Petiole Pro Blog

Applied AI for Soybean Breeders: Tools, Workflows and Use Cases from the 2026 Soybean Breeders' Workshop

Dr Maryna Kuzmenko's talk at the 2026 Soybean Breeders' Workshop, written up in full: the difference between AI and applied AI for soybean, which imaging tool fits each growth stage from germination to post-harvest QC, why drones cannot count soybean flowers, and where a smartphone beats expensive equipment.

Published on 1 March 2026 by Petiole Pro

Title slide reading Applied AI for Soybean Breeders: Tools, Workflows, Use Cases, by Dr Maryna Kuzmenko, Petiole, February 17th 2026, with soybean seeds and a Petiole Pro calibration plate.

Applied AI for plant breeding

On 17 February 2026, Dr Maryna Kuzmenko — founder of Petiole and Petiole Pro — spoke at the 2026 Soybean Breeders' Workshop, hosted by Brian Little and his team at the Center for Applied Genetic Technologies, University of Georgia. The talk was titled Applied AI for Soybean Breeders: Tools, Workflows, Use Cases, and it was delivered virtually from a rainy Peterborough, UK.

The core argument: most published lists of "AI for soybean" describe what AI could do. Applied AI is the much shorter list of what actually works in a breeding programme today — and nearly all of it is visual, smartphone-based, and stage-specific. Counting germinated seeds, measuring leaf area, quantifying leaf damage, counting flowers and pods, counting root nodules, and running post-harvest seed QC are real. Everything else is still a research direction.

This long-read reconstructs the full session: what separates AI from applied AI, which imaging tool belongs at each soybean growth stage, why drones cannot count soybean flowers, where satellite imagery stops being useful, what the Kansas State nodule-counting collaboration produced, and why the number on its own is never the deliverable.

Key takeaways

Ask a chatbot how to use AI in soybean research and you get five to ten directions. Try them and one, maybe two, actually work today. That gap is what "applied AI" means.
The tool follows the growth stage, not the other way round: RGB smartphone for germination, leaf area and seed QC; drones for emergence, lodging and maturity; thermal for water stress; satellite for almost nothing a breeder needs.
Drones cannot count soybean flowers or pods — they are hidden under the canopy, and white or pale flowers are hard for a computer vision model to separate from leaves even with good labelling.
An existing model is not a working model. The germination module trained on tree seedlings failed on soybean trays. Relevant training data is the bottleneck, not the algorithm.
A count with no context evaporates. Variety, sowing date, fertiliser, irrigation, soil test — the number only becomes data when it travels with the journey of growing.

The workshop, and why this talk happened

The 2026 Soybean Breeders' Workshop ran as a hybrid event out of the University of Georgia. The invitation arrived by an unusually honest route: Brian Little had been researching handheld devices that could scan and analyse things in the field, went down a YouTube rabbit hole, and found Petiole Pro.

Petiole is a UK-based company, nine years old, whose flagship free mobile application turns smartphone and drone imagery into AI-powered crop insights. The platform has been cited in more than 110 research papers, and the work has been supported by three grants from the UK and EU governments. Three of those papers reference Petiole Pro directly as the data-collection tool inside soybean experiments.

Slide titled Petiole Pro plus Soybean references, listing three research papers: Rodrigues Bueno et al. 2024, Dymytrov and Sabluk 2023, and Kurbanov et al. 2019. — The three soybean papers that cite Petiole Pro as a data-collection tool — spraying research, seed treatment with biopreparations, and breeding-process monitoring software.

AI vs. applied AI for soybean

Ask any search engine or chatbot "what is AI for soybeans?" and you will get an answer immediately. The answer will be fluent, structured, and mostly aspirational — a blend of summarisation, ideation, and a little dreaming. It will promise accelerated breeding, real-time pest detection, yield prediction with over 90% accuracy, and irrigation optimisation.

Screenshot of an AI Overview answer describing AI in soybean farming, listing accelerated breeding and genetics, pest and disease detection, yield prediction and seed analysis. — The generated answer to "AI for soybeans". Nothing here is false — but the distance between these bullet points and a working Tuesday morning in a breeding programme is the entire subject of the talk.

The problem is not accuracy. The problem is applicability. Ask the follow-up question — "how can I use existing tools for my soybean research?" — and you still get five or ten directions. Try to use them, and you end up with one, maybe two, genuinely working cases.

So this talk is not about general, all-purpose, all-directional AI for soybean breeding. It is about what is practical and applicable. Said precisely, the topic is applied, visual, smartphone-based AI for soybeans — because that is what nine years of work suggests holds the biggest promise for practical application.

