The simulation framework LandscapeDNDC has been developed to allow a flexible problem-specific model composition for the simulation of the water, carbon and nitrogen cycles in cropland, grassland and forest ecosystems . In this study, the microclimate within the forests is simulated using the sub-model CanopyECM according to . Vegetation related processes are modeled with the Physiological SImulation Model (PSIM) sub-model . Soil biogeochemical processes were represented with the MeTrx sub-model , while soil hydrological conditions are represented with the Ecosystem Hydrology (EcHy) sub-model .

PSIM enables the representation of a multiple species stand, where individual tree species can have their own characteristics and dimensions, while also having interactions with each other . Including multiple species in the simulation allows to consider the contributions of dominant trees, understorey and ground vegetation to carbon and water fluxes. In particular, the importance of understory in boreal forests has been highlighted before and is thus necessary to be included in the model .

The soil sub-model MeTrx accounts for carbon and nitrogen pools and fluxes and their responses to land use and forest management. Most relevant processes considered by MeTrx for this study are humification, mineralization, nitrification and denitrification as well as CH4 production and consumption . In MeTrx, various litter (Solutes, Cellulose and Lignin) and humus (Labile, Recalcitrant young and old) pools are differentiated according to their chemical structure and carbon to nitrogen ratio, respectively. These organic materials are decomposed by microbial pools in the nitrification and denitrification processes. Besides pool size and organic matter properties, processes depend on the availability of oxygen, pH, clay content, soil temperature and soil moisture. MeTrx also covers CH4 production as a final step at the end of decomposition under anaerobic conditions as well as CH4 transport from air to soil and CH4 oxidation. Oxygen content is dynamically calculated for each soil layer. Detailed descriptions of these processes can be found in , , and . Soil water dynamics are calculated by the EcHy module, which predicts a dynamic groundwater table (zgw) depending on the simulated water balance. In its conceptual design, EcHy is a one-dimensional (1-D) vertical soil column model. When applying its original version in this study, the simulation domain became waterlogged because precipitation exceeded evapotranspiration and percolation through the lower soil boundary. In reality, however, the site is drained by ditches that induce a lateral water flow component and counteract complete soil saturation. To account for this, a reference water table (Zgw) was introduced, which represents the depth below which no significant lateral flow occurs (e.g. the ditch depth), and the soil is assumed to be fully saturated. In addition, a parameter Ψ representing lateral groundwater flow velocity and ditch distance was implemented. This parameter determines the rate of lateral water loss whenever the simulated dynamic water table depth zgw is above Zgw. Together, these modifications introduce a lateral (2-D) flow component into an otherwise 1-D framework, enabling EcHy to better reflect local site conditions. Lateral groundwater movement q is given by, 1q=max⁡(0.0,|Zgw-zgw|)×Ψ×Ks zgw: depth of groundwater in m, Zgw: depth of reference groundwater in m, Ψ: Model parameter representing lateral groundwater gradient m-1, Ks: the saturated hydraulic conductivity of the last soil layer mmin-1.

In the model, soil organic matter is represented by three conceptual humus pools (labile, recalcitrant young, recalcitrant old) differing in turnover time and C:N ratio. The C:N ratio of the young recalcitrant pool is 1.5 times the bulk soil C:N ratio, which is the model's default parametrization for agricultural (mineral) soils. Peat soils in this study are characterized by systematically higher bulk C:N ratios than mineral soils. For peat, we interpreted the young recalcitrant pool as the main peat pool and therefore aimed to reduce the initial relative share of the old recalcitrant pool compared to mineral soils. To achieve this while conserving bulk soil C and N, we assigned a lower C:N ratio to the old recalcitrant pool than to bulk soil, which reduces the amount of C that is allocated to old recalcitrant pool. The C:N ratio of the recalcitrant old pool (CNrec) is scaled as a function of the bulk soil C:N ratio (CNbulk): 2CNrec=(0.9-max⁡(0,0.8-8×CNbulk))×CNbulk

When CNbulk very large: 3CNrec=0.1×CNbulk

When CNbulk≤10: 4CNrec=0.9×CNbulk

This formulation therefore reduces the size of the old recalcitrant humus pool with increasing soil C:N ratio. Changes to the model code can be found from .

