Assessing the long-term performance of terrestrial ecosystem models in northeastern United States: linking model structure and output
Abstract:Terrestrial biosphere models are being used to forecast ecosystem response to future climate change. However, the predictions of these models do not agree, and their ability to accurately represent decadal- and longer-scale ecological processes has rarely been tested.
Here we investigate how the structure of the terrestrial biosphere models affects their ability to accurately simulate vegetation dynamics over the past 1000 years. Six models and model variations are involved, including the Ecosystem Demography 2 model (ED2), the Community Land Model (CLM4.0, CLM4.5, CLM4.5-dynamic vegetation), and the Lund-Postdam-Jena models (LPJ-GUESS and LPJ-wsl). Using common paleoclimatic drivers and modeling protocols, we simulated vegetation changes in the northeastern US with these models for the past 1000 years. We compared the model outputs with paleoecological/historical benchmark datasets (paleo-vegetation data reconstructed from fossil pollen records and pre-EuroAmerican vegetation data from the Public Land Survey and Town Proprietor Surveys) to evaluate model performance. We characterized the models based on how they represented photosynthesis, water relations, soil, biogeochemical cycling, and vegetation dynamics.
Models with similar structures behaved more similarly over time; we found that model attributes of both fast (e.g. photosynthesis) and slow processes (e.g. vegetation dynamics) affect model long-term predictions. However, the effect of model attributes varies among regions and across time.
In addition to model structure, parameter uncertainty among models also appears to contribute to the differences in model output. We are currently working on assimilating paleoecological data into a subset of the terrestrial biosphere models to constrain the key model parameters and state variables, and in turn, to better evaluate the effect of model structure alone on model performance.