That framing matters for a specific reason. Researchers already know what is possible in front of very expensive equipment and very expensive cameras. The more useful question — for research work and for extension work alike — is what you can do with an ordinary smartphone in your hand.

Slide titled Applied AI for Soybean, showing a folder of soybean root images labelled Root_PI209331 and similar, and a phone screen titled Soybean Nodules with a Detect Nodules button. — Applied AI, made concrete: real soybean root scans from a breeding programme, and a phone that detects the nodules on them. This came from a researcher at Kansas State University — more on that below.

The technology stack behind every tool

Every tool in the talk sits on the same core stack, and it is worth naming the parts because the words get used loosely:

Computer vision — the technological eyes. How a computer sees an image and understands what is on it.
Machine learning / deep learning — deep learning is a part of machine learning. This is what takes the data from the imagery, processes it, and finds the patterns inside what has been collected.
Remote sensing (optional) — indices, most famously NDVI, used to understand deficiency in crops.
Sensors (optional) — a nice-to-have that gives you extra layers of data to support a claim: environmental monitoring, and how conditions map onto a specific stage.

Slide titled Tools and Tech - AI for Soybean, with a soybean life-cycle diagram showing bean seeds, germination, sprout, leaf growth, flowering, pod fill and mature seed pods, beside a list: machine learning / deep learning, computer vision, remote sensing (optional), sensors (optional). — The soybean life cycle on the left, the stack on the right. The rest of the talk walks the circle stage by stage, asking one question each time: which tool is applicable, and for what task?

Stage 1 — Germination

The job to be done at germination is to assess percentage germination, speed and vigour, and uniformity. In practice that breaks into two tasks: count the germinated seeds, measure radicle length, classify abnormal seedlings — and then build the germination curve automatically.

The tool is an ordinary RGB camera. A smartphone camera is an RGB camera, so nothing extra is needed. From the count you get germinated and non-germinated cells, and the germination rate falls out automatically.

The bonus layer is environmental: temperature and moisture sensors let you link germination outcomes to the conditions that produced them — which matters when trays sit in different places and your experiment conditions genuinely diversify.

Slide titled Germination listing the job to be done — assess percentage germination, speed and vigour, uniformity — beside a photograph of a seedling tray with dark plugs and green sprouts. — The germination slide. The tray shown is not soybean — it is used here as an analogy for how the same cell-by-cell counting would work on a soybean germination tray.

Here is what automation looks like in practice. The imagery below is tree seedlings, not soybeans — but the process runs identically on any tray-based crop, and it is shown to demonstrate what becomes possible for soybean germination trays. The beauty of it is not the algorithm. It is that nobody stands in front of the trays counting with their eyes and their fingers and writing into a notepad.

Two panels: a raw tray image of seedling plug trays, and the same trays with each cell classified by a green or orange dot, with a results summary reading total cells 240, germinated saplings 223, empty cells 17, germination rate 92.9 percent. — Raw tray image on the left, classified cells on the right. This is a tree-seedling nursery, shown as an analogy for the identical workflow on soybean germination trays: 240 cells, 223 germinated, 17 empty, germination rate 92.9% — computed automatically.

A grid of individual cropped tray cells each labelled 1.00 confidence, beside a large results summary: total cells 240, germinated saplings 223, empty cells 17, germination rate 92.9 percent. — Every cell is cropped and scored individually, so the summary is auditable rather than a black box. Again, tree seedlings used as an analogy for possible work with soybean plants — the digital proof of the experiment's status is the point, not the species.

And here is where the honesty comes in. One Petiole Pro user sent us their own soybean germination tray and asked whether our module could be applied to it. The answer was: you can try — but look at the two images side by side.

Two photographs side by side: on the left, dense grey plug trays of tree seedlings with fine green shoots; on the right, a tray of large dark plugs with sprawling bean seedlings and visible ungerminated seeds. — Left: the tree-seedling trays the model was trained on, included here as an analogy for the soybean case. Right: a real germination tray sent by a user. Even a human sees the difference instantly — different plug geometry, different seedling architecture, ungerminated seeds sitting visibly on the surface.

The computer vision algorithm sees that difference too — and it is lost. It does not understand what we want from it. The model already knows the concept: there are cells, there may be seedlings inside, no seedling means not germinated. The challenge is not the algorithm. It is providing relevant data to train and adjust it. As with every AI model, bring the relevant data, train the model, and performance and accuracy can be improved and boosted.

That distinction — an existing model versus a working model — is the single most useful thing a breeder can take from this section. If you want to go deeper on the mechanics of automated germination counting, we wrote it up separately in Beyond the Sprout: AI germination counting.