Lettosuo site, located in southern Finland (60°38′31′′ N, 23°57′35′′ E), covers a total area of 65 ha (Fig. A1). The site, extensively drained in 1969 using ditches that were 45 m apart, is classified as a nutrient-rich Vaccinium myrtillus type II forest (MtkgII, ). Lettosuo was originally a sparsely treed, mesotrophic pine-birch sedge fen rich in herbs . Fertilizers were used after drainage to help with the growth of existing pine (Pinus Sylvestris) trees. In the 1970s and later, some thinning was done to ensure better growth of pine, but those thinning information are unavailable. Over time Lettosuo became a mixed forest site, where pine and birch (Betula Pendula) formed the overstory and spruce (Picea Abies) formed the understory canopy. The forest floor consisted of various herbs, shrubs and graminoids .

The site was divided into three sections according to the applied management methods: a control stand where no management action had taken place; a Continuous Cover Forestry (CCF) stand where all the pine trees (75 % of the original total biomass) were harvested in March 2016 ; and a Rotational Forestry (RF) stand where all the trees were removed by clear-cutting, also in March 2016 .

In RF, the application of heavy harvesting machinery destroyed the forest floor vegetation, while in CCF, forest floor vegetation was only affected on logging trails. In RF, the harvest residues were left at the felling area, while in CCF, the harvest residues were placed on the logging trails. At RF, ditches were maintained after clear-cut and peat extracted from the ditches was used to form mounds of peat soil where new spruce seedlings were planted in 2017. Most of the spruce seedlings died during the drought in 2018 . Later naturally occurring birch seedlings became the dominant species at RF stand.

In the simulations, the vertical soil profile was divided into several layers of increasing thickness with increasing layer depth (Table 1). The carbon and nitrogen content, bulk density and pH of the soil layers were taken from the measured Lettosuo data previously reported by and . These soil variables and their prescribed initial levels are given in Table 1.

The simulations initialized with pristine peatland condition from 1969. During the first three simulation years fluctuations in litter and humus pools were stabilized. During this period, the Zgw was set to 0.01 m depth to stabilize the soil carbon pools for pristine peatland condition. In the case of pristine peatland simulation, to study the soil carbon changes, the Zgw was kept at 0.01 m until 2021. Otherwise, for forest stand simulations, Zgw was set to 0.62 m after initial three years to represent drained peatland conditions. The model then simulated WT dynamically with the modification described in Sect. 2.1.

The initialization of tree species and forest floor vegetation is presented in Table 2. The simulations began in 1969 with pine trees and the forest floor vegetation. Birch and spruce were added in 1971 and 1972, respectively. Management events (selective harvest and clear-cut) took place in 2016. Harvest residues were left in the simulation domain and stemwood biomass was removed after management events. All simulations were run until end of 2021. All the modules were run with a sub-daily time step (30 min) and sub-daily and daily simulation outputs were used for various investigations.

The sensitivity of NEE to soil decomposition was tested with separate one-year mature forest simulations for a range of soil parameters (Table A1) described in Sect. 2.6 and Appendix A2. For this study species parameters reported in , a study which simulated a mixed forest stand consisting of pine, spruce and birch trees on mineral soil in southern Finland, were used and adjusted to run the sensitivity test simulations and full forest simulations (Table A2). These parameters included e.g. photosynthesis related VCMAX (Maximum RubP saturated rate of carboxylation at 25 °C for sun leaves), KM20 (Maintenance coefficient at reference temperature), SLAMAX (Specific leaf area in the shade) and SLAMIN (Specific leaf area in the full light). Several sensitive soil parameters related to the decomposition of different humus and litter pools were later adjusted during full forest simulations (Table A3).

