Project title:
Transforming a 100-yr old discipline: PollenGeo, a first in digital palynology
Mentor name:
Location of internship. Will mentor be at this location?
CTPA, Panama. Mentor will be at this location
Project summary and objectives
Building an extant and fossil Neotropical palynological image training set that will revolutionize the use of pollen data for the Neotropics and ultimately serve as the model for microfossil data analysis globally
The problem to be solved: Palynology is a century-old practice. It is used in a variety of research areas, from pollination ecology to paleovegetation reconstructions and oil exploration. The core of any routine palynological analysis is the task of counting pollen grains. Millions of pollen grains need to be found, identified, counted, and compared to thousands of pollen grains from extant and fossil plants. How do we accomplish this? A palynologist will spend hundreds of hours scanning a microscope slide to find isolated pollen grains that need to be identified. This task is especially challenging in tropical settings because there are thousands of potential plant species which serve as pollen sources. It takes years of training to master tropical palynology and yet, the training only allows the recognition of a small number of species. This problem is compounded when dealing with the fossil record as there are thousands of fossil pollen types that belong to extinct species. Another major caveat is that individual identifications are rarely confirmed by other experts, as the coordinates of every grain on a particular microscope slide are not annotated. The reproducibility of counts is not evaluated, and identifications are often trusted at face value.
The proposed solution: Digital palynological slides + neural networks
Neural networks and the capability to digitize an entire microscope slide at multiple focal planes have made it possible for palynology to enter a new era. We demonstrated that this is feasible in a pilot study published in PNAS (Romero et al 2020) and we are scaling it up to the entire plant kingdom.
The PollenGeo scheme has three components: 1) A training set is produced. 2) A target unknown grain is digitized, and 3) A neural network algorithm compares the unknown grain with the training set to yield the probability that the unknown grain belongs to each of the taxa in the training set.
The most critical step in using neural networks is building the training set. Smithsonian is one of the few institutions that can provide the appropriate training set for Neotropical fossil and extant pollen. We have the largest palynological collection of the Neotropics in the world, the Graham Pollen Collection, with 17,800 species represented in ~25,000 slides (some species are represented by more than one slide). We also have the best collection of fossil pollen from the Neogene of Amazonia with nearly all the 700 fossil species already described. We also have decades of expertise studying tropical pollen, both recent and fossil, to properly train and validate a neural network.
Mentorship goals including benefits to the intern
Learning to digitize pollen and spores, and the analysis of microscope images
Intern’s role, desired background and expected products
Participate in a number of roles related to the digitization project, including helping with palynological boxes, finding grains at the microscope, using high-end microscopes to photograph pollen specimens, and using software (NIS, ZEN, Omero, Fiji) to optimize and analyse images
What are the regularly held occasions for group discussions, attendance at lectures, career counseling, and other educational and experiential opportunities for your interns?
Once a week
