You are here
Harvard Forest Symposium Abstract 2011
- Title: Assessing long term trends in carbon fluxes at Harvard Forest and the effects of parameter uncertainties in ecological forecasting through model-data fusion
- Primary Author: Trevor Keenan (Harvard)
- Additional Authors: Eric Davidson (University of Maryland - Center for Environmental Science); Antje Moffat (Max Planck Institute for Biogeochemistry); J. William Munger (Harvard University); Andrew Richardson (Northern Arizona University); Kathleen Savage (Woods Hole Research Center)
- Abstract:
Our ability to model terrestrial carbon fluxes, and forecast how those fluxes will respond to a changing climate, is not complete without a clear quantification of uncertainties in model predictions. Model-data fusion is a powerful framework by which to use both model and data uncertainties to give an integrated view of our confidence with regard to ecosystem processes and states, and our ability to model them. As we move into a data rich era in ecology, more researchers are moving towards model-data fusion as a means to address this issue.
In the presented study, we used 17 years of flux data and ancillary measurements from the Harvard forest in the northeastern United States to combine a forest carbon cycle model of intermediate complexity with 11 different data streams through Markov chain Monte Carlo model-data fusion. We present novel techniques that account for systematic biases in ecological observations, and allow for the exploration of an essentially infinite parameter space. An artificial neural network ensemble approach was used to benchmark the optimized model, assess model structural error, and separate biotic from climatic drivers of long-term trends in net ecosystem carbon uptake at Harvard forest. The optimized model and the related uncertainties were then used to quantify uncertainty in ecological forecasting, in particular under potential climate change scenarios.
Our multiple constraints approach constrained a large number of model parameters, and allowed for the identification of redundant parameters that had no effect on model performance. Modeled uncertainty fell well within the error range of each data stream. The combined model-data fusion artificial neural network approach successfully showed that recent trends of increased uptake at Harvard forest are largely independent of climate. Future projections of parameter uncertainty highlighted an extreme sensitivity of the long-term forest carbon cycle model predictions to initial conditions. Uncertainties in simulated fluxes and carbon stocks were relatively stable until around 2050. After 2050, however, model states and fluxes became effectively unpredictable. This casts a shadow of doubt over the ability of current terrestrial carbon cycle models to make reliable long-term predictions, and highlights further data needs to better constrain current models.
- Research Category: Ecological Informatics and Modelling
Forest-Atmosphere Exchange
Physiological Ecology, Population Dynamics, and Species Interactions
