Abstract
Grass (Poaceae) pollen is largely overlooked in investigations of grassland evolution because the pollen of most species cannot be differentiated using traditional optical microscopy. However, the combination of superresolution microscopy and deep learning enables the capture and quantification of distinct morphological variation across the pollen of grass species. Using a semisupervised deep-learning strategy, we trained convolutional neural networks (CNNs) on superresolution images of known modern and unlabeled fossil grass pollen. We used learned CNN features to estimate taxonomic diversity and discriminate between C
3
and C
4
grass species—applying Shannon entropy as a measure of morphological variability and a gradient-boosted decision tree as a classifier of photosynthetic type. We validated the method on modern grass pollen assemblages and then applied our trained models to fossil assemblages from a 25,000-y lake-sediment record from Mt. Kenya, correlating shifts in grass diversity through time with changes in atmospheric CO
2
concentration and proxy records of local temperature, precipitation, and fire occurrence. Our data show that the species diversity of these montane grasslands declined substantially between 21,500 and 16,000 years ago, coincident with the most severe regional cooling of the last ice age, and that the fraction of C
4
grasses gradually decreased after the late-glacial–Holocene transition, associated with higher temperatures and decreasing fire activity. Our results demonstrate that CNN features of pollen morphology can advance palynological analysis and enable robust estimation of changes in grass diversity and C
4
grass abundance in ancient grassland ecosystems.