Digital twin forests: how a carbon model is sharpening national climate accounts in Croatia and Slovakia

Every year, countries that signed the Paris Agreement must account for the greenhouse gases their landscapes absorb and emit, and few parts of that ledger are as consequential — or as incomplete — as the forest sector. A new study published in Regional Environmental Change shows how a process-based ecosystem model can close one of the most stubborn gaps in national carbon accounting, using Croatia and Slovakia as a real-world proving ground. The research team, led by Maša Zorana Ostrogović Sever of the Croatian Forest Research Institute together with Laura Dobor and Katarína Merganičová of the Czech University of Life Sciences Prague and Technical University Zvolen, built a fully reproducible workflow around the biogeochemical model Biome-BGCMuSo v6.2, and used it to estimate carbon stock changes in forest pools that most national inventories still cannot measure directly: deadwood, litter, and soil.

The problem the researchers set out to solve is structural rather than technical. Under the Land Use, Land-Use Change and Forestry (LULUCF) sector of the greenhouse gas inventory, countries must report net carbon dioxide emissions and removals for the subcategory of forest land remaining forest land. Aboveground biomass is generally well covered, because national forest inventories repeatedly measure tree growth. But deadwood, litter, and soil organic carbon are rarely measured systematically at intervals that allow carbon stock change factors to be derived empirically. Without those factors, the reported national net sink is methodologically incomplete — a gap that matters more as Europe’s forest carbon sink shows signs of saturation and faces mounting pressure from heat, drought, and disturbances.

Process-based models offer a way forward that differs fundamentally from the empirical growth-curve approaches many inventories rely on. Instead of extrapolating from inventory data alone, Biome-BGCMuSo simulates the daily cycling of carbon, nitrogen, and water through vegetation, litter, and soil, representing photosynthesis, respiration, carbon allocation, senescence, mortality, and decomposition mechanistically. The MuSo extension replaces the single-layer soil of the original Biome-BGC with a ten-layer soil profile, and the model can capture acclimatization to temperature, drought legacy effects on carbohydrate stores, and stomatal conductance that responds to atmospheric CO2. In this study the model was driven by the ERA5-Land meteorological reanalysis for 1951 through 2023, with the time series extended back to 1850 by random resampling so that even the oldest simulated stands, 140 years old, could be grown from bare ground through a spin-up phase to dynamic equilibrium.

The heart of the workflow is a stratification scheme that translates a whole country into a manageable set of representative stands. Environmental strata were built by crossing eight bioclimatic zones — from Mediterranean and sub-Mediterranean to lowland dry and wet, submontane, montane, subalpine, and azonal — with two soil types and two soil depths, yielding 18 combinations across the two countries. Forest strata combined six forest types (oaks, beech, pines, fir, spruce, and coppice), sixteen business-as-usual management regimes, and up to seven age classes, producing 54 strata for Croatia and 56 for Slovakia. Crossing the two yielded 486 modeling strata in Croatia and 616 in Slovakia, each simulated separately and then upscaled to the national level by multiplying results by the area each stratum represents. For the reporting period, the team assumed a uniform age-class distribution within each environmental zone and forest type, an application of the classic normal forest concept that smooths out artifacts from synchronized rotation events.

Validation was deliberately conservative. The researchers compared simulated aboveground wood carbon — converted to growing stock using country-specific conversion factors — against geo-referenced observations from the Croatian and Slovak forest management databases. Because the model assumes fully stocked stands while real forests include understocked ones, an uncorrected comparison revealed an overall overestimation of roughly 41 percent. The team therefore derived stocking correction factors from a quarter of the observational dataset, applied them only where they reduced predictions, and validated against the remaining 75 percent. After correction, the overestimation fell to about 4 percent, and the model achieved successful predictive performance, with a normalized root mean square error of 44 percent and normalized bias of 9 percent for Croatia, and 41 percent and 2 percent respectively for Slovakia. Performance was strongest for beech, oaks, and spruce, which explained 57 to 66 percent of the area-weighted variance, while pines, coppice, and fir — ecologically and managerially more heterogeneous groups — showed larger uncertainty.

