B117-06
Optimal model complexity for terrestrial carbon cycle prediction using model-data fusion
Wednesday, 16 December 2020: 05:50
Virtual
Caroline Famiglietti1, Thomas Luke Smallman2, Sophie Flack-Prain2, Rong Ge3, Paul Alexander Levine4, Shuang Ma5, Victoria Meyer6, Nicholas Parazoo5,7, Gregory Ross Quetin1, Andrew Revill2, Stephanie Grace Stettz5, Yan Yang5, Yuan Zhao2, Penghui Zhu2, A. Anthony Bloom5, Mathew Williams2 and Alexandra G. Konings1, (1)Stanford University, Department of Earth System Science, Stanford, CA, United States, (2)University of Edinburgh, School of GeoSciences, Edinburgh, United Kingdom, (3)Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences, Beijing, China, (4)NASA Jet Propulsion Laboratory, Pasadena, CA, United States, (5)Jet Propulsion Laboratory, California Institute of Technology, Pasadena, CA, United States, (6)School of the Art Institute of Chicago, Chicago, United States, (7)University of California Los Angeles, JIFRESSE, Los Angeles, CA, United States
Abstract:
Modeling and predicting the terrestrial net carbon balance is difficult due to the numerous processes driving variability of gross fluxes. Many approaches to reducing this model uncertainty have focused on model structure, by adding additional processes and increasing complexity. While adding processes may increase realism, the resulting models often rely by necessity on over-generalized parameters. It is not clear whether or to what extent carbon cycle predictability scales with structural complexity, or whether an intermediate level of complexity exists that may balance the costs of a low (more biased) or high (more variant) complexity model.
Here we demonstrate the extent to which the prediction accuracy of net ecosystem exchange (NEE) and leaf area index (LAI) scales with model complexity. To do so, we developed 16 structurally distinct carbon cycle models spanning a broad range of complexity and incorporated them into a model-data fusion system (CARDAMOM). We ran each model version at 6 globally-distributed eddy covariance sites representing a range of biomes and vegetation types under 42 different data scenarios (i.e., combinations of data constraints and observational error assumptions). We defined the complexity of each of the 4032 model runs based on its “inherent dimensionality”, computed using a principal component analysis that reduced the parameter space to its primary axes of variance. Though the range of complexity we evaluated is necessarily lower than that populated by land surface models, it reveals universal modeling elements that control performance. Specifically, increased complexity can improve forecast skill if parameters are adequately constrained (e.g., when NEE observations are assimilated); otherwise, increased complexity can degrade skill. This finding remains consistent regardless of whether NEE or LAI is predicted. This work highlights the importance of robust parametrization for land surface modeling.
