink. Over the 2001–2022 period, spring and autumn have played a growing role in shaping annual NEE, suggesting that these transitional seasons, along with winter, may become increasingly influential in the future, potentially counterbalancing summer carbon losses. However, the longevity of this balance remains uncertain. Some years, in our data, show a partial imbalance, indicating that the compensatory effect may not always fully buffer extreme conditions. Similar patterns have been observed in European forest ecosystems , where compensatory mechanisms were insufficient to maintain carbon balance; this provides a useful analogy for interpreting the partial signals we observe in our peatland data. While greening and seasonal compensation currently mitigate summer carbon losses, prolonged or intensified droughts in the future could challenge this balance and affect the peatland’s long term carbon sink function.

The significant increase in annual maximum temperatures about 8 °C over 64 years raises concerns about the vegetation's ability to withstand such extreme warming. Studies on potential shifts in plant composition under climate change could provide valuable insights into the future of these ecosystems . This further emphasizes the need to integrate a broader range of plant communities and their interactions into land surface models, to more accurately represent ecosystem dynamics and their role in the carbon cycle.