ORCHIDEE-PEAT is a land surface model (https://orchidee.ipsl.fr/, last access: 27 March 2026) with a module specifically developed for peatland. In general, the model comprises two main components: (1) Energy and water balance, and (2) Vegetation and soil carbon cycle. In each model grid cell defining the spatial unit of simulation, the water balance is simulated individually for each soil tile containing a different type of vegetation, while the energy balance is simulated for the whole grid cell (Xi et al., 2024). The soil hydrology is simulated using 11 vertical layers, considering incoming rainfall or snowfall, soil water evaporation, water infiltration in the soil profile every half-hour, surface runoff and drainage (also called subsurface runoff in similar models). One soil tile is specifically designated for peat with distinct soil hydrological parameters. To ensure that this peat soil tile keeps a high water content, bottom drainage is excluded and surface runoff from non-peat soil tiles is given to the peat soil tile at each time step, with a slab water layer that can be created above the peat soil surface of maximum thickness 10 cm. The vegetation and soil carbon cycle component calculates biogeochemical and biophysical variables for each plant function type (called PFT). A peatland tile can contain one or more specific PFTs that grow on it, with the possibility to prescribe a fraction of mosses, graminoids (= grasses + sedges) and shrubs.

The peatland percentage at a site is defined by the total fraction of peatland PFTs in the grid cell where the site is located. A more detailed description about the model can be found in Qiu et al. (2018, 2019). In addition, an improved routine for methane simulation for peatland was integrated into the model based on Salmon et al. (2022). This enables this study to analyze CO2 fluxes and CH4 emissions. The CO2 flux represented by Net Ecosystem Exchange (NEE) is calculated as the algebraic sum of gross primary productivity (GPP, negative sign), autotrophic respiration (AR, positive sign), and heterotrophic respiration (HR, positive sign). A simplification is that lateral export of Dissolved Organic Carbon (DOC) and erosion removal of Particulate Organic Carbon (POC) are not represented. Methane produced in the soil layers is transported to the soil surface via plant-mediated transport, ebullition, and diffusion (positive sign), and methanotrophy is simulated in each soil layer from the modeled O2 concentration.

Meteorological forcings used as input for the model (including precipitation, air temperature, air humidity, pressure, solar radiation, and windspeed) are extracted from the collection of CRUJRA forcing datasets (Friedlingstein et al., 2022). The 6 h time resolution of CRUJRA data is automatically interpolated to 30 min (default time step) by the ORCHIDEE model.

The idealized future drainage simulations will be implemented for 10 currently undisturbed peatland sites in temperate and boreal regions briefly described in Table A1 (Appendix A). The vegetation composition (graminoids, shrubs, and mosses) is taken from the literature when available; otherwise, it is estimated visually from satellite images.

The ORCHIDEE-PEAT model was first run for 100 years to reach the equilibrium state of soil thermal and hydrological conditions, followed by 10 000 years of spin-up to accumulate soil carbon with a fast module that computes only soil carbon changes from daily litter input and soil climate archived from the first simulation (Qiu et al., 2018), using repeated forcing data and a pre-industrial atmospheric CO2 concentration of 285 ppm. An additional 100 year simulation was then conducted before flux observations (Table A1) to simulate biogeochemical processes under observed atmospheric CO2 concentrations. The model was then calibrated for each site under present-day conditions following Liu et al. (2025), here calibrating the parameters controlling CH4 fluxes in addition to CO2 fluxes for sites that have CH4 flux observation. For CO2 fluxes, parameters governing photosynthesis rates, stomatal conductance, autotrophic respiration, soil organic carbon oxic decomposition rate and its sensitivity to temperature (Table B1 in Appendix B) were calibrated. For CH4, we calibrated parameters driving the methane production, oxidation, and transport (Table B1). The calibration was performed using the ORCHIDEE data assimilation system (ORCHIDAS, Peylin et al., 2016), a Bayesian statistical framework, employing a genetic algorithm to minimize a cost function between observation and simulated outcomes (Tarantola, 2005), i.e. here between observed and simulated daily NEE, and daily CH4 emission simultaneously. For sites without CH4 flux observations, only CO2 parameters can be calibrated, while parameters related to CH4 processes were assigned using the average calibrated values from all sites with available CH4 flux observations. After the calibration, to quantify the uncertainty from parameter values under current conditions affecting future predictions, we performed 10 Monte Carlo simulations by varying the temperature dependence of heterotrophic respiration (Q10) and the oxygen deficit enhancement of methanogenesis (O_2m), which were found to be most influential controlling CH4 emissions (Liu et al., 2025; Morel et al., 2019). Parameter uncertainty was then quantified as ±1 standard deviation across these 10 simulations.

