Stomata Count with AI: How to Measure Stomatal Density from Microscope Images in Seconds

What are stomata?

Stomata (singular: stoma) are microscopic pores in the epidermis of leaves and stems. Each stoma is a pore flanked by two guard cells that change shape to open or close it.

Stomata do two jobs at once:

• Gas exchange. When open, they let CO₂ into the leaf for photosynthesis and release O₂.
• Water regulation. The same pores let water vapour escape through transpiration. That cools the leaf and pulls water and nutrients up from the roots, but it also costs the plant water.

Guard cells respond to light, CO₂ concentration and the plant's water status. When water is plentiful they swell and the pore opens; when the plant is short of water they lose turgor and the pore closes. Every open stoma is a trade-off between carbon gained and water lost, and much of crop physiology comes down to managing that trade-off.

In grasses such as wheat, barley and maize, guard cells are dumbbell-shaped and sit in rows that run parallel to the leaf veins. That regular, row-by-row layout is clearly visible in the microscope images later in this article. Wheat leaves carry stomata on both surfaces, so the upper (adaxial) and lower (abaxial) sides are usually counted separately.

Why count stomata?

Stomatal traits are set during leaf development and respond to the environment the plant grows in, so they carry information about how a plant is built to handle water, heat and CO₂. People count stomata to:

• Breed for water-use efficiency. In bread wheat, Dunn and colleagues (2019) reduced stomatal density by over-expressing an epidermal patterning factor (EPF) gene. Plants with moderate reductions kept yields comparable with controls, including under drought and elevated CO₂, while their intrinsic water-use efficiency increased. Severe reductions cut yield. In barley, Hughes and colleagues (2017) found that lines with significantly fewer stomata had better drought tolerance and water-use efficiency with no loss of grain yield.
• Screen genotypes and treatments. Stomatal density varies between varieties and responds to treatments, so it is a useful trait for phenotyping trials. A 2025 field study of spring wheat in Estonia found that varieties with more stomata on the upper leaf surface tended to give lower grain yields.
• Study plant stress. Heat, drought, over-irrigation and CO₂ all affect stomatal behaviour. Linking stomatal counts to these variables turns them into a signal of how the plant is responding.
• Reconstruct past climate. In a classic 1987 study, Woodward compared herbarium leaves collected over the past two centuries and found that stomatal numbers fell as atmospheric CO₂ rose from pre-industrial levels. Fossil-leaf stomata are now used as a proxy for ancient CO₂.

What is stomatal density, and how is it calculated?

Two numbers describe how many stomata a leaf has.

Stomatal density (SD) is the number of stomata per unit of leaf area, usually per square millimetre:

Stomatal density (stomata/mm²) = number of stomata in the field of view ÷ area of the field of view (mm²)

For example, if a field of view covers 0.40 mm² and contains 30 stomata, SD = 30 ÷ 0.40 = 75 stomata/mm². The field-of-view area depends on your microscope and magnification, so measure it with a stage micrometer or a scale bar rather than assuming it.

Stomatal index (SI) is the proportion of epidermal cells that are stomata, introduced by Salisbury in 1927:

Stomatal index (%) = S ÷ (S + E) × 100

where S is the number of stomata and E the number of other epidermal cells in the same area. Because leaves expand by cells getting bigger as well as more numerous, SD changes with leaf expansion while SI is more stable. Researchers interested in development often report both.

Stomata counting AI is most directly useful for density, because detection and counting is exactly what object-detection models do well. Each detection also gives a position, so patterning can be analysed from the same output.

How are stomata counted by hand, and why is it a bottleneck?

The classic workflow has three stages.

1. Make an image of the leaf surface. The most common low-cost method is the leaf impression: paint a thin layer of clear nail varnish on the leaf, let it dry, lift the film with clear tape and stick it on a glass slide. Rothamsted Research's bioimaging guide describes this method and notes that some leaves are damaged by the solvent in the varnish. Alternatives include silicone dental impression material (useful for delicate leaves or repeated sampling of the same leaf), epidermal peels, and imaging the leaf directly with a microscope.
2. Photograph several fields of view per impression under the microscope.
3. Count the stomata in each image, usually by clicking each one in an image-analysis program or with a hand tally counter, then convert the counts to density in a spreadsheet.

The method is cheap and well proven. The problem is scale:

• The numbers add up quickly. Several fields of view per leaf, both leaf surfaces, several leaves per plant, several plants per plot, and several genotypes, treatments and dates soon means thousands of images and tens of thousands of stomata.
• Counting is tiring and drifts. Is a half-visible stoma at the edge of the frame counted? Is a blurred shape a stoma or an air bubble in the varnish? Different people decide differently, and so does the same person at the end of a long day.
• The data gets scattered. Images are sent by email or messaging apps, counts live in one spreadsheet, the trial design in another, and the conclusions in a PDF. Matching each image with the right plant, treatment and date becomes a job of its own.

This is why many groups measure fewer plants or fewer fields of view than they would like, which reduces the statistical power of the very trials stomatal traits are meant to inform.