Stage 2 — Sprout and emergence

The job to be done shifts to emergence percentage, stand count, gaps, and early stress. Two approaches: emergence counting at plot scale, and early vigour trends via indices.

You can use a smartphone. You can use a drone — RGB, or multispectral if you have it. But the most interesting method came from a researcher running a low-cost experiment: he attached a smartphone to a selfie stick, pointed it down at the plants, and walked along the row recording video. It is not innovation in any grand sense — nothing about it is new — and that is exactly why it works. It scaled, because walking scales, and the video processed cleanly into counts.

Slide titled Sprout (emergence) listing emergence counting at plot scale via smartphone on selfie stick or drone, and early vigour trends, beside a photo of soybean seedlings in dark soil with yellow detection boxes around each seedling. — Real soybean seedlings, detected individually along the row. Beyond presence, the later stage lets you assess the gaps *between* plants — a metric in its own right.

At this stage you can also link the germination and count data to weather-condition data and build the dependency. And you can start picking up broader trends — early disease signal, or the presence of chlorosis. But what we discovered is that the scope widens dramatically once leaf growth is present and there are at least a few leaves to work with. That is where it gets interesting.

Stage 3 — Leaf growth and the vegetative stage

Now the job to be done is canopy cover, vigour, LAI proxies, nutrient stress, weeds, and pest or disease onset. Two workhorses — remote sensing with drone multispectral plus drone RGB, and in-field sampling with a mobile application. Satellite imagery sits in the nice-to-have column, with three caveats attached: price, time, and accuracy.

Slide titled Leaf growth (vegetative) listing the job to be done as canopy cover, vigor, LAI proxies, nutrient stress, weeds, pests and disease onset, with remote sensing and in-field sampling as methods and satellite imagery as nice-to-have. — The vegetative stage is where the widest range of tools becomes applicable — and where the choice between them starts to matter financially.

RGB, multispectral, and the purple-light trap

Multispectral imaging can detect disease earlier than RGB because it captures wavelengths beyond visible light. It is worth seeing what that actually looks like, because there is a trap here that catches people out.

Three-panel comparison labelled RGB, Multispectral and LED: a green and yellow leaf on a white Petiole Pro calibration plate; the same kind of leaf rendered in purple tones by a multispectral lens; and a tomato plant under purple LED grow lights held against a calibration plate. — RGB, multispectral, and LED grow-light conditions compared. None of these are soybean leaves — they are apple and tomato, used as an analogy for the same imaging choice on soybean foliage. The middle and right panels look superficially similar and are completely different things.

Multispectral purple and vertical-farming LED purple are not the same. If you capture an image under LED grow lights, you are still processing it as ordinary RGB imagery — and you cannot get early disease detection in those conditions. Multispectral can. (Whether multispectral capture under LED is worth testing is an open and rather interesting question.)

Leaf area measurement

Once there are leaves, the first genuinely applicable smartphone task is leaf area measurement. This is not new science — it is a standard procedure. What changes is that a manual task becomes a digital one, which brings accuracy and, more importantly, scalability.

The scalability point is concrete. Manual leaf area assessment with ImageJ or grid paper can consume a day. With the mobile app, some researchers have taken thousands of photos, because they have a large selection of plants and need to measure all of them — and in that case more data directly means better accuracy of results. The method is simply: photograph the leaf in front of a calibration plate, or place the leaf on it.

Three Petiole Pro Leaf Analysis screens showing a ginkgo leaf measured at 57.42 square centimetres, a palmate leaf at 58.61 square centimetres, and a large leaf with a small calibration disc on it at 47.68 square centimetres. — Leaf Analysis on three different leaves. These are not soybean — they are shown as an analogy for identical work on soybean leaflets, and to make the point that leaf shape is not a constraint.

The calibration plates come in several sizes, driven by requests from growers. The smallest was developed specifically for a researcher in Colombia measuring leaf area in natural conditions high in the mountains — the app works offline, without internet, which is what made that possible at all. If you want the detail on why the plate is there and what happens without it, we covered that in Do I need a calibration plate for Petiole Pro?, and the capture technique in how to photograph leaves for accurate leaf area measurement in the field.

Two Petiole Pro Leaf Analysis screens showing a large irregular leaf outlined in purple on a calibration plate measured at 59.12 square centimetres, and the same leaf segmented on a black background at 60.28 square centimetres. — Capture and result. Not a soybean leaf — used as an analogy for possible work with soybean plants — but it shows the segmentation following a genuinely awkward leaf margin, which is the part that has to work.