Pre-drainage simulations were run for each site over a period of 100 years. We then simulate five virtual idealized drainage scenarios for each site over a subsequent 50 year future period in which the water table remains at its initial level to define a baseline and varies each day from hydrological balance simulations (no drainage), or is prescribed at 5, 10, 20, and 50 cm below the initial/baseline water level. For the 150 years of simulations, 6-hourly climate forcings of the flux-observed years are used by randomly selecting years. The same years are used for the undisturbed and drained simulations to ensure that the resulting differences in fluxes are not caused by differences in climate.

Performing drainage simulations requires a good representation of the baseline water table (WTD) before drainage. This is not straightforward in a model like ORCHIDEE where the numerical discretization of the soil into layers is coarser with increasing depth, e.g. a layer has thickness of 25 cm at 50 cm below the surface, which does not allow to position the water table in this layer accurately, i.e. the model can capture broad seasonal WTD trends, but often struggles to reproduce short-term fluctuation. Therefore, we developed a machine learning module, separate from the ORCHIDEE model, to simulate the accurate position of the water table as a function of simulated soil moisture in the soil layers, which leverages the WTD observation data that we have for 8 out of 10 sites. It is expected to reconstruct a continuous and more detailed WTD time series that aligns better with measured data, including finer temporal dynamics.

The module is trained to gap-fill the daily baseline water table observed at 8 sites and it is used to reproduce the missing water table measurements at 2 sites (CZ-Wet, DE-Akm; Table A1). Because the water table is closely related to soil moisture, our strategy is to extract the simulated soil moisture of all the layers to train a model defined by WTD =f(SM). Simulated soil moisture and observed water table time series are sampled by 7 d blocks to make up the training dataset for the model, with soil moistures as inputs and water table as output. Firstly, for a site where observed water table data is available but with missing values, its own training dataset is divided into an 80/20 ratio for training and testing, then the missing data are filled. This process is referred to as self-reconstruction. Secondly, a multi-site training dataset is created using all the sites that have water table observation with a training/testing ratio of 50/50 for four sites in Sweden and 80/20 for the remaining. The contribution of each Swedish site to the total training subset is reduced due to their proximity to each other, which potentially results in the over-representation of their relationships between water table and soil moisture when training the machine learning model. Once this multi-site model has been trained, it will take the simulated soil moisture (from ORCHIDEE-PEAT) of water-table-unknown sites as input to derive the baseline water table for these sites as output. In this study, we used a Support Vector Regression (SVR) model (Smola and Schölkopf, 2004) to learn and reconstruct the relationship between soil moisture and water table, because of its advantages such as robustness to outliers and effectiveness in high-dimensional space (Mohammed Rashid et al., 2022).

In order to simulate the effect of drainage on CO2 and CH4 fluxes, we assume that the water table is lowered by draining into ditches. We introduce into the model a new module with a ditch of which depth is the same as the desired lowered maximum water table depth (e.g., if the lowest water table under drainage is 40 cm, then ditch depth is 40 cm). In addition, runoff from other soil tiles to peat is assumed to be excluded during drainage implementation, reducing water supply to peat. Note that the simulated water table fluctuates continuously throughout the year from variable incoming rainfall, and so does the ditch water level. The difference in the hydrology simulations with and without drainage is shown in Fig. 1 and described below.

In the ORCHIDEE model, each soil column is separated into layers. We set soil layers below the water table to be saturated (θ=θs). Above the water table, in a volume of soil with horizontal dimension Δx, Δy and thickness Δz, without drainage, water flows through the soil in the vertical dimension from the infiltration of rainfall events and runoff given by other soil tiles in the same grid (Fig. 1a, blue arrows), a flux Qin entering and a flux Qout leaving the layer across the cross-section area ΔxΔy (Fig. 1c). These vertical fluxes are functions of hydraulic conductivity and diffusivity (Kv [ms-1] and Dv [m2s-1], respectively, with v for vertical) which are functions of soil moisture (θt+Δt).

