loading page

Parameter optimization for global soil carbon simulations: Not a simple problem
  • +2
  • Charles B Gauthier,
  • Joe R. Melton,
  • Gesa Meyer,
  • Raj Deepak S N,
  • Oliver Sonnentag
Charles B Gauthier
Universite de Montreal
Author Profile
Joe R. Melton
Environment and Climate Change Canada

Corresponding Author:[email protected]

Author Profile
Gesa Meyer
Environment and Climate Change Canada
Author Profile
Raj Deepak S N
University of Victoria
Author Profile
Oliver Sonnentag
Université de Montréal
Author Profile

Abstract

Accurate simulation of soil organic carbon (SOC) dynamics by terrestrial biosphere models is hampered by poorly constrained parameters and parameter equifinality, amongst other issues. To address this, we use Bayesian optimization to constrain the 16 SOC-related parameters in the Canadian Land Surface Scheme Including biogeochemical Cycles (CLASSIC). We employed a global sensitivity analysis (Sobol’) to develop four parameter sets based upon different sensitivity criteria. We then optimized each set against observed SOC (World Soil Information Service; WoSIS) and soil respiration (Soil Respiration Database; SRDB). Using two different loss functions; one focused on reproducing the observational mean value, and the other explicitly accounting for an estimated observational uncertainty. The best optimized parameter sets from each loss function had an average relative difference of 61%. Thus the choice of loss function impacts what parameter values are deemed optimal and should be considered carefully. The final selected optimal parameter set saw a 12% improvement against WoSIS and SRDB, had global SOC totals in line with literature estimates, and better simulated high-latitude SOC stocks evaluated against the Northern Circumpolar Soil Carbon Database (RMSD: 16.39 vs. 17.61; bias: -5.57 vs. -10.78 kg C m2) compared to the default CLASSIC parameters. However, some parameters were not well constrained, in particular those of needle-leaf deciduous trees which dominate Siberian boreal forests, a region relatively poorly observed in WoSIS and SRDB. Future work should apply further constraints on the optimization framework and address observational gaps.
19 Jul 2024Submitted to ESS Open Archive
22 Jul 2024Published in ESS Open Archive