The correlation coefficient (r) and Nash–Sutcliffe efficiency (NSE) were calculated from the daily aggregate of the available measurements and corresponding modeled data. NSE values can vary from -∞ to 1. Model output that produces an NSE value of 1 is described as the perfect simulation, thus NSE values closer to 1 are desirable.

NSE is defined as, 5NSE=1-∑i=1nQoi-Qmi2∑i=1nQoi-Qo‾2

Where, Qo represents the observed data at time i, Qm represents the modeled data at time i, Qo‾ is the mean of the observed data. n is the total number of observations.

We further used Root Mean Square Error (RMSE) to quantify the differences between the model and the observations for NEE and WT. In this analysis, seasons are defined as December–January (winter), March–May (spring), June–August (summer) and September–November (autumn).

Since a single simulation setup was used for all three forest stand simulations before the harvesting events, both measured and simulated data will be referred to as pre-harvest data (2010–2015). calculated the annual CO2 balances (NEE, GPP and TER) for 2016 from April 2016 to March 2017 and the modeled estimates for 2016 were also calculated similarly in the CCFpostharvest and RFpostharvest. To show the sensitivity of the simulated CH4 flux with WT after clear-cut harvest, the groundwater lateral gradient parameter was changed (28.7–38.7 m-1, Table A3) in the simulation to reproduce the lowest and highest observed WT at RF (Sect. 3.3).

The effect of drainage on the simulated soil carbon (SC) storage was compared to the SC loss reported by . The influence of harvesting, under different management methods, on SC dynamics was assessed by quantifying changes in SC following harvest. This included investigating the partition of different litter (solutes, cellulose and lignin) and humus (labile, recalcitrant young and old) pools. The evolution of the total carbon (soil carbon + carbon in vegetation) over the whole simulation period was also studied (Sect. 3.5). A pristine peatland simulation, without any drainage or harvesting, served as a reference. Data and Python codes used in this study can be found from .

Sensitivity analyses of soil parameters were conducted separately using mature forest stands in one-year simulations. The analyses were performed with the Python library SPOTPY and its Fourier Amplitude Sensitivity Test (FAST) algorithm . The sensitivity analyses were carried out sequentially, with the best-performing parameter values from each stage retained for subsequent analyses. The soil parameters were divided into three sets, each containing nine parameters, and the parameter ranges tested are presented in Table A1. The three parameter sets represented (1) decomposition rate constants for different soil pools, (2) factors describing the dependence of decomposition on environmental conditions and (3) humification factors for different soil pools. Rather than varying all parameters simultaneously, only one set was varied at a time, while the remaining parameters were fixed at either their default values or values determined from the preceding sensitivity analysis. The number of simulations performed for each parameter set was considered sufficient to provide robust sensitivity estimates. During all sensitivity analyses, species parameters were held constant using the values presented in Table A2.

For the first analysis, the nine decomposition rate related soil parameters in Set 1 were evaluated using 8649 model runs generated by FAST algorithm . Parameter values in Set 1 were sampled automatically within their predefined ranges, while all parameters in Sets 2 and 3 were fixed at their default values (Table A1). Following the first sensitivity analysis, the modeled NEE from all runs was compared against measured NEE to determine the parameter values for Set 1 to be used in subsequent analyses. The parameter values from the run that yielded the highest NSE were extracted (Table A1). These values were then fixed for the Set 1 parameters during the sensitivity test runs for Set 2. During the second sensitivity analysis, parameter values in Set 2 were again sampled by FAST algorithm and the parameters in Set 3 had default values. As in the first analysis, the parameter values from the run yielding the highest NSE were extracted in second analysis and subsequently fixed for the parameters in Set 2 when the sensitivity test runs for Set 3 was performed. The third sensitivity analysis was conducted by varying the nine parameters in Set 3 while keeping the selected parameter values from Sets 1 and 2 fixed. The parameter values corresponding to the highest NSE obtained in the third analysis are also presented in Table A1. The results of these analyses are presented in the Appendix (Table A1, Figs. A2–A4 and S1–S3 in the Supplement).