There are thousands, maybe millions, of scenarios where leaf area is a useful parameter: any agricultural input trial, any attempt to assess the impact of an environmental change on a specific variety. This is one method of automating it.

Early-stage change detection: the cucumber experiment

A farmer with six rows of plants applied a different treatment to each row. Visually he could tell that some rows were greener than others. His question was the right one: how can I prove it?

We set a greenness threshold and built a model around it. He shot video; the pipeline segmented each leaf out of the frames; each leaf was then compared against the threshold. He had never had numbers before. We returned six pages of metrics — per-leaf, per-row, against overall, against average — because once the data is good, extraction is not the hard part.

Slide titled Cucumber experiment: how healthy are the leaves? showing a table of treatment-vs-control differences from 14.6 percent down to 2.8 percent, three panels showing video from farmer, per frame processing and results, and totals: 4,228 total leaves, 3,801 healthy leaves, 89.9 percent healthy vs early chlorosis. — The cucumber experiment: 4,228 leaves counted, 3,801 healthy, 89.9% healthy versus 10.1% early chlorosis, and a 14.6% greenness gap between the best treatment and control. Cucumber, not soybean — included as an analogy for exactly this kind of treatment-comparison trial on soybean foliage.

Quantifying disease and pest damage

Leaf area is straightforward: count it. Disease is harder, because the question is not "is it there" but "how much, and which way is it moving?"

One of our UK government grants funded exactly this: an AI system for early-stage detection and monitoring of apple scab and apple canker for UK farmers. Two tasks — detect the lesions, which is difficult early, and then measure them. From the measurement you get a threshold, and from repeated measurement you get the dynamic: is the disease declining or not?

An apple leaf on a Petiole Pro calibration plate with red markers on individual lesions, beside a black panel showing each lesion cropped out and labelled with its percentage of leaf area: 0.3 percent, 0.0 percent, 0.6 percent, 0.6 percent, 0.4 percent, 0.1 percent, 0.0 percent. — Each lesion detected, cropped, and reported as a percentage of total leaf area. This is an apple leaf, shown as an analogy for how the same lesion-level quantification applies to soybean foliar disease.

A deliberate design choice: report percentages, not raw square centimetres. Farmers do not have time to work out whether 24 cm² is big or small. They love percentage — so deliver the number that is actually useful for the decision.

The same measurement method covers pest damage. There are two distinct kinds of damaged area, and they need different treatment: coloured or discoloured spots on the leaf, and physically missing, chewed-up area.

Slide showing a row of leaves with progressively more numerous brown ringed lesions labelled damaged area: coloured or discoloured spots on the leaves, and below, a petri dish of chewed leaves plus a grid of leaves with holes eaten through them, labelled damaged area: physically-missing or chewed-up area of the leaves. — Two types of damage, two different measurement problems. The leaves shown are not soybean — they illustrate, by analogy, both the lesion-severity series and the chewed-area case as they would apply to soybean.

For the chewed case, you can quantify the area that remains after the pest has eaten — and, more usefully, calculate which area has been eaten and report it as a percentage. That is the number that tells you about impact.

Two Petiole Spruce app screens: an analyze leaf capture showing three chewed leaves and a calibration plate on a backpack, and the detection result reading Detections: 3, confidence avg 0.95, min 0.92, max 0.97. — Three heavily chewed leaves detected at 0.95 average confidence in the Petiole Spruce app. Not soybean leaves — an analogy for pest-damage assessment on soybean, where quantifying the missing area is the goal.

We work with soft fruit growers counting pests on sticky yellow traps, but for soybean the more relevant experience is the damaged-area measurement above. The soybean case study is in the pipeline. If you want the full method behind damage quantification — segmentation, colour indices, and why percentages and absolute cm² are both worth having — it is written up in AI assessment of necrotic leaf area damage.

Stage 4 — Flowering, and why drones fail here

Flowering is where yield lives, so the job to be done is flowering date, uniformity, flower density, and stress impacts on set. In-field sampling with a mobile application comes first; drone RGB second; thermal imagery third. Phenology models — predicting the flowering window from weather plus planting date — are a nice-to-have for trial planning.

Slide titled Flowering (reproductive stage) listing in-field sampling, drone RGB and thermal imagery, beside a screenshot of a Strawberry Flower Counter tool showing strawberry plants with flowers circled and counted. — The Strawberry Flower Counter, shown as an analogy for soybean flower counting. Strawberries are not soybeans — but the counting problem and the yield-prediction motive are the same shape.

