Evaluating soil carbon in Earth system models: How good are the models and what drives model variability?

Background/Question/Methods:

The carbon cycle plays a critical role in how the climate responds to anthropogenic carbon dioxide. Recently, Earth system models (ESMs) have incorporated terrestrial carbon cycling. To evaluate how well these ESMs represent the carbon cycle, we examined predictions of below-ground carbon stocks. We compared the soil and litter carbon pools from 13 ESMs with data on soil carbon stocks from the Harmonized World Soil Database (HWSD). We used a nonlinear regression analysis to separate the effects of model input variation, including NPP, soil temperature and soil moisture, from the effects of model parameters on soil carbon predictions. The regression model represented one pool of soil carbon as a first-order linear function of NPP, soil temperature, and soil moisture.

Results/Conclusions:

We found a wide range in the total below-ground carbon predicted by the ESMs (578-3060 Pg-C), bracketing the HWSD 1580 Pg-C. However, none of the ESMs correlated well with the HWSD distribution (Pearson's correlation <0.45). All of the ESMs were first-order linear and multi-pool with outputs that were well described by our regression model (R² of 0.74-0.93). One exception was MPI-ESM-LR (R²=0.56) which used a different model structure. ESM input variables (NPP, soil temperature, and soil moisture) were consistent predictors of HWSD values in our regressions (RMSE of 7.8-10.2 with a coefficient of variation <0.1). In contrast, ESM input variables were inconsistent predictors of ESM soil carbon (RMSE of 8.3-17.5, coefficient of variation ~0.3). Similarly, the direct relationship between ESMs and HWSD soil carbon were highly variable (RMSE of 7.7-21.5, coefficient of variation ~0.4). These results imply that variability in ESM predictions are due to model parameterization and not differences in input variables. Furthermore, the inconsistent relationship between ESM input variables and HWSD soil carbon implies that changing the parameters of the ESMs would not substantially improve the models’ fit to data. Efforts to improve ESM predictive power should be directed to improving model structure or refining input variables.