Super-resolution microscopy and a trained model read differences invisible to a human eye in grass pollen from a Mt Kenya lake core, tracking diversity and the C3–C4 balance since the last ice age.

Pollen is the workhorse of palaeoecology. Grains preserve for tens of thousands of years in lake sediment, they are abundant, and their shape identifies the plant that produced them — which makes a sediment core a record of what grew nearby and when. The method has one conspicuous blind spot: grasses.

Grass pollen grains are nearly featureless under a standard light microscope and nearly identical to one another. In a grassland record — precisely where the vegetation history matters most — the pollen arrives in quantity and says almost nothing. Electron microscopy reveals more, but it is expensive and slow enough to be impractical at the scale a sediment core requires.

A study in the Proceedings of the National Academy of Sciences, led by University of Illinois Urbana-Champaign plant biology professor Surangi Punyasena, reports a way around the problem, and uses it to read 25,000 years of change in eastern African grasslands from a single lake-bottom core on Mt Kenya.

Seeing more, faster

The approach pairs super-resolution microscopy — optical techniques that push past the normal diffraction limit — with machine learning trained to detect differences among grains that a human observer cannot reliably see.

“You get close to electron microscopy quality, but the process is much faster, much easier,” Punyasena said. That trade is what makes the method usable: a core contains thousands of grains, and a technique that is marginally less detailed but an order of magnitude quicker changes what questions can be asked.

She frames the underlying constraint plainly. “As paleobotanists and paleontologists, we’re restricted to working with the morphology of pollen grains,” she said — “but the level at which we were able to analyze it before machine learning was limited by human ability.” The information was in the grains all along; the limit was in the eye reading them.

Marc-Élie Adaimé, formerly a doctoral student in Punyasena’s group and now a postdoctoral researcher at the Smithsonian’s Office of Digital and Innovation, describes what the model does without overstating it. “That’s essentially the definition of machine learning: for a computer model to be able to learn patterns and apply that learning to new data without being given explicit rules,” he said.

Diversity without identification

The important limitation is built into the method: the model cannot identify individual grass species. What it does is estimate species diversity in a mixed sample — how many distinct kinds are present, not which ones.

That is a weaker output than a species list, and it is still a substantial advance over the near-total silence grass pollen previously offered. The validation is the part that counts: tested on samples of known composition, the model’s diversity estimates closely tracked the actual diversity. A model that reproduces a known answer has earned some trust when applied to an unknown one.

What the core shows

Applied to the Mt Kenya record, the method produced two results that behave quite differently.

The first concerns diversity, and it tracks climate closely. “We saw how pollen diversity was substantially lower during the last ice age, especially between about 21,000 and 18,000 years ago,” Adaimé said. Diversity then increased, coinciding with rising carbon dioxide and temperatures. The window he names is the Last Glacial Maximum — colder, drier and lower in atmospheric CO₂ than any period since — and a grassland losing diversity under those conditions is what one would expect.

The second result is the one that does not fit. The balance between C3 and C4 grasses showed no obvious association with either CO₂ or temperature.

Why that is surprising

C3 and C4 refer to two different photosynthetic pathways. C4 photosynthesis concentrates carbon dioxide around the enzyme that fixes it, which makes C4 plants more efficient in hot, dry, low-CO₂ conditions. C3 plants do better where CO₂ is plentiful and temperatures are moderate. The relationship is well enough established that the C3–C4 balance in a fossil record is routinely treated as a proxy for past climate and CO₂.

So the direction of change here is roughly as predicted — C4 grasses were more prevalent in the final stretch of the ice age and declined gradually as C3 grasses took over. “So, the fraction of C4 grasses decreased, and C3 grasses took over as the climate got warmer and CO2 got higher,” Adaimé said.

What is absent is a clear statistical association with either variable. A gradual shift in the expected direction without a tight link to the drivers that should produce it suggests something else is also at work — local hydrology, fire regime, grazing pressure, or soil change, all of which shape grassland composition independently of global climate.

The researchers also raise a methodological possibility that would complicate the proxy itself: C4 grasses may invest fewer resources in individual pollen grains, which would mean the pollen record systematically under- or over-represents them relative to their abundance on the landscape. The release is explicit that more work is needed to confirm this. If it holds, it has implications well beyond this core, since C3–C4 ratios from pollen are used widely.

A recent ecosystem

The authors place grasslands in deep time. “Open grasslands are a relatively recent ecosystem in Earth history, with open-habitat grasses present in the Eocene, about 40 million years ago,” they write in their report.

Forty million years is recent on a geological scale, and grass-dominated ecosystems are younger still in their modern extent. They now cover a substantial fraction of the land surface, support most of the world’s grazing livestock, and in eastern Africa form the setting in which a good deal of human evolution took place. A method that makes their history legible addresses a gap that has been conspicuous precisely because the ecosystem is so important.

What is next, and what is disclosed

Adaimé’s closing note is about scope rather than this core. “It is satisfying to see that there is so much more information to be unlocked,” he said.

The obvious next step is other cores. One site on one mountain establishes that the method works and gives one region’s history; whether the diversity pattern and the loose C3–C4 relationship hold elsewhere in eastern Africa, or anywhere else, is unknown. The model would also need retraining on reference material from other floras, since it has learned the grasses of one region.

Funding came from the U.S. National Science Foundation, and the authors declare no competing interest. The release says Punyasena led the research and names Adaimé without specifying author order or a corresponding author, so the endnote below carries the paper title alone. It is dated 5 October 2026 against a publication date of 2 October.

The broader point is one that recurs as machine learning reaches older datasets: the limiting factor in a good deal of palaeontology has not been the fossils but the resolution of human pattern recognition applied to them. Collections sitting in drawers may hold more than anyone has been able to read.

---

Notes

1. “Deep learning of fossil pollen morphology reveals 25,000 years of ecological change in eastern African grasslands.” Proceedings of the National Academy of Sciences, 2 October 2026. DOI: 10.1073/pnas.260737212.
2. University of Illinois Urbana-Champaign. “Pollen study reveals the hidden history of grasses.” EurekAlert!, 5 October 2026. https://www.eurekalert.org/news-releases/1146438.