The sensitivity analysis of the first parameter set indicated that the decomposition rate constant for the recalcitrant young humus pool was the most influential parameter affecting modeled NEE (Fig. A2). Other highly sensitive parameters in this set were associated with the decomposition of labile humus, recalcitrant old humus, raw litter, and the effect of litter lignin concentration on decomposition. In the second parameter set, the most sensitive parameters were those describing the effects of temperature and pH on decomposition, as well as the reduction in decomposition under anaerobic conditions (Fig. A3). For the third parameter set, the most influential parameters were related to the C:N ratio of the old humus pool relative to the overall soil C:N ratio. Other sensitive parameters were the humification rate of recalcitrant young humus, dissolved organic carbon, active organic matter, and solutes (Fig. A4).

The sensitivity analyses were conducted solely to evaluate parameter sensitivity, and the resulting parameter values were not directly adopted for the full forest simulations. Instead, long-term simulations from seedling establishment to stand maturity were performed separately using a limited number of selected soil parameters, whose values were chosen to reproduce the observed tree species composition at the study sites (Table A3). The species-specific parameter values used in these simulations are provided in Table A2.

Particular attention was given to the calibration of the full forest simulations because relatively small changes in soil parameter values could alter competitive interactions among tree species and thus lead to forest compositions that differed from those observed in the field. As the primary objective of the study was to reproduce the observed forest structure, parameter values were selected to maintain realistic stand development and species composition. In contrast to the sensitivity analyses, where model performance was assessed solely using the NSE of simulated NEE, the evaluation of the full forest simulations considered multiple variables. These included NEE, soil moisture, WT, LAI, CH4 fluxes, and forest structure, all of which were compared with available measurements.

A comprehensive optimization procedure incorporating all evaluated variables would have been desirable; however, the computational demands of the full forest simulations were substantial. Consequently, parameter selection for the full simulations was based on an overall assessment of model performance across the observed ecosystem characteristics rather than on formal multi-objective optimization.

In the forest stand simulations, the model accurately reproduced the observed range of species-specific tree numbers at stand age 45 (Table 3). The height of the individual species was also close to the observed values. The modeled pine and spruce volumes were close to the observed volume ranges, while birch had 27 % less volume. The model successfully reproduced the reported pre-harvest LAI of the different tree species in 2014 (Fig. 1a). Forest floor vegetation had modeled LAI of 0.94 m2m-2 in August 2014, which was similar to the observed LAI of 1 m2m-2 reported by . The same literature reported LAI for above ground canopy (pine + spruce + birch) to be 4.66 m2m-2 and the corresponding modeled LAI was 4.68 m2m-2 (August 2014; Table 3).

Modeled pine and spruce LAI were diminishing from 2016 to 2021 for control simulation (Fig. 1b), which was not seen in the prior six years (Fig. S4 in the Supplement). The sum of pine and birch modeled LAI was closest to the satellite-based LAI (Fig. 1b), and this could be realistic since pine and a small percentage of birch were forming the upper-story canopy in the control stand. The gradual decline in modeled LAI from 2016 to 2021, evident in control simulation, was not seen in the CCF simulation. In fact spruce LAI continued to increase until 2021, likely a result of reduced competition between species owing to the removal of pine during harvest (Fig. 1c). Spruce LAI during summer months was closer to satellite-based LAI, while the sum of spruce and birch modeled LAI was always larger than the satellite-based LAI (Fig. 1c). The modeled LAI for birch and forest floor remained consistent from 2016 to 2021 for both control and CCF. After clear-cutting in RF, the modeled LAI was almost double the satellite-based LAI in 2016, whereas the discrepancy between the two decreased progressively after 2018 (Fig. 1d). Simulated birch and forest floor LAI were closest to satellite-based LAI in 2019–2021 (Fig. 1d).

