AI Analysis of 8 Million Herbarium Specimens Finds Flowering Times Shifting Fastest in the Tropics

A new report on the state of the world's plants and fungi from the Royal Botanic Gardens, Kew, drawing on contributions from 400 researchers in 40 countries, highlights how digitization and artificial intelligence are being used to analyze the world's herbarium collections. More than 145 million plant and fungi specimens have now been digitized from over 170 institutions worldwide, according to the report.

Researchers at Norway's NTNU University Museum, led by professor of plant ecology James Speed and postdoctoral machine-learning researcher David Williamson, trained a machine-learning model to recognize whether digitized herbarium specimens showed plants in flower. Applied to 8 million preserved specimens representing 200,000 species, the model completed in about a week work that Williamson estimates would have taken a human roughly 40,000 hours, or about 20 working years.

The analysis found that global flowering times have shifted by an average of 2.5 days per decade over the past century, both earlier and later, with the largest changes occurring in tropical regions — a result the researchers said surprised them, since the greatest temperature increases from climate change tend to occur in the far north. In the tropics, Speed said, flowering is more closely tied to rainfall, and a changing climate can shift the rainy season earlier, later, or cause it to fail altogether, which can create a mismatch between flowering plants and the insects that pollinate them.

Martin Cheek, a taxonomist at Kew and report co-author, said AI could speed up the identification of common species, freeing experts to focus on describing unknown species and conserving threatened ones, and on calculating the likelihood that an unrecorded species is extinct rather than simply undiscovered. Researchers cautioned that less than 16% of the world's herbarium specimens are digitally accessible, with the largest gaps in biodiverse countries where collections are understaffed, and that AI models require human experts to validate their results.