OOS 47-7
Reconciling uncertainty in long term data and uncertainty in long-term projections from models

Wednesday, August 12, 2015: 3:40 PM
314, Baltimore Convention Center
Jason McLachlan, Biological Sciences, University of Notre Dame, Notre Dame, IN
Michael Dietze, Earth and Environment, Boston University, Boston, MA
Christopher J. Paciorek, Statistics, University of California, Berkeley, Berkeley, CA
Jaclyn Hattala Mathes, Geography, Dartmouth College, Hanover, NH
Ann Raiho, Biological Sciences, University of Notre Dame
John W. (Jack) Williams, Geography, University of Wisconsin, Madison, Madison, WI

 Ecosystem responses to climate and land-use drivers at the macrosystem scale are dominated by "slow" processes including species range shifts, changing disturbance regime, and the slow turnover of carbon pools like wood and soil carbon. Although such processes occur at time frames beyond the typical planning horizon, they are difficult to reverse and thus have special policy importance.

Information for anticipating slow changes comes primarily from two sources: paleo/historical data, and mechanistic community/ecosystem models. Theoretically the strengths and uncertainties in these two approaches to understanding slow ecosystem dynamics should be complementary. In practice, due to both technical and cultural obstacles, inferences from the paleoecology and ecosystem modeling worlds are arrived at largely independently.

To illustrate the significance of this mismatch and to map a way forward for joint inference from paleodata and models, we examined the community and ecosystem history of southern New England over the past 3000 years using ecosystem models and paleoecological data. The models include the suite of land-surface models used in the CMIP5 global modeling effort and a local forest gap model, LINKAGES. The data are from a space-time statistical model of forest composition based on a network of fossil pollen sites. We asked: Which aspects of long-term community and ecosystem change are captured in the data and which are captured by the models? Are the trends in the data consistent with those of the models? Does joint inference from the data and the models provide better understanding of slow processes?


Our statistical model illustrates clear trends in composition over time, including a >20% increase in evergreen trees in some places. Such changes have significant biophysical and biogeoochemical consequences, but these trends are not captured in the models. Bayesian data assimilation of paleodata into ecosystem models provides joint inference from data and models. Data assimilation constrains model output to values more consistent with observed data, which will, in principle, help with improved forecasts. At this early stage of reconciling paleoecological data with long-term model runs, however, model-data fusion mainly serves to highlight the mismatch between our understanding of the mechanisms driving slow ecosystem dynamics and the empirical record of these dynamics.