Disentangling the Effects of Minerals and Other Environmental Factors on Soil Carbon Stocks, and Capacity

Soil organic carbon (SOC) is important for soil health and its accrual is discussed for carbon sequestration. The SOC fraction stabilized by mineral associations is of special interest, but limited reactive mineral surfaces comprise a natural boundary. The potential upper limit of soils to store SOC as mineral-associated organic carbon (MAOC), the mineralogical capacity, cannot be directly measured as MAOM formation is the result of a complex interplay between mineral properties, plant litter input, and microbial growth and transformation. Accordingly it, depends on a variety of environmental drivers. Current approaches use boundary line regression to identify the dependency of the mineralogical capacity on texture, mineral type, and other environmental conditions and thereby suffer from data sparsity and neglect interactions among the different drivers.
 To exploit multiple sources of data and combine them via common and expert knowledge, we developed a parameter learning framework that combines machine learning and mechanistic modeling. The spatial distribution of parameters (e.g. mineralogical capacity or litter decomposition rates) of a mineral and microbial explicit mechanistic model is inferred using a hybrid neural network, where the mechanistic model forms the final layer. The neural network learned the mechanistic parameters from observations of SOC and MAOC, using environmental covariates like texture, climatological and vegetational conditions as inputs. Influences from mineral properties and other environmental conditions can thereby be separated in an informed way.
 Bootstrapping and analyzing the distribution of mechanistic parameters revealed that relying solely on SOC observations from the Land Use and Land Cover Survey (LUCAS) is insufficient for stable results. Thus, the output space was further constrained by penalizing unrealistic predictions, using MAOC and other sparse observations, and restricting the degrees of freedom in the framework. The posterior parameter combinations per site were thereby limited, which reduces equifinality and assures physical consistency of all model parts.
 Results of the distribution of the mineralogical capacity, steady state MAOC/POC and sensitivities on mineral and environmental conditions can inform the carbon sequestration and soil health community about areas of interest. As rates of change and respective sensitivities are also of high interest, the framework should be extended in the future with a dynamic mechanistic model.