Here is the part that is worth saying plainly, because it cuts against the direction of travel in a lot of phenotyping conversations. Drones cannot properly capture soybean flowers. Two reasons, and they compound:

The flowers are hidden. They sit below the leaf coverage. A camera looking down sees leaves.
Colour separation is hard. If the soybean has purple flowers, you at least have a difference to work with. If they are white or yellowish, it is a genuinely difficult problem for a computer vision algorithm even with good labelling — labelling being the step where you tell the model, on the photo, this is a flower and this is a leaf.

Add drone motion blur to that. To reduce it you need slower speed and lower altitude — but fly lower and you still cannot see the flowers under the top leaves. It is not commercially viable and it is not practically viable to use an unmanned aerial vehicle for counting soybean flowers.

What works instead: handheld devices collecting imagery at plant level, not above it. Or a smartphone attached to a quad bike or tractor, collecting at canopy height. Data from the level of the plants is more insightful than data from the top. And on a huge field, sample — then scope the count with statistics.

Get the raw flower count first with computer vision. Then add phenology models on top for a richer outcome. But the count is the foundation; everything else is layered onto it.

Stage 5 — Pod fill and maturity

Pod fill inherits the flowering problem exactly: the pods are the same colour as the foliage, and they sit below the leaves. The job to be done is pod set success, stress during fill, biomass, and yield proxies — and once again in-field phenotyping outperforms remote-sensing phenotyping.

Slide titled Pod Fill listing remote sensing and thermal imagery as methods and lodging detection and disease severity mapping as nice-to-haves, beside a video frame of chilli pepper plants with dozens of red detection boxes labelled pepper with confidence scores, captioned Fruit count Chilli Pepper. — Chilli peppers counted in a Kenyan greenhouse. Not soybean pods — but the analogy is close: the shape is similar, and green varieties pose the same colour-separation problem. This count came from a grant where we were actually hunting a pest hidden inside the peppers; counting every pepper was the prerequisite, so the count arrived as a by-product.

Thermal imagery earns its place at this stage and earlier. From the first leaves onward, thermal cameras on drones detect water stress and flag areas of interest. Drones cannot count your pods or your flowers — but they can find your stress, and that method transfers to other crops.

Slide titled Mature Seed Pods listing the job to be done as maturity timing, dry-down, shattering risk and harvest readiness, with in-field sampling, drone RGB and thermal imagery as methods, and moisture sensing plus predictive models to forecast the best harvest date window as nice-to-have. — At maturity the signal is colour change and harvest readiness — a computer vision classification task. Ripe or not; ready to harvest or not. From sampling, you build the expectation calendar for when to harvest.

Stage 6 — Post-harvest seed QC

Everything so far has been in-field or in-greenhouse. Post-harvest is different, and it is one of the clear success stories — because it happens in a controlled environment. You are not dependent on temperature or weather. You have one lighting setup and one data-collection protocol.

The job to be done: count, size distribution, cracks, discoloration, mould, impurities, varietal purity, and viability proxies. Two routes, which stack rather than compete:

Mobile application, visual QC. Count, size, presence or absence of defects, visible cracking, and any problems on the sample.
NIR / hyperspectral / X-ray. For internal state. Computer vision and RGB imagery are good at what we can see; they say nothing about the chemical profile of a seed or its internal problems. Proteins and oils are discoverable this way — but the equipment is expensive.

At production level, this becomes automated QC machines and sorters, again built on computer vision.

Slide titled Post-Harvest QC of Bean Seeds listing mobile application visual QC and NIR / hyperspectral / X-ray, beside a screenshot of soybean seeds on a calibration plate with each seed marked, and results reading total seeds detected 68, average seed area 0.55 square centimetres, standard deviation 0.08, min area 0.40, max area 0.75. — Real soybean seeds, counted and measured: 68 seeds, average area 0.55 cm², standard deviation 0.08 cm². The distribution is the point — not just how many, but how they vary.

Three workflows: satellite, drone, smartphone

Satellite — and why it barely featured

Satellite imagery came up once, briefly, and mostly to explain its absence. Plant breeders are interested in very specific results from specific experiments. Satellite is for policy and for understanding the overall scope of production. The resolution simply does not deliver the insights a breeding programme needs.

Slide titled Workflow - Satellite Imagery comparing three views of the same soybean field: drone imagery at 200ft AGL at 0.75 inch per pixel, Planet satellite imagery at 5 metre per pixel, and Sentinel satellite imagery at 10 metre per pixel, with the field detail becoming progressively blockier. — A poor soybean stand in mid-June 2021, seen three ways. At 5 m/pixel the issue is still identifiable but much less pronounced; at 10 m/pixel the detail is lost and the utility for decision support is diminished. Credit: Rob Austin, via soybeanresearchinfo.com.

