Industries / Food, farming and forestry / Agriculture and farming / Seed production
Seed testing
Testing a small sample from a seed lot in a laboratory to find how much of it will grow, how pure it is and how strong it is, by written rules that let any lab repeat the result.
6
Milestones achieved, from 1869 to 2025.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1869
A station tests seed
Friedrich Nobbe set up a seed testing station at Tharandt in Saxony, in today's Germany, and in August 1869 published a "Statute concerning the testing of agricultural seeds". The International Seed Testing Association (ISTA), today's world standards body for seed tests, calls the statute the starting point of seed testing and the Tharandt station the first seed testing station.
Credit
Friedrich Nobbe
1908
North America agrees methods
Seed analysts from experiment stations, agricultural colleges, the US Department of Agriculture (USDA) and the Canada Department of Agriculture met at USDA in Washington, D.C., in 1908 and formed the Association of Official Seed Analysts (AOSA). Its aim was "the perfecting and promulgating of uniform methods for seed testing", and its 1910 constitution added "uniformity and accuracy in methods, results, and reports".
Credit
Association of Official Seed Analysts
1931
One rulebook across borders
The European Seed Testing Association, formed in Copenhagen in 1921, gave way to ISTA, founded at a 1924 meeting in Cambridge, England, that 26 countries attended. In 1931, at the request of the seed trade's international federation (FIS, now ISF), ISTA adopted the first International Rules for Seed Testing and set up a scheme of ISTA certificates for reporting results. Since 1996 ISTA has also audited and accredited the labs that issue them.
Credit
ISTA
1949
A red stain shows life
In the tetrazolium test, seeds are cut and soaked in a colorless salt that living tissue turns red, so an analyst can see which seeds are alive without growing them; a 2019 review gives less than 20 hours for most crops. George Lakon of the seed testing station at Stuttgart-Hohenheim, Germany, built the test on 2,3,5-triphenyl tetrazolium chloride after dropping an earlier selenium stain as too toxic. It entered Germany's official seed testing rules in 1949, ISTA's rules in 1966 and AOSA's practice, with a handbook, in 1970.
LessonA fast test still needs trained eyes; by the same review an experienced analyst reads only about four to five samples an hour.
Credit
George Lakon; Kühn and Jerchel first reported the reaction in living tissue in 1941
2001
Vigor enters the rules
A germination test is run in ideal conditions, and a lot that passes it can still come up poorly in cold or wet soil. Vigor tests try to catch such weak lots. In 1967 S. Matthews and W. T. Bradnock published a method for peas that measures the electrical conductivity of the water seeds soak in: weaker seeds repair their cell membranes more slowly as they take up water, so more of their contents leak out. In 2001 this test for garden peas became the first vigor test in ISTA's International Rules.
LessonIn João Marcos-Filho's 2015 review, germination and tetrazolium results "might not be closely related with field seedling emergence", especially when conditions are unfavorable.
Credit
S. Matthews and W. T. Bradnock; ISTA
2025
Rules for camera analysts
ISTA's guideline TCOM-P-12, approved on 2 May 2025 and in force from 15 May 2025, sets out how an ISTA-accredited lab may use a computer vision tool in place of an analyst: to sort seeds in a purity test, find seeds of other species, or judge seedlings in a germination test. Before use, the tool must pass 30 blind samples (samples whose contents the lab staff overseeing it do not know), with at least 90% correct where the right answer is known, or match at least three human analysts on 30 samples by a set statistical test where it is not.
Credit
ISTA's Advanced Technologies Committee and its purity, germination, vigor and statistics committees
Watch for
Seedlings measured, not judged
In a germination test, analysts sort each seedling as normal, abnormal or dead by eye. Syngenta's SABIMA tool instead measures the total root and shoot area in photos of rolled paper-towel tests with machine learning. Syngenta reports that its measurements explained 90.2% of the variation in manual evaluation for wheat (104 tests) and 91.8% for corn (16 tests), and that shoot area predicted 82% of the variation in corn field emergence (33 lots). NIAB, a UK crop research body, produced the manual comparison data; the tool is patent-protected.
Achieved whenan independent ISTA-accredited lab verifies it under ISTA's computer vision guideline or reports using it in routine testing.
