Spatial hybrid modeling of soil organic carbon processes: testing common assumptions using multivariate, dynamic data with simple models

The land surface accounts for a large share of variability in the global carbon cycle. Although increasing atmospheric CO₂ concentrations have led to higher net primary production and increased land carbon stocks, vegetation carbon stocks appear largely constant, implying that changes in land carbon are primarily driven by soil organic carbon (SOC). As SOC represents the largest active carbon pool, its dynamics are critical for land–atmosphere feedbacks. However, strong spatial heterogeneity and measurement limitations result in sparse and mostly static SOC data, complicating the identification of dominant processes.

Recent studies address this limitation by assimilating soil carbon models to spatial SOC and covariate datasets using neural networks (hybrid modeling). The resulting spatial parameter fields are then interpreted in terms of underlying mechanisms. These approaches typically rely on three key assumptions: steady-state conditions, adequate process representation by the assimilated SOC model, and the sufficiency of bulk SOC data to infer processes. In this study, we explicitly test these assumptions.

We use the Europe-wide LUCAS dataset, which provides spatially resolved physical and chemical soil data at multiple time points. A subset of the dataset includes SOC subfractions, including mineral-associated organic carbon and microbial biomass carbon. Several simple SOC models were assimilated in their steady-state form in the hybrid framework, while accounting for differences in model flexibility. This allowed exclusion of specific modeling assumptions. Comparisons across time steps were used to assess the validity of the steady-state assumption. In addition, first results obtained with a dynamic SOC model are presented.