基于知识引导型机器学习模型的黑龙江省天然林固碳潜力预测OA

Accurate quantification of carbon stocks and their dynamic changes in forest ecosystems is essential for assess-ing regional carbon sink potential and addressing climate change.However,conventional machine learning models often suffer from reduced extrapolation ability in long-term predictions due to the absence of ecological process constraints.To overcome this limitation,this study established a knowledge-guided machine learning(KGML)framework for natural for-ests in Heilongjiang Province by integrating fixed-plot observations from 1976-2015 with historical and future climate and CO₂ concentration datasets.The CO₂ fertilization effect was incorporated and optimized into the forest carbon seques-tration(FCS)model,and its simulated results were used as prior features input into forest carbon sequestration model(XGBoost)to achieve the fusion of process knowledge with data-driven learning.Results showed that the optimized FCS model significantly improved prediction accuracy(R² increased from 0.85 to above 0.90),while KGML model demon-strated superior simulation accuracy of carbon density acrossvarious forest types compared to machine learning and pro-cess models,while also exhibiting greater stability and extrapolation capability in long-term predictions(≥10 years).Scenario simulations indicated that natural forests under SSP126 would reach the highest sequestration potential by 2060,with carbon stocks of 1 379.26-1 439.02 Tg C.This study demonstrates the effectiveness and potential for promo-tion of knowledge-guided machine learning in forest carbon prediction and provides a new path for improving the accu-racy of carbon sink assessment and achieving the dual-carbon goals.