Because validating deadwood, litter, and soil against field data was not feasible at the stratum level, the team evaluated those pools against published plausible ranges for European forests. Simulated annual net ecosystem exchange stayed mostly within the published range of minus 460 to plus 60 grams of carbon per square meter per year, with the expected exception of young stands, which acted as carbon sources for roughly 10 to 15 years after final harvest. Carbon stocks in all three pools remained within literature ranges throughout 1990 to 2023, and country totals compared credibly with national forest inventory figures: 214 versus 248 tonnes of carbon per hectare for Croatia in 2008, and 250 versus 277 tonnes per hectare for Slovakia in 2016.

The payoff came in the form of model-derived carbon stock change factors ready for use in the Common Reporting Tables of national inventories. Rather than reporting year-to-year values, which can jump abruptly when a large stratum reaches its rotation age in simulation, the team used the slope of each pool’s carbon time series over the reporting period as a proxy for the average annual change. Simulated soil organic carbon increased modestly in both countries, litter stocks were essentially stable, and deadwood showed slightly negative trends — a conservative result the authors attribute to model limitations, including the inability to represent historical wood extraction during spin-up and an assumed 90 percent extraction rate during salvage logging that likely overestimates real removals. Applying the new factors, and adding an estimate for Croatia’s heterogeneous forests out of yield, raised the previously reported national net CO2 removals for 2023 by 86.11 kilotonnes, or 1.6 percent, in Croatia, and by 138.61 kilotonnes, or 2.1 percent, in Slovakia.

A sensitivity and uncertainty analysis of the workflow itself revealed which methodological choices matter most. The representation of forest management caused the largest variation in growing stock estimates, exceeding 30 percent, followed by the stocking correction at roughly 25 percent. Environmental stratification and carbon-to-growing-stock conversion factors contributed moderate variability of 12 to 16 percent, while statistical uncertainty propagated through the upscaling step added only about 5 percent. The message is clear: for process-based models deployed at national scale, accurate management history and explicit treatment of stand structural variability are more decisive than the fine details of parameterization, and future model development should integrate stand structure directly rather than correcting for it afterward.

The authors are candid about the simplifications that remain. Tree species composition was held constant over the 34-year reporting period — defensible, they argue, because post-disturbance management in Croatia typically preserves composition and Slovakia’s observed shift from spruce to beech has occurred mainly in young stands with modest carbon impact. Age structure within each stratum was represented by a single fixed age, and mortality was kept constant. Disturbances other than drought were not represented, despite evidence of rising disturbance levels in both countries, which is one reason the simulated deadwood trends should be interpreted with caution. Even so, the workflow is conceptually transferable to other process-based models that provide comparable carbon pool outputs, and the modest size of the reported changes — 1.6 and 2.1 percent — sits comfortably within the workflow’s own estimated uncertainty range.

What makes this study resonate beyond two Central European countries is its timing. Europe’s forests have long been counted on as a reliable climate ally, yet recent research warns of sink saturation and growing disturbance pressure, making the accuracy of national LULUCF accounts a matter of continental consequence. By demonstrating that a transparent, validated, process-based modeling chain can supply country-specific carbon stock change factors for the very pools that field measurement programs have left uncovered, the team has turned a methodological gap into an actionable pathway. The increase in reported removals is small, but the strengthening of methodological completeness is not: it gives inventory compilers a defensible, reproducible estimate where previously there was silence, and it offers other nations a tested template for making their own forest carbon accounts whole.

Subject of Research: Country-scale process-based forest carbon modeling to derive carbon stock change factors for national LULUCF greenhouse gas inventory reporting in Croatia and Slovakia

Article Title: Country-scale forest carbon modeling to support national greenhouse gas inventories in Croatia and Slovakia

Article References: Country-scale forest carbon modeling to support national greenhouse gas inventories in Croatia and Slovakia. (n.d.). https://doi.org/10.1007/s10113-026-02705-9

Image Credits: AI Generated

DOI: 10.1007/s10113-026-02705-9

Keywords: forest carbon, Biome-BGCMuSo, greenhouse gas inventory, LULUCF, carbon stock change, Croatia, Slovakia, soil organic carbon, deadwood, process-based modeling, carbon sink, climate mitigation