Multi-decadal and regional validation of the AMG model at county and grid scales unravels the role of crop residue inputs in increasing soil organic carbon stocks in the Tuojiang River Basin, China

Understanding and predicting long-term soil organic carbon (SOC) dynamics from regional to continental levels is key to evaluating the effects of various environmental variables and management practices on the SOC budget. To that end, accurate and reliable SOC models are decisive. However, SOC models remain poorly evaluated for long-term and large-scale applications. The AMG model, named after its creators Andriolo, Mary, and Guérif, is 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. In the present study, we applied AMG model to simulate cropland SOC dynamics both at a gridded 1 km spatial resolution and at the county-level scale in the Tuojiang River Basin, China, from 1911 to 2017. We validated its performance against SOC stocks measurements and map reconstruction derived from soil monitoring campaign in 1980 and 2017. By doing so we aimed to: (i) test the model accuracy for large-scale and long-term applications, (ii) elucidate the uncertainties associated with the scale resolution of the model, and (iii) clarify the role of different environmental factors and management practices in decadal variations in SOC stocks. The validation test against SOC stock measurements from the national soil monitoring campaign in 1980 revealed that the AMG model run at the grid-scale resolution accurately reproduced the SOC stocks in 1980 (R2 = 0.76). The validation tests against model simulation of the change in SOC stocks from 1980 to 2017 was performed both at the grid-scale (R² = 0.35) and county-level scale (R² = 0.44) and revealed slightly better performance at the latter scale. This result highlights that implementing the SOC model at the county-level using reliable input datasets derived from inventory data is more reliable than using spatially explicit input datasets reconstructed from the same inventory data. This suggests that the reliability of model simulations is predominantly influenced by the quality of the input datasets rather than by the spatial resolution of the model implementation, so that increasing spatial resolution of model simulation is only beneficial if the reliability of the input dataset is not impoverished by the downscaling. These model validations allowed to investigate the main drivers of the century scale SOC dynamics by performing counterfactual scenarios. These indicated that the 50% increase in SOC stock simulated over the past century can mainly be attributed to an increase in crop residue inputs, particularly after the green revolution in the mid-1960 s. Overall, our study provides the first SOC model validation at the regional level against SOC stock time-series observations.