COS 41-1 - Benchmarking predictive models against long term ecological data

Tuesday, August 8, 2017: 8:00 AM
B113, Oregon Convention Center


Eliandro Tolentino Tavares, Boston University; Istem Fer, Boston University; Michael Dietze, Boston University


Long term ecological research (LTER) data provide us with a broad range of knowledge about ecosystem functioning and services and are widely considered to be some of the most valued and influential in the study of ecology. However, such data are not being used in any systematic manner to calibrate or validate the ecosystem models that test our current understanding and predictions based on probability. In this study, we use Hubbard Brook LTER site data to validate and calibrate over a dozen ecosystem models using the Predictive Ecosystem Analyzer (PEcAn), an informatics framework for integrating data with models. PEcAn provides tools that facilitate data ingest and allow users to set up automated benchmarks that are easily repeatable and extensible to other models. This allows models to be compared and improvements to be tracked through time. In contrast to conventional multi-model / data inter-comparisons, we also leverage PEcAn’s tools for uncertainty analysis and propagation to assess model confidence, and PEcAn’s Bayesian calibration tools to better isolate structural and calibration errors.


As a first benchmark, we evaluated the models' ability to reproduce Hubbard Brook’s classical 1965 watershed deforestation experiment. Here we used 40-plus years of biomass data to compare model performances in predicting ecological growth patterns of both control and experimental stands. To reduce biases due to model parameterization differences, we calibrated the models using the Hubbard Brook `bird study’ forest plot data. The post-calibration performances of the models showed an improved agreement with the data while the uncertainties around model predictions were also reduced. Work in progress is focused on identifying the structural features associated with better performance and extending validation beyond biomass trajectories to watershed discharge and changes in species composition. Our study highlights the importance of using long term ecological data to improve our model predictions of how forest diversity and ecosystem services will respond to the interactive effects of climate change and forest management.