With drainage, a new horizontal flux across the cross-section area ΔyΔz is added to represent water running out of the soil to the ditch (Fig. 1c, orange arrow). Assuming that ditches are placed perpendicularly to the flow of groundwater, this additional drainage flux is a function of horizontal hydraulic conductivity (Kh) and hydraulic gradient (i). The change of moisture in a time step Δt is given by: 1ΔθΔtΔxΔyΔz=(Qin-Qout)ΔxΔy-(Khi)ΔyΔz Removing ΔxΔy in both sides, Eq. () becomes: 2ΔθΔtΔz=(Qin-Qout)-(Khi)ΔzΔx with Δx now playing a role as ditch spacing. Using a numerical method with boundary conditions at top layer where Qin=(precipitation-soil evaporation) and at bottom layer where Qout=free drainage=zero, a tridiagonal matrix system is constructed to solve θt+Δt (Ducharne et al., 2018). Note that in the ORCHIDEE model, vertical water fluxes are simulated only in the downward direction – from the surface layer to lower soil layers. Upward water movement due to capillary forces is not currently represented in the model. Lateral subsurface water fluxes are also not represented.

The results of the model calibration under present conditions are shown in Fig. 2, with CO2 fluxes (Net Ecosystem Exchange, NEE) and CH4 emissions calibration for sites that have observations of both fluxes (Fig. 2a–h), and CO2-only calibration for other sites (Fig. 2i and j). The CH4 fluxes of all eight sites were well simulated, with RMSE ranging from 0.017–0.025 gCH4m-2d-1. The calibrations of NEE also showed good performance, with RMSE smallest at Sweden sites (from 0.585–0.691 gCO2m-2d-1).

Figure 3 presents the reconstruction of water table depth from soil moisture with the SVR model trained for individual sites (Fig. 3a–h, left column). The RMSE of modelled and observed water table depth ranged from 2.76 cm (FI-Sii) to 6.11 cm (US-Los) and the model captured in general the seasonal variations very well, though less so the short-term variations. At the Swedish sites, the peat is frozen between mid-October to early May, during which time the WTD is not fluctuating in reality. However, due to the lack of WTD measurements during these frozen periods, the self-reconstructed WTD from the SVR model for these times were accepted as the best available approximation. With the SVR model trained using all the sites together, SVR was less effective because it must compromise the patterns between all the sites: RMSE increased up to 6.76 cm but still captures the main seasonal variations. The water table depth reconstructed using SVR trained from all other sites and applied at the two sites where direct WTD observations were missing including CZ-Wet, DE-Akm (Fig. 3i and j, right column) was used in this study as the baseline water table depth.

Drainage reduced the water content in the soil and increased oxygen concentration (Fig. 4). With more oxygen in the soil and less moisture, heterotrophic respiration was increased. On the contrary, methane production was suppressed in aerated soil, leading to less methane emission when the water table gets deeper. Figure 4 presents the monthly average of soil water content, oxygen, NEE and CH4 emissions during the first three years after drainage, for different WTD levels. It shows that the inclusion of a ditch in the model could lower soil moisture effectively and that flux simulation responded as expected with decreased NEE uptake or NEE switching to a net CO2 source, and decreased CH4 emissions. Note that there could be a slight inconsistency between the pre- and during-drainage simulations. In the pre-drained simulation, soil moisture was calculated by the model, whereas in the drainage simulation, all the soil column below the ditch depth was forced to be saturated, which introduces a different water content of the deep layers. For instance, a deep layer that was near-saturated in the pre-drainage simulation was set to become saturated in the drainage simulation. The impact of this inconsistency on gas fluxes, however, was small because deep layers did not contribute much to changes in CH4 and CO2 fluxes (they contain less labile organic carbon) and they were already near-saturation or at full saturation before drainage. The 50 year time series of NEE and CH4 emissions are shown in Fig. C1 (Appendix C).