Drone RGB — where drones genuinely win

Drones are excellent, and the talk was clear about where: emergence counts, lodging, and maturity. Lodging in particular is a case where handheld simply cannot compete — you need to go up to see the affected areas at the later stage. Maturity, being a colour-difference problem across a field, is also a good fit.

Slide titled Workflow - Drone RGB listing emergence counts, lodging and maturity, beside an aerial image of soybean rows with yellow boxes marking identified plant gaps and orange boxes marking rows with no gaps over the threshold distance. — Plant gap analysis from drone RGB on soybean rows: gaps over a user-specified threshold distance are flagged as an indirect measure of plants per acre. Credit: Rob Austin, North Carolina State University Extension, via soybeanresearchinfo.com.

Mobile devices — the through-line

Germination. In-field scouting: plant health, flower and yield count, pest and disease pressure. Post-harvest seed visual QC. And then a bonus category that nobody asks for and everybody needs: operational routine.

Slide titled Workflow - Mobile Devices listing germination, in-field scouting (plant health, flower and yield count, pest/disease pressure, operational routine) and post-harvest seed visual QC, beside a phone screen titled Soybean Seeds showing thousands of seeds segmented in blue, purple and red by standard deviation, with counts 2072, average 1.10, standard deviation 0.09. — 2,072 soybean seeds counted on one screen, with diameter distribution segmented by standard deviation — blue within 1 SD, purple 2 SD, red 3 SD. Not just the count: the allocation.

That standard-deviation view is worth pausing on. We do not just count and show the count — we show how the seeds are distributed: how many sit in the average range, how many at the smallest, how many at the biggest. It is the difference between a number and a characterisation.

Three panels: a Soybean Seeds Counter web tool showing seeds on a calibration plate with detected seeds marked and a seed area distribution histogram; a phone screen with 2,072 seeds segmented by diameter; and a CREATE Soybean Seeds screen with fields for Variety #1, Count 2072 and Experiment #2. — The same soybean seed workflow across surfaces — web counter with the seed-area histogram, the on-phone segmentation, and the record that ties the count to a variety and an experiment.

Operational excellence: the part nobody asks for

This was the quiet heart of the talk. Years of work with breeders and growers taught us that getting the number is not the task. You always need to attach something to the number: which variety, when it was sown, when fertiliser or any agricultural input was applied. It is never a story of one number. It is a journey of growing, and the context has to attach smoothly — otherwise the count is done, and then it disappears, because the next stage has already started.

Three Petiole Spruce app screens: a soybean canopy photo with an instant action button reading CREATE Soybean Plant, a CREATE Soybean Plant form with Variety and Agricultural Input fields showing Fertiliser #1, Fertiliser #2 and No Fertilizer options, and a Soybean Plants record showing agricultural input, plot and variety properties. — Capture on the go. The result is not a photo — it is an image carrying the fields you need for your report. Come back next week, photograph the same plant, fill the same fields, and the progress is already structured.

The pain this solves is embarrassingly ordinary. A grower with three experiments and four fertiliser applications works in the field, comes back to the office, and tries to transfer his notes into an Excel table. It is time-consuming. Sometimes the notes are lost. Sometimes they have been out in the rain and are not very good. And the historical record ends up somewhere in Google Drive, to be hunted for later.

Three app record screens: Crop Protection Data showing 3.0 pesticide applications and Pesticide #2 for Experiment #2, Soil Test Results attached to Experiment #2, and Soybean Irrigation Data showing a water meter photo, the note water meter should be checked, Experiment #2 and Plot #3. — Irrigation, soil tests, crop protection — each attached to an experiment and a plot, each with a photo as digital proof. Two years later, when the experiment closes, none of it has evaporated.

There is a second reason to keep it together beyond not losing it. Once all the data is in one place, you can feed it into an AI algorithm that provides more insight — because sometimes we miss things, and summarising data is exactly what these systems are good at.

Two app screens labelled All data in one app and On-demand AI models: an Objects list showing Crop Protection Data, Soil Test Results, Soybean Irrigation Data, Stomata Count, Soybean Plants and Soybean Seeds records from February 17 2026, and an AI Library showing installable models for Blueberry Fruit, Stomata, Individual Leaf, Apple Half and Cells Germination. — All the day's records in one place, and an AI library where models are installed on demand rather than shipped as one monolithic app.

