Midwest U.S. Croplands Determine Model Divergence in North American Carbon Fluxes

W. Sun¹, Y. Fang², X. Luo^3,4, Y. Shiga¹, Y. Zhang⁴, A.E. Andrews⁵, K. Thoning⁵, J.B. Fisher⁶, T.F. Keenan^7,4 and A. Michalak¹

¹Carnegie Institution for Science, Department of Global Ecology, Stanford, CA 94305; 650-325-1521, E-mail: wsun@carnegiescience.edu
²Bay Area Air Quality Management District, San Francisco, CA 94109
³National University of Singapore, Department of Geography, Kent Ridge, Singapore
⁴Lawrence Berkeley National Laboratory, Climate and Ecosystem Sciences Division, Berkeley, CA 94720
⁵NOAA Global Monitoring Laboratory (GML), Boulder, CO 80305
⁶NASA Jet Propulsion Laboratory, California Institute of Technology, Pasadena, CA 91109
⁷University of California at Berkeley, Department of Environmental Science, Policy, and Management, Berkeley, CA 94720

Terrestrial biosphere models are key to understanding continental-scale patterns of biospheric carbon fluxes. However, individual models disagree on whether temperate forests or croplands dominate the growing-season net carbon uptake over North America. Evidence from novel photosynthetic proxies has cast doubt on the “strong forest, weak cropland” carbon uptake patterns simulated by most models. To glean robust space-time patterns of North American carbon fluxes from widely divergent model estimates, we need to leverage regional-scale (10³–10⁵ km²) constraints from carbon dioxide (CO₂) atmospheric observations in model evaluation. Here, we evaluate a large suite of terrestrial biosphere models and data-driven models by examining how well the atmospheric signals resulting from their carbon flux estimates capture the space-time variability in atmospheric CO₂, as is observed by a network of tall towers over North America.

Models with gross or net carbon fluxes that are consistent with the observed CO₂ variability share a salient feature of strong growing-season carbon uptake in Midwest U.S. croplands. Conversely, the remaining models place most growing-season uptake in boreal or temperate forests. Differences in model explanatory power depend mainly on the simulated timing of peak uptake in croplands rather than the distribution of seasonal cycle amplitudes of fluxes across biomes. Our results suggest that improved model representation of cropland phenology is crucial to robust, policy-relevant estimation of North American carbon exchange.

Figure 1. Fraction of variance (R²) in atmospheric CO₂ explained by modeled carbon fluxes (GPP and NEE), solar-induced chlorophyll fluorescence (SIF) data products, and the absorbed photosynthetically active radiation (APAR) during 2007–2010, with error bars showing 95% confidence intervals obtained from bootstrapping. For each model with both GPP and NEE estimates, data points that denote R² values of GPP and NEE of that model are linked by a thin dashed line. The vertical lines denote R² values of three references (from left to right): APAR and monthly and three-hourly geostatistical inverse estimates of NEE (GIM NEE).

Figure 2. The June–July–August (JJA) mean patterns of the extracted first principal component (PC1) among (a) GPP and SIF models of high R², (b) GPP models of low R², (c) NEE models of high R², and (d) NEE models of low R², where models of high R² and of low R² are separated using the R² value of APAR (0.327) as a threshold. Percentage numbers represent the fraction of variance in each group explained by the PC1, whereas N indicates the number of models in each group. For both GPP and SIF (a) and NEE (c), models with high R² show common features of strong summer uptake in the mid-continental cropland region.