Prediction of soil organic carbon storage and future carbon sequestration potential in grassland and cropland under different climate scenarios: an integrated method combined improved AMG model and random forest

Soil organic carbon (SOC) is a cornerstone of global carbon cycling, ecosystem health, and climate regulation. However, accurately predicting SOC storage (SOCS) and its sequestration potential under varying climatic scenarios remains a major challenge, particularly in high-altitude, climate-sensitive regions like the Qinghai-Tibet Plateau. Grasslands and croplands in this region are pivotal for carbon management, yet their dynamics remain insufficiently understood. This study addresses two core scientific questions: (1) How can SOCS dynamics be modeled accurately across large spatial scales and diverse ecosystems under future climate scenarios? (2) How to comprehensively evaluate potential of SOC sequestration and effectively guide development of targeted carbon management strategies?

Andriolo, Mary and Guérif developed a simple first-order kinetic model that relies on key controlling input data, which is ideal for application across large spatial and long temporal scales. This study improved the traditional AMG model by using NPP as the core carbon input indicator, replacing the traditional crop harvest index (HI), which is more suitable for grassland ecosystems. In addition, the model dynamically adjusted the carbon mineralization rate parameter 𝑘 to reflect the effects of temperature, precipitation and soil properties on SOC dynamics. The improved AMG model (I-AMG) generates time series data as input variables for random forest (RF) model by simulating the SOC dynamics of grassland and cropland. We further combine historical SOCS and environmental variables such as terrain and vegetation indices for training and prediction. The RF-AMG integrated model combines the process simulation capability of I-AMG model with nonlinear fitting capability of  RF algorithm, which can capture complex environmental variable interaction effects and significantly improve prediction accuracy.

We used the global SOC content data provided by the Harmonized World Soil Database (HWSD) in 1980 to estimate the baseline SOCS at a resolution of 1km. The SOCS data in 2020 was provided by the National Qinghai Tibet Plateau Science Data Center using a grid dataset of soil carbon pools created through field surveys and machine learning, and was used as an observation to evaluate the simulation prediction accuracy of our improved AMG model. Meanwhile, we predicted the SOCS of cropland and grassland in Qinghai Province over the next 40 years under mild (RCP4.5) and extreme (RCP8.5) climate scenarios. And further proposed a four-quadrant method to evaluate the energy storage potential of SOC, dividing the carbon sequestration potential level of Qinghai Province in the next 40 years into four different categories based on SOC saturation deficit and change rate. This method identifies the spatial characteristics of SOC sequestration potential in Qinghai over next 40 years, which can help decision-makers gain a detailed understanding of regions with different carbon management priorities.

This study demonstrates the strength of combining process-based modeling with machine learning to address complex environmental challenges. This novel framework can be used for assessing soil carbon sequestration potential of natural ecosystems, and practical guidance for policymakers to develop tailored strategies for soil conservation, sustainable agriculture and ecosystem restoration. These efforts support global carbon neutrality goals and provide valuable insights into climate-smart land management practices.