Maximizing the carbon sink function of paddy systems in China with machine learning

Developing low-carbon agriculture and alleviating the "carbon crisis" requires optimizing strategies that fully leverage the carbon sink function of paddy systems. Accurate assessment of the effects of various agricultural management practices (AMPs) on the carbon sink function of paddy systems is crucial to this end. Here, we have presented a soil organic carbon sequestration rate (SOCSR) database of paddy systems in China based on 1388 groups of experimental data from 143 peer-reviewed publications. We analyzed the impact trend of different AMPs on SOCSR, compared two traditional regression models, four classic machine learning models and two deep learning models, and established a data-driven SOCSR prediction model to quantify the impact of AMPs on SOCSR and provide the optimal strategies. Our model (Random Forest) had the characteristics of high accuracy (R² = 0.71, RMSE = 0.53 Mg ha^-1), strong flexibility, low time cost with a certain degree of interpretability for the regional scale of China. We found that inorganic N fertilizer, inorganic K fertilizer, organic fertilizer, tillage and residue management are relatively important AMPs for improving SOCSR. The carbon sink function of paddy systems would reach saturation when the application rate of inorganic N fertilizer, inorganic K fertilizer and organic fertilizer reached around 80 kg N ha^-1, 40 kg K ha^-1 and 2200 kg C ha^-1, respectively. Compared to half residue returning and conventional tillage, full residue returning and no-tillage increased SOCSR by 39.8 % and 9.2 %, respectively. Our optimal combination of strategies could achieve SOCSR of 1.179 Mg ha^-1 in paddy systems of China. Our work enables swift and precise evaluation of SOCSR in paddy systems, provides a new idea for assessing SOCSR of paddy systems on a regional scale, and serves as an essential part for the accurate assessment of the carbon footprint of rice production.