Figure 5 shows the average emission factor (EF) across all peatland sites for 50 years since starting the idealized drainage scenarios. The EF is defined by the average GHG flux per unit of area over a given period of time under drainage minus the flux at the same site under undisturbed conditions. Qiu et al. (2021) implemented similar calculations using the ORCHIDEE-PEAT model for estimating carbon emission factors after peatland cultivation. Their emission factors calculated over two decades following conversion of peatland to cropland was 16 gCO2m-2d-1 as a median, ranging from 5 to more than 50 gCO2m-2d-1. Our estimates are lower than their values, with a median of 2.56 gCO2m-2d-1 (95 % CI: 0.35–10.71) for 80 cm drainage (Fig. 5). Our lower median EF values than Qiu et al. (2021) are because deep soil layers remain saturated in our simulations, preserving their carbon from decomposition, while Qiu et al. (2021) assumed that the entire soil profile became drained and under-saturated when converting to cropland.

Regarding the temporal variation of emission factors, we found a decrease in CO2 EF between the first and the fifth decade (Fig. 5) from 3.82 gCO2m-2d-1 (0.36–12.76) down to 1.65 gCO2m-2d-1 (0.29–4.57) for 80 cm drainage, a similar trend than the findings of Qiu et al. (2021). As GPP and autotrophic respiration (AR) was found to be little affected during the drainage period (Fig. C2f and g), the change in emission factor over time was primarily driven by the increase of heterotrophic respiration (HR) under drainage compared to undrained conditions (Fig. C2e). HR includes one part from litter and one from soil. In the early decade, the decomposition of litter accelerated in the case of drainage, leading to the quick decrease in litter amount (Fig. C2a). Later, the reduced availability of litter for decomposition caused the heterotrophic respiration from litter in drained peatlands to approach the same levels than in undrained peatlands (Fig. C2b) where litter accumulated gradually, or decomposed at a much slower rate, and maintained a stable and low heterotrophic respiration rate. On the other hand, while SOC in undrained peatlands accumulated gradually, SOC in drained peatlands increased rapidly in the early decades due to a large input of carbon from litter but then slowed down as litter carbon input declined (Fig. C2c). Consequently, the difference in soil heterotrophic respiration (HR) between drained and undrained peatlands narrowed over time (Fig. C2d). These declining trends of both soil and litter HR in drained peatlands explain why the simulated EF decreased by time. A small decreasing trend in EF over time following drainage was determined experimentally by Truskavetskii (2014) based on soil measurements from chronosequence data. Rojstaczer and Deverel (1993) reported a trend in the subsidence of organic soil over time using periodic leveling surveys, suggesting decreased EF as well, although our model did not explicitly include subsidence and compaction effects. For CH4 emissions, on the contrary, the decrease of EF amplifies with time, going from -0.07 gCH4m-2d-1 in the first decade down to -0.10 gCH4m-2d-1 in the last decade in the case of 80 cm drainage as shown in Fig. 5 (negative EF in this Figure indicating less CH4 emissions in the drained than in the undrained).

To evaluate our drainage simulation results for GHG fluxes, we used observed changes of CO2 and CH4 fluxes collected from multiple sources through two meta-analyses of drained wetlands experiments (Huang et al., 2021b; Zou et al., 2022), the network of field flux measurements collected in Europe by Evans et al. (2021), and measurement collected in the dataset of Ecosystem Services publicly provided by the WET HORIZONS project (https://www.wethorizons.eu/resources/#database, last access: 27 March 2026). Water table depths were provided along with the fluxes in both meta-analysis studies, while in Evans et al. (2021), an “effective” water table depth was used, defined as the smaller value between the measured water table depth and the measured peat depth. The data collected by Huang et al. (2021b) provided measurements of undisturbed fluxes at control sites, as well as 77 measurements of fluxes after <1 year, 42 after 1–10 years, and 78 after >10 years of drainage. Since measurements taken within 10 years of drainage seem to dominate, we compared 10 years average values of our simulation with these data. Figure 6a and b shows the distribution of simulated and observed NEE and CH4 fluxes for different water table levels, regardless of the site and drainage level, excluding CH4 fluxes from CZ-Wet and DE-Akm, which were not calibrated during the pre-drainage period. Overall, our simulation results lie within the large range of the observed data, although our modeled CH4 emissions are in the upper range of the observed distribution. In Evans et al. (2021), a linear relationship was obtained between effective water table depth and net ecosystem production (NEP, the sum of NEE and biomass removal). Retaining only sites within boreal and temperate zones from their study (the yellow points in Fig. 6a and b), we found that our model, despite its simplified hydrology, can reproduce such a strong linear correlation between effective WTD and NEE (NEE=-0.062×WTD-1.425, R2=0.66, p<0.001). Our results (the red points in Fig. 6a and b) were in good agreement with Evans et al. (2021), yet a smaller slope between WTD and NEE was obtained in the model (NEE=-0.041×WTD-1.060, R2=0.64, p<0.001). Meta-analysis data mixing different sites, experiment duration and conditions showed a weak correlation (R2=0.04 in Huang et al., 2021b, and R2=0.18 in Zou et al., 2022), suggesting that the Evans et al. (2021) field measurements with continuous flux and WTD records at the same site may be a better benchmark of the model results.

