Wed, Aug 04, 2021:On Demand
Background/Question/Methods
Understanding how plant communities respond to rapid warming after the Last Glacial Maximum (c. 21,000 years ago) provides critical insight to predict plant community responses to ongoing climate change. Fossil pollen data from sediment cores makes it possible to infer past distribution and relative abundance of plant taxa. Increasingly large and standardized online repositories of fossil pollen records provide data from many sites to estimate tree occurrences over large geographic areas and temporal domains. Traditionally, spatiotemporal pollen abundance is modeled using generalized linear or additive modeling approaches. However, these approaches are often computationally inefficient, such that modeling multiple taxa across large regions and time periods while accounting for uncertainty is not possible. In this study, we estimate, with uncertainty, pollen relative abundance for 12 tree taxa across eastern North American over the past 21,000 years. This novel modeling approach allows for the formalization of spatiotemporal dependence, making it possible to infer relative abundance of pollen, even at locations and times for which we do not have pollen count data, by borrowing information across the full network of pollen records.
Results/Conclusions We model pollen relative abundances for eastern North American tree taxa using a novel Bayesian hierarchical spatiotemporal model for multinomial count data. The sampler yields a reduced computational burden compared to other Bayesian spatiotemporal approaches by taking advantage of Gibbs sampling. We found that pollen distributions shifted contemporaneously with climate during warming events, particularly between 15k and 10k years ago. Overall, the model predicts changes in pollen abundances consistent with traditional estimates of pollen distribution and relative abundance. However, the spatiotemporal dependency in the model yields novel pollen occurrences in locations and time periods not observed in previous pollen models. We discuss the implications of these discrepancies in terms of model behavior and pollen record biases. Additionally, the model reveals important spatiotemporal trends in uncertainty. As expected, uncertainty increased with increasing distance from a pollen sample in space and time. Addressing uncertainty in estimates of plant community characteristics is critical to further our understanding of historic biogeographic processes, such as climate-induced tree migration. Furthermore, complete spatiotemporal abundance maps can be incorporated into landscape or ecosystem models by providing information on past vegetation composition, along with robust uncertainty estimation, across large regions and tens of thousands of years into the past.
Results/Conclusions We model pollen relative abundances for eastern North American tree taxa using a novel Bayesian hierarchical spatiotemporal model for multinomial count data. The sampler yields a reduced computational burden compared to other Bayesian spatiotemporal approaches by taking advantage of Gibbs sampling. We found that pollen distributions shifted contemporaneously with climate during warming events, particularly between 15k and 10k years ago. Overall, the model predicts changes in pollen abundances consistent with traditional estimates of pollen distribution and relative abundance. However, the spatiotemporal dependency in the model yields novel pollen occurrences in locations and time periods not observed in previous pollen models. We discuss the implications of these discrepancies in terms of model behavior and pollen record biases. Additionally, the model reveals important spatiotemporal trends in uncertainty. As expected, uncertainty increased with increasing distance from a pollen sample in space and time. Addressing uncertainty in estimates of plant community characteristics is critical to further our understanding of historic biogeographic processes, such as climate-induced tree migration. Furthermore, complete spatiotemporal abundance maps can be incorporated into landscape or ecosystem models by providing information on past vegetation composition, along with robust uncertainty estimation, across large regions and tens of thousands of years into the past.