As an important global carbon (C) reservoir, soils store large amounts of organic matter and play a key role in regulating global carbon-nitrogen (C-N) cycling in the context of climate change. The first-order kinetic model which divides SOM into multiple C pools with specific decomposition rates has been developed and widely applied in fitting laboratory soil incubation experiment data. In Earth System Models (ESMs) with first-order kinetic assumption, the potential first-order decomposition rates are used to be globally constants to simplify model parameterization. However, this simplification might introduce bias in global-scale simulation, leading to unrealistic global C cycle projections. Here, we compile first-order kinetic parameters (pool sizes and decomposition rates) derived from laboratory incubation experiments with soils from diverse climate zones and ecosystems. By using linear regression models (LR) and machine learning models (ML), we attempt to explore the relationship between these kinetic parameters and environmental factors and further predict the global patterns of SOM decomposition kinetics.
Results/Conclusions
Results show that the Gradient Boosting Machine (GBM), a machine learning algorithm, performed much better than the LR model in predicting both the pool sizes and associated decomposition rates. The GBM-based variable importance analysis indicates that soil texture (clay and sand content) dominates the decomposition of the fastest degrading SOM pool, whereas pH is the most important factor determines the turnover rate of the slowest SOM pool. The global prediction of decomposition rates depicts remarkable regional variation and change dramatically with latitude. Our machine-learning-based approach provides global patterns of SOM decomposition kinetic parameters, which could be an important reference for global-scale C modeling.
