Quantifying uncertainty in spatio-temporal forest composition changes inferred from fossil pollen records

Background/Question/Methods

Vegetation interacts with the environment at multiple spatial and temporal scales. Vegetation is affected by changing climate, but can also drive regional climatic change. Understanding past compositional changes provides insight into the relationships between climate and vegetation at time scales greater than those described by modern ecological surveys. Vegetation reconstruction relies predominantly on fossil pollen data obtained from sedimentary lake cores, where stratigraphic changes in pollen proportions allow us to infer changes in species composition over time. However, system uncertainty arises due to the complex relationship between pollen and surrounding vegetation, deposition rates, the uneven distribution of fossil pollen sites, and the age-depth model.

To account for this uncertainty, we use a Bayesian hierarchical model to link vegetation composition to fossil pollen data via a dispersal model. Model calibration relies on the use of Public Land Survey forest composition data, and a corresponding pre-settlement pollen sample determined through expert elicitation. Parameters governing the processes linking pollen and vegetation are estimated in the calibration phase, and are subsequently used in the prediction phase to generate spatially explicit maps of species distributions across the upper Midwest over the last 2500 years, with robust uncertainty estimates.

Results/Conclusions

From the calibration model, we obtain estimates of differential production and dispersal values for each pollen taxon. These estimates are then used to estimate pollen assemblages at each lake with uncertainty, and improve upon the assumption that the abundance of nearby vegetation is representative of deposited pollen. Preliminary results show that the uncertainty associated with the pre-settlement calibration sample selection is less than that associated with the age-depth model, implying the need for improved age-model construction and reporting practices. Estimates of the process parameters (and their uncertainty) obtained from the calibration model are subsequently used in the prediction phase to estimate spatio-temporal changes in vegetation composition, and the uncertainty associated with these changes. Initial results identify compositional shifts in vegetation over the last 2000 years, challenging the assumption of pre-settlement forest stationarity often used in ecosystem modelling. In particular, spatial maps indicate changes in the relative abundance of oak in southern Minnesota, as well as changes in the distributions of many other key taxa including pine, birch, maple, and spruce. Our novel spatio-temporal composition and abundance estimates will be used to improve the forecasting capabilities of ecosystem models.