How does AI count stomata?

Automatic stomata counting is an object-detection problem: find every instance of one type of object, a stoma, in an image. The process breaks down into a few steps.

1. Train a detection model. A deep-learning model is trained on microscope images in which people have marked every stoma. It learns the visual pattern of a stoma (the pair of guard cells around a pore, in the context of the surrounding epidermis) and learns to ignore look-alikes such as cell walls, hairs, debris and air bubbles. Petiole's model was fine-tuned on microscope images of wheat stomata.
2. Detect every stoma in a new image. When you give the model a new image, it proposes a location for every stoma it finds, along with a confidence score between 0 and 1: how sure it is that the detection is a real stoma.
3. Show the detections. Each detection is drawn on the image, so you can see at a glance what the model counted and what it missed.
4. Count, filter and summarise. Detections above a chosen confidence threshold are counted. The result is reported as a total, a breakdown by confidence level and, with the image area, stomatal density.

What does the AI output look like?

Here is a real result from one wheat micrograph, as shown in our March 2025 demo.

Average confidence: 0.78.

Two-thirds of the detections (56 of 83) have a confidence of 0.8 or more. These are clear, textbook stomata. The rest are where human judgement still helps.

This is also why the confidence threshold matters. With these detections:

• counting everything gives 83 stomata;
• counting only detections at 0.5 or above gives 69;
• counting only detections at 0.7 or above gives 61.

None of these is automatically “right”. The best threshold for your material is the one whose counts agree best with careful hand counts of the same images, which is why validation is part of the workflow.

Reviewing detections, not counting them

Every detection is also shown as a small crop, sorted from highest to lowest confidence.

The crop gallery changes the job. Instead of clicking every stoma in every image, you scroll down to where the confidence drops and decide whether those few detections are real stomata (for example, ones that are slightly out of focus) or look-alikes. That keeps a person in control of the result while removing most of the repetitive work.

Stomata counting on your phone, offline

The obvious objection to mobile stomata counting is that “labs aren't run on phones”. That's true, but many decisions about plants are made in the field and the greenhouse, not in the lab. If a stomata count only arrives weeks after sampling, the moment to act on it has often passed.

That is why the stomata counting module we are building for the Petiole Pro mobile app runs on the device:

• Edge AI. The detection model runs on the phone or tablet's own processor. There is no upload to the cloud and no waiting for a server.
• No internet required. Many fields, glasshouses and growth facilities have no reliable mobile signal. The count works anyway.
• Data stays with you. Images from sensitive breeding or commercial trials don't have to leave the device.
• One visit: capture, analyse, act. You can compare genotypes or treatments the same day, flag plants that look stressed before wilting is visible, and decide which plots to irrigate, protect or prioritise.

The images still come from a microscope. The phone does the analysis and keeps the results with the context.

From counting to connected data

Counting is only the first step. The module is designed so that each image stays linked to its metadata: plant variety, treatment, trial, date, location and experimental design. The count, the image and the context stay together, ready for analysis, instead of being spread across emails, chat groups, spreadsheets and PDFs.

The bigger goal is to connect stomatal density with the other variables in an experiment or production cycle, such as heat stress, water deficit or excess, CO₂ concentration and other environmental changes, so that stomata become a measurable signal of how the plant is responding.

How do you get images that count accurately?

Most counting errors start with the sample or the image, not the model. A checklist:

1. Sample consistently. Use the same leaf (for example the flag leaf or the youngest fully expanded leaf), the same position on the leaf (often mid-way between base and tip, avoiding the midrib) and the same leaf surface for every plant you compare. Record which surface you sampled.
2. Make thin, even impressions. A thin coat of clear varnish gives sharper detail than a thick one. Let it dry completely and avoid trapping air bubbles, which can look like stomata.
3. Keep the impression flat on the slide, without wrinkles or folds in the tape.
4. Focus carefully. Stomata out of focus get lower confidence scores and are more likely to be missed. If your microscope allows it, focus on the plane where the guard cells are sharpest.
5. Use the same magnification and settings for all images in a trial. The model and your density calculations both rely on a consistent scale.
6. Know your field-of-view area. Calibrate it once with a stage micrometer for each objective, and use that area to convert counts to stomata per mm².
7. Take several fields of view per sample and average them. Stomatal density varies across a leaf, so one image is rarely representative.
8. Use a consistent rule for stomata on the edge of the image, for example: count those touching the top and left edges, but not those touching the bottom and right. Apply it the same way whether the counting is done by a person or by the software.

How accurate is AI stomata counting, and what are its limits?

Published work shows that automatic stomata counting can match people closely, as long as the model has seen material like yours.

• StomataCounter (Fetter and colleagues, New Phytologist, 2019) was trained on 4,618 images from more than 700 plant species. It identified stomata with 94.2% accuracy on species it had not been trained on, with a precision of 0.99 and a recall of 0.93 on the combined datasets. Its counts on poplar images correlated very closely with human counts (r² = 0.96), while agreement was weaker on more varied image collections.
• StomaAI (Sai and colleagues, New Phytologist, 2023) measures stomatal pores as well as density in Arabidopsis and barley, and produces measurements consistent with those of human experts in a fraction of the time.

