labs_title

Toward “optimal” integration of terrestrial biosphere models

C.R. Schwalm, D.N. Huntzinger, J.B. Fisher, A.M. Michalak, K. Bowman, P. Ciais, R. Cook, B. El-Masri, D. Hayes, M. Huang, A. Ito, A. Jain, A.W. King, H. Lei, J. Liu, C. Lu, J. Mao, S. Peng, B. Poulter, D. Ricciuto, K. Schaefer, X. Shi, B. Tao, H. Tian, W. Wang, Y. Wei, J. Yang and N. Zeng

Combining the output of multiple Earth system models is an established method for improving prediction quality. Here we show that model estimates of key descriptors of the carbon cycle (gross and net productivity as well as biomass) are insensitive to alternate weighing of combining models that were examined. Treating all models as equal (arithmetic mean) versus weighting models based on how well they reproduce the historical record results in the same, within statistical uncertainty, answer. This highlights the importance of robust uncertainty quantification of both model outputs and the reference datasets used to judge model skill.


Figure: Difference between weighted and unweighted model combinations. Left column: Histograms (gray), fitted normal distribution (black line), unweighted (blue line), weighted values (dark red line), and uncertainty bounds for weighted values. Right column: Difference map of weighted and unweighted values.

Abstract

Multimodel ensembles (MME) are commonplace in Earth system modeling. Here we perform MME integration using a 10-member ensemble of terrestrial biosphere models (TBMs) from the Multiscale synthesis and Terrestrial Model Intercomparison Project (MsTMIP). We contrast optimal (skill based for present-day carbon cycling) versus naïve (“one model-one vote”) integration. MsTMIP optimal and naïve mean land sink strength estimates (−1.16 versus −1.15 Pg C per annum respectively) are statistically indistinguishable. This holds also for grid cell values and extends to gross uptake, biomass, and net ecosystem productivity. TBM skill is similarly indistinguishable. The added complexity of skill-based integration does not materially change MME values. This suggests that carbon metabolism has predictability limits and/or that all models and references are misspecified. Resolving this issue requires addressing specific uncertainty types (initial conditions, structure, and references) and a change in model development paradigms currently dominant in the TBM community.

Schwalm, C.R., D.N. Huntzinger, J.B. Fisher, A.M. Michalak, K. Bowman, P. Ciais, R. Cook, B. El-Masri, D. Hayes, M. Huang, A. Ito, A. Jain, A.W. King, H. Lei, J. Liu, C. Lu, J. Mao, S. Peng, B. Poulter, D. Ricciuto, K. Schaefer, X. Shi, B. Tao, H. Tian, W. Wang, Y. Wei, J. Yang, N. Zeng (2015), "Toward “optimal” integration of terrestrial biosphere models", Geophysical Research Letters, 42 (11), 4418–4428. doi:10.1002/2015GL064002.