Evaluation and improvement of a global land model against soil carbon data using a Bayesian Markov chain Monte Carlo method

Long-term land carbon (C) cycle feedback to climate change is largely determined by dynamics of soil organic carbon (SOC). However, most evaluation studies conducted so far indicate that global land models predict SOC poorly. This study was conducted to evaluate predictions of SOC by the Community Land Model with Carnegie-Ames-Stanford Approach biogeochemistry submodel (CLM-CASA'), investigate underlying causes of mismatches between model predictions and observations, and calibrate model parameters to improve the prediction of SOC. We compared modeled SOC to the SOC pools from a globally gridded observation data product and found that CLM-CASA' on average underestimated SOC pools by 65% (r(2) = 0.28). We applied data assimilation to CLM-CASA' to estimate SOC residence times, C partitioning coefficients among the pools, and temperature sensitivity of C decomposition. The model with calibrated parameters explained 41% of the global variability in the observed SOC, which was substantial improvement from the initial 27%. The SOC and litter C feedback to changing climate differed between models with original and calibrated parameters: after 95 years of climate change (2006-2100), soils released 48 Pg C less in the calibrated than in the noncalibrated model and litter released 6.5 Pg C less in the calibrated than the noncalibrated model. Thus, assimilating estimated soil carbon data into the model improved model performance and reduced the amount of C released under changing climate. To further reduce the uncertainty in the soil carbon prediction, we need to explore alternative model structures, improve representation of ecosystems, and develop additional global datasets for constraining model parameters.