Comparing model-data fusion methods for parameter estimation and state variable prediction using a forest gap model

Models represent our hypotheses of real ecological systems, and consequently, they may incompletely or incorrectly quantify ecosystem dynamics, leading to imprecise predictions. One way to improve model performance is via model-data fusion, a class of methods that combine model predictions and observations. One model-data fusion method is parameter optimization, in which model simulations are conducted with different parameter combinations and predicted state variables are compared post hoc to data. The parameter combinations that lead to the most accurate predictions are sselected as the “true” parameter values. Another method is data assimilation, which involves iteratively adjusting an ensemble of model predictions with data. When parameters are varied in each simulation and updated along with the state variables via ensemble weighting, both parameters and state data assimilation are performed simultaneously.

We will show results from employing two model-data fusino methods using the forest gap model LINKAGES and reconstructions of taxon-level aboveground biomass (AGB). We compare parameter optimization and data assimilation in terms of their ability to reasonably estimate model parameters and forecast AGB. We hypothesize that parameter optimization will make more accurate forecasts of AGB than data assimilation because the model parameters are forced to take on values that are consistent with the reconstructed AGB, regardless of the hypothesized covariance structure between parameters represented in the process model. On the other hand, we hypothesize that data assimilation will yield more ecologically realistic parameter combinations than parameter optimization because data assimilation uses the hypothesized covariance in model parameters to maintain realistic parameter space.