The lesson from both is the same: accuracy is highest on material similar to the training data and drops on unfamiliar species, image types or magnifications. Situations to watch for:

• A new species or leaf type. A model fine-tuned on wheat will perform best on wheat and other grasses with similar stomata. Leaves with randomly oriented stomata, such as many broadleaf crops, look different.
• Look-alikes: air bubbles and debris in the varnish, hairs (trichomes), dust on the slide and torn impressions.
• Blur and poor focus, which push real stomata to low confidence scores.
• Crowded or overlapping stomata, which can merge into one detection.
• Partial stomata at image edges, which need the same counting rule as above.
• Changes in microscope, camera or magnification between training and use.

How to validate it for your setup: choose 20–30 images that cover the range of your material: different genotypes, treatments, leaf surfaces and image quality. Count them carefully by hand, ideally with two people working independently. Then compare the hand counts with the AI counts at a few confidence thresholds. Plot AI count against hand count: a slope close to 1 and a high r² mean the model is reliable for your material at that threshold. If the differences are systematic, such as consistently missing stomata on one leaf surface, adjust the sampling or imaging, or send us examples so we can improve the model.

Who uses AI stomata counting?

• Plant physiologists and researchers studying gas exchange, water-use efficiency, drought, heat and CO₂ responses.
• Plant breeders and pre-breeding teams screening many lines for stomatal traits linked to water use and yield.
• Agronomists and trial managers comparing treatments such as irrigation regimes, biostimulants or crop protection products in the field and glasshouse.
• Greenhouse and controlled-environment growers exploring data-driven irrigation and climate control.
• Universities and teaching labs for whom stomatal density is a classic practical exercise that is quicker to teach when counting is automated.
• Palaeobotanists and climate scientists using stomatal density and index in fossil or herbarium leaves as a record of past CO₂.

Stomata counting in Petiole Pro

Petiole's stomata counter started as the Leaf Stomata Counter web demo shown above and is being built into the Petiole Pro mobile app as a module, alongside our other image-based phenotyping tools. It is not publicly available yet. The planned workflow is:

1. Make a leaf impression (or prepare an epidermal sample) and photograph it under the microscope.
2. Load the image into Petiole Pro on your phone or tablet.
3. The on-device AI marks every stoma and returns the count with confidence scores, with no internet needed.
4. Review the low-confidence detections in the crop gallery.
5. Keep the image, the count and the trial metadata together, and export the results for analysis.

Petiole Pro is designed as a modular toolkit: you choose the measurements you need. Stomata counting works on microscope images, while modules such as leaf area work on ordinary photos of whole leaves next to a calibration plate. When the stomata module is released, access will be free for research organisations and universities; for commercial organisations we will agree terms to suit the project.

About the author

Dr Maryna Kuzmenko is the founder of Petiole, a UK company building AI and computer-vision tools that turn images 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, publishes the YouTube series AI in Agriculture: One Idea a Day and teaches the Udemy course AI in agriculture.

References

1. Fetter K. C., Eberhardt S., Barclay R. S., Wing S., Keller S. R. (2019). StomataCounter: a neural network for automatic stomata identification and counting. New Phytologist. doi:10.1111/nph.15892
2. Sai N., Bockman J. P., Chen H., Watson-Haigh N., Xu B., Feng X., Piechatzek A., Shen C., Gilliham M. (2023). StomaAI: an efficient and user-friendly tool for measurement of stomatal pores and density using deep computer vision. New Phytologist 238(2), 904–915. University of Adelaide repository
3. Dunn J., Hunt L., Afsharinafar M., Meselmani M. A., Mitchell A., Howells R., Wallington E., Fleming A. J., Gray J. E. (2019). Reduced stomatal density in bread wheat leads to increased water-use efficiency. Journal of Experimental Botany 70(18), 4737–4748. doi:10.1093/jxb/erz248
4. Hughes J., Hepworth C., Dutton C., Dunn J. A., Hunt L., Stephens J., Waugh R., Cameron D. D., Gray J. E. (2017). Reducing stomatal density in barley improves drought tolerance without impacting on yield. Plant Physiology 174(2), 776–787. doi:10.1104/pp.16.01844
5. Higher adaxial stomatal density is associated with lower grain yield in spring wheat (2025). bioRxiv preprint. doi:10.1101/2025.01.07.631641
6. Woodward F. I. (1987). Stomatal numbers are sensitive to increases in CO₂ from pre-industrial levels. Nature 327, 617–618. doi:10.1038/327617a0
7. Salisbury E. J. (1927). On the causes and ecological significance of stomatal frequency, with special reference to the woodland flora. Philosophical Transactions of the Royal Society of London B 216, 1–65.
8. Rothamsted Research Bioimaging (2017). Measuring stomatal density using nail varnish. PDF