Drivers of ecosystem variability across temporal scales: Insights from a multi-model comparison of forest dynamics for the past millennium
Terrestrial ecosystem models are used to predict changes in ecosystem properties such as composition, biomass, and energy fluxes in response to environmental perturbation such as climate change. A large community of models with differing structures and underlying assumptions has been applied to forest ecosystems in regions such as the northeastern United States. These models often operate at high, sub-hourly temporal resolution and have been validated at annual or decadal scales. However, the primary driver of ecosystem change at the centennial time scales at we are making predictions of future forest dynamics may be different than the controls of ecosystem variability at decadal scales. To bridge this gap and understand the temporal scalability of models and model inferences at annual, decadal, and centennial time scales, we have run a suite of six models at six locations in the northeastern and upper-midwestern regions of the United States for the past 1200 years. We evaluate how variation in environmental forcing variables including temperature, precipitation, and CO2drive changes in a range of ecosystem properties such as aboveground biomass, net primary productivity, soil carbon, and soil moisture in both model output and empirical observations for the same region.
Ecosystem fluxes such as net primary productivity generally displayed greater agreement among models than pools such as above ground biomass. Preliminary runs from the six ecosystem models showed a 3.7-fold range in net primary productivity among models with most estimates fluctuating around the range of modern observations. In contrast, mean biomass varied over a 6.2-fold range across models. At a single site where modern biomass is estimated at 10.8 kgC/m2, mean modeled biomass ranged from 4.0 to 21.6 kgC/m2. Change in ecosystem state and the primary driver of change at annual, decadal, and centennial time scales varied by model. Differences in model structure and parameterization regarding disturbance and dynamic vegetation may contribute to lack of coherent, cross-model patterns in drivers of variability across temporal scales. Ongoing modeling efforts have used these results to update model inputs such as plant parameterization and meteorological driver bias corrections with the goal of increasing model agreement in ecosystem states and predicted changes through ti