Ecology Meets Machine Learning to Forecast China’s Forest Carbon Future

A new modeling study published in Annals of Forest Science reports that embedding ecological process knowledge into machine learning substantially improves the reliability of long-term forest carbon predictions. Researchers led by Lukun Geng and Bin Wang developed a knowledge-guided machine learning, or KGML, framework for the natural forests of Heilongjiang Province in northeastern China, and found that it maintained strong predictive performance over 10- to 15-year horizons during which a conventional machine learning model degraded markedly. The work, published open access on 10 September 2026, addresses one of the most persistent weaknesses of data-driven environmental modeling: the tendency of algorithms trained on historical observations to fail when asked to extrapolate into conditions they have never seen.

The motivation stems from a well-recognized limitation. Machine learning tools such as Random Forest and XGBoost excel at capturing complex nonlinear relationships in large datasets, and they have been widely used to model forest biomass and carbon stocks. Yet these models are effectively black boxes: they lack explicit representation of the ecological processes that govern tree growth, offer limited interpretability, and depend heavily on the quality and representativeness of their training data. When the prediction horizon extends beyond the range of the observations, accuracy can collapse. For carbon accounting under climate change, where projections decades into the future are exactly what policymakers need, that fragility is a serious problem.

The team’s solution combines a process-based model with a data-driven algorithm. The foundation is the forest carbon sequestration model, a semi-process-based, empirical model grounded in ecosystem succession theory that describes vegetation growth with a logistic equation. In its original form, the model treats carbon accumulation as a function of stand age, mean annual temperature, and mean annual precipitation. What it omits is the fertilizing influence of rising atmospheric carbon dioxide on plant growth, an omission the authors argue introduces substantial uncertainty in long-term projections. The researchers therefore modified the model’s intrinsic growth rate with a carbon dioxide adjustment term, calibrated against plot-level climate and atmospheric concentration data.

The optimization paid off measurably. Incorporating the carbon dioxide fertilization effect raised the coefficient of determination of the process model from approximately 0.85 to above 0.90. Across the three forest types examined, the refined model achieved an R-squared of 0.9010 for coniferous forest, 0.9306 for broadleaf forest, and 0.9091 for mixed conifer-broadleaf forest, with root-mean-square errors between roughly 14 and 17 tonnes of carbon per hectare. The authors caution, however, that the fertilization term is a simplified representation: nutrient availability, nitrogen limitation, soil moisture, and species-specific physiology were not explicitly modeled, so the adjustment should be read as an approximation rather than a complete ecological description.

The knowledge-guided framework then integrates this optimized process model into XGBoost at two levels. First, in feature engineering, the process model’s predictions are fed to the machine learning algorithm as an additional input feature, supplying physical priors such as biomass upper limits and climate response patterns. Second, in training, a hybrid loss function blends the usual mean-squared data error with a physics term measuring deviation from the process model’s predictions. A dynamic weighting strategy applies the physics constraint only where the process model proves more reliable than the data alone at a given sample, preserving fitting accuracy where observations are trustworthy.

The evidence for the framework’s advantage comes from comparisons across forest types and prediction intervals using forest fixed-plot observations from 1976 to 2015. At the 5-year interval, all models performed well, with the KGML model reaching R-squared values of 0.9586, 0.9755, and 0.9573 for coniferous, broadleaf, and mixed forests respectively, only marginally ahead of standalone XGBoost. The picture changed dramatically at longer horizons. Extending predictions to 10 and 15 years, conventional XGBoost lost an average of roughly 0.24 in R-squared, and in mixed conifer-broadleaf forests its 15-year R-squared fell to just 0.02, an extrapolation failure. The KGML model stayed above 0.7 across all forest types at 15 years, with the smallest average decline of the three approaches.

Ablation experiments clarified which component does what. Adding process-model predictions as input features alone yielded little short-term benefit but markedly improved long-term stability, lifting the 15-year mixed-forest R-squared from 0.0214 under XGBoost to 0.6727 under the full KGML model. The hybrid loss function, by contrast, mainly boosted short- and medium-term accuracy and contributed little on its own over long horizons. The optimal balance between the two loss components also varied by forest type, with the best weighting parameter ranging from 0.1 for broadleaf forests, which had the most training data, to 0.4 for mixed forests. The combined framework performed best overall, indicating that feature priors and training constraints act synergistically rather than redundantly.

Applied to future scenarios, the model projects that Heilongjiang’s natural forests hold their greatest carbon sequestration potential under the low-emission SSP1-2.6 pathway. Depending on model configuration and forest type, projected total carbon stocks in 2060 range from 1371.37 plus or minus 422.04 to 1442.38 plus or minus 316.77 teragrams. Under SSP1-2.6, carbon density averages 123.09 tonnes per hectare by 2060, and the fastest growth occurs between 2025 and 2030 at 2.468 tonnes per hectare annually. Across the other three scenarios, growth peaks earliest, between 2020 and 2025, then slows as stands approach their potential maximum biomass. The projections carry wide uncertainty bands and, as the authors note, remain subject to uncertainty from climate scenarios, process-model assumptions, inventory coverage, and the unmodeled effects of disturbances and management.

Spatially, the simulations place the highest carbon densities in 2015 in the southern Lesser Khingan Mountains and the Zhangguangcai Ridge, expanding southward across southern Heilongjiang by 2060, while the Greater Khingan Range in the north remains lowest, a gradient the authors link to species composition and climate, since cold temperatures and short growing seasons constrain coniferous biomass accumulation there. Because broadleaf forests showed higher prediction accuracy than the other types, spatial predictions in broadleaf-dominated regions are likely associated with lower uncertainty. The study also projects that maximum carbon density declines slightly after 2040 in places, consistent with maturing stands nearing carbon saturation, a pattern the authors say underscores the importance of stand-age structure in long-term carbon balance assessments.

The broader significance lies in methodology rather than in any single number. The framework relies on widely available forest inventory, climate, and atmospheric carbon dioxide data, and the authors suggest it could be transferred to other forest ecosystems after regional calibration, though planted forests would require additional parameter adjustment because management practices shape their growth. Future work, they write, should incorporate disturbances such as fires, insect outbreaks, and tree mortality, and pursue formal uncertainty quantification. For now, the Heilongjiang results offer a concrete demonstration that ecological theory, far from being superseded by machine learning, can be the ingredient that keeps algorithms honest when they are asked to peer decades ahead.

Subject of Research: Knowledge-guided machine learning for predicting long-term carbon sequestration in natural forests

Article Title: Knowledge-guided machine learning for predicting the carbon sequestration potential of natural forests in Heilongjiang Province

Article References: Geng, L., Bian, S., Ma, X., Fu, Q., Zhou, Y., & Wang, B. (2026). Knowledge-guided machine learning for predicting the carbon sequestration potential of natural forests in Heilongjiang Province. Annals of Forest Science, 83(1), Article 42. https://doi.org/10.1186/s13595-026-01359-1

Image Credits: AI Generated

DOI: 10.1186/s13595-026-01359-1

Keywords: knowledge-guided machine learning, forest carbon stocks, carbon sequestration, Heilongjiang, CO2 fertilization, XGBoost, process-based modeling, climate scenarios, forest inventory, long-term prediction, SSP scenarios, ecological modeling