An unexpected one: nozzle output

One of the soybean papers citing Petiole Pro was nominally about leaf area — but the researchers' real task was assessing nozzle setup on a sprayer. They used water-sensitive paper and needed to quantify the footprint of each nozzle at each setting. That is not a breeding application, but applying agricultural chemicals correctly is an important job, and the same measurement machinery works: quantify the print output instead of eyeballing it, and get at least a percentage.

Three panels: a sheet of paper with dark spray droplets beside a Petiole Pro calibration plate; the app measuring one droplet at 0.35 square centimetres; and a reference figure of yellow water-sensitive paper covered in purple droplets, credited to Tadic et al. 2014, with a results comparison table of planimeter, scanner and app measurements. — Spray droplets measured on paper. Not a plant at all — included as an analogy for how the same area measurement transfers from soybean leaves to spray-coverage assessment, with the app validated against planimeter and scanner methods.

Slide showing a phone and tablet running a germination detection screen labelled Mobile App — Android (free to download) plus iOS (on-demand) — beside a Petiole Pro web dashboard showing seedlings scored today, time per 100 scores, crops, tasks and recent scores, labelled Web App — Windows OS. — Two surfaces for two jobs. Field scouts collect on mobile; managers reading the overall situation sit in front of a desktop, so there is a Windows application too. You can read more about Petiole Pro and the newer Petiole Spruce app on their product pages.

Nodules, stomata, and models trained on demand

Two final pieces of science, both of which started as somebody's problem rather than our roadmap.

Stomata counting was brought to us by wheat breeders trying to assess their success, for whom manual counting was punishing. It is not a soybean task, but it may be useful — and it now exists as both a web and a mobile application.

Two panels: a Leaf Stomata Counter web tool showing a microscope image of leaf epidermis with 64 stomata circled in green and yellow by confidence, and a phone screen titled Count Stomata reading Detections: 60, confidence avg 0.82, min 0.41, max 0.89. — Stomata counted automatically from a microscope image, with per-detection confidence. Not soybean — this came from wheat breeders — and included as an analogy for the same microscopy workflow on soybean leaf epidermis.

Nodule counting is a soybean success story, and it came from a researcher at Kansas State University. He got in touch with a fair complaint: you measure leaves, but I work with roots, and counting these nodules takes my whole day when I am busy with other tasks — can you help? It was an interesting task, so we took it. That was around three years ago, before the AI boom, when computer vision was nothing like as popular an application as it is now.

Two panels: a Soybean Nodules Counter web tool showing a soybean root scan on a blue background with detected nodules marked, and a large zoomed view of the same root system with 30 nodules each marked by a green box. — The Soybean Nodules Counter. Real soybean root scans, with each nodule detected individually — the day's work that started this collaboration.

Nodule Count Results panel reading total nodules detected 30, average confidence 0.86, above a grid of thirty individual cropped nodule images. — 30 nodules, 0.86 average confidence, and every nodule cropped out for inspection. The count is checkable, which is what makes it usable in a paper.

This is what "AI on demand" means in practice, and it closes the loop back to the germination failure. The germination model did not work on soybean because it had not seen soybean. But that is a solvable problem: if something is already developed, we can go deeper on your imagery and build it together.

Watch the full talk

The complete session — all six stages, the tool comparisons, the failure cases, and the audience questions — is on YouTube.

Dr Maryna Kuzmenko's full talk at the 2026 Soybean Breeders' Workshop, University of Georgia.

If the basics underneath all of this are what you actually want — what machine learning is, how deep learning differs from it, what separates traditional farming from smart farming — there is a free introductory course on Udemy. AI moves fast; agriculture moves at the speed of seasons and biological cycles, which is not the same speed at all, and it is worth reminding ourselves of the fundamentals.

Udemy course page for AI in Agriculture: Practical Introductory Course by Maryna Kuzmenko, rated 4.7 from 494 ratings with 2,598 students, listed as free with one hour of on-demand video. — *AI in Agriculture: Practical Introductory Course* — free, one hour, and deliberately basic.

Four themes from the wider programme

The talk was one session in a full programme, and what stayed with us was not only our own. Four directions felt especially relevant across the workshop:

Speed and throughput are no longer a nice-to-have. Winter nurseries, smarter greenhouse workflows, technician-led efficiency — that is how cycle time is measured in reality. Otherwise routine eats the time.
High-throughput phenotyping was discussed mainly through drone imaging for maturity and field performance. We were pitching handheld tools — but with genuine (and mutual) affection for drones, it is good to see this data-collection method treated as a repeatable signal for better in-field decisions.
Durable resistance is gaining power. Soybean Cyst Nematode and disease pressure remain relentless, and there was a lot of practical thinking around stacking, validation, and keeping resistance useful over time rather than winning a single season. Everyone is actively scouting novel resistance sources, including wild soybean resources.
Quality traits are moving earlier in the pipeline. Oil profile, protein, composition analytics as usual — but now scaling fatty-acid profiling (FAME/GC) and other compositional screens earlier, not only at the end.