The distributions of the changes of fluxes compared to no drainage are presented in Fig. 6c and d for different drainage levels (ΔWTD). For this comparison, we kept only samples for 5, 10, 20, 50, and 80 cm (±10 %) drainage levels in the reference data. The numbers of measurements for each drainage level differs and different drainage levels can include different peatland sites and types. The results indicate that the modeled distribution of ΔNEE falls within the observed distributions from Huang et al. (2021b) and Zou et al. (2022), although all of our simulations show positive ΔNEE values ranging from 0 to 7.70 gCO2m-2d-1, i.e., more CO2 is emitted or less CO2 absorbed after peatlands are drained, whereas the meta-analysis data exhibit either increases or decreases of NEE (-7.33 to 17.80 gCO2m-2d-1 in Huang et al., 2021b, and -11.75 to 12.26 g CO2m-2d-1 in Zou et al., 2022). For methane, all ΔCH4 values are negative in our model, ranging from -0.11 to 0 gCH4m-2d-1, i.e. less CH4 is emitted when peat is drained, while the meta-analysis data shows a negative median value of the distribution, yet with negative or positive changes of ΔCH4 (-0.16 to 0.22 gCH4m-2d-1 in Huang et al., 2021b, -1.27 to 0.40 gCH4m-2d-1 in Zou et al., 2022). These discrepancies mainly come from the fact that the meta-analysis studies took into account emission variations for different WTDs which can come from natural climate, microtopographic conditions and laboratory experiments, while our study focuses on ditch drainage conditions. For example, Huang et al. (2021b) included a comparison between a lawn with near-surface WTD and an adjacent hollow filled with water, which gave a negative ΔNEE because the hollows leached dissolved or particulate organic carbon, facilitating oxidation, thus releasing more CO2 than the lawn (Villa et al., 2019). A decrease in NEE may also result from anomalously cold weather, which can reduce respiration more significantly at the drained site than at the undrained site (Renou-Wilson et al., 2016), given that dry soil is more susceptible to air temperature fluctuations, and low temperatures inhibit soil decomposition. Additionally, inconsistencies in the data may contribute to their large range of ΔNEE, as they include different drainage levels and varying durations. Potential errors in the meta-analysis studies, such as inconsistencies in data collection, may also play a role in estimating negative ΔNEE or positive ΔCH4; for instance, our inspection of meta-analysis data found cases where emission data reported for different depths were from different locations.

In this section, we analyzed the sensitivity of GHG fluxes defined by the change of NEE or CH4 emission per cm of lowered WTD computed from our simulations, for different starting WTD levels at the same site, after 10 years of drainage. For the observations, we calculated the sensitivities from changes in GHG emissions reported by Huang et al. (2021b) and Zou et al. (2022) and changes in WTD between 5 and 80 cm. The distributions of the resulting sensitivities are shown in Fig. C3. Averaged across all sites, the modeled mean NEE sensitivity (0.07 gCO2m-2d-1cm-1; 95 %CI: 0.01 to 0.28) is close to the observation-based values derived from the data of Huang et al. (2021b; 0.12 gCO2m-2d-1cm-1; -0.23 to 0.67), but higher than the data of Zou et al. (2022), although within their 95 % confidence interval (0.01 gCO2m-2d-1cm-1; -0.37 to 0.43). The simulated methane emission sensitivity (-1.43 mgCH4m-2d-1cm-1; -6.73 to 0.0) is also comparable to the empirical values (-1.82 mgCH4m-2d-1cm-1; -15.75 to 0.17 of Huang et al., 2021b, and -1.71 mgCH4m-2d-1cm-1; -19.47 to 5.02 of Zou et al., 2022). Kwon et al. (2022) used ORCHIDEE for 100 years drainage simulations at six arctic peatland sites with WTD lowered by 5 to 50 cm and reported average sensitivities of 0.01±0.02 gCO2m-2d-1cm-1 for NEE, and of -1.10±1.10 mgCH4m-2d-1cm-1 for CH4 emissions. Our modeled NEE sensitivity is more positive and our modeled CH4 emission sensitivity is slightly more negative than Kwon et al. (2022), despite using almost the same model. This may be attributed to the fact that we did simulation across a larger range of climate while Kwon et al. (2022) only simulated drainage of arctic wet peatlands where winter respiration is absent and high rainfall and snowmelt inputs maintain relatively more stable hydrological conditions, thus their NEE and CH4 emissions are less sensitive to lowered WTD.

