Typological and morphometric analyses of phytoliths in modern plants: Developing tools for archaeological and palaeoecological research

Phytoliths are microscopic silica bodies produced by plants that can persist in soils and sediments long after plant decay. Their durability makes them a powerful line of evidence for reconstructing past vegetation, environments, and human–plant interactions. Yet their diagnostic value, their taxonomic and ecological signals, remains difficult to assess due to morphological similarity across taxa and limited reference data. In this context, taxonomic signal refers to the extent to which phytolith morphometry reflects plant identity (e.g. subfamily, tribe, or genus), while ecological signal refers to whether variation in phytolith shape or size corresponds to environmental conditions such as temperature, precipitation, or tree cover. Assuming that these signals can be reliably detected and quantified in modern plants, they can then be applied to fossil assemblages to infer past vegetation composition and habitat conditions. This thesis addresses these challenges by developing phytolith reference datasets and applying morphometric analysis (measuring phytolith shape and size) and machine learning approaches to evaluate and improve phytolith classification across both modern and fossil contexts.

Three main objectives guided this work: (1) establishing a modern reference framework for temperate–cool dicotyledonous angiosperms, a taxonomically underrepresented group in phytolith studies; (2) developing and testing morphometric and classification approaches for diagnostic morphotypes from palms, grasses, and cereals; and (3) applying these approaches to fossil assemblages spanning the Eocene to Miocene (~56–5 Ma) from North America, Argentina, and Turkey to infer taxonomic identity and palaeoecological conditions.

Across five papers, this thesis integrates typological phytolith reference work, 2D and 3D morphometric analysis, and machine learning to assess taxonomic and ecological signal in a range of phytolith morphotypes (SPHEROID ECHINATE; PETASOID ECHINATE; various Grass Silica Short Cell Phytoliths; ELONGATE DENDRITIC). The results show that taxonomic signal varies across plant groups (e.g. grasses, palms, dicotyledons) and taxonomic levels (e.g. subfamily, subtribe, tribe, genus, species, subspecies). Morphometric analysis and classification methods enable subfamily- and tribe-level classification for grasses using GSSCP, species-level resolution in cereals using ELONGATE DENDRITIC, and more limited taxonomic resolution in palms using SPHEROID ECHINATE and PETASOID ECHINATE.

Ecological signal is morphotype-dependent: some morphotypes show correlations with modern climate data, while others do not. Notably, CRENATE forms exhibited clear ecological signal.

Fossil analyses demonstrated the feasibility of applying trained models to deep-time assemblages. However, fossil predictions were strongly influenced by model architecture—that is, the specific design and structure of the classification approach, including the type of algorithm used (e.g. linear discriminant analysis, random forest), the input features selected, and how classes are represented. This variation highlights the importance of careful methodological control and expert oversight. Overall, this thesis demonstrates both the potential and the limitations of automated approaches, and provides a foundation for future studies to build upon—both within and outside of phytolith research.