OOS 50-4 - Hierarchical Bayesian calibration of a dynamic vegetation model using inventory data

Friday, August 11, 2017: 9:00 AM
Portland Blrm 256, Oregon Convention Center
Istem Fer1, Sean McMahon2, Jacqueline E. Mohan3, James S. Clark4 and Michael Dietze1, (1)Earth and Environment, Boston University, Boston, MA, (2)Smithsonian Environmental Research Center, 647 Contees Wharf Road, Edgewater, MD, (3)Odum School of Ecology, University of Georgia, Athens, GA, (4)Nicholas School of the Environment and Earth Sciences, Duke University, Durham, NC

Dynamic vegetation models (DVMs) greatly extend our ability to understand climate impacts on terrestrial biosphere through their process-based representations. Yet, the parameterization of these models continue to govern their performance in capturing empirical observations, and persist to be a source of uncertainty in their predictions. Bayesian calibration of DVM parameters have drawn recent attention for its flexibility in assimilating multiple data-sets and partitioning uncertainties. However, such studies are limited with single site, simple Bayesian modeling applications whereas the real strength of Bayesian approach comes from its ability to incorporate information from multiple sites in a robust framework, and allow and account for stochasticity at different levels. In this study, using a hierarchical Bayesian framework we calibrate a sophisticated process-based model, Ecosystem Demography model version 2 (ED2) with demographic data from a network of inventory plots, and generalize this framework within the Predictive Ecosystem Analyzer (PEcAn) ecoinformatics toolbox. Our sites cover a broad range in the climatic space including a variety of biomes and functional types. We fitted ED2 to each site individually, jointly, and hierarchically considering across site random effects.


As expected, model predictions showed the best match to site specific data with reduced uncertainties after the site-level calibration. The joint fitting of the model resulted in falsely over-confident predictions where the uncertainties were greatly reduced due to more information fed into the calibration, but the model-data agreement decreased with site specific data. Finally, the hierarchical fitting produced the most robust calibration results, meaning, while neither the model-data agreement was not as good as site-level calibration nor the confidence interval around the model prediction was not tightened as much as the joint fitting, model performance was improved on both ends. The vigor of hierarchical Bayesian approach became apparent during the out-of-sample validation where we applied the (individually/jointly/hierarchically) calibrated model to a new site whose data has not been used in its calibration. The hierarchical fitting was the only approach where the model predictions were able to capture the field data at the new site. Future work will focus on analyzing the differences between the sites to determine what the model is missing in its representations that cause it to perform differently across sites.