University of Tasmania
Brown_whole_thesis.pdf (10.8 MB)

Old plants, new tricks : machine learning and the conifer fossil record

Download (10.8 MB)
posted on 2023-05-27, 18:47 authored by Brown, MJM
The palaeobotanical record contains a wealth of information on the evolution and ecology of species, as well as the palaeoenvironment. However, accessing and analysing this information can be challenging. It is not possible to study the ecology of fossils directly, but there are two main inferential approaches to palaeoecology: nearest living relative (NLR) techniques, and physiognomic methods. In NLR techniques, the taxonomic identity of the fossil is key ‚ÄövÑv¨ once the nearest living relatives of a fossil have been identified, we can study the ecology of living species in order to make inferences about the fossil species. However, there are instances where the nearest living relatives of co-occurring fossils are climatically incompatible, which suggests that some extant taxa inhabit different climates to their fossilised relatives. To date, no study has quantitatively analysed this phenomenon in fossils older than the Quaternary (the last 2.6 million years), possibly because of a lack of suitable methodologies. Alternatively, physiognomic methods seek to extract environmental signal that is encoded in the fossil morphology. This approach has been widely used for macroscopic leaf traits, but comparatively less for epidermal characters (in particular, the shape and arrangement of epidermal cells). The epidermis is the interface between the plant and its environment and is responsible for many functions, including gas exchange and mediation of transpiration (via stomata), so there is good reason to believe that there will be links between epidermal traits and environment. However, the calibration of epidermal physiognomic proxies has been hampered by the non-feasibility of undertaking large multivariate studies where each character is extremely time-consuming and laborious to measure, as well as the complex relationships between genetic and plastic variation. In this dissertation, I explore how we can use novel computational techniques (including machine learning) to glean new insights from the fossil record, with a focus on southern conifers (Podocarpaceae, Araucariaceae, Callitroideae). I present two new computational methods (with accompanying R packages) and their palaeoecological applications. In the first chapter, I provide an overview of some of the analytical challenges in palaeoecology and why machine learning techniques are well-suited to solve these problems. I also provide a short review of existing machine learning approaches in palaeoecology. In the second chapter, I present ‚ÄövÑv=hyperoverlap‚ÄövÑv¥, an R package that uses a novel application of a machine learning classifier to evaluate multi-dimensional overlap between point clouds (e.g. occurrence records in climate space). This chapter is published. In the third chapter, I use ‚ÄövÑv=hyperoverlap‚ÄövÑv¥ to quantitatively examine the fossil record of southern conifers and to identify no-analogue associations (those pairs of fossils for which the nearest living relatives inhabit disparate climatic conditions). By quantitatively analysing the climatic overlap in fossil communities, I found that there is significant lability in the thermal niches of southern conifers, but extreme stability in the precipitation niche, implying that future changes to rainfall regime may pose more of a threat to southern conifers than thermal shifts. This chapter is under review, after revision. In the fourth chapter, I present ‚ÄövÑv=epidermalmorph‚ÄövÑv¥, an R package that automates the extraction of leaf epidermal traits from images. As well as trait measurement, this package includes tools for pre-processing, estimations of trait reliability (for any study system) and optimising sampling effort. In the fifth chapter, I use ‚ÄövÑv=epidermalmorph‚ÄövÑv¥ to assess the degree of climatic adaptation in the epidermal cells of Podocarpaceae. I found some evidence for adaptive significance of stomatal index and cell wall undulation, but there were no viable proxies for either tree height or climatic conditions, suggesting that the functional variability in Podocarpaceae leaves is more likely to be plastic, rather than hard-coded. In the final chapter, I summarise this thesis, discuss the challenges and limitations to applying sophisticated computational techniques to the palaeobotanical record and suggest potential avenues for future research.



School of Natural Sciences

Publication status

  • Unpublished

Rights statement

Copyright 2022 the author

Repository Status

  • Restricted

Usage metrics

    Thesis collection


    No categories selected


    Ref. manager