A machine learning gap-filling approach for daily forest soil CO2 flux based on environmental factors and eddy covariance variables.

Accurate gap-filling of daily soil CO₂ flux is essential for understanding forest carbon cycling and supporting the achievement of carbon neutrality goals. Using observations from the Ailaoshan forest site (2010-2013), we evaluated five machine-learning models with inputs including environmental variables, eddy covariance products such as ecosystem respiration and gross primary productivity, and a seasonal phase indicator. Seasonal patterns remained stable during the 4-year period, and no long-term trend was detected after quality control. After preprocessing, multiple machine learning models were compared and evaluated through rolling-window cross-validation. Model robustness and interpretability were further examined using complexity curves, SHAP analysis, and ablation experiments. The empirical Q10 × SWC model served as a baseline, and independent data from the Xishuangbanna site (2003-2008) were used for external validation. CatBoost achieved the best balance between accuracy and complexity and remained stable across rolling windows. Incorporating GPP, RE, and cos_doy improved test performance, raising R2 from ~ 0.85 to ~ 0.90 and reducing RMSE from ~ 0.66 to ~ 0.57. Compared with Q10 × SWC, CatBoost achieved an average Skill score of 0.18, corresponding to a 3-32% reduction in RMSE. External validation at Xishuangbanna further confirmed model robustness (Skill ≈ 0.81). The proposed framework provides an effective tool for generating continuous and reliable daily soil CO₂ flux datasets, offering methodological support for forest carbon cycle research and the realization of "dual carbon" goals.