Fire plays an important role in terrestrial ecosystems by regulating biogeochemistry, biogeography, and energy budgets, yet despite the importance of fire as an integral ecosystem process, significant advances remain to improve its prognostic representation in carbon cycle models. To recommend and to help prioritize model improvements, this study investigates the sensitivity of a coupled global biogeography and biogeochemistry model, LPJ, to observed burned area measured by three independent satellite-derived products, GFED v3.1, L3JRC, and GlobCarbon. Model variables are compared with benchmarks that include pantropical aboveground biomass, global tree cover, and CO 2 and CO trace gas concentrations. Depending on prescribed burned area product, global aboveground carbon stocks varied by 300 Pg C, and woody cover ranged from 50 to 73 Mkm 2 . Tree cover and biomass were both reduced linearly with increasing burned area, i.e., at regional scales, a 10% reduction in tree cover per 1000 km 2 , and 0.04-to-0.40 Mg C reduction per 1000 km 2 . In boreal regions, satellite burned area improved simulated tree cover and biomass distributions, but in savanna regions, model-data correlations decreased. Global net biome production was relatively insensitive to burned area, and the long-term land carbon sink was robust,~2.5 Pg C yr
À1, suggesting that feedbacks from ecosystem respiration compensated for reductions in fuel consumption via fire. CO 2 transport provided further evidence that heterotrophic respiration compensated any emission reductions in the absence of fire, with minor differences in modeled CO 2 fluxes among burned area products. CO was a more sensitive indicator for evaluating fire emissions, with MODIS-GFED burned area producing CO concentrations largely in agreement with independent observations in high latitudes. This study illustrates how ensembles of burned area data sets can be used to diagnose model structures and parameters for further improvement and also highlights the importance in considering uncertainties and variability in observed burned area data products for model applications.