AI-Powered Germination Count from a Photo: How It Works, What It Measures and How to Get Accurate Results

What is a germination count?

A germination count is the number of seeds that have germinated out of the number sown, usually expressed as a percentage, the germination rate or germination percentage:

Germination rate (%) = germinated cells ÷ total cells × 100

For example, if 223 out of 480 cells across two trays have a seedling, the germination rate is 223 ÷ 480 × 100 = 46.5%.

The word “germination” means slightly different things in different settings, and it helps to be precise:

When we talk about a germination count from a photo of a tray, we are really counting seedling emergence. That is the number growers and propagators work with day to day, and it is what this article focuses on.

Other germination metrics worth knowing

If you photograph the same trays on several days, a simple count becomes a germination curve, and you can calculate speed and uniformity as well as the final percentage:

• Final germination percentage: the plateau of the curve, once no new seedlings appear.
• Mean germination time (MGT): the average time it takes a seed to germinate. MGT = Σ(n_i × t_i) ÷ Σn_i, where n_i is the number of newly germinated seeds counted on day t_i.
• T50: the time taken to reach 50% of the final germination.
• Uniformity: how tightly germination is clustered in time. Uneven emergence means uneven plants at transplanting.

Two seed lots can reach the same final percentage while one gets there three days later and far less evenly, and in a greenhouse schedule that difference matters.

Why is counting germination by hand a bottleneck?

Counting seedlings by eye is simple, but it does not scale:

• The numbers add up quickly. A single 240-cell tray is manageable. Twenty trays is 4,800 cells, and a seed trial with several treatments, replicates and count days can mean tens of thousands of individual decisions.
• Counts drift between people and over time. Is a hook-shaped seedling just breaking the surface “germinated”? Different counters answer differently, and the same counter answers differently at 4 pm than at 9 am.
• Mistakes are hard to audit. A number written on a clipboard can't be checked later. A photo can.
• Detail gets lost. In practice, a hand count usually ends up as one number per tray. Which rows failed, and where in the tray, is rarely recorded.

This is why official laboratory tests are so structured: they use replicates (under the US Federal Seed Act regulations, at least 400 seeds tested in replicates of 100 or fewer) and set tolerances for how far replicates may disagree. Outside the lab, in a propagation house or a trial greenhouse, that level of rigour is usually not affordable by hand. A camera makes it affordable.

How does AI count germination from a photo?

The process breaks down into a few steps, and it is worth understanding each one, because it explains both why the method works and where it can go wrong.

1. Find the tray. Square fiducial markers (printed black-and-white codes) are placed in the four corners of the tray. They are easy for software to detect, and because their positions are known, they tell the system exactly where the tray is, even when the photo was taken at a slight angle or with other trays crowding the frame.
2. Straighten the image. Using the four marker positions, the image is geometrically corrected so the tray looks as if it had been photographed perfectly from above. This step is what makes a handheld phone photo usable.
3. Lay the grid over the tray. The user selects the tray format (for example 12 × 20 cells), and the software divides the corrected tray into exactly that many cells. Because each cell's position is now known, the AI no longer has to search the whole image for plants. It only has to look inside each cell.
4. Classify every cell. A trained model decides, cell by cell, whether a seedling is present. In the results, each cell gets a coloured mark: green for germinated, yellow or orange for not germinated, so you can see at a glance what the AI decided and check it against the photo.
5. Add it up. Germinated cells, empty cells and the germination rate are calculated for the tray, for each row, for each photo and for the whole trial.

Which tray sizes can be counted?

Seed trays come in many formats, so the grid has to match the tray. Petiole Pro has five default tray sizes:

Other formats can be added: the 288-cell (12 × 24) tray in the per-row example below is one of them.

What does per-row germination analysis show?

An overall germination rate tells you how much germinated. A row-by-row breakdown tells you where, and that is often the more useful information.

In the example above, the tray as a whole germinated at 9.0%. On its own, that figure suggests a poor seed lot. The row analysis tells a different story: the first 17 rows have no seedlings at all, while rows 18–24 hold every seedling in the tray, at 25% to 50% per row (row 21 is the one exception, with none). A pattern that clean is far more likely to come from something physical, such as part of the tray sown later, a seeder skipping, a dry section on the bench, or a different treatment applied to one end, than from random seed failure.

Per-row and per-tray results help you to:

• spot sowing faults, such as a blocked needle on a seeder that leaves a whole line empty;
• find environmental gradients, such as cold or dry edges, uneven misting or shading;
• separate treatments when one tray holds several, as in many seed trials;
• decide on gapping, which means knowing exactly which trays need empty cells refilled before dispatch.

Can you count many trays at once?

Yes, and for anyone running trials, this is where the time saving really adds up. In Petiole Pro, photos are grouped into a trial, and several photos (five at a time in the current version, a number that could be increased) can be processed together. Results are then reported per photo and combined for the whole trial.

Grouping by trial also makes comparisons straightforward: “First trial” vs “Second trial”, seed lot A vs seed lot B, treated vs untreated.

How are results reported and shared?

A count is only useful if it reaches the person who makes the decision: the head grower, the customer or the trial sponsor. Results can be shared as a web report showing the trial name and date, total cells, germinated and non-germinated cells with percentages, the grid size and the photos used.

Reports can also be generated in CSV, PDF or other formats for spreadsheets, trial records or customer documentation.

