65/100

Moderate evidence

Cite
More
EmailLinkedInX

AI flags flowers on 11.1 million of 22 million plant specimens, but only 2.9 million are usable

Study title: Petal to the metal: The slow road to automating large-scale phenology labeling for herbarium specimens

Evidence type
Observational study
Sample
n = 22,000,000
Sources
1 sourcePeer-reviewed
Published
1 min read

Study title: Petal to the metal: The slow road to automating large-scale phenology labeling for herbarium specimens

By October 2026, researchers described a machine learning pipeline using an ensemble modeling approach to detect flowers on herbarium specimens, addressing challenges from diverse floral morphologies and variable specimen and image quality. After finding expert-curated training data essential, they applied the model to 22 million filtered records and labeled 11.1 million as having flowers present.

Read the study abstract

Premise Herbarium specimens represent critical historical records of plant phenology, yet automating annotation of reproductive structures remains challenging given the diversity of floral morphologies, specimen age and quality, and image quality. Methods Here, we present a machine learning pipeline that uses an ensemble modeling approach to detect flowers on herbarium specimens and deliver these data to the phenology research community. After testing multiple strategies for generating training data, we found th…

Petal to the metal: The slow road to automating large-scale phenology labeling for herbarium specimens
Alocasia Sanderiana W. Bull (Kew Herbarium Catalogue Specimen K000400154) by Alocasia sanderiana W.Bull Collected in Philippines by Royal Botanic Gardens, Kew (licensed under https://creativecommons.org/licenses/by/4.0/ ). CC BY 4.0 · https://creativecommons.org/licenses/by/4.0

59Gain side

How the two sides scoreClose call

56Problem side

The gain

An AI pipeline labeled 11.1 million of 22 million herbarium specimen records as having flowers present, which can widen the plant and time coverage of the Phenobase database.

The problem

Only 2.9 million of the 11.1 million flower-labeled records had the complete metadata needed for downstream research, and uncertainty remains about missed flowers (false negatives).

A fixed formula sets both scores from the evidence and its sources. Reader votes don’t move them.

What the scores mean

G is the gain side, out of 100. It goes up when the gain is specific, when it measurably helped people and when the sources behind it are strong.

P is the problem side, out of 100. It goes up when the problem is specific, when it measurably hurt people and when the sources behind it are strong.

In brief

The work matters because herbarium specimens are critical historical records of plant phenology, and machine-labeled data integrated into Phenobase can expand taxonomic and temporal coverage for large-scale analyses. Uncertainty remains around false negative rates and the fact that only 2.9 million labeled records had complete metadata needed for downstream research.

Main points

  1. Pipeline used an ensemble modeling approach to detect flowers on herbarium specimens after testing multiple training data strategies.
  2. In-house expert-curated annotations were found essential for producing reliable results.
  3. Final filtered dataset was 22 million records, with 11.1 million labeled as flowers present.
  4. Only 2.9 million labeled records contained complete metadata necessary for downstream phenology research.

The full story

Researchers tested multiple strategies for generating training data for flower detection on herbarium specimens and concluded in-house expert-curated annotations were essential for reliable results.

Expert validation assessed the ensemble on a filtered final image dataset of 22 million records, resulting in 11.1 million records labeled with flowers present.

The authors note only 2.9 million of those labeled records contained complete metadata necessary for downstream phenology research, pointing to the need for full label digitization efforts, and demonstrate integration into Phenobase to expand coverage.

What this doesn’t fix

Performance was constrained by diversity of floral morphologies, specimen age and quality, and image quality, with moderately high false negative rates and dependence on expert-curated training data and complete label digitization.

Sources

1 on file
  1. Applications in Plant Sciencesdoi:10.1002/aps3.70085
    Peer-reviewed2026-10-01

Talk it over

Three questions to start on, two more if you have time
The pipeline labeled 11.1 million of 22 million specimen records as having flowers, but only 2.9 million had complete metadata. How do you weigh the wider coverage against that gap?

Start with a question

AI flags flowers on 11.1 million of 22 million plant specimens, but only 2.9 million are usablehttps://truvace.com/item/petal-to-the-metal-the-slow-road-to-automating-large-scale-phenology-labeling-fo

  1. The pipeline labeled 11.1 million of 22 million specimen records as having flowers, but only 2.9 million had complete metadata. How do you weigh the wider coverage against that gap?

    Answer in the thread
  2. The authors report moderately high false negative rates. How could missed flowers change what researchers conclude about plant phenology across places and years?

    Answer in the thread
  3. The researchers found in-house expert-curated annotations essential. Who should pay for and do the label digitization the authors call for, and why?

    Answer in the thread
If you have more time
  • Expert validation was run on the filtered set of 22 million records. What would you ask to learn how errors differ between old and recent specimens?

    Answer in the thread
  • Phenobase now includes the machine-labeled records. What would you want shown on each record so later researchers know a machine, not a person, labeled it?

    Answer in the thread

What readers say