Improved half-hourly carbon flux simulation of China's forest ecosystems: An ensemble machine learning approach with limited meteorological data.

Accurate quantification of forest carbon fluxes is essential for improving regional carbon budget assessments and supporting China's carbon neutrality goals under climate change. However, global data-driven products rely heavily on FLUXNET sites with sparse coverage across China, and existing machine learning studies in China predominantly focus on coarse daily or monthly scales. Thus, high-temporal-resolution carbon flux simulations remain critically limited for China's forests. To address these gaps, we developed a fusion ensemble model (denoted as the Fusion model) that combines Random Forest (RF) and eXtreme Gradient Boosting (XGBoost) to simulate half-hourly net ecosystem CO2 exchange (NEE) across 28 forest flux sites in China. The results show that the Fusion model outperformed the standalone models at most sites, yielding higher R2 values at 96.43% of the sites and lower RMSE values at 92.86% of the sites. Model performance varied among forest types, with higher accuracy in evergreen needleleaf, deciduous broadleaf, and mixed forests (R2 ranging from 0.6332 to 0.6468), while evergreen broadleaf forests (EBF) exhibited lower predictability (R2 = 0.4192). Notably, the fusion model showed the greatest relative improvement in the most challenging forest type (EBF). Feature importance and Shapley Additive Explanations (SHAP) identified downward shortwave radiation (DR), air temperature (Ta), and vapor pressure deficit (VPD) as dominant drivers of NEE dynamics. By achieving reliable half-hourly carbon flux simulations using a limited set of meteorological variables, this study proposes a machine learning fusion framework for carbon flux estimation, providing important support for regional carbon budget assessment and forest ecosystem management.