And the cherry on top — or better, a bunch of beautiful mature soybean pods: digitalisation is slowly but confidently wrapping more and more aspects of soybean breeding. Innovation rarely arrives as a big reveal. It arrives as dozens of small, disciplined, stacked improvements until the whole pipeline moves faster.

If you work in wheat, barley, oilseed rape, pulses, vegetables or berries, the parallel themes hold. The tools differ and the constraints differ, but the directions are shared.

Petiole Pro for soybean breeding

Bring your imagery, and let's see what works

The Petiole Pro mobile app is free on the Play Store and works offline — measure leaf area, count seeds, and assess damage from your phone. If you have a task that needs a model trained on your own soybean imagery — germination trays, nodules, pods, flowers — that is exactly the conversation we want. Write to us at [email protected].

Building a grant application that needs phenotyping, seed QC, or damage measurement? We are a UK-based company with Innovate UK plant-science grants behind us and are glad to join as a consortium partner.

The closing slide. Dr Maryna Kuzmenko writes the AI in Agriculture newsletter on LinkedIn, including several editions focused specifically on soybeans.

Frequently asked questions

What is the difference between AI and applied AI for soybean breeding?

AI for soybean, as usually described, is a list of possible directions — accelerated breeding, yield prediction, pest detection, irrigation optimisation. Applied AI is the much shorter list of what works in a breeding programme today. Ask a chatbot how to use existing tools for soybean research and you get five to ten directions; try them and one or two are genuinely usable. Applied AI for soybean is overwhelmingly visual, smartphone-based, and tied to a specific growth stage.

Can AI count soybean germination from a photo of a tray?

Yes in principle — an RGB smartphone camera is enough to count germinated and empty cells and compute the germination rate automatically. But a model trained on another crop's trays will fail on soybean: in the workshop demo, a module that scored 240 tree-seedling cells at 92.9% germination could not read a soybean tray, because the plug geometry and seedling architecture differ. The model concept exists; the bottleneck is relevant soybean training data.

Can drones count soybean flowers and pods?

No, not practically. Soybean flowers and pods sit below the leaf canopy, so a downward-looking camera mostly sees leaves. White or yellowish flowers are also hard for a computer vision model to separate from foliage even with good labelling, and drone motion adds blur. Handheld devices at plant level — or a smartphone mounted on a quad bike or tractor — are more insightful than imaging from above.

What are drones actually good for in soybean?

Emergence counts, lodging, and maturity. Lodging in particular is hard to see with handheld devices — you need altitude to spot the affected areas late in the season. Maturity is a colour-difference problem across a field, which suits drone RGB well. Thermal cameras on drones are also effective for water-stress detection from the first leaves onward.

Is satellite imagery useful for soybean breeding?

Rarely. Satellite imagery suits policy work and understanding the overall scope of production. Breeders need specific results from specific experiments, and the resolution does not deliver them: a poor soybean stand visible in drone imagery at 0.75 inch per pixel remains identifiable at 5 m/pixel but with much less detail, and at 10 m/pixel the detail is lost and the utility for decision support diminished.

What can a smartphone measure on soybean that expensive equipment usually does?

Leaf area (with a calibration plate, in real cm², offline), germination counts, greenness and early chlorosis against a threshold, disease lesion severity as a percentage of leaf area, pest-chewed area, flower and pod counts at plant level, root nodule counts, and post-harvest seed count with size distribution and defect screening. What it cannot do is see inside a seed — chemical profile, protein and oil content need NIR, hyperspectral or X-ray.

Why report leaf damage as a percentage rather than in cm²?

Because growers do not have time to work out whether 24 cm² is a lot or a little. Percentage is immediately interpretable and drives the decision. For scientific work, both are worth having: the percentage gives you the relative severity, and calibrated cm² gives you the absolute measurement.

Why does a count need context like variety and fertiliser attached to it?

Because a number alone disappears. Working with breeders and growers showed that a count only becomes data when it travels with the journey of growing — variety, sowing date, agricultural inputs, irrigation, soil tests. Field notes transferred later into Excel get lost, rained on, or scattered across drives. Keeping everything in one place also makes it possible to feed the whole record into an AI system for summarisation and further insight.