For generations, the fossil record of Earth's grasslands lay locked inside pollen grains too similar to tell apart — a library written in a script no one could read. Now, by pairing super-resolution microscopy with machine learning, researchers at the University of Illinois and the Smithsonian have taught algorithms to see what human eyes could not, decoding 25,000 years of African grassland history from sediment pulled from the floor of a Kenyan lake. The work is less a technological novelty than a philosophical reckoning: the past was always speaking; we simply lacked the grammar to listen.
Machine learning unlocks 25,000 years of grass evolution hidden in pollen
Suddenly you have a scalable way to read the history of grasslands globally.
Mark
Why does it matter that we can now tell grass pollen apart? Isn't that pretty specialized?
Mimi
It matters because grass pollen is everywhere in sediment cores—you can have millions of grains in a tiny sample. But until now, scientists couldn't reliably say what species they were looking at, so all that data was essentially noise.
Luke
But the model doesn't actually identify individual species, right? It estimates diversity. That's different from knowing what's there.
Mimi
True. It's a statistical estimate based on patterns the machine learning model learned. But when they tested it on samples with known compositions, the estimates tracked the actual diversity very closely.
Mark
So what did they actually learn from the Mount Kenya core?
Mimi
That grass diversity crashed during the coldest part of the last ice age, when CO2 was extremely low. Then it recovered as temperatures and CO2 rose. And C4 grasses—the more efficient ones—dominated during the ice age but gradually lost ground to C3 grasses as the climate warmed.
Luke
That's interesting, but I want to be careful here. They're looking at one site over 25,000 years. That's a single data point, geographically speaking. You can't generalize that to all grasslands.
Mimi
No, you can't. But it's a proof of concept. If this method works on Mount Kenya, it can work on cores from other places. That's the real value—suddenly you have a scalable way to read the history of grasslands globally.
Mark
What's the hypothesis about why C4 grasses declined?
Mimi
They think C4 grasses might invest less in individual pollen grains than C3 grasses do, which could give them an edge in harsh, dry conditions. But Adaimé was careful to say that's still speculative. The evidence supports it, but they need more work to confirm it.
Luke
And the machine learning model itself—how confident should we be in it?
Mimi
It was trained on images from identifiable grass species, so it learned real patterns. But it's only as good as the training data. If there are grass species with unusual pollen morphology that weren't in the training set, the model might miss them.
Mark
What comes next?
Mimi
They want to refine the method and eventually apply it to pollen from all land plants, not just grasses. Imagine being able to read 25,000 years of plant history from a single core.
O Pulso
- Grass pollen grains are so visually uniform that decades of fossil records have been scientifically stranded — ecologists could see the archive but not read a word of it.
- Super-resolution microscopy revealed previously invisible surface textures and cell-wall differences in pollen grains, giving the AI something real to learn from.
- A machine-learning model trained on those microscopic signatures can now distinguish grass species and even separate C3 from C4 photosynthetic types — a distinction that tracks climate history directly.
- Applied to a 25,000-year sediment core from Mount Kenya, the method exposed a clear ecological pulse: diversity collapsed during the Last Glacial Maximum, then rebounded as CO2 and temperatures rose.
- The team is now moving to extend the approach to all land plant pollen and spores, a step that could rewrite the entire field of paleobotany.
For generations, the fossil record of Earth's grasslands lay locked inside pollen grains too similar to tell apart — a library written in a script no one could read. Now, by pairing super-resolution microscopy with machine learning, researchers at the University of Illinois and the Smithsonian have taught algorithms to see what human eyes could not, decoding 25,000 years of African grassland history from sediment pulled from the floor of a Kenyan lake. The work is less a technological novelty than a philosophical reckoning: the past was always speaking; we simply lacked the grammar to listen.
For decades, paleobotanists studying ancient sediment cores ran into the same wall: grass pollen grains look nearly identical under a standard microscope. Unlike other flowering plants, whose pollen carries distinctive shapes and textures, grass pollen offered almost nothing to distinguish one species from another — leaving vast fossil archives effectively unreadable.
Surangi Punyasena of the University of Illinois Urbana-Champaign and her then-doctoral student Marc-Élie Adaimé, now at the Smithsonian, attacked the problem from two directions at once. Using super-resolution microscopy — which reconstructs laser-scattered light to reveal surface detail at near-electron-microscope clarity — Adaimé spotted subtle patterning differences and variations in cell-wall thickness that previous researchers had missed. He then trained a convolutional neural network on those images, teaching it to recognize what distinguished one grass type from another and to estimate the overall diversity of ancient grassland communities.
The model proved capable of something especially valuable: separating C3 and C4 grasses, which differ in how they fix carbon dioxide during photosynthesis. The researchers hypothesized that C4 grasses, being more metabolically efficient, might invest less in individual pollen grains — producing thinner walls and simpler surfaces — and the evidence lends support to that idea.
