DECIPHERING THE IDENTITY OF EOCENE GRASSES USING AUTOMATED, QUANTITATIVE IMAGE ANALYSIS OF FOSSIL PHYTOLITHS

Fossil grasses (Poaceae) have the potential to be key indicators of the types of ecosystems in which they lived. Many extant grass clades have strong niche specialization (i.e., bamboos in forests, C4 lineages in warm, arid conditions). Untangling the Cretaceous-Cenozoic biogeographic history of Poaceae and the evolution of grass-dominated ecosystems has long relied on fossil pollen and scant macro-remains, most of which have been difficult to place within the family. Recently, plant silica microfossils (phytoliths) preserved in paleosols and sediment have been found to constitute a rich, alternative source of data on ancient grasses. For example, analysis of phytolith assemblages from the Americas and eastern Mediterranean has helped fill gaps in the early Cenozoic history of the group. This work has shown that, prior to the Oligocene, grasses were relatively rare and consisted primarily of forest-associated taxa. However, as these identifications are based on traditional classification schemes that suffer from being semi-quantitative, inherently subjective, and not strongly linked to phylogeny, they remain untested.

Our work seeks to solve these problems and refine identification of ancient grass phytoliths—and our ability to interpret the paleoecological data they represent—through quantitative methods. Specifically, we have developed new methods for quantifying and analyzing detailed 3-D surfaces of isolated phytoliths through geometric morphometrics. We have also produced an extensive 2D image library of modern grass phytoliths to classify fossils using artificial intelligence. Both approaches can successfully and robustly classify fossil phytoliths into grass subclades (e.g., subtribe), although precision varies depending on subclade and method. We therefore use a mixed methods approach to data analysis which combine our new quantitative methods with traditional morphotype-based analysis. Further, we employ phylogeography to improve inference. We report on new applications of our 3-D methods to Eocene grass phytoliths from central North America and Turkey that were previously described as likely forest grasses. Results from this analysis are used to better understand taxonomic diversification of the grasses as well as Eocene paleoenvironments.