Instead of fitting a linear regression between all the flux changes and all the WTD changes in the simulations (for Fig. 6c and d), we calculated the sensitivity as the change of flux divided by the change of WTD (ΔfluxΔWTD) for each possible pair of flux and negative water tables changes, i.e. when WTD becomes deeper, like in Huang et al. (2021b). The “starting” water table depth in the x axis of Fig. 7 is the shallower WTD in ΔWTD used to calculate sensitivities. This starting water table level can be taken from a drained simulation, and thus differs from the “initial” water table which corresponds to undisturbed conditions. The results shown in Fig. 7 indicate that the more shallow is the starting WTD, the higher are the sensitivities of NEE and CH4 emissions fluxes when the WTD is further lowered. This is because when drainage is applied to a site with a shallow starting water table, it affects upper soil layers that were previously saturated and causes large emissions changes. Conversely, if drainage is applied to a site with a deep starting water table, upper layers that were already exposed to oxygen experience minimal change and deeper layers dry out but the SOC that they contain decomposes more slowly (they contain more slow and passive carbon with slower turnover time in the model), resulting in a smaller sensitivity of fluxes. The sensitivities of NEE is a non-linear function of starting WTD (non-linear AIC =-824< linear AIC =-820, with AIC definition in Appendix C). CH4 emissions sensitivity also depends on a non-linear manner of the starting WTD (non-linear AIC =-1935< linear AIC =-1928 for CH4 emission) and tends to saturate when drainage starts from a deep water table. The sensitivity of CH4 emissions increases from negative values to almost zero when the starting WTD is deeper than ≈40 cm.

Figure 8 shows the simulated sensitivities (ΔfluxΔWTD) of the 10 sites after 10 years of drainage, together with other model variables that could plausibly explain their values. For clarity, the sensitivities are shown in rank from the highest to the lowest for NEE (Fig. 8a). The regression between sensitivities and each of the selected model variables (Fig. 8c–h) are shown in Fig. 9.

Firstly, we analyzed whether the sensitivities depend on the magnitude of the initial fluxes, based on the hypothesis that a site with a higher baseline flux before drainage would experience more drastic changes during drainage. Among all sites, DE-Akm and CZ-Wet had the highest sensitivities of NEE (0.14 and 0.12 gCO2m-2d-1cm-1, respectively, Fig. 8a). These two sites, along with JP-BBY, also exhibited the greatest CH4 emissions sensitivities (-3.17, -2.85, and -2.37 mgCH4m-2d-1cm-1, Fig. 8b). These sites were all characterized by the largest initial CO2 sinks and CH4 emissions before drainage. Conversely, the site with the smallest pre-drainage CH4 emissions, CA-SCB, had the smallest CH4 sensitivities. SE-Srj has the smallest NEE sensitivities with the smallest initial NEE. Overall, we found that the CH4 sensitivities show a very strong linear relationship with initial CH4 emissions (R2=1, p<0.001) across sites. Similarly, the NEE sensitivities are correlated with the initial NEE, but the relationship is weaker (R2=0.71, p=0.002) (Fig. 9, top row).

