RECONSTRUCTING FOSSIL GRASS POLLEN DIVERSITY USING SUPERRESOLUTION MICROSCOPY AND MACHINE LEARNING

Grass-dominated ecosystems cover one-third of Earth’s land surface and influence key biogeochemical processes. Understanding the environmental factors controlling grass diversity in Earth’s history is useful for projecting the response of grasslands to future environmental change. Unfortunately, the utility of fossil grass pollen is often limited by low taxonomic resolution, necessitating the use of alternative proxies such as δ¹³C, phytoliths, and dispersed cuticles. To increase the taxonomic information that can be derived from fossil grass pollen, we applied two methods new to palynology: structured illumination superresolution microscopy (SR-SIM) and convolutional neural networks (CNNs). SR-SIM can image taxonomically diagnostic nanoscale features on and within the grass pollen wall and CNNs are a powerful new class of machine-based image classifiers. To train our CNN, we imaged pollen from 58 species of modern grasses, mostly from Eastern Africa. We assessed the ability of our classification system to capture grass diversity by comparing grass pollen diversity in lake surface sediments to known grass diversity along an elevation gradient of Mt Kenya, an area with robust paleovegetation and paleoclimate records. Preliminary results indicate the CNN can successfully classify pollen to species with a ~62% accuracy from averaged predictions of fixed-size axial crops and grain size measurements. Marginal improvement is observed when combining two different levels of axial context. From this we suspect performance gains from more levels of spatial context as well as additional sample metadata. The effectiveness of features learned for species discrimination suggests they may be one of the best proxies for quantifying the taxonomic diversity of recent (i.e., <25,000 yr BP) fossil grasslands. We will apply this method to the reconstruction of changes in grass pollen diversity in a 25,000 yr sediment core from Lake Rutundu and compare these shifts in diversity to that of the dispersed cuticle, terrestrial biomarker, and grass pollen δ¹³C records from the lake.