Germination counting in the Petiole Pro app

Germination counting sits in the Greenhouse section of Petiole Pro, alongside our other vision-based phenotyping tools. You can start with the example photos to see how it works before photographing your own trays. Both free and paid options are available.

In short, the workflow is:

1. Place a marker in each corner of the tray.
2. Take one photo from directly above.
3. Choose the tray size.
4. Get the germinated count, not-germinated count and germination rate, per tray, per row and per trial.
5. Share the report or export the data.

How do you take a photo that counts accurately?

Most counting errors start with the photo, not the model. A checklist:

1. Shoot straight down. Hold the phone parallel to the tray, centred over it. Corner markers correct a small tilt, not a steep angle.
2. Keep all four markers in the frame, uncovered by fingers, labels or leaves.
3. Fill the frame with the tray. More pixels per cell means smaller seedlings are easier to see.
4. Use even, diffuse light. Bright overcast light or shade is ideal. Avoid hard shadows from benches and harsh sun reflecting off wet trays and plastic.
5. Count at the right stage. Photograph once cotyledons are clearly visible and before leaves spread into neighbouring cells. Too early and tiny emerging seedlings are missed; too late and one plant can appear in two cells.
6. Be consistent. Same time of day, same stage, same distance. Consistency matters more than perfection when you compare trays or days.
7. Label trays visibly. A tray label in the photo, like the treatment labels in our examples, makes later checking much easier.

How accurate is AI germination counting — and what are its limits?

A well-set-up system is consistent, but no image-based count is perfect. These are the situations to watch for:

• Very small or just-emerging seedlings can be missed, especially fine-leaved crops at the hook stage.
• Overlapping leaves from large seedlings can spill into an empty neighbouring cell and make it look occupied.
• Things that aren't seedlings, such as weeds, moss, liverwort, bark or perlite, can confuse the model if they look green or plant-like.
• Glare and deep shadow hide what is in a cell.
• “Not germinated” is not the same as “not sown”. A camera sees an empty cell. It can't tell whether a seed failed or was never sown, so seeder performance and seed quality need to be separated by other means if that distinction matters.
• One seedling per cell is assumed. Where several seeds are sown per cell, the count tells you how many cells are occupied, not how many seedlings there are.

How to validate it for your setup: take five to ten trays covering your typical crops and growth stages, count them carefully by hand (ideally two people independently), and compare those counts with the AI result cell by cell. The coloured dots make disagreements easy to find. If the differences are small and random, you can rely on the method. If they are systematic (for example, always missing seedlings of one crop at an early stage), adjust when you photograph or tell us so that the model can be improved.

Where else is AI used to count germination?

Tray-based counting is one part of a wider field of image-based germination analysis:

• Laboratory germination assays. Systems such as SeedGerm, developed by the Earlham Institute, the John Innes Centre, Syngenta and NIAB and published in New Phytologist in 2020, image seeds on germination media over time and use machine learning to score germination and its timing across several crop species.
• Deep learning on Petri dishes. Genze and colleagues (2020, Plant Methods) trained object-detection models to tell germinated from non-germinated seeds of maize, rye and pearl millet in Petri-dish images. Their models reached a mean average precision of about 94–98% depending on the species, and the germination indices calculated from their predictions were more accurate than those from manual counts.
• Field emergence (stand counts). Drone and ground-camera images are used to count emerged plants in rows, so that establishment can be measured and gaps found early.

The principle is the same in each case: turn an image into a reliable count and keep the image as a record. The plug tray is one of the easiest settings for it, because the grid already tells you where every plant should be. For what we learned running it in practice, read 5 surprising lessons from using AI for germination counting, and for counting seeds and fruit outside a tray, see counting round objects with a smartphone.

Who uses AI germination counting?

• Young-plant producers and propagators: checking tray fill before dispatch, planning gapping and documenting quality for customers.
• Seed companies and seed testing teams: comparing seed lots and running germination and vigour checks at scale.
• Plant breeders: screening many lines for emergence and uniformity.
• Researchers and trial managers: comparing seed treatments, biostimulants, substrates or growing conditions, with photo records for every replicate.
• Vertical farms and controlled-environment growers: tracking germination batch by batch as part of a tight production schedule.

About the author

Dr Maryna Kuzmenko is the founder of Petiole, a UK company building AI and computer-vision tools that turn smartphone photos into plant measurements for growers, breeders and researchers. She hosts a weekly LinkedIn Live on understanding agricultural AI, writes a LinkedIn newsletter on AI in agriculture and teaches the Udemy course AI in agriculture.

References

1. Colmer J., O'Neill C. M. et al. (2020). SeedGerm: a cost-effective phenotyping platform for automated seed imaging and machine-learning based phenotypic analysis of crop seed germination. New Phytologist. doi:10.1111/nph.16736; press release, John Innes Centre.
2. Genze N., Bharti R., Grieb M., Schultheiss S. J., Grimm D. G. (2020). Accurate machine learning-based germination detection, prediction and quality assessment of three grain crops. Plant Methods 16, 157. doi:10.1186/s13007-020-00699-x
3. US Code of Federal Regulations, 7 CFR Part 201 — Federal Seed Act Regulations, Germination tests in the administration of the Act (§§ 201.54–201.58). eCFR
4. International Seed Testing Association (ISTA). International Rules for Seed Testing. seedtest.org
5. Aravind J. et al. germinationmetrics: Seed Germination Indices and Curve Fitting (R package; definitions of MGT, T50 and related indices). CRAN