Researchers apply machine learning to study fossil pollen
They aim to improve the accuracy of fossil pollen identification and discover links to modern plants.
In a study published online this month in Proceedings of the National Academy of Sciences, the researchers explained how they developed and trained three convolutional neural network models – artificial neural networks most commonly used to analyze visual images – to identify fossil pollen specimens from an unknown group of legumes.
“This exciting breakthrough in high-resolution pollen identification improves efficiency in analyzing large datasets and greatly enhances their use in ecological and evolutionary research,” says study co-author Oboh-Ikuenobe, who is interim associate dean of S&T’s College of Engineering and Computing.
Pollen grains preserved in sediments and sedimentary rocks provide records of how different groups of plants have evolved and the environmental factors that have played a role in this evolution. Their identification is based on observations under the microscope by plant scientists and palynologists, like Oboh-Ikuenobe, who study ancient pollen. Correctly measuring and identifying the shape and structure of a pollen grain can be incredibly difficult because there is no identification record for many of these ancient pollens, the researchers say.
The new approach allows researchers to train machine classification models using pollen from living plants and then confirm the plants’ fossil relatives, iteratively learning from each identification to differentiate among specimens that closely resemble one another. For the first time, this method allowed the team to recognize genera within a larger morphological grouping of fossil legume pollen. The trained models classified fossil specimens from western Africa and northern South America dating back to the Paleocene (66-56 million years ago), Eocene (56-34 million years ago) and Miocene (23-5.3 million years ago). The most accurate model used a combination of images from both the exterior and interior of the pollen grain and correctly identified samples with 90.3% accuracy.
Traditional methods such as scanning electron and transmission electron microscopy destroy the sample and are very labor- and time-intensive. Airyscanning, by comparison, is a light microscopy method that can see below the diffraction limit of light and can be used to collect cross-sectional images from inside and outside of the pollen grain without destroying the grain.
“This new method is very useful for samples that are not abundant,” says study co-author Dr. Carlos Jaramillo, a paleontologist at the Smithsonian Tropical Research Institute who earned a master of science degree in geology and geophysics from Missouri S&T. “You can mount the grains on a slide and image them efficiently without damaging the sample.”
Researchers from the University of Illinois at Urbana-Champaign are leading the project, which also includes researchers from the University of California Irvine, Carnegie Mellon University, Smithsonian Tropical Research Institute, Université de Montpellier, University of New Brunswick, and Universidade Federal de Mato Grosso, Brazil. The research was funded by the National Science Foundation.