With the method validated, they applied it to a sediment core from a lake on Mount Kenya, spanning 25,000 years. The record was vivid. Grass diversity plummeted during the Last Glacial Maximum as CO2 and temperatures bottomed out, then climbed again as the climate warmed. The C3-to-C4 ratio shifted in tandem: C4 species dominated the cold, low-CO2 ice age, then gradually ceded ground to C3 grasses as conditions changed — a competitive reshuffling written in pollen and now, finally, legible.
The stakes extend well beyond academic ecology. Grasses gave humanity wheat, rice, maize, and millet, and understanding how they evolved under past climate pressures matters for anticipating their future. Punyasena and Adaimé plan to refine their approach and eventually extend it to pollen and spores from all land plants — a prospect that could fundamentally transform how scientists reconstruct the planet's botanical past.
For decades, paleobotanists have faced a frustrating wall: grass pollen grains look nearly identical under a microscope. Unlike pollen from other flowering plants, which display distinctive shapes, spikes, or grooves that make identification straightforward, grass pollen offers almost no visual clues to distinguish one species from another. This meant that scientists studying ancient sediment cores—which can contain thousands or millions of pollen fossils in a single cubic centimeter—were unable to reliably classify what they were looking at. The result was a vast archive of data that remained largely inaccessible, a historical record of grassland evolution locked away in plain sight.
Surangi Punyasena, a plant biology professor at the University of Illinois Urbana-Champaign, and Marc-Élie Adaimé, then her doctoral student and now a postdoctoral researcher at the Smithsonian's Office of Digital and Innovation, set out to crack this problem using two complementary technologies. They began with super-resolution microscopy, a technique that uses lasers to illuminate pollen grains point by point, then algorithmically reconstructs the scattered light to reveal surface details with nearly the clarity of electron microscopy—but far faster and cheaper. When Adaimé examined the resulting images, he noticed something previous researchers had missed: subtle variations in the surface patterning of different grass pollen grains, along with differences in cell wall thickness.
Armed with these observations, Adaimé trained a machine-learning model on images from several identifiable grass species, teaching it to recognize the patterns that distinguish one type from another. The model used convolutional neural networks, mathematical structures loosely inspired by how the human brain processes visual information. Once trained, the model could analyze new pollen samples and estimate not just the presence of different species, but the overall diversity of the grassland community that had produced them. When tested on samples with known compositions, the estimates closely matched reality.
The breakthrough extended beyond simple species counting. The model could also distinguish between two fundamentally different types of grass: C3 and C4 species, which differ in how they concentrate carbon dioxide during photosynthesis. C4 grasses are more efficient at using CO2 and water, a trait that gives them advantages in certain climates. Adaimé and Punyasena hypothesized that this efficiency trade-off might be reflected in pollen structure—that C4 grasses might invest less in individual pollen grains, resulting in thinner walls and simpler surface patterns. The evidence supports this idea, though more work is needed to confirm it.
With their method validated, the researchers applied it to a sediment core from the bottom of a lake on Mount Kenya, a record spanning 25,000 years. The results told a clear story of ecological change tied to climate. During the Last Glacial Maximum, roughly 21,000 to 18,000 years ago, when Earth's climate was at its coldest and atmospheric CO2 levels plummeted, grass diversity in the region dropped sharply. As temperatures rose and CO2 rebounded in the millennia that followed, diversity climbed again. The proportion of C4 to C3 grasses showed a different pattern: C4 species dominated during the final stretch of the ice age, then gradually gave way to C3 grasses as the climate warmed and CO2 increased—a shift that suggests C3 grasses gained competitive advantage under the new conditions.
The implications ripple outward. Grasses are among humanity's most important crops—wheat, rice, maize, barley, sorghum, and millet all depend on understanding how these plants evolved and adapted. Open grasslands themselves are relatively young in Earth's history, having emerged around 40 million years ago, and grasses were among the first plants domesticated, roughly 12,000 years ago. By unlocking the information hidden in pollen morphology, Punyasena and Adaimé have opened a new window onto that deep history. Their method can now convert vast archives of fossil pollen into reliable ecological data, revealing how grasslands responded to past climate shifts. The team plans to refine the approach and eventually extend it beyond grasses to pollen and spores from all land plants, a prospect that could fundamentally reshape how scientists read the planet's botanical past.
Citações Notáveis
Within a small cubic centimeter of sediment, you could have thousands, potentially millions of pollen fossils. But the level at which we were able to analyze it before machine learning was limited by human ability.— Surangi Punyasena, University of Illinois Urbana-Champaign
It is satisfying to see that there is so much more information to be unlocked. Now we have a way to begin unraveling the history of grasslands, and perhaps even the deep evolutionary history of grasses, from clues hidden in subtle differences among their pollen grains.— Marc-Élie Adaimé, Smithsonian Office of Digital and Innovation
Quer a matéria completa?
Leia o original em EurekAlert! Science News Releases ↗