Understanding the Dynamics of Soil Carbon in CMIP5 Models
Abstract:Soil carbon stocks have the potential to be a strong source or sink for carbon dioxide over the next century, playing a critical role in climate change. These stocks are the result of small differences between much larger primary carbon fluxes: gross primary production, litter fall, autotrophic respiration and heterotrophic respiration. There was little agreement on predicted soil carbon stocks between Earth system models (ESMs) in the most recent Climate Model Intercomparison Project. Predicted present-day stocks ranged from roughly 500 Pg to over 3000 Pg and predicted changes over the 21st century ranged from -70 Pg to +250 Pg). The primary goal of this study was to understand why such large differences exist.
We constructed four reduced complexity models to describe the primary carbon fluxes, making different assumptions about how soil carbon fluxes are modelled in ESMs. For each of these reduced complexity models we statistically inferred the most likely model parameters given the gridded ESM simulation outputs. Gross primary production was best explained by incoming short wave radiation, CO2 concentration, and leaf area index (global GPP comparison of simulation vs reduced complexity model of R2>0.9 (p < 1e-4) with slopes between 0.65 and 1.2 and intercepts between -13 and 67 Pg C yr-1). Autotrophic respiration was best explained as a proportion of GPP (R2 > 0.9 (p < 1e-4) with slopes between 0.78 and 1.1 and intercepts between -15 and 14 Pg C yr-1). Flux between the vegetation and soil pools were best explained as a proportion of the vegetation carbon stock (R2 > 0.9 (p < 1e-4) with slopes between 0.9 and 2.1 and intercepts between -65 and 25 Pg C yr-1). Finally heterotrophic respiration was best explained as a function of soil carbon stocks and soil temperature (R2 > 0.9 (p < 1e-4) with slopes between 0.7 and 1.5 and intercepts between -40 and 15 Pg C yr-1).
This research suggests three main lines of decomposition model improvement: 1) improve connecting sub-models, 2) data integration to improve parameterization, 3) modification of model structure. The implied variation in RCM parameterization suggests that data integration could constrain model simulation results. However, the similarity in model structure may lead to systematic biases in the simulations without the introduction of new model structures.