The model reproduced soil moisture dynamics well during the growing seasons of 2015–2017 (May–September), but deviations from the measurements were evident in 2018 (Fig. 2). The model captured the dynamics of 10 cm soil moisture for 2016 and 2017 best. However, the model slightly underestimated soil moisture at a depth of 10 cm in 2015 and overestimated it in 2018 relative to the measurements. At a depth of 20 cm, agreement between simulated and measured soil moisture was strongest during 2015–2017, whereas the model overestimated soil moisture from July 2018 onward.

Simulated WT during the pre-harvest period (2010–2015) was in good agreement with the observations at the CCF stand (Fig. 3a). Deviations between the simulated and measured mean WT were most pronounced during the summer of 2016 and 2017 (see also Fig. S5 in the Supplement). From 2018 onward the simulated WT in CCF stand was again similar to the observed WT. The model also predicted the control stand WT within the observed range and reflected the observed dynamics (Fig. 3b). Deviations between simulated and observed WT were most pronounced in the control stand during the autumn of 2018 and 2019 (see also Fig. S6 in the Supplement). At the RF stand, the simulated WT closely followed the observed temporal dynamics and showed relatively small deviations from the measurements (Fig. 3c; see also Fig. S7 in the Supplement). Consistent with the observations, the model predicted a rise in WT after clear-cutting at the RF stand as a result of reduced transpiration following biomass removal. In contrast, the effect of selective harvesting on WT at the CCF stand was minor (Fig. 4), and the differences between simulated and observed WT following harvest, which were evident in 2016 and 2017, largely disappeared in later years.

Manual chamber measurements indicated that the RF stand became a CH4 source during the summer months following clear-cutting, whereas the initial RF simulation suggested that the site remained a CH4 sink during the same period. Although temporal and spatial variability in CH4 fluxes can be attributed to the dynamics of WT, soil temperature and vegetation communities and in dry conditions to the wind speed and soil temperature , only the role of WT was investigated further in this study. The model results showed that CH4 fluxes were highly sensitive to WT variations after clear-cutting. To examine this sensitivity, two alternative WT scenarios were simulated by modifying the groundwater lateral gradient parameter (Ψ) while keeping the reference WT (Zgw) unchanged. Altering Ψ had little effect on simulated WT before clear-cutting but resulted in distinct WT dynamics after harvest. No additional modification to the model setup were required to produce this shift in WT. The two WT scenarios generated in the CH4 sensitivity analysis were also consistent with the lowest and highest WT values observed at the RF stand (Fig. 5). Simulation with the original value (Ψ=28.7 m-1, Table A3), produced the lowest WT, led to a higher sink of CH4. While the modified parameter value (Ψ=38.7 m-1) produced a higher WT and resulted in smaller sink of CH4. In the high-WT simulation, the site even acted as a CH4 source on some occasions. In both RF simulations, modeled CH4 fluxes showed only minor difference before harvest and represented a weaker CH4 sink than that observed in the measurements (Fig. 5).

For both control and CCF stands, modeled CH4 fluxes fell within the range of the automatic chamber measurements (Fig. 6). The measurements indicated that both stands functioned primarily as CH4 sinks over the 3.5 year measurement period. At the seasonal scale, the model generally simulated CH4 uptake at the upper end of the observed range during spring and summer, while underestimating uptake during autumn and early winter. However, unlike in the other years, the tendency of the model to underestimate CH4 uptake during late summer was less pronounced in 2016 for both stands. At the site, reported that CH4 fluxes were influenced by soil temperature at the beginning of summer and by WT from the mid-July.

The model reproduced pre-harvest NEE more accurately (r=0.88, NSE=0.75, slope=0.92) compared to NEE in either the CCFPostharvest and RFPostharvest simulations (Fig. 7a–c). Performance for modeled GPP was similar for both pre-harvest and CCFpostharvest, while RFpostharvest had a lower NSE (0.75) and higher slope (1.26) to the fitted line (Fig. 7d–f). A similar comparison for TER showed that the modeled TER performed well in all three simulations (r=0.94–0.96, NSE=0.87–0.92, Fig. 7g–i).

