Confronting terrestrial biosphere models with forest inventory data

Efforts to test and improve terrestrial biosphere models (TBMs) using a variety of data sources have become increasingly common. Yet, geographically extensive forest inventories have been under‐exploited in previous model–data fusion efforts. Inventory observations of forest growth, mortality, and biomass integrate processes across a range of timescales, including slow timescale processes such as species turnover, that are likely to have important effects on ecosystem responses to environmental variation. However, the large number (thousands) of inventory plots precludes detailed measurements at each location, so that uncertainty in climate, soil properties, and other environmental drivers may be large. Errors in driver variables, if ignored, introduce bias into model–data fusion. We estimated errors in climate and soil drivers at U.S. Forest Inventory and Analysis (FIA) plots, and we explored the effects of these errors on model–data fusion with the Geophysical Fluid Dynamics Laboratory LM3V dynamic global vegetation model. When driver errors were ignored or assumed small at FIA plots, responses of biomass production in LM3V to precipitation and soil available water capacity appeared steeper than the corresponding responses estimated from FIA data. These differences became nonsignificant if driver errors at FIA plots were assumed to be large. Ignoring driver errors when optimizing LM3V parameter values yielded estimates for fine‐root allocation that were larger than biometric estimates, which is consistent with the expected direction of bias. To explore whether complications posed by driver errors could be circumvented by relying on intensive study sites where driver errors are small, we performed a power analysis. To accurately quantify the response of biomass production to spatial variation in mean annual precipitation within the eastern United States would require at least 40 intensive study sites, which is larger than the number of sites typically available for individual biomes in existing plot networks. Driver errors may be accommodated by several existing model–data fusion approaches, including hierarchical Bayesian methods and ensemble filtering methods; however, these methods are computationally expensive. We propose a new approach, in which the TBM functional response is fit directly to the driver‐error‐corrected functional response estimated from data, rather than to the raw observations.