Secondly, we examined if the vegetation composition of each site affects the sensitivities when the site is drained. To quantify this response, we calculated a moss index (MI) defined by the ratio moss-grass-shrubmoss+grass+shrub so that it is equal to -1 in absence of moss and to 1 with full moss cover, and found a negative correlation between MI and NEE sensitivity (R2=0.75, p<0.001), and a positive correlation between MI and CH4 emission sensitivity (R2=0.65, p=0.01) (Fig. 9a and b, last items). We identified two mechanisms in the model that explain this response of sensitivities to the moss fraction. The first mechanism is that a higher moss fraction is associated with less decomposable organic carbon. Among the three types of peatland vegetation, soil organic carbon decomposes the slowest for mosses and the fastest for shrubs, with residence times of 2 years for moss, 1 year for grass/sedges, and 200 d only for shrubs at >30°, for the active soil organic carbon pool of the model. Accordingly, we found that the sites with higher fractions of shrubs and lower fractions of mosses display higher sensitivities of NEE to drainage (Fig. 8). The methane emission sensitivities depend on vegetation cover in a different way. Peatland sites with mosses have shallower effective root depths (1–5 cm), reducing plant-mediated methane transport in the model, and the opposite is true for grasses/sedges and shrubs (root depth ≈30–50 cm). Most of the sites in this study have an initial mean WTD already below 5 cm, i.e. deeper than the depth at which mosses typically facilitate methane transport. When the WTD gets deeper, plant-mediated transport of methane by moss is thus not strongly affected in our model. Sites with more mosses (e.g. CA-SCB) therefore show smaller methane emission sensitivities. The second mechanism is that the moss fraction regulates the initial amount of SOC, which partly explains the differences in sensitivities shown in Fig. 8. Note as well that moss-dominated sites have a higher potential for oxidation of CH4 due to a symbiosis between mosses and methanotrophic bacteria (Larmola et al., 2010), but our methanotrophy module does not simulate this effect. Sites dominated by shrubs and grasses/sedges (DE-Akm, CZ-Wet) accumulate more SOC before drainage. When drainage exposes the upper peat soil layers to oxygen, accelerating soil respiration and limiting methane production, these sites with larger initial SOC pools show a larger increase in CO2 emissions and a larger decrease of CH4 emissions compared to moss-dominated sites (CA-SCB, FI-Sii).

Thirdly, we analyzed how the sensitivities depend on the initial WTD across all the sites. We found that the shallower the initial water table, the higher the sensitivities of NEE and CH4 emissions. This result is qualitatively consistent with the data shown in Fig. 7 where all sites are displayed together, but it additionally indicates that the site to site differences of sensitivities are partly explained by the variability of initial water table values (R2=0.36, p<0.001 for NEE; R2=0.54, p=0.02 for CH4 emissions) (Fig. 9a and b, middle left items). The initial WTD controls which soil layers are affected by drainage, which is in turn related to the amount and lability (active, slow, passive) of SOC exposed to oxygen, thereby influencing the sensitivities.

Finally, air temperature was also found to have a correlation with sensitivities. Sites in warmer climates tend to have more sensitive NEE (R2=0.73, p=0.002) and CH4 emission (R2=0.86, p<0.001) responses (Fig. 9a and b, middle right items). However, it remains uncertain whether these correlations are causal or reflect co-variations between temperature and other factors, given the small number of points in the regressions shown in Fig. 9.

It is impractical to run a complex, process-calibrated model like ORCHIDEE-PEAT for calculating the sensitivities of GHG fluxes to drainage over a whole region, for instance, to derive spatially variables emission factors that could be used in GHG accounting studies (IPCC Guidelines Tier 3, Eggleston, 2006) and decision support tools or meta-models like (e.g. FAO, 2021). Therefore, following the analysis of the driving factors of the sensitivities in the previous section (Fig. 9), we developed an emulator of the model sensitivities based on multilinear regressions. Given the limited number of sites and the acknowledged limitations of the underlying process-based model, the emulator presented here is intended as an exploratory, proof-of-concept approach rather than a fully validated predictive tool. To do so, we first fitted all possible regressions of the simulated sensitivities (SNEE and SCH4) against the six variables as shown in Fig. 9, including initial NEE (NEEinit), initial CH4 emission (CH4init), initial water table (WTDinit in cm), soil organic carbon content (SOC in kgCm-2), air temperature (T in K), and moss index (MI). Then, the best regression model was selected based on the minimum corrected Akaike information criterion (AICc, Appendix C), constrained by the cutoff of variance inflation factor (VIF) >5 to avoid multicollinearity among predictors (Fig. C4): 3SNEE=0.022-0.013×NEEinit-0.025×MI(R2=0.91,p≤0.001)4SCH4=0.045-14.426×CH4init(R2=0.998,p≤0.001)

To evaluate the robustness of the selected model, a leave-one-out cross-validation was performed. The model was re-fitted multiple times, each time excluding one site from the data, and the performance metrics (R2 and p-value) were recorded for the withheld site (Table C1). The small variation in R2 (+SD of R2) and p-value (+SD of p-value) among all re-fitted models suggests that the selected regression relationships are internally consistent across sites, within the limits of the available sample.