The modeled NEE improved after WT calibration compared with the simulation using the reference WT (Fig. S8a in the Supplement). Overall, the calibrated modeled NEE agreed well with the EC based NEE in both magnitude and temporal dynamics, although some minor discrepancies were evident during pre-harvest period. The highest RMSE for modeled NEE (9.5gCO2m-2d-1) was in summer 2010 (Fig. S9 in the Supplement). Relatively high RMSE values were also observed during spring, summer and autumn of 2014. The model reproduced the seasonal dynamics of both GPP and TER similarly to the EC-based estimates (Fig. S8b and c in the Supplement, respectively). The largest discrepancy in modeled GPP was observed in summer 2014, when the RMSE was 17.1gCO2m-2d-1 (Fig. S9 in the Supplement). TER was generally simulated well, with the exception of pronounced respiration peaks during the summers of 2010 and 2013. In addition, the model produced slightly higher winter respiration than the EC-based estimates in early 2013 (Fig. S8c in the Supplement). The highest RMSE values for modeled TER were observed during the summers of 2010 and 2014 (Fig. S9 in the Supplement).

Following the selective harvest in 2016, the model reproduced NEE well in 2017, 2020, and 2021. A similar level of agreement was observed during late summer and autumn of 2018 and 2019, although the model simulated greater uptake during spring and early summer in those years (Fig. S10a in the Supplement). In 2016, notable discrepancies in NEE occurred during summer and autumn, coinciding with an overestimation of GPP and an underestimation of TER by the model (Fig. S10b and c in the Supplement, respectively). During spring and early summer of 2018 and 2019, both GPP and TER were overestimated. Summer RMSE values for modeled NEE ranged from approximately 6–8gCO2m-2d-1, whereas RMSE values in the other seasons were generally below 6gCO2m-2d-1 and most often below 4gCO2m-2d-1 (Fig. S11 in the Supplement). The highest RMSE for modeled GPP was observed during summer 2018, while the highest RMSE for modeled TER was observed during summer 2019.

After clear-cutting, modeled NEE captured the gradual transition from a strong daily net CO2 source during the first three years after harvest to a weaker source, and occasionally even a sink during the summer season, in the following three years (Fig. S12a in the Supplement). Agreement between modeled NEE and the measurements was highest during 2019–2021 and, to a lesser extent, during autumn 2018. From May to August in 2016 and 2017, modeled NEE was underestimated, indicating a smaller net CO2 source than observed. Modeled GPP agreed well with EC-based GPP during 2016–2019, whereas the model slightly overestimated GPP during late spring and early summer of 2020 and 2021 (Fig. S12b in the Supplement). Similarly modeled TER was lower than EC-based TER during the summers of 2016 and 2017 but was overestimated during late spring and early summer of 2020 and 2021. In contrast, TER in 2018 and 2019 was reproduced well by the model (Fig. S12c in the Supplement). The highest RMSE for modeled GPP and TER was observed in 2021 (Fig. S13 in the Supplement).

On the annual scale, modeled NEE was generally consistent with the EC-based NEE under both pre-harvest and post-harvest conditions, with a few notable exceptions. During pre-harvest period, the largest discrepancy occurred in 2014, followed by 2012 (Fig. 8a), Whereas in both CCF and RF the greatest differences were observed during the first post-harvest year (April 2016–March 2017; Fig. 8b and c). Following selective harvesting modeled NEE indicated that the CCF stand remained a net CO2 source for the first three years and returned to a net CO2 sink from the fourth year onward, consistent with EC-based NEE. In the RF stand, both modeled NEE and EC-based NEE indicated that the ecosystem was a net CO2 source throughout the six post-harvest years following clear-cutting, although annual CO2 emissions decreased over time.

During the pre-harvest period, modeled GPP was generally similar in magnitude to EC-based GPP, with the largest discrepancy occurring in 2014 when modeled GPP was 35 % higher (Fig. 8d). In CCFpostharvest, modeled GPP followed a similar temporal pattern to EC-based GPP, although the largest difference was observed in 2018, when modeled GPP exceeded EC-based GPP by 44 % (Fig. 8e). In RFpostharvest, modeled GPP ranged from 21 % lower than EC-based GPP in 2018 to 24 % higher in 2021 (Fig. 8f).