Claimed by
Tahl Paran and colleagues, Syngenta, in Seed Testing International, March 2026
Camera finds dodder seeds
Dodder (Cuscuta) is a parasitic weed whose seeds are banned in some forage seed trade. Checking a sample of alfalfa, lotus or white clover seed for it means searching 200 to 500 g of seed, about 6 to 7 person-hours per sample by hand. MachVision's Smart Analyzer finds dodder seeds by camera with a neural network trained on 100,000 images per crop, and its developers report 98.3% to 99.5% accuracy on their own held-out test images. They have proposed a 30-sample blind validation under ISTA's guideline.
Achieved whenthe tool passes the guideline's 30 blind samples at 90% or better, checked by ISTA's committees or an accredited lab.
Claimed by
Victorio Costa, MachVision, with Rayen Laboratorios, in a presentation to ISTA, 2026
Open
01
Open seedling evaluation images
A public set of photos of standard germination tests in wheat and maize, with every seedling labeled normal, abnormal or dead by at least two trained seed analysts following ISTA's rules, and a baseline model trained on it.
The briefCredit · unclaimed02
Blind-test an open purity checker
An open-source camera tool for one purity job, finding seeds of another crop mixed into a crop seed sample (for example rye in wheat), and the first public run of it through the 30 blind samples in ISTA's computer vision guideline.
The briefCredit · unclaimed
Open 01 · the brief
Open seedling evaluation images
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of photos of standard germination tests in wheat and maize, with every seedling labeled normal, abnormal or dead by at least two trained seed analysts following ISTA's rules, and a baseline model trained on it.
Why it is still open
The largest public germination image set, from Genze and colleagues (2020), labels only whether each maize, rye or pearl millet seed has sprouted: 23,797 images of 2,449 seeds, republished on Hugging Face by the AgML project. Brutas and colleagues (2024) trained models to tell normal, abnormal and ungerminated seeds of pole sitao, a kind of cowpea, but no public image set with seedling labels of that kind was found. The one tool found that replaces seedling judgment, SABIMA, is patent-protected. ISTA's Advanced Technologies Committee told its 2026 meeting that imaging tools need large image sets that labs lack, and it was still drafting a consortium agreement for a shared, paid image store.
Built with what exists
a phone camera on a copy stand or a flatbed scanner; Label Studio, an open labeling tool; an open object detection model such as YOLO; ISTA's TCOM-P-12 guideline as the test protocol.
First step
Run ten paper-towel germination tests on one untreated wheat lot, photograph every towel at each count, and ask one seed lab analyst to label 500 seedlings.
Done looks like
at least 5,000 seedlings from at least 30 seed lots, each labeled by two analysts, with their agreement rate published; and a baseline model whose normal-seedling percentage, on 30 held-out samples scored by three analysts, passes the paired t-test and agreement limits that TCOM-P-12 sets out.
Open 02 · the brief
Blind-test an open purity checker
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An open-source camera tool for one purity job, finding seeds of another crop mixed into a crop seed sample (for example rye in wheat), and the first public run of it through the 30 blind samples in ISTA's computer vision guideline.
Why it is still open
The guideline took effect in May 2025, and the only tool found going through it, MachVision's Smart Analyzer, is proprietary. Public seed image sets exist but have not been put through a blind spiked-sample test: BiSID-5k (Kukushkin and colleagues, 2025) holds 5,000 paired color and hyperspectral images of single seeds of 10 species found in oilseed rape lots and excludes any image without exactly one seed; Ahmed and colleagues (2024) detected 19 weed species in 485 microscope images of mixed seeds from soil, not seed lots.
Built with what exists
a flatbed scanner or a phone on a copy stand; the BiSID-5k images (CC BY 4.0) for pretraining; an open detection model such as YOLO; Label Studio.
First step
Buy untreated wheat and rye seed, photograph 2,000 single seeds of each, and train a first model to tell them apart.
Done looks like
30 samples of one crop, each spiked by a second person with 0 to 20 seeds of another crop at levels the operator does not know, run through the tool, with at least 27 of 30 (90%) classified correctly, and the code, images and tally published.