Evaluation of Site and Continental Terrestrial Carbon Cycle Simulations with North American Flux Tower Observations

Terrestrial carbon models are widely used to diagnose past ecosystem-atmosphere carbon flux responses to climate variability, and are a critical component of coupled climate-carbon model used to predict global climate change. The North American Carbon Program (NACP) Interim Regional and Site Interim Synthesis activities collected a broad sampling of terrestrial carbon model results run at both regional and site level. The Regional Interim Synthesis Activity aims to determine our current knowledge of the carbon balance of North America by comparing the flux estimates provided by the various terrestrial carbon cycle models. Moving beyond model-model comparison is challenging, however, because no continental-scale reference values exist to validate modeled fluxes. This paper presents an effort to evaluate the continental-scale flux estimates of these models using North American flux tower observations brought together by the Site Interim Synthesis Activity. Flux towers present a standard for evaluation of the modeled fluxes, though this evaluation is challenging because of the mismatch in spatial scales between the spatial resolution of continental-scale model runs and the size of a flux tower footprint. We compare model performance with flux tower observations at monthly and annual integrals using the statistical criteria of normalized standard deviation, correlation coefficient, centered root mean square deviation and chi-squared. Models are evaluated individually and according to common model characteristics including spatial resolution, photosynthesis, soil carbon decomposition and phenology. In general all regional models are positively biased for GPP, Re and NEE at both annual and monthly time scales. Further analysis links this result to a positive bias in many solar radiation reanalyses. Positively biased carbon fluxes are also observed for enzyme-kinetic models and models using no nitrogen limitation for soil carbon decomposition. While the former result is likely influenced by model sensitivity to positively biased radiation the latter finding is more likely a result of differences in model structure. Not surprisingly, site level models outperform regional level models for almost all statistical criteria suggesting: (1) locally observed driver data is important for capturing local carbon fluxes and (2) biases in regional driver data adversely effect regional model output. The inability of regional models to reproduce NEE inter-annual variability in combination with limited improvement between site and regional level models in terms of inter-annual variability, suggests that model structure remains a limiting factor in overall performance. Monthly correlations between models and data are much higher than annual correlations. Regional models show similar performance to their site level counterparts in terms of correlation of monthly fluxes. EC-MOD, a model using assimilated flux tower data, performs the best overall. CASA-GFEDV2 and VEGAS2 also perform well.