Furthermore, we calculated the product of the regression coefficients for each variable in Eq. (3) by their standard deviations between sites to quantify the “effect” of each variable on the sensitivity (Jung et al., 2017). For the sensitivity of NEE, the effect of NEEinit is -0.012, comparable to that of MI, at -0.011.

To assess the combined climate forcing of CO2 and CH4 flux changes due to drainage, we used the Global Warming Potential (GWP) metric to convert methane emissions into CO2-equivalents (IPCC AR6). Figure 10 shows the resulting CO2-equivalent flux changes averaged across the 10 sites for 20 and 50 year periods of drainage as a function of the WTD drawdown. We used GWP100=27 for the 50 year period and GWP20=79.7 (IPCC, 2023) shown in different colors in Fig. 10. At the site scale, the simulated CO2-equivalent flux changes exhibit substantial uncertainty, as reflected by the error bars in Fig. 10. These uncertainties arise from both parameter uncertainty and uncertainty associated with vegetation cover changes induced by drainage, the latter quantified as the range across alternative vegetation scenarios explored in Appendix D.

Averaging our model results for all 10 sites at 80 cm drainage leads to a slight warming of 0.05 gCO2-eq m-2d-1 using GWP100 after 50 years of drainage, and to a cooling of -2.46 gCO2-eq m-2d-1 using GWP20 after 20 years. In comparison, from the data collected by meta-analysis studies (Huang et al., 2021b; Zou et al., 2022), retaining only sites where both ΔNEE and ΔCH4 flux are available, we calculated for all sites a mean cooling effect of -5.46 gCO2-eq m-2d-1 with a huge spread (95 % CI: -73.06 to 17.64) using GWP20, and a mean cooling of -0.06 gCO2-eq m-2d-1 (-25.60 to 11.82) using GWP100, with the occurrence of cooling and warming effects being almost equal among all the sites (Fig. C5). This result is rather consistent with our model simulations. In contrast, pooling all sites and WTD levels together, thus combining ΔCO2 and ΔCH4 from different locations in an inconsistent manner, from the data of Zou et al. (2022) excluding their reported N2O emissions changes, a net cooling effect of -0.30 gCO2-eq m-2d-1 using GWP100 was estimated. Similarly, using a GWP100 value of 25, Huang et al. (2021b) estimated across their sites a net warming effect of 0.033 gCO2-eq m-2h-1 (0.009 to 0.057). This suggests that the averaging of sites from meta-analysis data can lead to assessing either a cooling or a warming, depending on whether only sites with both measured ΔCO2 and ΔCH4 are used, or all sites are used.

Our model simulations also show that different drainage periods and the choice of a different GWP time horizons lead to distinct warming or cooling effects (Fig. 10). For 9 sites out of 10 (i.e. except CA-SCB), however, we simulated a larger cooling on a 20 year horizon compared to a 100 year horizon, due to the stronger radiative forcing impact of reduced CH4 emissions in the short term. In addition, most of the sites show a larger radiative forcing change when the WTD is deeper, specifically a higher cooling from reduced CH4 emissions on a 20 year horizon. Secondly, the magnitude and the sign of changes, whether warming or cooling, varies significantly between sites. CA-SCB is the only site having warming effects (from 0.4 to 1.5 gCO2-eq m-2d-1) regardless of time scales and drainage levels. On the other hand, CZ-Wet, DE-Akm, JP-BBY always have cooling effects (down to -6.5 gCO2-eq m-2d-1) because their CH4 emissions are reduced more than their CO2 emissions are increased. Other sites have both warming and cooling effects depending on time scale considered and drainage level. Neutral effects on climate were observed in some cases, primarily when considering 50 years of drainage (green line), such as FI-Sii with a 20 cm drainage and US-Los at a 80 cm drainage level.