Models for ecological succession are widely used to predict changes in forest structure and composition far into the future. But these long term predictions are often calibrated with short term data, and data are not often incorporated into the ecological forecasting workflow itself. Data assimilation takes the forecasted model state, updates it with data, then forecasts from the updated state. This workflow keeps the model in sync with the data. We hypothesize that assimilating long term data into a model for ecological succession will improve long term predictions. At Harvard Forest, it was hypothesized that red maple (Acer rubrum) would come to dominate stand-level biomass because of its ability to be both an early and late successional species. However, modern day Harvard Forest is dominated by red oak (Quercus rubra) and red maple has begun to decline. We implemented data assimilation within the PEcAn (Predictive Ecosystem Analyzer) framework to make our workflow generalizable to many types of models and data. Here, we used a forest gap model (LINKAGES) and two different long term datasets for data assimilation, tree ring estimated species biomass and fossil pollen estimated fractional species composition.
Results/Conclusions
By assimilating tree ring estimated species biomass at Harvard Forest into LINKAGES, we improved modeled species biomass and composition. Without data assimilation, LINKAGES predicted that yellow birch (Betula alleghaniensis) and white pine (Pinus strobus) would dominate with high above ground biomass (~ 100 kgC/m2). With long term data assimilation, we were able to constrain LINKAGES to correctly forecast red oak as the dominant species with more realistic biomass (~ 25 kgC/m2). Our assimilation workflow allows us to estimate a species covariance matrix that quantifies the interactions misrepresented by LINKAGES. By assimilating fossil pollen into the same model, we discovered that species lag climate changes at different rates and that species that are highly correlated during succession are not necessarily correlated over millennia. This finding suggests that the processes that drive competition between species in long term forecasts may not be well constrained by models that are calibrated with relatively short term (succession-scale) data.