Spatiotemporal Pattern of Ecosystem Respiration in China Estimated by Integration of Machine Learning With Ecological Understanding

Abstract Accurate estimation of regional and global patterns of ecosystem respiration (ER) is crucial to improve the understanding of terrestrial carbon cycles and the predictive ability of the global carbon budget. However, large uncertainties still exist in regional and global ER estimation due to the drawbacks of modeling methods. Based on eddy covariance ER data from 132 sites in China from 2002 to 2020, we established Intelligent Random Forest (IRF) models that integrated ecological understanding with machine learning techniques to estimate ER. The results showed that the IRF models performed better than semiempirical models and machine learning algorithms. The observed data revealed that gross primary productivity (GPP), living plant biomass, and soil organic carbon (SOC) were of great importance in controlling the spatiotemporal variability of ER across China. An optimal model governed by annual GPP, living plant biomass, SOC, and air temperature (IRF‐04 model) matched 93% of the spatiotemporal variation in site‐level ER, and was adopted to evaluate the spatiotemporal pattern of ER in China. Using the optimal model, we obtained that the annual value of ER in China ranged from 5.05 to 5.84 Pg C yr −1 between 2000 and 2020, with an average value of 5.53 ± 0.22 Pg C yr −1 . In this study, we suggest that future models should integrate process‐based and data‐driven approaches for understanding and evaluating regional and global carbon budgets.