During the pre-harvest period, annual modeled TER was similar to EC-based TER (Fig. 8g). In CCFpostharvest, modeled TER was overestimated by 28 % in 2018 and by 26 % in 2019, while the differences in other years were negligible (Fig. 8h). In RFpostharvest, modeled TER ranged from 16 % lower than EC-based TER in 2017 to 22 % higher in 2021 (Fig. 8i).

Following drainage, simulated SC storage showed a continuous decline until the implementation of the management events (Fig. 9). In contrast, the pristine, non-drained peatland exhibited a slight accumulation of SC under the assumption that ground vegetation covered 40 % of the total area. The development of SC after the harvest varied among the managements and depended on the management method. In control, where no trees were removed, SC continued to decrease until the end of the simulation. In CCF, selective harvesting led to a temporary increase in SC, likely driven by carbon inputs from harvest residues. Thereafter, SC began to decline, although it remained higher than in the control throughout the simulation period. In RF, complete tree removal resulted in a temporary increase in SC, again driven by carbon inputs from harvest residues. Although SC began to decrease after several years, it remained higher than in both the control and CCF simulations. Annual SC loss for 2010–2015 was 233 gCm-2yr-1. SC loss for the period of March 2016–2021 was 221 gCm-2yr-1 in control, 95 gCm-2yr-1 in CCF, and SC gain 79 gCm-2yr-1 in RF.

The small difference between total carbon (TC) and SC in the pristine peatland was due to the relatively low carbon storage in vegetation (Fig. 9). Following drainage, TC declined across all forestry simulation scenarios, as soil carbon losses were greater than the carbon sequestered by vegetation. Continued forest growth led to greater carbon accumulation in tree biomass, resulting in a stabilization of TC between 2000–2010 and an increase thereafter as carbon uptake by vegetation exceeded soil carbon loss. In control, TC continued to increase until the end of the simulation. However, in CCF, TC decreased for a few years after the selective harvest and once the remaining vegetation rebounded TC started to increase again. In contrast, in RF, the accumulation of carbon by the forest stopped after the mature forest had been removed. Furthermore, carbon accumulation in the newly established stand was insufficient to compensate for soil carbon losses, leading to a continued decrease in TC.

The litter pool comprising soluble compounds (solutes), which has a rapid turnover rate, received a large input immediately after harvest but declined rapidly thereafter (Fig. 10a). Following this initial increase, solute levels in RF were even lower than those in the control and CCF during 2016–2019, as the remaining vegetation contributed little fresh litter input. However, solute inputs increased again from the fourth year after clear-cutting as the regenerating birch stand began contributing litter. In both CCF and RF, cellulose and lignin pools dominated total litter storage after harvest, whereas the contribution of solutes remained relatively small (Fig. 10b and c). The cellulose pool reached its maximum in the third year after harvest in RF and the second year in CCF before beginning to decline. In contrast, the lignin pool remained elevated during the first three post-harvest years and began to decline in the fourth year after harvest.

In the new model version, a larger fraction of carbon entering the humus pools was allocated to the recalcitrant young pool (Fig. 10e) than in the previous version (Fig. S15 in the Supplement). Consequently, heterotrophic respiration increased in the new model version (Fig. S16 in the Supplement) because the larger recalcitrant young pool decomposed more rapidly. Following management, inputs from decomposing harvest residues also caused an initial increase in the storage of the fast-decomposing labile humus pool (Fig. 10d). This effect was stronger in RF than in CCF because of the greater harvest-residue input. The labile humus pool reached its peak earlier in RF than in CCF. Decomposed harvest residues also contributed to an increase in the recalcitrant young humus pool (Fig. 10e), whereas only minor effects were observed in the recalcitrant old humus pool after harvest (Fig. 10f).